An Original GIS and Remote Sensing Protocol to

Detect Agricultural Drought Effects on Rainfed Agro

Ecosystems in Semi-Arid Developing Regions: A Case

Study for Central Mexico

Andres Sierra-Soler

Bioresource Engineering

Faculty of Agricultural and Environmental Sciences

McGill University, Montreal 2

A thesis submitted to McGill University in partial fulfillment of the requirements of

the degree of Masters in Science

©2013 3

Contents

Acknowledgements ...... 8 Abstract ...... 9 Chapter 1 ...... 12 1.1 Introduction ...... 12 1.2 Objectives ...... 17 1.3 Thesis Outline ...... 18 Chapter 2 ...... 21 Literature Review and Selection of Methods ...... 21 2.1 What is a Drought? ...... 21 2.2 What Causes Drought? ...... 22 2.3 Classification of Droughts ...... 24 2.4 Overview of Drought Indices ...... 25 2.4.1 Data driven drought indices ...... 25 Palmer Drought Severity Index (PDSI) ...... 26 Standardized Precipitation Index (SPI) ...... 26 Crop Moisture Index (CMI) ...... 27 PDSI vs. SPI ...... 28 2.4.2 Drought indices derived from Satellite Data...... 28 Normalized Difference Vegetation Index (NDVI) ...... 28 Vegetation Condition Index (VCI) ...... 30 The Normalized Difference Water Index (NDWI) ...... 31 Tasseled Cap Transformation Wetness (TCW) ...... 32 Temperature Condition Index (TCI) ...... 34 Vegetation Health Index (VHI) ...... 35 Standardized Vegetation Index (SVI) ...... 35 2.5 Comparing Drought Indices ...... 36 2.6 The Use of Land Use Land Cover (LULC) Maps for Detecting the Effects of Droughts ...... 38 2.7 Why Study Drought? Social Vulnerability to Drought ...... 39 2.7.1 Social vulnerability to drought ...... 40 2.7.2 Droughts as a natural hazard ...... 40 4

2.7.3 Droughts and human activity ...... 41 2.7.4 The importance of maize in Mexico...... 42 2.7.5 Overview of land tenure processes in Mexico ...... 42 2.8 Selection of Methods ...... 46 2.8.1 Meteorological Analysis ...... 46 2.8.2 Remotely Sensed Data ...... 48 2.8.3 Land Use Land Cover Maps ...... 50 2.8.4 Spectral indices to detect agricultural drought ...... 52 2.9 Summary and Conclusions ...... 55 Preface to chapter 3 ...... 55 Chapter 3 ...... 56 Assessing Agricultural Drought at a Regional Scale Using LULC Classification, Standardized Precipitation Index (SPI), and Vegetation Indices...... 56 Abstract ...... 56 1. Introduction ...... 57 1.1 Study Area ...... 62 1.1.1 Climate ...... 63 1.1.2 Agriculture ...... 63 2. Materials and Methods ...... 64 2.1 Materials ...... 64 2.1.1 Monthly Precipitation Data ...... 64 2.1.2 Open Access SPI software ‘SPI_SL_6.exe’: ...... 65 2.1.3 Satellite Imagery ...... 65 2.1.4 Ground truth data ...... 67 2.1.5 Climatic and topographic maps ...... 68 2.1.6 GIS software ...... 68 2.1.7 Remote sensing software ...... 68 2.2 SPI Analysis ...... 68 2.3 Land Use Land Cover (LULC) Classification ...... 70 5

2.3.1 Preprocessing images ...... 70 2.3.2 Extraction of training sampling locations ...... 71 2.3.3 Land Use Land Cover (LULC) Analysis: Saadat et al., (2011) simplified method ...... 72 2.3.4 Assessing accuracy in remotely sensed data ...... 74 2.4 Spectral Indices for Detecting Drought ...... 75 2.4.1 The normalized difference vegetation index (NDVI) ...... 75 2.4.2 The vegetation condition index (VCI) ...... 78 2.5 Change Detection ...... 79 2.5.1 Vegetation types and phenomenology ...... 80 3. Results and Discussion ...... 81 3.1 SPI Results ...... 81 3.2 LULC Maps ...... 83 3.3 NDVI and Vegetation Cover Change ...... 85 3.4 NDVI Change Detection ...... 87 3.5 VCI and the Vegetation Cover Change ...... 91 4. Conclusion ...... 91 Appendix 1: Confusion Matrices ...... 97 Appendix 2: Accuracy Totals ...... 106 Appendix 3: Kappa Index ...... 118 Chapter 4 has been submitted to the Journal of Agriculture and Environment for International Development. The manuscript has been co-authored by Andres Sierra-Soler, Jan Adamowski, Hossein Saadat, Zhiming Qi and Santosh Pingale...... 126 Chapter 4 ...... 127 High Accuracy Land Use Land Cover (LULC) Maps for Detecting Agricultural Drought Effects in Rainfed Agro-Ecosystems in Central Mexico...... 127 Abstract ...... 127 1. Introduction ...... 128 2. Study Area ...... 133 3. Materials and Methods ...... 135 6

3.1 Materials ...... 135 3.2 LULC Classification ...... 136 3.2.1 General description ...... 136 3.2.2 Preprocessing images ...... 136 3.2.3 Extraction of a training sampling location ...... 138 3.2.4 Supervised classification of the images into LULC classes...... 138 3.2.5 Step 4: Image segmentation and zonal statistics ...... 139 3.2.6 Assessing the accuracy of remotely sensed data ...... 142 3.2.7 Spectral analysis for detecting drought effects on vegetation ...... 143 3.2.8 Normalized difference vegetation index (NDVI) ...... 144 3.2.9 The normalized difference water index (NDWI)...... 145 3.2.10 Tasseled cap transformation wetness (TCW) ...... 146 3.2.11 Change detection ...... 147 4. Results ...... 148 4.1 Land Use Land Cover (LULC) Maps ...... 148 4.2 NDVI Results...... 150 4.3 NDWI and Soil/Vegetation Moisture ...... 156 4.4 TCW and Soil/Vegetation Moisture ...... 158 5. Conclusion ...... 160 Appendix 1: Confusion Matrices ...... 164 Appendix 2: Accuracy Totals ...... 168 Appendix 3: Kappa Index of Agreement ...... 170 Appendix 4: Photographic samples of each LULC class in the study area ...... 171 References ...... 173 Chapter 5 ...... 181 5.1 Summary and General Conclusions ...... 181 5.2 Contributions to Knowledge ...... 184 5.3 Suggestions for Future Research ...... 184 Bibliography ...... 186 7 8

Acknowledgements

The authors gratefully acknowledge Dr. Margaret Kalacsk and Dr. Pablo Arroyo for their support in the data analysis process. Data analysis was carried out at the High Performance Computer

Laboratory at the Geographic Information Centre in the Department of Geography, at McGill

University. Dr. Sergio Cortina-Villar for reviewing this work and for his insights in Land Use and

Management in the rural landscape of Mexico, Dr. Irma Saucedo Vaca for her insights in the dynamics of irrigated and non-irrigated agriculture in Mexico, Dr. Francisco Ramirez Diaz for his opinions in food security, and Juan Montes for his help during the development of the field work carried out in Jan-March 2013. This research was supported by the National Council of

Science and Technology of Mexico (CONACYT) and the State Commission of Water and

Sewerage of the State of Hidalgo (CEAAH in Spanish). Funding was also provided by NSERC

Discovery Grant and a CFI grant held by Jan Adamowski, as well as a McGill startup grant held by Dr. Zhiming Qi. 9

English Abstract

Drought is a silent and pervasive phenomenon, it creeps up over weeks, months, and even years often without any warning, affecting profoundly ecosystems and human activity on a global scale. Globally droughts are one of the most devastating natural hazards in terms of the people affected and inflicting directly or indirectly economies, societies and ecosystems. As any natural hazard, the degree of exposure and the ability of systems to be resilient is directly related with the vulnerability of communities. Agricultural drought represents a great threat to poor farmers in semi-arid regions developing regions. For farmers who depend rainfed agricultural production for self-sufficiency a drought can result in great suffering, thus there is need to understand how droughts disturb the landscape in such regions.

Satellite observations of the Earth have the potential in proving insights on vegetation conditions, crop yield and furthermore to monitor the impact of droughts. The relationship between spectral properties of vegetation and their biomass has been recognized since the first spectrometric field experiments in the 1970s. Satellite remote sensing provides a synoptic view of the land and a spatial context for measuring drought impacts which have proved to be a valuable source of spatially continuous data with improved information on monitoring vegetation dynamics.

The following thesis had the objective of proposing a new methodology to monitor the development of different vegetation covers in the presence of detected abnormally dry conditions and to be compared with vegetation development in periods with normal precipitation with a particular interest on rainfed agricultural lands. In synthesis the analysis of the impacts of droughts to vegetation was performed with the use remote sensing techniques to detect its effects on vegetation. An original protocol for performing a Land Use Land Cover (LULC) classification that combines climatic, topographic and reflectance information from 18 Landsat ETM+ images was applied to subsequently distinguish drought effects in different classes through the selected years. The achieved LULC classification produced overall classification accuracies ranging from 87.88% to 92.42%. Spectral indices for vegetation and soil/vegetation moisture were then used to detect anomalies in vegetation development caused by drought and 10 furthermore, the area of water bodies was measured and compared to detect changes in water availability for irrigated crops. The methodology was tested in Central Mexico to determine which growing season period to detect the evolution of a drought occurred in the year 2005 and could be applied to other semiarid regions.

French Abstract

La sécheresse est un phénomène silencieux et omniprésent, il glisse au fil des semaines, des mois, voire des années souvent sans aucun avertissement, affectant les écosystèmes et les activités humaines à l'échelle mondiale. Sécheresses à l'échelle mondiale sont l'une des catastrophes naturelles les plus dévastatrices en termes de personnes touchées et infliger directement ou indirectement économies, les sociétés et les écosystèmes . Comme toute catastrophe naturelle, le degré d'exposition et la capacité des systèmes de résilience sont directement liés à la vulnérabilité des communautés. La sécheresse agricole représente une grande menace pour les agriculteurs pauvres dans les régions en développement des régions semi-arides. Pour les agriculteurs qui dépendent de la production agricole pluviale pour l'autosuffisance une sécheresse peuvent entraîner de grandes suffering. C’est nécessaire de comprendre comment la sécheresse perturbent le paysage dans ces régions. Les observations satellitaires de la Terre ont le potentiel de fournir des indications sur les conditions de végétation, le rendement des cultures et, en outre de surveiller l'impact des sécheresses . La relation entre les propriétés spectrales de la végétation et la biomasse a été reconnue depuis les premières expériences sur le terrain spectrométriques dans les années 1970 . La télédétection par satellite offre une vue synoptique de la terre et un contexte spatial pour mesurer les impacts de la sécheresse qui se sont révélés être une source précieuse de données spatialement continues avec une meilleure information sur le suivi de la dynamique de la végétation . Cette thèse avait pour objectif de proposer une nouvelle méthodologie pour suivre l'évolution des différents couverts végétaux en présence d' détectés conditions anormalement sèches et de la comparer avec le développement de la végétation dans les périodes avec des précipitations normales avec un intérêt particulier sur les terres agricoles pluviales . L'analyse des impacts de la sécheresse sur la végétation a été réalisée avec l'utilisation des techniques de télédétection . Un protocole original 11 pour effectuer l'utilisation des terres de la couverture terrestre (OUS ) classification qui combine des informations climatiques, topographiques et de réflexion de 18 ETM + Landsat images ont été appliqués à distinguer la suite des effets de la sécheresse dans les différentes classes dans les années sélectionnées. La classification LULC (en anglais) atteint produite précision de la classification globale allant de 87,88 % à 92,42 %. Indices spectraux pour la végétation et humidité du sol ont été ensuite utilisés pour détecter des anomalies dans le développement de la végétation causée par la sécheresse et, en outre la région des masses d'eau a été mesurée et comparée à détecter les changements dans la disponibilité en eau pour les cultures irriguées . La méthodologie proposée a été testée dans le centre du Mexique pour déterminer la période à laquelle la saison de croissance pourrait être utilisé pour détecter l'évolution de la sécheresse (dans ce cas, l'année 2005) et qui pourrait être appliqué à d'autres régions semi-arides . 12

Chapter 1

1.1 Introduction

Globally, droughts are one of the most devastating natural hazards, affecting millions of people and inflicting widespread economic consequences (Wilhite 2007). On a global scale, the direct economic costs of drought have been estimated to have been almost US$80 billion over most of the 20th century (EM-DAT 2009). Economic and social development have historically been strongly linked to the development of water resources, and the breakdown of this connection might have contributed to the fall of great civilizations (Sheefield and Wood 2011).

The impacts of droughts are commonly classified in economic, environmental, and social terms.

The early 1970s were a turning point in global awareness about the need to better understand the drought phenomenon, its causes and consequences, and to develop mitigation strategies to cope with its consequences (Glantz 1994).

Crop failures and social suffering caused by droughts are often blamed only on rainfall deficits, while in reality the causes are more complex (Johan, 2003). As with all natural hazards, the economic impacts of drought vary within the sectors and geographic regions and produce a complex web of impacts. This complexity is largely caused by the dependence of so many sectors on water for producing goods and providing services (Wilhite, 2000). The risk droughts pose to society and ecosystems depend on a combination of its physical nature and the degree to which a population or activity is vulnerable to the effects of periods of atypical water shortage.

So what exactly is a drought? Drought is a silent and pervasive phenomenon, it creeps up over weeks, months, and even years often without any warning, affecting profoundly ecosystems 13 and human activity on a global scale. On many occasions, it is perceived too late, when the after effects of falling crops, dying livestock, fried-up fiver beds, and severe economic and social damage has already been presented. Thus, the simplest definition of drought is a deficit in water relative to normal circumstances in any particular system. However, one characteristic of drought that distinguishes it from other natural hazards its lack of a universal definition. Hundreds of definitions exist, largely because drought must be defined according to the characteristics of each climatic regime and the specific impact sector or application to which the definition is being applied (Wilhite 2007). Differences in hydro-meteorological variables and socioeconomic factors as well as the stochastic nature of water demands in different regions around the world have become an obstacle to having a precise definition of drought (Mishra and Singh 2010, 2011).

Yevjevich (1967) stated that widely diverse views of drought definitions are one of the principal obstacles to investigations of droughts. Therefore, droughts have been classified depending on different fields of study that have intended to analyze their nature.

In the assessment of droughts, freshwater planning and management is of primary importance. This requires understanding historical droughts in the region as well as impacts of droughts during its occurrences. Droughts have had great impacts in the 20th century. The occurrence of severe droughts throughout Africa and India, North America, China, the former

Soviet Union, Australia, and Western Europe underscored the vulnerability of developed and developing societies to drought (Glantz 1994).

During the last two decades the impacts, recurrence and severity of droughts in North

America have increased significantly (Wilhite 2007). Over the years 1980 to 2003, in the United

States as a whole, droughts accounted for 10 of the 58 weather-related disasters (Mishra and 14

Singh 2010). In Canada, most regions have experienced droughts; however, the Canadian

Prairies are more susceptible mainly due to their high variability of precipitation in both time and space (Mishra and Singh 2010). Over much of the Prairies, several consecutive seasons of below average precipitation have led to one of the most severe prairie droughts on record, devastating many water dependent activities in 2001 and 2002 (Mishra and Singh 2010).

In the Mexican context, it is agricultural drought that poses the greatest threat to smallholder self-sufficiency rainfed agricultural producers, who are the most vulnerable sector of the Mexican society. Agricultural drought, usually, refers to a period with declining soil moisture and consequent crop failure without any reference to surface water resources (Mishra and Singh

2010). Thus, drought vulnerability can be determined by a combination of variables, including the biophysical context, the condition of human settlements, infrastructure, public policy and administration, organizational abilities, social inequalities, gender relations, and economic patterns (Shahid and Behrawan 2008). Clearly, biophysical processes refer to the ecological or natural systems that characterize a geographical context, while the social systems refer to how humans interact with their activities and institutions in such environments. In human ecosystems, there is such a direct influence of both social and ecological processes that attempting to study these systems and ignoring these interactions, the results would offer a very limited vision of reality. According to Liverman (1999), biophysical vulnerability to drought is greatest in the northern and central regions of the country, where rainfall is most variable, and the timing of rainfall and the extent of the mid-summer drought are critical. For a country with a millennial agricultural tradition, Mexico’s climate is not particularly favorable for agricultural production.

Thus, as much as 46% of Mexico’s land area is classified as arid, which has the effect of limiting 15 agricultural land use to grazing or irrigated farming (Reyes-Castañeda 1981). Mexico’s agriculture is also limited by large areas where the poor quality of the soil, paucity of water resources, and complex topography makes only rainfed agriculture viable (Eakin 2000).

On the other hand, social vulnerability to drought in Mexico varies greatly by region and social group, and has altered over time as a result of technological, economic and demographic changes (Liverman 1999). Indigenous and non-indigenous or mestizo farmers have lived in the most precarious conditions since the Spanish colonization in the 16th until today; constantly harassed by discrimination or ignored in isolated conditions and marked by historical massive expropriations of land. However, in one of the most comprehensive empirical studies of agro ecology of small scale Mexican farmers (Wilken 1987), it has been documented that farmers use numerous micro-ecological adjustments to control environmental variability, including application of mulches, alterations in tillage practices, and sophisticated management of trees and shrubs as shade barriers. Using traditional and empirical agricultural techniques and systems, farmers have been able to reduce the impacts of drought (Eakin 2000).

In January 2012, the Secretary of Social Development in Mexico declared a state of emergency due to the most severe drought in 71 years. This drought devastated croplands in nearly half of the country and left two million people without access to water (Rodriguez 2012;

Zabludovsy 2012). NASA’s Earth Observatory reported that the 2011-2012 drought clearly slowed plant growth in both United States and Mexico, and declared that the 2010-2011 La Niña likely contributed to the drought’s severity (NASA Earth Observatory 2012). The most affected areas were reported in the northern and central territories. According to Romero-Polanco (2012), an estimated 1.4 million hectares suffered damage from adverse weather conditions and lost 3.2 16 million tons of maize, 600,000 tons of bean, and 60, 000 livestock died. Hundreds of farmers who had lost everything, marched to Mexico City to draw the attention of the government on the severity of this extreme environmental event (Jornada 2012; Prado 2012).

In Mexico, drought monitoring at the national scale is currently performed by North

America Drought Monitor (NADM), which is a cooperative effort between drought experts in the United States, Mexico, and Canada. The NADM program was initiated in April 2002 and is part of a larger effort to improve the monitoring of climate extremes on the continent. The data produced by NADM focuses only on three data driven indexes: the SPI, the PDSI, and the

Percent of Average Precipitation. Other studies in Mexico have reported qualitative drought impacts documenting smallholder rainfed maize production and climatic risk (Eakin 2000), climate change impacts on food security (Appendini and Liverman 1994), estimates of the effects of El Niño Southern Oscilation (ENSO) on crop yield (Adams et al. 2003) and vulnerability and adaptation to drought in (Liverman 1990). However, very little has been done in the region in terms of using remotely sensed data in combination with meteorological analysis to estimate the effects of drought on rainfed agricultural production.

Given the adverse consequences of drought, and the vulnerability of rainfed agriculture in semi-arid developing regions, this study will attempt to detect drought effects on agricultural areas in Central Mexico. The results have the potential to be used as background information for decision makers such as the Water Council of the State of Hidalgo (CEAAH in Spanish) or the

Water National Commission (CONAGUA in Spanish) to identify areas that are more prone to be vulnerable to droughts because of the meteorological and environmental context they are in. The present study aimed to develop an original combination of methods to detect agricultural drought 17 in Central Mexico. The approach was made through (a) a meteorological analysis of a long time series of monthly precipitation data to find dry and normal periods of time to (b) use remote sensing techniques to develop Land Use Land Cover (LULC) classification then (c) calculate vegetation condition and soil/vegetation moisture indexes to monitor the monthly effects of droughts at regional scales related to agriculture in developing regions. Finally, (d) change detection was performed for each class individually and for each of the four indexes previously calculated.

1.2 Objectives

The objective of this study was to monitor the behavior of different vegetative covers in the event of abnormally dry conditions with a particular interest on rainfed agricultural lands in

Central Mexico. To do so, an original combination of methods was settled after a profound literature review. This was done through a meteorological analysis, and subsequently through the use remote sensing techniques to detect its effects on vegetation. In this section, a discussion of the reviewed methods is presented to further justify the selected methodology. Specific objectives include:

1. Use monthly precipitation time series since 1980 to derive the Standard Precipitation

Index and find three growing seasons that are characterized by: a) one year with

normal precipitation, b) one year with drought conditions, and c) the 2011-2012

growing season that was reported to be one of the most severe droughts in Mexico’s

recent history.

2. Develop accurate Land Use Land Cover (LULC) multitemporal maps using Landsat

images to rainfed agriculture. The accurate delineation of agricultural lands and 18

specifically where rainfed farming production is practiced was of great importance.

Saadat et al. (2011) original protocol for LULC classification was used. This protocol

was designed for classifying large areas using single date Landsat ETM+ images and

incorporating ancillary topographic and climatic data (Saadat et al. 2011). This

methodology has proven to be effective in classifying semiarid landscapes in Iran

with very high accuracy. Ground truth data consisting of samples taken in the field

were collected to test the accuracy of the LULC maps. This methodology has never

been used in the past for monitoring drought.

3. Use satellite imagery to process spectral indices to analyze photosynthetic activity

and therefore vegetation vigor and soil/vegetation moisture. Several indices have

been developed in the past to achieve these tasks.

6. The previously mentioned methods were tested in the Tortugas-Tepezata watershed

located in the higher Moctezuma Watershed in the State of Hidalgo Mexico.

1.3 Thesis Outline

The present thesis was written in 5 chapters. The first chapter had the purpose of introducing the general concepts of drought and with particular interest of its effects on rainfed agricultural lands in the Mexican context and finally stating the specific objectives of the present work.

The second chapter reviewed and discussed the current and relevant approaches in the literature that have studied drought in the past. This literature review had the purpose of presenting descriptions of several techniques used in the past to quantify drought effects. First, some definitions of drought were presented. A description of several techniques to quantify 19 disturbance in environments was described. Next, a discussion of the social components that induce social vulnerability was presented to justify the necessity of studying drought in semiarid developing regions. In that subsection, a summary of the historic Mexican rural context was presented with the intention of reviewing complex social processes that should not be ignored when studying the environment. Finally, the chapter finishes with a selection of the methods that were used for this study.

The third chapter is the first attempt to detect agricultural drought in central Mexico. This chapter aimed to develop an original combination of methods to detect drought through i) a meteorological analysis of a long time series of monthly precipitation data to find dry and normal periods of time to ii) use remote sensing techniques to develop Land Use Land Cover (LULC) classification and then iii) calculate two vegetation condition indices to monitor the monthly effects of droughts in regional scales related to agriculture in developing regions, and analyze trends in vegetation conditions. Finally iv) change detection was performed for each vegetation class individually and for each of the four indexes previously calculated.

In this chapter, a simplified protocol of the coauthors of this study Saadat et al. (2011) is presented. It was used for classifying large areas using 18 Landsat ETM+ images and incorporating ancillary topographic and climatic data (Saadat et al. 2011). Finally the

Normalized Vegetation Index (NDVI) and the Vegetation Condition Index (VCI) were processed to detect vegetation conditions on a monthly basis of the three years 2000, 2005, and 2011. The

LULC classes were used to extract the resulting processed layers to discern the response of the different vegetation covers.

Chapter 4 had the objective of going deeper into the accuracy of LULC maps 20 corresponding to months that in Chapter 3 were found to be key to the study of agricultural drought for the time period selected. Four images were used in the LULC methodology that produced overall accuracies ranging from 87.88% to 92.42%. Spectral indices for vegetation and soil/vegetation moisture were used to detect anomalies in vegetation development caused by drought and furthermore, the area of water bodies was measured and compared to detect changes in water availability for irrigated crops. Finally change detection was completed. Finally, Chapter

5 is a discussion and the conclusion of the study that finished by highlighting the contributions of this study to the science of drought. 21

Chapter 2

Literature Review and Selection of Methods

This chapter reviews and discusses the current relevant approaches that have studied environmental drought in the past. First, some definitions of drought are presented. Then a description of several techniques to quantify the disturbance in the environments are described.

Next, a discussion of the social components that induce social vulnerability was presented to justify the necessity of studying drought in semiarid developing regions. In that subsection, a complete summary of the historic Mexican rural context is presented with the intention of reviewing complex social processes that should not be ignored when studying the environment.

Finally, the selection of methods is presented that was used to base a new methodology to study drought using geographic information systems and remote sensing.

2.1 What is a Drought?

The hydrological cycle is a series of inflows (e.g., precipitation and soil moisture) and outflows (e.g., evapotranspiration and drainage) that are balanced by a set of stores the hydrological water balance. The distribution of stores, inflows, and outflows in the water cycle vary through space and time and they are shaped by the weather, the climate, and the characteristics of the land surface (elevation, slope, vegetation, land use, and water bodies). The basic spatial distribution of climate is a result of the general circulation of the atmosphere, which represents the major flows of water and energy around the world (Sheffield and Wood 2011). As well as varying spatially, the climate varies across multiple timescales, from hourly, to daily, to seasonal, to centuries and longer spans. The climate responds naturally to external factors, 22 forcing such as solar radiation, and atmospheric composition. Over time, if the inflows become less than the outflows, the store will decrease, resulting in a water deficit better known as a drought (Sheffield et al. 2011).

Every geographical context in the planet has diverse meteorological and socioeconomic variables and different water demands, and for this reason many authors have agreed that it has been impossible to have a universal definition of drought. However adopting a general definition as a reference is important to have a frame of reference.

The following general definition of drought was adopted in this study because it highlights the following: i) drought is a transient phenomenon, ii) it is difficult to predict due to its nature, iii) the way it affects ecosystems, and iv) this definition can be applied to any context.

2.2 What Causes Drought?

The onset and the end of a drought are very difficult to determine, the impacts increase slowly and often accumulate over a period of time. The impacts of droughts are non-structural, spread over large areas, and can be directly triggered by human activities (Mishra et al. 2010).

Drought results due to extremes in climate. It may also be exacerbated or dampened by anthropogenic influences (groundwater over pumping, diverting rivers or cultivating agricultural land). Droughts have been studied in detail regionally and globally in the past and much more in the last decade in terms of their occurrence and forcing mechanisms (Sheffield and Wood 2011).

The variation in climate and a spectrum of temporal and spatial scales is the main driver of drought. Typically, droughts are initiated by atmospheric circulation and weather systems that conspire to cause lower precipitation and/or higher evaporation than normal in a region

(Sheffield and Wood 2011). Droughts occur when the usual pattern of weather changes its 23 seasonal timing, its location, or persists for longer than normal. Thus, there is a chain of processes, ranging from the multiple-year timescales and large space scales, with the ocean being the main driver to regional and local scales, where seasonal and daily weather processes and land features play major roles.

Although droughts are generally driven by a period of low precipitation, there are other factors that modulate their occurrence, depending on the characteristics of the region and how droughts propagate through the hydrological system (van Lanen et al. 2004). These characteristics include the endemic climate, soil type, and stratigraphy, elevation and slope, vegetation cover and physiology, the groundwater system, and its connections with surface hydrology and neighboring regions.

Sheffield and Wood (2011) argue that drought generally has its initial stage, which is meteorologically driven by low precipitation, but can also be initiated by atmospheric demand through increased evapotranspiration. Low precipitation will impact in different ways, depending on the season, antecedent conditions and characteristics of the region. This is the linked to the storage of water in the soil column. A decrease in precipitation will lead to a decrease in soil moisture. This will reduce evapotranspiration, with subsequent impacts on the vegetation leading to ecological drought. If this occurs during the dry season, then recharge to groundwater will be small and hydrological drought will not emerge. However, if the depletion of soil moisture is high, it will need to be restored in the wet season, thereby reducing subsurface flow and recharge.

High potential evapotranspiration can also cause a reduction in soil moisture. Potential evapotranspiration is the maximum possible level of evapotranspiration that can be reached, 24 given unlimited water supply and current atmospheric conditions. Higher temperatures and radiation, as well as stronger and drier winds, can increase the effects.

If the climate is humid, the reduction may not be significant to plants, but may eventually reach wilting point (the amount of water below which a plant starts to wilt). In dry regions, the recharge is low anyway and the reduction in soil moisture does not matter. The reduction in soil moisture is not indefinite, as the soil moisture store will eventually deplete. There are also feedback mechanisms with the vegetation that may shut down transpiration if soil moisture becomes too low, although the vegetation may not be affected in the long term. If low soil moisture persists near the wilting point for a period of time, then the vegetation may wither and dry. This depends on the type of vegetation and its strategy for coping with drought.

2.3 Classification of Droughts

In the literature, there is no consensus on one definition of drought; however, the phenomenon has been classified and studied depending on the system it disturbs. Mishra et al.

(2010, 2011) developed a review of drought concepts and models and described five types of drought in the literature:

1. Meteorological drought occurs with a lack of precipitation over a region for a period

of time. This classification approach is based on considering precipitation deficit with

respect to average values.

2. Hydrological drought appears after a period of inadequate surface and subsurface

water resources for established uses of a given water resource management system.

3. Socio-economic drought occurs when the demand of an economic good exceeds

supply as a result of a weather-related shortfall in water supply. 25

4. Groundwater drought: When groundwater systems are affected by droughts, the direct

effects are lower groundwater heads and a decrease in groundwater flow to riparian

areas, springs, and streams.

5. Agricultural drought occurs when there is not enough availiable plant soil moisture in

the root zone (Johan 2003). The capacity of soils to retain and release water depends

on factors such as soil texture, depth, structure, organic matter content, and biological

activity (Bot 2005). Agricultural drought is generally characterized by two key

factors: the estimated water demand and expected water supply. The formulation of

these key factors for a region largely depends on the agro-climatic conditions (Yurekli

and Kurunc 2006). Land degradation reduces the capacity of agricultural systems to

absorb environmental shocks. Land degradation itself is related with population

growth, poverty, weak land polices and other social variables (Johan 2003).

2.4 Overview of Drought Indices

Several drought indices have been developed in recent decades with the purpose of detecting drought impacts that quantify intensity, duration, severity, and spatial extent. Each of these indices is able to quantify drought in different time scales for which time series analysis is essential (Mishra and Singh 2010). In this section, we will describe drought indices found in the literature are described along their proposed methods for drought detection and some specific results of different studies. These indices are then compared.

2.4.1 Data driven drought indices

The following section discusses commonly used data driven drought indices. Data driven drought indices use meteorological variables to detect and predict periods of abnormal water 26 deficit in the hydrological cycle. Such indices use precipitation, either as the only input or in combination with other meteorological elements like temperature and evapotranspiration.

Palmer Drought Severity Index (PDSI)

Developed by Palmer (1965) with the main variables for defining a water budget including precipitation, evapotranspiration, runoff, and soil water content, which can be calculated by the water balance method or a hydrological model (Xiaofan 2012). PDSI is the most widely used index of meteorological drought in the United States (Tang and Piechota

2009). The advantage of this drought index is that it can be computed for different scales. For this reason, it has been widely used and adopted by US agencies as a drought trigger. However, this drought index presents some disadvantages, since values may lag behind emerging droughts by several months, and it does not handle frequent climatic extremes well (Sheffield and Wood

2011).

Standardized Precipitation Index (SPI)

The Standardized Precipitation Index is based on the long-term precipitation record for a desired period. This record is then fit into a probability distribution, then transformed to a normal distribution, so that the mean SPI for the location and desired period is zero (McKee 1993). In the literature, there are many approaches when calculating the SPI, and many different probability distributions have been used with different purposes. The SPI may be computed on shorter or longer time scales, which reflect different lags in the response of water cycle precipitation anomalies. When the time scale is short, for instance 3 months, the SPI reflects the seasonality of the data and is more appropriate to identify drought impacts on agriculture. As the time scale increases, the SPI responds more slowly to changes in precipitation, and results for the 27

12-month time scale identify dry periods of long duration, which relate to the global impacts of droughts on hydraulic regimes and water resources of a given region (Moreira et al. 2006,

Moreira et al. 2008). SPI is a very flexible drought index, since it can be computed for different timescales. SPI has been adopted by the World Meteorological Organization (WMO) as a standard meteorological index (Sheffield and Wood 2011). The wide acceptance, flexibility, and capacity of this index to reflect drought impacts on agriculture makes it very attractive to use for the purposes of this study. Furthermore, the National Drought Mitigation Center (NDMC) in the

United States has developed a downloadable pre compiled program to calculate SPI derived from monthly precipitation data for long periods of time (30 minimum years is recommended by the

NDMC). This program significantly facilitates obtaining SPI results and gives the user the opportunity to use such results for different time scales (3 months, 6 months, 9 months, and 12 months) depending on the desired time scale of drought effects.

Crop Moisture Index (CMI)

If soil water availability is reduced due to climatic conditions, vegetation and soil properties play a substantial role in triggering the occurrence of drought (Zierl 2001). CMI is used to evaluate short-term moisture conditions (week to week).

The Crop Moisture Index principles have been used in models such as the WAWAHAMO

(WaldWasserHaushalts-Model, 2001). This model computes the main components of the water balance aiming to quantify drought stress of forest stands. On a daily basis, it predicts soil moisture content, evaporation and transpiration, interception, snow cover, drainage, radiation budget, and phenology assuming a fixed bucket size defined by the available soil water capacity and requiring only a moderate set of input data (Zierl 2001). This drought index can be easily 28 computed from precipitation and temperature data; however, it is not applicable for long term droughts and suffers from the same general disadvantages as the PDSI (Sheffield and Wood

2011).

PDSI vs. SPI

There has been much comparison between SPI and PDSI for monitoring droughts (e.g.,

Mishra and Singh 2010). Some additional differences include i) spatial characteristics of PDSI vary from site to site while those of SPI do not, ii) SPI is more representative of short-term precipitation than PDSI and thus is a better indicator of soil moisture variation and soil wetness

(Sims et al. 2002), iii) SPI is a better predictor of crop production, as it represents the moisture state of soil better (Quiring and Papakyriakou 2003), iv) SPI provides a better spatial standardization than PDSI for extreme drought effects, and v) SPI detects the onset of a drought earlier than PDSI (Hayes et al., 1999).

2.4.2 Drought indices derived from Satellite Data.

Since the 1970s, several studies have used satellite imagery to monitor vegetation dynamics over growing seasons. Satellite remote sensing provides a synoptic view of continuous data with improved information on monitoring vegetation dynamics over large areas. Remote sensing of vegetation is accomplished by using the strong coupling between reflected visible and near infrared radiation with the physiological condition of leaves and their density (Anyamba

2005).

Normalized Difference Vegetation Index (NDVI)

There are many examples in the literature of vegetation monitoring using Normalized

Difference Vegetation Index (NDVI) derived from Advanced Very High Resolution Radiometer 29

(AVHRR) used to detect vegetation response to short-term weather conditions (Peters et al.,

2002). NDVI is based on the fact that healthy vegetation has a low reflectance in the visible portion of the electromagnetic spectrum due to absorption by chlorophyll and other pigments and high reflectance in the Near Infrared (NIR) because of the internal reflectance by the mesophyll spongy tissue of a green leaf (Campbell, 1987). NDVI is calculated as the ratio of the red (RED) and the Near Infrared (NIR) bands of a sensor system and is represented by the following equation:

[1]

NDVI values range from -1 to +1. Because of high reflectance in the NIR portion of the electromagnetic spectrum, healthy vegetation is represented by high NDVI values between 0.05 and 1. Higher NDVI indicates a greater level of photosynthetic activity (Sellers 1985; Tucker et al. 1991). Conversely, non-vegetated surfaces such as water bodies yield negative values of

NDVI. Bare soil areas represent NDVI values close to 0 due to higher reflectance in both the visible and the NIR portions of the electromagnetic spectrum (Lillesand and Kiefer 1994;

Rahimzadeh Bajgiran, Darvishsefat, Khalili, and Makhdoum 2008).

The NDVI has been used successfully to identify stressed and damaged crops and pastures, but interpretive problems arise when results are extrapolated over non-homogeneous areas. It can be said that NDVI has two components: ecology and weather. Therefore, when

NDVI is used for the analysis of weather impact on vegetation, the weather component must be separated from the ecosystem component (Rahimzadeh Bajgiran et al. 2008). It has also been found that NDVI correlates with net primary production, biomass, vegetation fraction, and yield 30

(Goward et al. 1987; Maselli et al. 1992; Rasmussen 1992; Quarmby et al. 1993; Hayes and

Decler 1996; Korgan 1997; Unganai and Korgan 1998).

Vegetation Condition Index (VCI)

Since the 1970s, several studies have used satellite imagery to monitor a variety of dynamic land surface processes (Peters et al. 1993; Reed 1993; Mishra and Singh 2010). Satellite remote sensing provides a synoptic view of the land and a spatial context for measuring drought impacts, which have proved to be a valuable source of information on monitoring vegetation dynamics over large areas. The vegetation condition index is computed from satellite advanced very high resolution radiometer (AVHRR) radiance (visible and near infrared) data, and is primarily useful for the summer growing season. The VCI was developed by Kogan (1995) and has been used to estimate the weather impact on vegetation. The weather-related NDVI envelope is linearly scaled to 0 for minimum NDVI and 1 for the maximum for each grid cell. The VCI varies from 0 to 1, corresponding to changes in vegetation condition from extremely unfavorable to optimal. NOAA-AVHRR (Advanced Very High Resolution Radiometer) data has been extensively used for drought monitoring in various areas of the world and good correlations have been established between NDVI or VCI and precipitation (Karabulut 2003; Wang et al. 2003;

Bayarjargal et al. 2006; Bhuiyan, Singh, and Kogan 2006; Hartmann, Di Bella, and Oricchio

2003; Amin Owrangi 2011). The VCI allows for the detection of drought and measurement of the time of its onset and its intensity, duration, and impact on vegetation. However, since VCI is based on vegetation, it is primarily useful for the summer growing season. It has limited utility for the cold season when vegetation is largely dormant (Mishra and Singh 2010). The VCI 31 involves the detection of drought and measurement of the time of its onset and its intensity, duration, and impact on vegetation. The VCI is calculated as follows (Kogan 1990):

[2]

The Normalized Difference Water Index (NDWI)

The normalized difference water index (NDWI) is a more recent satellite-derived index from the NIR and short wave infrared (SWIR) channels that reflect changes in both the water content and spongy mesophyll in vegetation canopies. NDWI is calculated as follows (Gao

1996):

[3]

Where NIR is the Near Infrared band and MidIR is the Mid Infrared Band.

Because NDWI is influenced by both desiccation and wilting of vegetative canopy, it may be a more sensitive indicator than NDVI for drought monitoring (Mishra and Singh 2010).

NDVI and NDWI sense similar depths through vegetation canopies. However, NDWI is less sensitive to atmospheric effects than NDVI. NDWI does not remove completely the background soil reflectance effects, similar to NDVI because the information about vegetation canopies contained in the 1.24µm channel is very different from that contained in the red channel (Gao

1996). For this reason, NDWI should be considered complementary but not a substitute for

NDVI.

NDWI was calculated for the 18 Landsat ETM+ images. Then, in the same way as previously done for NDVI, the NDWI output layers where divided by the LULC classes; in this way each class could be analyzed separately and a NDWI time series per class was plotted with 32 the purpose of discerning the different NDWI values per class in the different stages of the growing seasons in the two dry and normal years.

Gu et al. (2007 and 2008) evaluated the relationship between NDVI and NDWI using

MODIS 500-m satellite imagery in Oklahoma and Kansas, USA. They found good correlation between the two indices and furthermore NDWI was found to be more sensitive to drought than

NDVI. Rhee et al. (2010) identified a drought index that had the possibility to be used for agricultural drought monitoring in arid/semi-arid regions as well as humid/sub-humid regions

(also using MODIS data). They found that NDWI has a better response in arid/semiarid regions than in humid regions when detecting drought because of good correlations with other precipitation and temperature indices. It is convenient to detect drought at larger scales using coarse resolution imagery such as MODIS; however, this approach cannot provide detailed land cover response to precipitation anomalies.

Tasseled Cap Transformation Wetness (TCW)

Numerous methods have been developed for transforming available information from multispectral sensors to deriving features to interpret characteristics of the land surface. Such methods include the three indexes previously described that are based on ratio and differences of bands. The Tasseled Cap Transformation of Landsat Multispectral Scanner and Thematic mapper

(Kauth and Thomas 1976; Crist and Cicone 1984) offer a way to optimize data viewing for vegetation studies.

The different bands in a multispectral image can be visualized as defining an N- dimensional space where N is the number of bands. Each pixel, positioned according to its data file value in each band, lies within the N-dimensional space. This pixel distribution is determined 33 by the absorption/reflection spectra of the imaged material. For viewing purposes, it is advantageous to rotate the N-dimensional space such that one or two of the data structure axes are aligned with the View X and Y axes. In particular, the axes that are largest for the data structure produced by the absorption peaks of special interest for the application (Crist and Kauth

1986). Research has produced three data structure axes that define the vegetation information content. This option can show these three axes (or layers) as a degree of brightness, greenness, and wetness, as calculated by the Tasseled Cap coefficients used. Layer 1 (red) outputs the brightness component and indicates areas of low vegetation and high reflectors, layer 2 (green) is the greenness component and indicates vegetation status, and finally layer 3 (blue) is the wetness component that indicates water and moisture in the scene.

The Tasseled Cap Transformation Wetness (TCW) was used to determine the amount of moisture being held by the vegetation or soil, thus termed wetness pointing to the vegetation and the brightness of soil (Fadhil 2009). TCW images were derived from ETM images of the study area using the tasseled cap transformation algorithm with ER Mapper according to the following equation (Jin and Sader 2005):

Where B, G, R, NIR, SWIR, and SWIR2 are the Landsat ETM+ bands excluding the thermal bands and the panchromatic band.

The tasseled cap transformation (TCT) has been used widely for vegetation mapping and monitoring land cover change (Oeyyer et al. 2001; Junb and Sader 2005; Fadhil 2011). The TCT of Landsat thematic mapper consists of six multispectral features, all of which could be potentially differentiated in terms of stability and change in a multitemporal data set (Jin, 2004). 34

The first three features have been labeled brightness, greenness, and wetness (band 1, 2, and 3, respectively). The third feature, wetness, has been shown to be sensitive to soil plant moisture

(Jin and Sader 2004). Tasseled cap wetness (TCW) contrasts the sum of the visible and the near- infrared bands with the sum of the shortwave bands (Jin and Sader 2004).

The NDWI and TCW have both been used in studies to detect drought or disturbance in ecosystems. Jin and Sader (2004) used a time series of both NDWI and TCW derived from

Landsat to compare forest disturbances in Maine, USA. They found high correlations (>0.95) between the two indices. Fadhil (2011) used NDWI and TCW to detect drought effects on vegetation in the Iraqi Kurdistan region. He derived both spectral indices from two Landsat images from consecutive years to calculate five vegetation and soil/vegetation moisture indices and perform change detection. This study did not include a classification of the vegetation so it is unclear how different kinds of vegetation were affected by the drought.

Temperature Condition Index (TCI)

To remove the effects of cloud contamination in the satellite assessment of vegetation condition, Kogan (1995) proposed the Temperature Condition Index (TCI). TCI is derived from

BT and reflects the opposite of the NDVI vegetation’s response to temperature (high temperature is not favorable for vegetation).

VCI and TCI have been shown to have good correlation with corn in semiarid areas

(Unganai and Kogan 1998). However, VCI allows for the detection of drought, and measurement of the time of its onset and its intensity, duration, and impact on vegetation (Mishra and Singh

2010). Furthermore, VCI is based on vegetation, it is very useful for the summer growing season which is of interest to this study. 35

Vegetation Health Index (VHI)

The Vegetation Health Index (VHI) was developed by Kogan (1995, 1997) who successfully applied it in many different environmental conditions. VHI is a composite index joining the VCI and the Temperature Condition Index (TCI). The TCI algorithm is similar to VCI but relates to the brightness temperature estimated from the thermal infrared band, channel 5 of

AVHRR (Rojas, Vrieling, and Rembold 2011). This index was proposed to remove the effects of cloud contamination in the satellite assessment of vegetation condition due to the fact that the

AVHRR channel 4 is less sensitive to water vapor in the atmosphere compared with the visible light channels. It is the additive combination of VCI and TCI for week i. In some studies different weights (w1 and w2) are assigned to VCI and TCI (Rojas et al. 2011).

Standardized Vegetation Index (SVI)

Peters et al. (1993, 2003) based their study on transforming a probability of an observed greenness into an index using AVHRR 1-km NDVI biweekly data to monitor areas of drought/ vegetation conditions called “Standardized Vegetation Index” (SVI). The SVI allows for the visualization of “greenness probability” for each 1-km/2 pixel location through the use of a probability estimate. Their research began with 14-day maximum value NDVI composite images. The SVI is based on calculation of a “Z” score for each AVHRR pixel location. The Z score is a deviation from the mean in units of the standard deviation, calculated from the NDVI values for each pixel for each week during 12 years. This per pixel probability is an estimate of the probability of occurrence of the present vegetation condition. SVI’s were grouped into 5 classes. They found that SVI maps can be a very useful tool that is capable of proving near real 36 time indicators of the onset, extent, intensity and duration of vegetation stress at a resolution of

1km.

2.5 Comparing Drought Indices

The following chart compares the different popular drought indices, highlighting advantages and disadvantages of every index. 37

Drive Description Advantages Disadvantages Precipitation Calculated as the actual Simple and effective for The mathematical meaning of percentage of precipitation divided by the single locations and seasons. normal is different to the general normal average annual value. concept of normal weather. Rainfall Declines Stratifies precipitation into Gives an accurate statistical Long records of data are required to deciles. measure of precipitation. obtain accurate values. Standardized Based on the probability of Can be computed for Values may change as new data are Precipitation Index precipitation. different timescales. incorporated. Reflects only (SPI) Adopted as the World precipitation. Meteorological Organization ( W M O ) s t a n d a r d meteorological index. Palmer Drought Calculated as the departure Takes into account Value may lag behind emerging Severity Index of moisture from normal precipitation inputs and droughts by several months; does (PDSI) using a simple water outputs from evaporation not account for snow; does not balance model. Used by and run-off. handle frequent climatic extremes; many US agencies as a complex drought trigger Crop Moisture Reflects short-term Easily computed from Not applicable to long-term drought. Index (CMI) moisture supply for crop p r e c i p i t a t i o n a n d Suffers from the same general regions. Derived from temperature data. disadvantages as PSI. Palmer Indices.

N o r m a l i z e d D i f f e r e n c e b e t w e e n A measure of general Requires data form airborne or space Difference maximum absorption of vegetative condition. borne sensors. Difficult to discern Vegetation Index visible and near infrared Satellite data provide large other influences on vegetative (NDVI) radiation. areal coverage and high health. spatial resolution.

Vegetative condition Normalized version of As for NDVI. As for NDVI. Index (VCI) NDVI NDWI D i f f e r e n c e b e t w e e n A measure of water particles Vegetation growth is dependent maximum absorption of the in vegetation and soil. upon diverse environmental factors near infrared and the mid A better response in arid/that could be misinterpreted such as infrared radiation. semiarid regions. nutrient availability, disease, insect infestation, etc. TCW Wetness component that It has been found to have Sometimes it has been found to i n d i c a t e s w a t e r a n d good correlation with have overestimated the effects of moisture in the scene. NDWI. It is used as a wet conditions and presented weak wetness index because of the correlations with meteorological enhancement of visible and data. infrared bands in Landsat images. 38

2.6 The Use of Land Use Land Cover (LULC) Maps for Detecting the Effects of Droughts

Land use land cover (LULC) classification is one of the most widely used applications in remote sensing. Many approaches have been used to correlate satellite images with vegetation cover. There are several distinct growth periods during a growing season and classification of vegetation characteristics based on the presence or absence of vegetative cover and the condition of that vegetation of the acquired image (Saadat et al. 2011).

Guerschman et al. (2002) explored the use of multi-temporal Landsat TM data from the same growing season for the classification of land cover types in the Argentine Pampas. They recommended the use of at least two scenes for satisfactory classification, and the addition of more images has the potential to improve the capability to separate classes and embrace the shift between winter and summer crops. Oetter et al. (2001) used five images from a single year to successfully characterize 20 classes of agricultural and related land cover in western Oregon.

They found that the major strength of their approach for mapping came from the interpretable spectral information available for multi-seasonal Tasseled Cap imagery, especially for crop development. Lucas et al. (2007) used time series of Landsat images and integrated ancillary datasets for image segmentation and the development and implementation of ecological rules for the classification of vegetation types. They concluded that the use of multi-temporal images allowed for sampling of temporal trends in land cover radiance and NDVI facilitating discrimination of bracken agricultural lands and grasslands at various levels of improvement.

Maxwell et al. (2004) presented a methodology for identifying corn resulting from multi- temporal Landsat imagery. Their approach was to enable automation for the classification 39 approach of likelihood of the vegetation cover to be classified as corn (highly likely corn, likely corn or unlikely corn).

Saadat et al (2011) proposed a new protocol for LULC classification for large areas based on readily available ancillary information and analysis of three single date Landsat ETM+ images, and demonstrated that successful mapping depends on more than just analysis of reflectance values. It was found in this research that incorporating climatic and topographic conditions helped to delineate what was otherwise overlapping information.

This protocol has proven classification steps to accurately classify LULC types. The protocol enhances initial reflectance based classification with the use of image segmentation and supplemental segregation via the use of ancillary information (climatic and landform).

2.7 Why Study Drought? Social Vulnerability to Drought

Understanding how drought affects systems is of primary importance for freshwater planning and management. To successfully implement good management strategies, the use of scientific research is of great value. However, environmental research often excludes social dynamics that can be directly or indirectly linked with society’s ability to respond to environmental hazards. In this section, social dynamics were considered as a background to understand how droughts impact society, and furthermore what have been the historical land use dynamics in the area of interest. This required an understanding of the historical context of the dynamics of social structures, and their interactions with the environment.

The first part of this section reviews in general how droughts pose a threat to societies and how human activities can also trigger droughts. Then the chapter explores in the rural

Mexican context, the importance of maize in Mexico and the evolution of farmer’s conditions 40 since before the Spanish colonization to now. This understanding is essential for future applications of environmental science combined with social studies to implement more integrated approaches for water management and planning for drought prevention.

2.7.1 Social vulnerability to drought

According to Shahid and Behrawan (2008), drought vulnerability is determined by a combination of several factors, including the condition of human settlements, infrastructure, public policy and administration, organizational abilities, social inequalities, gender relations, and economic patterns. By definition, it is different for different individuals (Shahid and

Behrawan 2008). A system could decrease its vulnerability if there is i) a shift in well-being function that decreases the sensitivity to critical stressors, ii) a change in the position relative to a threshold of damage, and iii) a modification in the system’s exposure to stressors of concern

(Luers et al. 2003).

2.7.2 Droughts as a natural hazard

A natural hazard is a threat of a naturally occurring event that will have a negative effect on people or the environment (Mishra and Singh 2011). Drought is slow in onset, insidious in nature, and is often well established before it is recognized as a threat, taking months or years to develop (Pongracz, Bogardi, and Duckstein 1999).

The reasons for the occurrence of droughts are complex, since they depend not only on the atmosphere but also on hydrologic processes, which feed moisture to the atmosphere. The lower the relative humidity, the less probable the rainfall becomes, as it will be harder to reach saturation conditions (Mishra and Singh 2010). Droughts rank first among all natural hazards when measured in terms of number of people affected (Obasi 1994; Hewitt 1997). 41

2.7.3 Droughts and human activity

Throughout history, humans have endeavored to tame water flows to reduce hazards, enable navigation and transport, and store water for regular supply. Some examples of human interference are land-use change, irrigation of farmland, reservoir building, river diversion, wetland drainage, and groundwater extraction. These activities may be able to reduce the occurrence of drought, as in the case of reservoirs, but they may also inadvertently increase the frequency of droughts.

Land-use change has been significant in many parts of the world, especially for the purpose of cultivating crops, driven by the demands of a growing population and more recently the expansion of urban areas (Foley et al., 2005). Deforestation of rain forests, conversion of grassland to agriculture, and urbanization have all changed the landscape hydrology. The impact of the changes are direct, in terms of the changes to the interception of precipitation by the vegetation itself, and the related impacts of soil infiltration, and overland flow.

Urbanization can result in reduced evapotranspiration, increased overland flow, and quick flow through the storm-water drainage network. Many of these changes lead to higher run-off rates and flooding (Villarini et al., 2009), but at the same time reduce infiltration into the soil and recharge into groundwater, which increases the potential for drought.

The largest direct human impact on drought most likely stems from the engineering of water resources. These activities include diversion river courses, drainage of lakes and wetlands, construction of reservoirs and dams, pumping of ground water, and irrigation of dry lands.

Although these changes are often implemented to relieve drought and provide year-round access 42 to water, they can have a considerable impact on remote downstream drought regimes, especially where no consideration is given to ecosystems.

2.7.4 The importance of maize in Mexico.

Millennial tradition is profoundly impregnated in the practice of agriculture in Mexico; from the production systems known as the milpa1 system, to the vast consumption of maize, which was the foundation of pre-Hispanic culture and Mexican cuisine. The Mesoamericans believed themselves to have been molded from maize and in turn they created innumerable varieties and uses that are still objects of study and admiration (Barkin, 2002). These ancient civilizations had domesticated a perennial plant to produce tortillas in a wide variety of colors and shapes; maize had become not only the foundation of the pre-colonial empire’s economies but also the basis of their culture and religion. Today, maize is still a staple of the Mexican diet and its role in society has become more complex than ever due to globalization and neoliberal policies that have profoundly affected agricultural production.

2.7.5 Overview of land tenure processes in Mexico

The agricultural system of pre-colonial Mexico was based on a land tenure institution, which consisted on a mixture of family-allocated agricultural and living units and commonly held lands. Indigenous land tenure did not distinguish between private and public goods

(Herrera-Rodriguez, 2011). This cosmology of common property has prevailed until present time and shaped the land tenure policies through the history of rural areas in the country. The tenancy, in which rural villagers continue to support their traditional and social productive organizations

1 The milpa is complex agriculture production system where a large variety of associated crops (corn, beans, squash and by products like huitlacoche which is an edible fungus that grows in corn). The milpa usually requires a high degree of cooperation among community members for the construction and maintenance of rods, agricultural terraces and water management, for clearing fields and harvesting the crops (Barkin, 2002). 43 with their own resources, is evidence of the currency of their unique vision of society (Barkin,

2002). In 2007, the Mexican National Institute of Geography, Statistics and Informatics (INEGI) reported that 68.5% of all productive units in the country are communally owned ‘ejidos’.

However, for this to become reality, a long and complex process with not so beneficial results had to take place before poor farmers were again granted land.

Farmers have been abused, ignored, and expelled from their lands for centuries in Mexico since the Spanish colonization in 1521 (Herrera-Rodriguez 2012). Independence from Spain in

1810 was followed by ideas of modernization in the nineteenth century that took positivism and capitalism from Europe as their philosophic and economic basis. This ideology promoted systematic campaigns for destroying indigenous communities (who managed their land communally) with the belief of promoting ‘progress’. Communally held indigenous lands were seen as backwards and as obstacles for economic development. Further laws enacted by liberal presidents and especially during the dictatorship of Porfirio Diaz (1876-1910) accelerated the accumulation of lands among a few rich ‘latifundistas’ (land owners) and increased the number of landless farmers living under slave-like conditions within the ‘haciendas’ (a large plantation within a dwelling house). Even though Profirio Diaz’ ideas led to great economic development and modernization of the country, the social cost was immense and this led to social instability, and the regime ended with the Mexican Revolution of 1910. The aftermath of the Diaz regime demanded deep structural changes regarding the distribution of agricultural lands and other social justice issues (Herrera-Rodriguez 2011). Emiliano Zapata was one of the most important leaders of the agrarian resistance; under the famous motto of “Tierra y Libertad” (Land and

Freedom), Zapata sought to institute an agrarian model that had the institution of the pre- 44

Columbian common land tenure at its base. He demanded the prompt restitution of lands to its original owners and the redistribution of land to landless farmers (Herrera-Rodriguez, 2011). The agrarian reform promulgated in 1917, following the Mexican Revolution, was a dramatic and effective program for promoting food self-sufficiency and rural wellbeing during its first years.

Common land granted by the government were classified as ejidos2 or as comunidades agrarias3.

The post-revolutionary governments conceived these lands as a transitional land tenure system designed to allow farmers to get used to private property (Stavenhagen, 1975). However, usually the government granted marginal lands that were characterized for their aridity and lack of accessibility to water and transportation routes. Furthermore, it was not until 1930 where Article

27 of the constitution was modified and the government was allowed to grant ejidos to farmers upon request.

It was not until the government of Lázaro Cardenas (1936-1940) where land reform was launched as a systematic process where the real interests of the revolution were represented. This land reform was the most successful program to raise farmer welfare by liberating important forces that would encourage and enable them to increase production. It was also one of the most successful land reforms implemented in the twentieth century in the whole world (Barkin 2002).

This model was so successful that Mexico was one of the few countries that was able to regain food self-sufficiency in its staple foods during the twentieth century. In spite of the absence of any effective structure to deliver technical assistance and credit to farmers, they were able to double yields of maize planted on rainfed lands, demonstrating their ability to apply traditional

2 Ejidos are territories granted to groups of landless farmers upon demand (Herrera-Rodriguez 2011).

3 Comunidades agrarias are lands given back to indigenous peoples based on demonstrated historic ownership. (Herrera-Rodriguez 2011). 45 knowledge and harness their collective experience to mutual benefit. This stage of history is often referred to as the “Mexican Miracle.” After 1940, the Mexican government adopted development-oriented policies with the intention of promoting the modernization of the agricultural sector with implementation of chemical fertilizers and pesticides and with this, the so-called ‘Green Revolution’. These policies supported and expanded the efficient agribusiness sector of Mexican agriculture while greatly limiting access to loans and subsidies augmenting debts and costs of ejidos (Herrera-Rodriguez, 2011).

Once food self-sufficiency was achieved in the early 1960s, the price support level for maize was frozen, and with time it became very difficult for farmers to make a living from their agricultural products, forcing them to migrate to cities or to the United States in search of employment opportunities. Huge maize imports started during the end of 1960s and 1970s and political and social issues of great magnitude followed. An important program of subsidies was used in the 1980s to regain food sufficiency, but the 1982 debt crisis ended the program for good

(Barkin 2002). Furthermore, soils were eroded and waters polluted by the intensive use of agrochemicals in agricultural activities, and productivity became very low. The government often blamed the lack of productivity on the land tenure policies, which ironically the same political party had promoted years earlier.

The neoliberal policies since the 1980s further restricted access to credit and exposed the communal sector to a disadvantageous competition to set product prices in the market for their agricultural products. With its accession to the North American Free Trade Agreement (NAFTA) in 1994, Mexico substantially increased its maize imports, reorienting public policy to stimulate 46 maize production and agro-exports in irrigated areas while discouraging small-scale farmer production completely (Barkin 2002).

The Zapatista uprising in Chiapas 1994 is the most visible effect of the distress and frustration of indigenous farmers living in extreme poverty in the face of government tolerance for steep and persistent rural inequalities (Brown 1997; World Bank 2001).

By 2007, the collapse of the neoliberal agricultural food-provision model became clear when the price of tortillas increased by 66% in 2 months in Mexico City, and by 150% in the rest of the country (Herrera-Rodriguez 2011). Given that tortillas represent up to 60% of the average daily caloric intake, these price increases resulted in mass protests in Mexico City, where practice controls where demanded in the most important staple foods (Oxfam, 2008).

2.8 Selection of Methods

The objective of this study was to monitor the behavior of different vegetative covers in the event of abnormally dry conditions, with a particular interest on rainfed agricultural lands in

Central Mexico. To do so, an original combination of methods was selected after a detailed literature review. The methodology presented in this thesis merged some pre-existing techniques of detecting drought; first through a meteorological analysis and subsequently through remote sensing techniques to detect its effects on vegetation. In this section, a discussion of the reviewed methods is presented to further justify the selected methodology.

2.8.1 Meteorological Analysis

With the objective of differentiating years with different precipitation conditions, monthly precipitation data since 1980 were analyzed. The aim was to detect and compare two dry years with one normal year. As mentioned in previous sections, the SPI has been often compared with 47 the also commonly used PDSI. PDSI is an effective tool for detecting drought, and for this reason it is used by many international agencies that monitor global environmental extremes (e.g.

North American Drought Monitor). However, PDSI has some disadvantages, since the results may lag behind emerging droughts by several months, and therefore it does not handle frequent climatic extremes well (Sheffield and Wood 2011). PDSI also requires additional data of evapotranspiration, runoff, and soil moisture for its calculation (Xiaofan 2012), which are often not available. For the Standardized Precipitation Index (SPI), only precipitation data is required to detect drought occurrence. The drought definition based on standard precipitation is simply the difference of precipitation from the mean for a specified time period divided by the standard deviation where the mean and the standard deviation are determined from past records (McKee

1993). A disadvantage of this simple method is that precipitation is typically not normally distributed for accumulation periods of 12 months or less, but this can be overcome by applying a transformation to the distribution.

SPI was chosen as the most appropiate method for this study because it has been shown to yield successful drought monitoring by only depending on precipitation data input.

Furthermore, this data driven index has been so widely used that the National Drought

Mitigation Center (NDMC 2012) and the University of Nebraska designed an open source free access program (named SPI_SL_6.exe) to calculate SPI for different monthly intervals (i.e. 1, 3,

4, 6, 12-month SPI). All of the above made the SPI the most convenient and appropiate method for detecting lower than normal and normal precipitation years to be analyzed. The SPI was executed using monthly data since January 1980 through December 2011. The results were 48 analyzed then one recent year with normal conditions and two years with dry conditions were selected (2000, 2005, and 2011). With this, a time frame was established for the study.

2.8.2 Remotely Sensed Data

With the objective of further investigating vegetation development, a search for available satellite imagery was performed. By selecting a relatively small area of approximately 1095 km2, the study attempted to use imagery with the best resolution available at the most affordable cost, given the focus of this study on developing an approach that is suitable for developing countries context. The accurate observation of vegetation development relies on the quality of the input data. The higher the resolution of the satellite images, the better the understanding of how crops develop and are prone to fail in the event of drought.

NOAA-AVHRR and MODIS imagery has been widely used in the literature and is freely available with very high temporal resolution. However, these products offer a spatial resolution of 1 km and 500 m (AVHRR and MODIS respectively). The Satellite Pour l’Obervation de la

Terre (SPOT), IKONOS, and RapidEye offer very good resolution imagery ranging from 1m to

39m; however, this imagery is sold by square kilometer and is beyond what could be afforded by most organizations in developing countries. The Landsat program has provided the longest continuous space-based record of the Earth’s surface. Images are available since 1972 and offer a spatial resolution of 30m in the multispectral bands and 10m for the panchromatic band. Landsat

7 ETM+ was launched in April 1999 and has an orbital frequency of 16 days (USGS, 2012). This means that the temporal resolution allows the user to have access to approximately two images of the same site each month as long as persistent cloud cover does not limit the clear view of the

Earth’s surface. Landsat has been widely used for land use land cover classification (Oetter et al. 49

2001; Guerschman et al. 2002; Maxwell et al. 2004; Lucas et al. 2007; Saadat et al. 2011) and furthermore to detect drought (Lenney 1996; Oeyyer et al. 2001; Anyamba 2005; Junb and Sader

2005; Rhee 2010; Fadhil 2011; Vanderpost et al. 2011).

This study used as many images as there were available for the three years to be analyzed, which resulted in 18 satellite images. Each scene selected included the entire surface of the study area so there was no need for mosaicking images. The scenes were selected for different months of different years with the purpose of comparing vegetation development in seasons with different precipitation conditions. Since the spring-summer cycle represents 85.5% of rainfed maize sowing at a national level, which yields approximately 70% of the total grain production (Financiera Rural 2011), the temporality of the chosen satellite images was done according to the agricultural calendar in central Mexico for two months of each of the selected years due to saturated cloud cover (see table 2). This, however, does not represent a significant limitation because there are sufficient images that can be compared to acknowledge the effects of drought. Fadhil (2009), for example, used one Landsat 7 image (from the same months but different years) to successfully map drought in Iraq.

A limitation of the acquired data was the data gaps in the images caused by a malfunction of the Scan Line Corrector (SLC) of the instrument on May 31, 2003 (USGS, 2012). The SLC- off effects are most pronounced along the edge of the scene and gradually diminish toward the center of the scene. The middle of the scene contains very little data loss, and this region of each image is very similar in quality to previous ("SLC-on") Landsat 7 image data (see figure 1). The selected study area was located near the center of the image. The area of the striped data gaps 50 was calculated for the study area (1095 km2) and it represented approximately 9% of the total area of each image.

2.8.3 Land Use Land Cover Maps

The 18 Landsat ETM+ scenes were processed to obtain the Land Use Land Cover

(LULC) maps where the cover of interest were classified (Forest, Urban Areas, Rainfed

Agriculture, Irrigated Agriculture, Water Bodies, and Pastureland). It was assumed that the images are representative for each month of the growing season of each respective year.

Land use land cover (LULC) classification is one of the most widely used applications in remote sensing. Many approaches have been used to correlate satellite images with vegetation cover. There are several distinct growth periods during a growing season and classification of vegetation characteristics based on the presence or absence of vegetative cover and the condition of that vegetation of the acquired image (Saadat et al. 2011). Two authors of the current paper developed a methodology whereby incorporating climatic and topographic conditions as ancillary data helped to delineate classes that could otherwise result in overlapping information

(Saadat et al. 2011). Saadat et al.,(2011) merged various techniques for developing high accuracy maps by incorporating climatic and topographic conditions as ancillary data that helped delineate classes that otherwise could result in overlapping information. This methodology is superior to others because it blends several techniques to obtain high accuracy classifications. It applies the concept used by Oetter et al (2001) and Guerschman et al (2002) by using several multi-temporal images to create one LULC. It also applied Lucas et al.’s (2007) notion of using auxiliary datasets to apply image segmentation and zonal statistics. Furthermore, it uses maximum likelihood classification (Maxwell 2004) to determine the likelihood of pixels to belonging to 51 one class. For classification of Landsat ETM+ images the approach proposed by Saadat et al.,

(2011) proceeded in five general steps: (i) preprocessing of the images, (ii) random extraction of a training sampling location: an unsupervised classification and two digital ancillary layers served in identifying potential LULC areas to aid in identifying sampling points, (iii) supervised classification of the image into LULC classes (iv) enhancement of the LU classification via image segmentation and zonal statistics, and (v) LC classification via NDVI and climatic zones and creation of a final LULC map. These steps are reviewed in depth in the following chapters.

Given the complexity of the remote sensing techniques used in our study for classifying vegetation, there was a need to assess the reliability of the results. In the literature, there are several techniques (such as creating an error matrix), that have been recommended by many researchers and that should be adopted as the standard reporting convention for classification analysis (Congalton, 1991).

Perhaps the simplest descriptive statistic is the error matrix can be used as a starting point for descriptive and analytical statistical techniques. It is calculated by dividing the total correct pixels (i.e., the sum of the major diagonal) by the total number of pixels in the error matrix. In addition, both user’s and producer’s accuracy are calculated.

Traditionally, for producer’s accuracy the total number of correct pixels in a category is divided by the total number of pixels of that category as derived from the reference data. This accuracy measure indicates the probability of a reference pixel being correctly classified and is really a measure of omission error (Congalton, 1991). For user’s accuracy, the total number of correct pixels in a category is divided by the total number of pixels that were classified in that category, then this result is a measure of commission error. User’s accuracy is indicative of the 52 probability that a pixel classified on the image actually represents that category on the ground

(Story and Congalton 1986). The Kappa Index of agreement is a powerful technique in its ability to provide information about a single matrix as well as to statistically compare matrices. This test determines whether the results presented in the error matrix are significantly better than a random result (Congalton 1991). Once LULC classification was performed and a good accuracy was achieved, four remotely sensed spectral indices were selected in this study to monitor the effects of drought in vegetation.

2.8.4 Spectral indices to detect agricultural drought

Remotely sensed approaches have been used to detect agricultural drought, and often associate image reflectance using spectral indices to indicate vegetation condition such as the normalized difference vegetation index (NDVI), the vegetation condition index (VCI), and soil/ vegetation wetness such as the normalized difference wetness index (NDWI) and the tasseled cap wetness (TCW). These indices were chosen in this study because of their individual characteristics and ability to detect the effects of drought.

The normalized difference vegetation index (NDVI) is a spectral index that has been shown to be highly correlated with parameters associated with plant health and productivity, and has been widely used for drought detection (Lenney, 1995; Kogan, 1995; Rahimzadeh, 2002;

Peters et al. 2002, Barbosa et al. 2006). The NDVI has been used successfully to identify stressed and damaged crops and pastures, but interpretive problems arise when results are extrapolated over non-homogeneous areas. It can be said that NDVI has two components: ecology and weather. The need of multi temporal NDVI analysis for detecting drought is that in any single

NDVI image in a given growing season, barren fields may be indistinguishable from temporarily 53 fallow healthy fields (Wallace et al., 1993), and immature crops with low density cover can be confused with stressed crops (Lenney et al. 1996). Therefore, when NDVI has been used for the analysis of weather impact on vegetation, the weather component must be separated from the ecosystem component (Rahimzadeh-Bajgiran et al. 2008). It has also been found that NDVI correlates with net primary production, biomass, vegetation fraction, and yield (Goward et al.

1987; Maselli et al. 1992; Rasmussen 1992; Quarmby et al., 1993; Hayes and Decler 1996;

Korgan, 1997; Unganai and Korgan, 1998; Rahimzadeh-Bajgiran et al. 2008). NDVI is one of the most frequently used spectral indices used in the literature and for this reason was chosen to be the main index used for this study.

The VCI was developed by Kogan (1995), and has been used to estimate weather impact on vegetation. The weather-related NDVI envelope is linearly scaled to 0 for minimum NDVI and 1 for the maximum for each grid cell. The VCI varies from 0 to 1, corresponding to changes in vegetation condition from extremely unfavorable to optimal. NOAA-AVHRR (Advanced Very

High Resolution Radiometer) data have been extensively used for drought monitoring in various areas of the world and good correlations have been established between NDVI or VCI and precipitation (Karabulut 2003; Wang et al. 2003; Hartmann et al., 2003; Bayarjargal et al. 2006;

Bhuiyan, Singh, and Kogan 2006; Amin Owrangi 2011). The VCI was chosen because it complements NDVI and provides further insights of vegetation stress related to the weather component.

The NDWI proposed by Gao (1996) can be obtained using the near infrared (NIR) and shortwave infrared (SWIR) channels. Gao found that NDWI is a measure of liquid water molecules in vegetation canopies that interacted with the incoming solar radiation and is less 54 sensitive to atmospheric scattering effects than NDVI. Gu et al., (2007 and 2008) evaluated the relationship between NDVI and NDWI using MODIS 500-m satellite imagery in Oklahoma and

Kansas, USA. They found good correlation between two indices and furthermore NDWI resulted to be more sensitive to drought than NDVI. Rhee et al. (2010) identified a drought index that had the possibility to be used for agricultural drought monitoring in arid/semi-arid regions as well as humid/sub-humid regions also using MODIS data. They found that NDWI has a better response in arid/semiarid regions that in humid regions when detecting drought. It is convenient to detect drought at larger scales using coarse resolution imagery such as MODIS; however, these approaches cannot provide detailed land cover response to precipitation anomalies.

The tasseled cap transformation (TCT) has been used widely for vegetation mapping and monitoring land cover change (Meyer et al., 2001; Jin and Sader 2005; Fadhil 2011). The TCT of

Landsat thematic mapper consists of six multispectral features, all of which could be potentially differentiated in terms of stability and change in a multitemporal data set (Jin, 2004). The first three features have been labeled brightness, greenness, and wetness (band 1, 2, and 3, respectively). The third feature, wetness, has been shown to be sensitive to soil plant moisture

(Jin and Sader 2004). Tasseled cap wetness (TCW) contrasts the sum of the visible and the near- infrared bands with the sum of the shortwave bands (Jin and Sader 2004). This spectral index was used as a complement to NDWI with the purpose of analyzing and comparing its responsiveness to vegetation/soil moisture. The NDWI and TCW have both been used in studies to detect drought or disturbance in ecosystems. Good correlations between the two indices have been found. Fadhil (2011) used NDWI and TCW to detect drought effects on vegetation in the

Iraqi Kurdistan region. 55

2.9 Summary and Conclusions

Agricultural drought has been studied by a wide range of methods and different techniques such as data driven models or remotely sensed driven methods. In the present study, a new combination of methods was used to detect the effects of droughts. This methodology first involved the analysis of meteorological data to detect years with different precipitation regimes.

Then for the years detected, satellite imagery was used for developing high accuracy land use land cover maps. Finally, the application of different spectral indices that reveal vegetation health and vegetation/soil moisture was used.

Preface to chapter 3

The review of literature (chapter 2) shows that in order to study drought a time series analysis is required in order to distinguish periods of abnormally lower values of either precipitation or spectral vegetation indices that indicate the effects of such an event. Thus, chapter 3 focuses on the analysis of precipitation data since Jan 1980 to Dec 2011 in order to establish a time scale for the study. Then the analysis used Landsat ETM+ satellite imagery to detect the effects on vegetation caused by drought.

This chapter has been intended to be presented and published by the International Journal of Water Resources Development 2014. The manuscript has been co-authoired by Andres Sierra- Soler, Jan Adamowski, Hossein Saadat, Zhiming Qi and Santosh Pingale. 56

Chapter 3

Assessing Agricultural Drought at a Regional Scale Using LULC Classification,

Standardized Precipitation Index (SPI), and Vegetation Indices.

A.Sierra-Soler, J. Adamowski, H.Saadat, Z. Qi, S. Pingale

Department of Bioresource Engineering, Macdonald Campus, McGill University, 21,111

Lakeshore, Ste-Anne-de-Bellevue, QC, Canada, H9X 3V9

[email protected], [email protected], [email protected],

[email protected]

Abstract

The relationship between spectral properties of vegetation and their biomass has been recognized since the first spectrometric field experiments in the 1970s. For this reason, satellite observations can provide insights on vegetation conditions, crop yield, and furthermore to monitor the impact of droughts. Rainfed crops develop from sowing to harvest primarily as a function of meteorological variables are then enriched by farming practices that in developing communities are restricted by the access to agricultural assets. For farmers who depend on this activity for self-sufficiency, a drought can result in great human suffering, thus there is a great need to understand how droughts disturb the landscape in such regions. This paper has the objective of using remote sensing to assess the phenomenological impacts of two isolated droughts and finally distinguishing the response of different vegetation covers in semiarid developing regions where rainfed agriculture is widely spread. Using the Standardized

Precipitation Index (SPI), one normal and two dry years where selected (2000, 2005, and 2011, 57 respectively). Then, an original protocol for Land Use Land Cover (LULC) classification that combines climatic, topographic, and reflectance information from 18 Landsat ETM+ images was applied to subsequently distinguish drought effects in different classes through the selected years.

Finally, two vegetation indexes (NDVI and VCI) were calculated to detect drought severity impacts over the different LULC classes. This approach was tested in Central Mexico, and this original combination of methods provided accurate information on the location of and extent of areas affected by drought. The proposed approach has the potential to be used as a system for drought risk management in semi-arid developing regions.

Key Words: Drought, Rainfed Agriculture, Developing Regions, SPI, LULC maps, NDVI, VCI,.

1. Introduction

The onset and the end of a drought are very difficult to determine, their impacts increase slowly and often accumulate over periods. Droughts are often perceived too late, when the after effects of falling crops, dying livestock, and severe economic and social damage has already occured. Although droughts are generally driven by a period of low precipitation, other factors modulate their occurrence, depending on the characteristics of the region and how water scarcity propagates through the hydrological system (Van Lanen et al. 2004). The variation in climate and a spectrum of temporal and spatial scales are the main drivers. Typically, droughts are initiated by atmospheric circulation and weather systems that conspire to cause lower precipitation and/or higher evaporation than normal in a region (Sheffield 2011). Furthermore, the aftermath is non- structural and many times it is spread over large areas and can be directly triggered by human activities.

Droughts rank first among all natural hazards when measured in terms of the number of 58 people affected (Obasi 1994; Hewitt 1997; Wilhite 2000; Mishra et al. 2010). In developing regions, agricultural drought poses a great threat to smallholder rainfed producers. Rainfed agriculture in many cases is practiced mainly for producing self-sufficiency crops and people depending on this economic activity are vulnerable upon the inevitable environmental variability.

In the literature, there is no consensus on one definition of drought; however, the phenomenon has been classified and studied depending on the system it disturbs (e.g., meteorological, hydrological, groundwater, agricultural, and socio-economical). The assessment of droughts is of primary importance for planning and management of water resources. This requires the understanding of how water scarcity impacts the systems of interest.

Several indices have been developed in recent decades with the purpose of detecting drought disruption on different systems. Depending on the data input, drought indices are separated in two groups: data driven indices and satellite driven drought indices (Mishra and

Singh 2010).

A frequently used data driven index is the Standardized Precipitation Index (SPI). The

SPI is based on a long-term precipitation record (minimum 30 years) fitted into a probability distribution, then transformed to a normal distribution, so that the mean SPI for the location and desired period is zero (McKee 1993). The SPI has been often compared with the also commonly used Palmer Drought Severity Index (PDSI). However, PDSI presents some disadvantages since the results may lag behind emerging droughts by several months, and therefore does not handle frequent climatic extremes (Sheffield and Wood 2011). PDSI also requires additional data of evapotranspiration, runoff, and soil moisture for its calculation (Xiaofan 2012), which often are not available, especially in developing countries. 59

Satellite driven indices have been derived from a wide variety of satellite systems orbiting the Earth since the 1970s. These systems have permitted the observation of Earth’s surface reflectance and revealed insights on vegetation status and change (e.g., NOAA-AVHRR,

MODIS, SPOT, and Landsat). Such systems have provided imagery products that vary in spatial and in temporal resolutions. AVHRR and MODIS have been frequently used for global drought detection for their high temporal resolution but this data has the disadvantage of low spatial resolution. On the other hand, medium resolution products like Landsat offer the potential of generating detailed vegetation classification for understanding effects of drought in specific classes even though it has lower temporal resolution.

Various studies have shown the effectiveness of satellite imagery in classifying heterogeneous vegetation cover (Joshi et al. 2006; de Asis Omasa 2007; Focadi et al. 2008), forest status (Labrecque et al. 2006; Sivanpillai et al. 2007) crop yields (Cohen and Shoshany

2002; Wardlow et al. 2007) and the effects of droughts on agricultural crops (Lenney 1996;

Anyamba 2005; Fadhil 2011).

These remotely sensed approaches have been used to detect agricultural drought and are often associated with image reflectance of the visible and infrared wavelengths with vegetation characteristics. One example is the normalized difference vegetation index (NDVI) that has been shown to be highly correlated with parameters associated with plant health and productivity and has been widely used for drought detection (Lenney 1995; Kogan 1995; Rahimzadeh 2012;

Peters et al. 2002, Barbosa et al. 2006).

In semiarid developing regions, medium resolution remotely sensed data have been used to assess the impacts of drought. Vanderpost et al. (2011) used Landsat imagery to assess long 60 time conditions of rangeland in semiarid areas of Botswana. By calculating vegetation indices, he found significant degradation in vegetation corresponding to the droughts between the time period of 1984 and 2000. The technique used was limited by the many gaps in the long-time coverage and hence lack of continuity in terms of change. Fadhil (2011) used only two Landsat images from consecutive years to calculate five vegetation and soil/vegetation moisture indices in the Iraqi Kurdistan region. This approach lacks the continuity of time series analysis and therefore it was unable to anticipate the evolution of drought.

In Mexico, drought monitoring at the national scale is performed by North America

Drought Monitor (NADM), which is a cooperative effort between drought experts in the United

States, Mexico, and Canada. The NADM program was initiated in April 2002 and is part of a larger effort to improve the monitoring of climate extremes on the continent. The data produced by NADM focuses only on three data driven indexes: the SPI, the PDSI, and the Percent of

Average Precipitation. Other studies in Mexico have reported qualitative drought impacts, documenting smallholder rainfed maize production and climatic risk (Eakin 2000), climate change impacts on food security (Appendini and Liverman 1994), estimates of the effects of El

Niño Southern Oscilation (ENSO) on crop yield (Adams et al. 2003) and vulnerability and adaptation to drought (Liverman 1990). However, very little has been done in the region for using remotely sensed data in combination with meteorological analysis to estimate the effects of drought on rainfed agricultural production, making this study very original and useful.

The present study aimed to develop an original combination of methods to detect drought through i) a meteorological analysis of a long time series of monthly precipitation data to find dry and normal periods of time, to ii) use of remote sensing techniques to develop Land Use 61

Land Cover (LULC) classification, then iii) calculate vegetation condition and soil/vegetation

moisture indexes to monitor the monthly effects of droughts at regional scales related to

agriculture in developing regions, and finally iv) to use change detection for each class

individually and for each of the four indexes previously calculated.

The meteorological analysis consisted in the computation of the Standardized

Precipitation Index (SPI), which was based on the long-term monthly precipitation record for the

period of 1980-2011. Medium and long term SPI trends were chosen to be analyzed over short

term SPI tends (6-month, 9-month, and 12-month SPI) with the purpose of effectively showing

seasonal drought conditions on agriculture. SPI was calculated from 1980 to 2011 with the

purpose of selecting two dry years and a normal year and subsequently using remote sensing

techniques to detect vegetation development. Eighteen Landsat 7 ETM+ satellite images were

used and a land use land cover (LULC) classification was performed to study rainfed agriculture

status through the different years selected.

An original protocol for LULC classification was used. This protocol, which was

developed by the second and third authors of this paper (Saadat et al. 2011), was designed for

classifying large areas using 18 Landsat ETM+ images and incorporating ancillary topographic

Figure 1: Location of the Tortugas-Tepezata Sub-Watershed and climatic data (Saadat et al. 2011).

Finally, the Normalized Vegetation Index

(NDVI) and the Vegetation Condition

Index (VCI) were processed to detect

vegetation condition on a monthly basis for

three years 2000, 2005, and 2011. The 62

LULC classes were used to extract the resulting processed layers to discern the response of the different vegetation covers.

The LULC classification presented in this paper is a simplified process of the original approach by Saadat et al. (2011), and it is shown in this paper to be successful in classifying an average of 80% accuracy in short time with 18 medium resolution multispectral satellite images in a short period of time. Moreover, the geographical scale of the study permitted a focus on the regional landscape and the combination of methods here described is original in the in the science of drought monitoring in developing regions including Central México.

The imminent climatic variability and the future uncertainty demand a deep understanding and creation of up-to-date analysis of drought impacts and this study aimed to address that purpose. Moreover, there is no study in the region that has attempted to assess the impacts drought on rainfed agricultural crop development using remote sensing techniques in marginal developing territories making this study the first of its kind for central Mexico.

1.1 Study Area

For a country with a millennial agricultural tradition, Mexico’s climate is not particularly favorable for agricultural production. Thus, as much as 46% of Mexico’s land area is classified as arid, which has the effect of limiting agricultural land use to grazing or irrigated farming

(Reyes Castaneda 1981). According to Liverman (1999), biophysical vulnerability to drought is greatest in the northern and central regions of the country where rainfall is most variable, and the timing of rainfall and the extent of the mid-summer drought are critical.

With an area of approximately 1095 km2, the Tortugas-Tepezata sub-watershed (between 20° 27’

03” N 98° 46’ 23” W and 20°04’50” N 98°13‘53” W) is located in the high Moctezuma 63

Watershed in the state of Hidalgo in Central Mexico. The Moctezuma Watershed belongs to the

High Panuco Hydrological Region that flows into the Gulf of Mexico in the Northeastern State of Tamaulipas. The Tortugas-Tepezata sub-watershed is located between two regions known as

The Lower Sierra and The Tulancingo Valley and it covers 9 municipalities (CONANP, 2003).

The main feature of the sub watershed is the river canyon that flows along 100 km and it is named according to the region, identifying three major sections. The first entry to the South

Canyon is with the Rio Grande Tulancingo, the second is at the junction with the San Sebastian

River and Venados River, and the third is at the Irrigation District-8 named Metztitlán River, north of the sub-basin.

1.1.1 Climate

The climatic conditions in the region are dry and semi-dry climates, characterized for having an average annual precipitation between 300 to 500mm. In the region, there are two temperature regimes; the warm regime with an annual average temperature above 18 °C and the cold regime with lower than 18 °C (SEMARNAT, 2010).

1.1.2 Agriculture

Most of the agricultural lands owned by smallholder farmers in the area involve self- sufficiency rainfed agriculture with the milpa system. The milpa is a complex traditional agriculture production system where a variety of associated crops (corn, beans, squash, and by products such as huitlacoche, which is an edible fungus that grows in corn) are produced (Barkin

2002).

3 -1 Mexico City and its metropolitan region discharge 60 m s of raw wastewater that has been used to irrigate agricultural land in the Mezquital Valley and in Hidalgo for nearly 100 64 years (Prieto-Garcia et al. 2005). Its annual agricultural production includes diverse agricultural products like alfalfa, corn, forage oat, bean, marrow, of green chili, and tomatoes (green plus red varieties), among others (INEGI 1999). However, only a small area (about 5% of the study area) of the Tortugas-Tenejapa sub Watershed is irrigated using water from Mexico City, according to

LULC maps developed by SEMARNAT 2010. The growing season calendar for rainfed agriculture in Central Mexico starts with the rain season in April-May and it takes between 175 to 200 days for native maize to develop to maturity.

2. Materials and Methods

2.1 Materials

2.1.1 Monthly Precipitation Data

Monthly precipitation data was provided by the National Meteorological Service (SMN in Spanish) from the six available meteorological stations inside the study area that have been active since 1975 to the year 2013. The monthly precipitation data were retrieved from 1980 to

2012 for each of the individual stations. Precipitation data was used to process the Standardized

Precipitation Index (SPI) to identify years with dry conditions. Table 1 shows the name of each station and location.

Table 1: Name and Location of Climatic Stations inside the study area (SMN, 2012)

# Station Name Municipality Latitude Longitude Altitude Agua Blanca de 1 Agua Blanca 20o09’00” 98o27’0” 2100 m Iturbide 2 Alchoyola Acatlán 20o13’32” 98o32’10” 2060 m 3 Atotonilco Atotonilco el Grande 20o17’ 98o40’ 2110 m 4 El Zembo Huasca de Ocampo 20o10’13” 98o20’46” 1800 m 5 Huasca Huasca de Ocampo 20o01’4” 98o20’46” 1600 m 6 Santa María Amajac Atotonilco de Tula 20o02’00” 99o12’00” 2490 m 65

2.1.2 Open Access SPI software ‘SPI_SL_6.exe’:

Developed by the National Drought Mitigation Center in the University of Nebraska

(NDMC, 2012), this open source program was designed to easily calculate SPI monthly intervals

(i.e. 1, 3, 4, 6, 12-month SPI) using input from monthly precipitation data.

2.1.3 Satellite Imagery

Eighteen Landsat ETM+ scenes (Table 2) were downloaded freely from the NASA and

USGS Earth Explorer (2013). These 18 images were chosen because they were all the available images clear from clouds for the anticipated intervals of time. Each scene selected included entire surface of the study area. The scenes where selected for different months of different years with the purpose of comparing vegetation development in seasons with different precipitation conditions. Since the spring-summer cycle represents 85.5% of rainfed maize sowing at a national level, which yields approximately 70% of the total grain production (Financiera Rural

2011), the temporality of the chosen satellite images was done according to the agricultural calendar in central Mexico.

Table 2: Temporal Resolution of selected Landsat ETM+ images corresponding to the agricultural calendar.

Note that SLC-ON is for all images in year 2000 and SLC-off is for all images in 2005 and 2011.

Scan Line Month/ Corrector Year Apr May Jun Jul Aug Sep Oct Nov SLC-On 2000 22 13 25 03 28 NA NA 16 SLC-Off 2005 04 22 07 09 10 NA NA 14 SLC-Off 2011 NA 23 08 NA 11 28 30 15

The accurate observation of vegetation development relies on the quality of the input data. The higher the resolution of the satellite images the better the understanding of how crops develop and are prone to fail in the event of drought. However, while coarse-resolution images

(e.g. AVHRR and MODIS with resolution of 1 km and 250 m respectively) are often freely 66 available, up to date, higher resolution images are expensive. For this study, Landsat products were chosen because they offer up to date medium resolution (30 m for multispectral 1-7 bands and 10 for panchromatic band).

Landsat 7 was launched in April 15, 1999 and has an orbital frequency of 16 days

(USGS, 2012). This means that the temporal resolution allows the user to have access to approximately two images of the same site each month, as long as persistent cloud cover does not limit a clear view of the Earth’s surface. In this study, there was lack of availability of satellite images for two months of each of the selected years due to saturated cloud cover (see table 2). This, however, does not represent a significant limitation because there are sufficient images that can be compared to acknowledge the effects of drought. Fadhil (2009), for example, used one Landsat 7 image (from the same months but different years) to successfully map drought in Iraq.

Another limitation of the acquired data is the data gaps in the images caused by a malfunction of the Scan Line Corrector (SLC) of the instrument that occurred on May 31, 2003

(USGS, 2012). The SLC-off effects are most pronounced along the edge of the scene and gradually diminish toward the center of the scene. The middle of the scene contains very little data loss, and this region of each image is very similar in quality to previous ("SLC-on") Landsat

7 image data (see figure 1). The selected study area was located near the center of the image. The area of the striped data gaps was calculated for the study area (1095 km2) and it represented approximately 9% of the total area of each image.

Figure 2: Without an operating SLC, the Enhanced Thematic Mapper Plus (ETM+) line of sight traces since

2003 a zig-zag pattern along the satellite ground track (USGS, 2012) 67

These 18 Landsat ETM+ scenes were processed to obtain the Land Use Land Cover

(LULC) maps where the covers of interest were classified (Forest, Urban Areas, Rainfed

Agriculture, Irrigated Agriculture, Water Bodies, and Pastureland). It was assumed that the images are representative for each month of the growing season of each respective year.

2.1.4 Ground truth data

Ground truth data were collected during field site visits once a week to the study area during a period of three months (January, February, and March 2013). During field work, it was necessary to obtain samples of the classes that characterized the region. A topographic road map

(1:250 000), a digital camera to provide evidence of each sample, and a GPS were needed to visit each location. To obtain the best possible results, these sampling locations were visited with local people with knowledge of the local conditions. The visits were performed accompanied by a hired native guide and driver familiar with the area, who had a good understanding of the local agricultural practices. Farmers, residents, and local authorities were engaged in conversation with the intent of investigating agricultural practices, past natural disasters (such as droughts, fires, and floods) and land use changes in the past 10-15 years. Such information, complemented with photographs and the geo-referenced locations, were recorded. Derived from this effort, 132 samples were collected across all the study areas where roads permitted access. Each sample comprised cataloguing different land use and land cover classes that were to be used to create 68 ground truth maps for assessing the accuracy of the supervised classification performed by the remote sensing software explained below.

2.1.5 Climatic and topographic maps

Two digital ancillary layers were also collected to assist in the interpretation and classification of the remotely sensed data. These were a 1:25 000 topographic map, and a 1:50

000 scale climatic zone map both made available by CONABIO (2005). These maps were used to follow Saadat et al., (2011) methodology for the development of the LULC maps.

2.1.6 GIS software

ArcGIS 10.1® is a complete system for designing and managing solutions through the application of geographic knowledge. It enables the user to perform deep analysis, gain a greater understanding of data, and make more informed high-level decisions. This software was used to choose the study area and to manage all the thematic layers in the study.

2.1.7 Remote sensing software

ENVI® and ERDAS Imagine® were used to process the LULC classification and the

NDVI analysis.

2.2 SPI Analysis

In arid and semiarid areas, rainfall is the principal determinant of primary production and has been found to be highly correlated with vegetation dynamics analysis. The Standardized

Precipitation Index (SPI) computation is based on the long-term precipitation data for the desired time step. It is calculated by taking the difference of the precipitation from the mean for a particular time step, and then dividing it by the standard deviation (Sonmez et al. 2005). The SPI is a dimensionless index where negative values indicate dry conditions and positive values 69 indicate wet conditions. Drought intensity, magnitude, and duration can be determined, as well as the historical database probability of emerging from a specific drought.

The hydro-meteorological analysis of this study consisted of processing monthly precipitation data since January 1980 through December 2011. The SPI values were determined using the SPI to detect one year with normal precipitation and two years with precipitation deficit. The precompiled program for calculating SPI was used to obtain values for different monthly scales (6, 9, 12 months).

The National Drought Mitigation Center (NDMC 2012) and University of Nebraska designed an open source free access program (named SPI_SL_6.exe) to calculate SPI for different monthly intervals (i.e. 1, 3, 4, 6, 12-month SPI). The program is designed to compile the SPI as input files in a 3-column format. These columns of data must be ordered in Year,

Month, and Monthly Precipitation Value. Since precipitation values must not include decimals, the values were entered in millimetres. Special attention had to be made in the column spacing and missing data issues. As recommended by the software developers, missing precipitation data

(which consisted only of 17 values out of 2160 of all stations) was substituted for zeros because missing data flag or -9999 would not compile.

The SPI was calculated for three different monthly intervals (6, 9, and 12-month SPI) using the precipitation values of the 7 stations inside the study area. The results were plotted for analysis. Medium and long term SPI trends were chosen to be analyzed over short term SPI tends

(1 and 3-month SPI), with the purpose of effectively showing seasonal drought conditions affecting agriculture. According to the NDMC (2012) in the 9-month SPI, values below -1.5 for these time scales are usually a good indication that fairly significant impacts are occurring in 70 agriculture and may be showing up in other sectors as well.

The monthly interval SPI results were classified according to McKee (1993) in the following categories: [0 to -0.99] Mild Drought, [-1.00 to -1.49] Moderate Drought, [-1.5 to

-1.99] Severe Drought and [≤ -2.00] Extreme Drought. After careful analysis of the results, two recent dry years with values smaller than -1.5 (years 2005 and 2011) and one normal year with values above 0 (year 2000) were selected. The purpose of selecting two dry years and one normal year was to be able to compare the different effects on vegetation for different time periods. It should be noted that additionally there was an interest in studying the year 2011, since a state of national alert for drought was announced in January 2012.

2.3 Land Use Land Cover (LULC) Classification

2.3.1 Preprocessing images

Remote Sensing Science requires that solar radiation reflected from the Earth’s surface passes through the atmosphere before it is collected by the instrument. For this reason, when attempting quantitative analysis of surface reflectance removing the influence of the atmosphere is a critical pre-process step. It is atmospheric effects such as the amount of water vapor, distribution of aerosols, and scene visibility that affect the raw imagery. Such effects must be eliminated so that images taken in different times can be compared accurately. Because direct measurements of most atmospheric properties are rarely available, the Remote Sensing Software can infer from the imprint on radiance data. The Fast Line-of-sight Atmospheric Analysis of

Spectral Hypercubes (FLAASH) atmospheric correction module in the ENVI software

(developed by Spectral Sciences, Inc.) was used for atmospheric correction to retrieve spectral reflectance data from the multispectral Landsat ETM+ images. FLAASH is a first-principles 71 atmospheric correction tool that corrects wavelengths in the visible through near-infrared and shortwave infrared regions. It also includes correction for the adjacency effect (pixel mixing due to scattering of surface-reflected radiance), an option to compute a scene-average visibility

(aerosol), cirrus and opaque cloud classification, and adjustable spectral polishing for artifact suppression (Adler-Golden et al. 1999).

Georeferencing raster data is the process in which the user defines its local map coordinates and assigns it to the coordinate system of the data frame. All 18 Landsat ETM+ images were subjected to geo-referencing and image to image registration was performed. This process was verified by importing the tracks recorded from the GPS that were obtained during field work as ground control points. After preprocessing of all images was complete, the study area was clipped from each scene.

2.3.2 Extraction of training sampling locations

Training samples are regions of interest where all the pixels in the sample represent fully one class each of the land cover classes in our image. The classification of a particular surface shows the spatial distribution of identifiable features of the surface. The ISODATA unsupervised classification was used and consists of the calculation of initial class means distributed in the data space, which then iteratively clusters the pixels into the nearest class. It is possible to extend this limited information over an entire surface to obtain a potential Land Use Land Cover map.

Different classes of vegetation types when a statistically significant difference exists between measured irradiances (Focardi et al. 2008). Since each driven class has similar characteristics

(similar topography, climate zone, and spectral range), climate and topographic maps created by

CONABIO (2005) were used as supportive tools for identifying appropriate training sampling

Figure 3: Flow Chart for the development of Supervised Classification.

Clipped and Sampling Pont 72 locations across the study area. Based on the stratified random sampling procedure described by

Stenham (1999), 150 training sampling locations were extracted. However, due to inaccessibility and time available for the field work, only 132 locations were sampled, covering most regions in the study area.

2.3.3 Land Use Land Cover (LULC) Analysis: Saadat et al., (2011) simplified method

Land Use Land Cover (LULC) classification involves sorting pixels into a finite number of individual classes, or categories of data, based on similar reflectance values. If a pixel satisfies a certain set of criteria, the pixel is assigned to the class that corresponds to those criteria.

Processing LULC maps has been one of the most used applications in Remote Sensing science and many approaches and methods have been developed in the past. Such maps can play an important role in watershed management as a whole and help for example in deciding what sort of lands are capable of sustaining agriculture (Cihlar 2000; Renschler and Harbor 2002). In a previous paper, two authors of the current paper developed a methodology whereby incorporating climatic and topographic conditions as ancillary data helped delineate classes that otherwise could result in overlapping information (Saadat et al., 2011). For classification of

Landsat ETM+ images, the approach proposed by Saadat et al. (2011) proceeded in five general steps: (i) preprocessing of the images, (ii) random extraction of a training sampling location: an unsupervised classification and two digital ancillary layers served in identifying potential LULC areas to aid in identifying sampling points, (iii) supervised classification of the image into LULC classes (iv) enhancement of the LU classification via image segmentation and zonal statistics, and (v) LC classification via NDVI and climatic zones and creation of a final LULC map. 73

Upon completion of all these processes, the accuracy of the classifications were evaluated for each imaging date and comparisons made. In the methodology presented in this paper, Saadat et al., (2011) was simplified to process 18 Landsat ETM+ for obtaining the highest accuracy possible. The core notion of using ancillary climatic and topographic data to aid in generating maps with homogeneous areas was used. However, step 4 previously mentioned was omitted with the purpose of expediting the process of 18 Landsat ETM+ images. The aim was to get reasonably accurate results to acquire a comprehensive idea of how each class behaved in different precipitation conditions through time.

The supervised classification was processed using training sites chosen using stratified random sampling procedure that were homogeneous and that fully represented each of the LULC classes to be segregated. Six different classes that characterize the rural landscape in Central

Mexico were designated to be used for supervised classification and to generate a LULC map;

Rainfed Agriculture (RA), Irrigated Agriculture (IA), Forests (F), Pastureland (P), Shrubland

(SH), and Water Bodies (W).

To analyze crop evolution of rainfed agriculture in different precipitation contexts, the 18

LULC maps were processed. The total area of the Rainfed Agriculture (RA) was calculated, transforming raster layers to polygons using GIS software.

Subsequently, a supervised classification (figure 3) was performed. The Maximum

Likelihood Classification algorithm assumes that the statistics for each class in each band are normally distributed and calculates the probability that a given pixel belongs to a specific class. Each pixel is assigned to the class that has the highest probability and that is the maximum likelihood per pixel. 74

2.3.4 Assessing accuracy in remotely sensed data

Given the complexity of the remote sensing techniques used in the study to detect agricultural drought and execute LULC maps, there is a need to assess the reliability of the results. In the literature, there are several techniques (such as creating an error matrix), that have been recommended by many researchers and that should be adopted as the standard reporting convention for classification analysis (Congalton 1991). In this section, we will discuss the methods that were developed for assessing accuracy of the remotely sensed results.

Perhaps the most widely used descriptive statistic used in the literature is the error matrix, which can be used as a starting point for descriptive and analytical statistical techniques. It is calculated by dividing the total correctly classified pixels (i.e., the sum of the major diagonal) by the total number of pixels in the error matrix. The overall accuracy (OA) was calculated by summing all of the numbers within the matrices diagonal (D) correctly identified samples and dividing by the sum of all the errors (N) or numbers found outside the diagonal (Jensen 2005).

[1]

In addition, users and producers accuracy (PA) was calculated (see Appendix 2). The producer’s accuracy is a statistic that specifies the probability of a ground reference datum being correctly classified and it is a measure of the omission error. This statistic is calculated because the producer may want to know how well an area can be classified. The producer’s accuracy is calculated by dividing the diagonal number from a class’s column by the sum of the entire column, including the number found within the diagonal (Jensen 2005; Story and Congalton 1986).

[2]

Where xii is the total number of correct cells in a class and x+i is the sum of cell values in the column. 75

The user’s accuracy is a measure of the commission error. This statistic indicates the probability of how well the classified sample represents what is found on the ground. This measure is calculated by dividing the diagonal of a class by the sum of the numbers within the row of that class

(Jensen 2005; Congalton 1991).

[3]

Where xii is the total number of correct cells in a class and xi+ is the sum of cell values in the row.

The kappa coefficient (k) is another measure of the accuracy of the classification. It is calculated by multiplying the total number of pixels in all the ground truth classes (N) by the sum of

the confusion matrix diagonals (xkk), subtracting the sum of the ground truth pixels in a class times the sum of the classified pixels in that class summed over all classes, and dividing by the total number of pixels squared minus the sum of the ground truth pixels in that class times the sum of the classified pixels in that class summed over all classes. The Kappa index of agreement is a powerful technique given its ability to provide information about a single matrix as well as to statistically compare matrices. This test determines whether the results presented in the error matrix are significantly better than a random result (Congalton 1991).

[4]

Where r is the number of rows in the matrix, xii is the total number of correct cells in a class, x+i is the sum of cell values in the column and xi+ is the sum of cell values in the row.

2.4 Spectral Indices for Detecting Drought

2.4.1 The normalized difference vegetation index (NDVI)

The Normalized Difference Vegetation Index (NDVI) is one of the most successful tools to identify the condition of vegetated areas. NDVI has been shown to be closely associated with parameters linked to productivity, vegetation density and cover, green leaf biomass, and crop condition (Lenney et al. 1996; Peters et al. 2002; Huete et al. 2006; Rahimzadeh et al. 2008; 76

Xiaofan et al. 2012). Thus, NDVI can be used as a general indicator of vegetation cover and vigor. NDVI is calculated as the ratio of the red (RED) and the Near Infrared (NIR) bands of a sensor system and is represented by the following equation (Kogan 1995):

[5]

NDVI values range from [-1  +1]. Because of high reflectance in the NIR portion of the electromagnetic spectrum, healthy vegetation is represented by high NDVI values between 0.5 and 1.

In a year with normal precipitation conditions, agricultural fields have high NDVI values during periods of peak plant growth as well as low NDVI values after harvesting, during the early stages of growth, and when crops are stressed. In the occurrence of droughts, photosynthesis is reduced on account of rainfall deficit. This decreases total dry matter accumulation and yields and therefore results in lower NDVI values. Furthermore, NDVI values of non-vegetated areas and urban settlements are constantly low.

The key to using the NDVI to monitor and assess droughts is thus to have accurate time series satellite data for long periods of time; below normal values indicate the occurrence of droughts (Boken et al. 2005). The need of multitemporal NDVI analysis for detecting drought is that in any single NDVI image in a given growing season barren fields may be indistinguishable from temporarily fallow healthy fields (Wallace et al. 1993) and immature crops with low density cover could be confused with poor crops (Lenney et al. 1996). Alternatively, if scenes are observed at multiple dates, certain characteristic trends in crop development can be observed.

Furthermore, useful statistics have been used in the past for describing trends of NDVI values over a number of dates. The maximum and the range values of NDVI are multi-temporal features 77 that enable one to have an insight of the vegetation development during the time of interest. Low maximum NDVI values are indicative of non-productive lands, areas under constant stress and non-vegetated surfaces. High maximum values represent areas that contain healthy vegetation at least in one image (Lenney et al. 1996).

The NDVI was calculated for all 18 Landsat ETM+ images. Subsequently, to recognize the different NDVI characteristics of each LULC class previously created, the NDVI maps were masked and an individual NDVI layer per class was extracted. This resulted in different NDVI outcomes for each vegetated class (Forest-NDVI, Irrigated Agriculture-NDVI, Pastureland-

NDVI, Rainfed Agriculture-NDVI, and F-NDVI). This process resulted in 108 NDVI individual layers, representing all the classes throughout the 18 Landsat ETM+ images.

The first attempt to discern the properties of NDVI of each class was to show the distribution in a scatterplot of the multitemporal features of range NDVI versus maximum

NDVI. A time series of the NDVI per class was also plotted to distinguish the different patterns of NDVI through the growing season of the three different years.

Even though NDVI has been successfully used to identify stressed and damaged crops and pastures, interpretive problems may arise when these results are extrapolated over non- homogeneous areas (Singh et al. 2003). According to Singh et al. (2003), the NDVI has two components: ecology and weather. For vegetated regions, the integrated area of the weather component is smaller than the ecosystem component. The weather-related NDVI fluctuations are therefore not easily detectable (Kogan et al. 1987; Singh et al. 2003). Therefore, when NDVI is used for analysis of weather impact on vegetation, it should be complemented with other indices such as the Vegetation Condition Index (VCI). 78

2.4.2 The vegetation condition index (VCI)

The VCI was developed by Kogan (1990) and has been used to estimate weather impact on vegetation. The weather-related NDVI envelope is linearly scaled to 0 for minimum NDVI and 1 for the maximum for each grid cell. The VCI varies from 0 to 1, corresponding to changes in vegetation conditions from extremely unfavorable to optimal. The VCI is used in detection of drought and measurement of the time of its onset and its intensity, duration, and impact on vegetation. The VCI is calculated as follows (Kogan, 1990):

[6]

The Vegetation Condition Index in this study was calculated using the NDVI statistics extracted per class. Each individual extracted NDVI class was used to calculate the VCI. Once

more, this lead to 108 VCI classes Figure 4: NDVI for June 25, 2000 derived from Landsat ETM+ corresponding to the 6 classes in

the 18 Landsat ETM+ images. This

was done to compare the different

condition of the different classes in

vegetation to be compared.

The literature argues that VCI is

primarily useful for the summer

growing season and it has limited utility for cold seasons when vegetation is largely dormant (Mishra and Singh 2010). For this reason, two other indices that facilitated the observation of moisture in soils and vegetation were calculated in this study. 79

2.5 Change Detection

To find the change for selected land classes, it was necessary to first merge all polygons of the same classes in the LULC shapefile into one polygon for each class and for each image analyzed. After merging polygons selected LULCs were exported to individual shapefiles and then imported into the ENVI software to use as Regions of Interest (ROI). This allowed for computation of NDVI and VCI statistics for each ROI/LULC classes.

Remotely sensed based change detection is the process through which changes in the state of an object or a phenomenon are identified by observing it over repeated time intervals

(Beland et al. 2006). The change detection technique was used to map and assess land cover changes between each overlapping months of the selected years (e.g. May 2000, May 2005, and

May 2011). The year 2000 was used as a reference because of its normal seasonal precipitation.

In this way, the produced vegetation and water satellite-based indices were used to evaluate and to map the impacts of drought on the vegetative cover and soil/vegetation moisture.

Given the availability of overlapping dates of the satellite images, 6 periods in time were compared for 2000-2005 vegetation and wetness change (for the months of April, May, June,

July, August, and November) and 4 periods in time for the comparison of 2000-2011 (May, June,

August, and November). Change detection was based on the produced vegetation and wetness satellite-based indices of the different Land Use Land Cover (LULC) classes. Thus, change was detected for the same classes in different periods of times. For example, change of Pastureland-

NDVI 2000 vs. Pastureland-NDVI in 2005 and then Pastureland-NDVI 2000 vs. Pastureland-

NDVI 2011 and so on for all classes was accessed. This simple but elaborated class-change detection has not been reported in the literature of drought monitoring for developing regions. 80

Understanding the way LULC classes behave is important for decision making regarding drought prevention and relief.

Change detection was performed using the Erdas IMAGINE zonal change detection tool.

This tool requires the following inputs to generate the output image difference file:

1. Input Before Image: This image is the earlier of the two images.

2. Input After Image: This image is the more recent of the two images and reflects

change over time. The process performs an unsupervised classification on the Input

before Image and uses that to compute a probability of change between the Input

After Image to generate the Output Image Difference File.

3. Output Image Difference File: This image is the probability of change between the

Input Before Image and the Input After Image. This is a single band, floating point

image with values in the range 0.0 to 1.0.

4. Probability Threshold: To compute Areas and Percentages of change per zone, the

Output Image Difference File must be the threshold. Since the Output Image

Difference File represents probability of change, this threshold is a probability. The

threshold used was 10% so if any pixel shown more than 10% in change, it was

showed in red.

2.5.1 Vegetation types and phenomenology

Taking into account phenomenological changes in the different vegetation covers is key to understanding how vegetation responds in the event of droughts, which is not to be confused with inter annual variability of vegetation. Land cover types differ in vegetation type and density

(Fig. 4) throughout the study area. Furthermore, in semiarid regions, opportunistic annual species 81 green up rapidly in response to wetting of the soil surface and their vigor is primarily related to recent rainfall events (Zaitchik et al. 2006). In the case of rainfed agricultural land, the time of plantation can result in different values of NDVI because of different plantation conditions. With the objective of assessing statistical significance difference between changes due to plantation conditions and drought effects, a two tailed t-test was conducted for the areas classified as rainfed agriculture. One hundred random points were generated, constrained to the areas inside polygons classified as rainfed agriculture. NDVI pixel values were extracted to the set of point features. NDVI pixel values were extracted for same locations for each monthly paired image for the different years (dry and normal). The null hypothesis was that there would not be a significant difference in NDVI values between two images with different plantation conditions corresponding to the same month. Otherwise, the effect of a drought would be palpable.

Statistical difference was tested for all pairs of images corresponding to the normal and dry years, with an α of 0.01 statistical significance tested for the same months in different years.

3. Results and Discussion

3.1 SPI Results

According to McKee (1993), the Standardized Precipitation Index (SPI) is a dimensionless index where negative values represent precipitation deficits while positive values represent normal and wet conditions. The SPI was processed using monthly data from January

1980 to December 2011. The 9-month SPI compares the precipitation for that period with the same 9-month period over the historical record, and it is a good indicator of seasonal conditions affecting agriculture. According to the SPI results of this study, several years resulted in very dry conditions and also very wet conditions. The years 1982-1983, 1987, 2005, and 2011 had 82 moderate to severe drought conditions (SPI values of < -1). The year 1998-1999 resulted in values lower than -1.5, representing severe dry conditions. Moreover, this same period registered values of around 1.5, representing very wet conditions. With the objective of presenting more up- to-date data and imagery, the study focused on the results of the last decade.

The year 2000 was observed to have normal wetness conditions because it presented consistent positive values within approximately one standard deviation in the 9-month SPI values that increased in April, peaking in the month of June and then diminishing over the consequent months. The year 2000 presented a mean value over the 6 stations of 0.3437, a maximum of

1.93, a minimum of -1.55, and a standard deviation of 0.913. The year 2000 was chosen as the year with normal conditions to be used as a reference. The year 2005 persistently showed consistently dry conditions throughout the 6 meteorological stations inside the study area (see

Fig. 5). This year presented values ranging from mild drought (0 to -0.99) during the first months of the year to extreme drought in May, June, and July, according to the SPI results recorded from the Sta. Maria Amajac Meteorological Station. This year was chosen as a dry year to be used in

Figure 5: 9-Month SPI results for the six meteorological stations inside the study area. the study. There was a special interest to study the year 2011, since in January 2012, the 83 government of Mexico declared a state of emergency due to the most severe drought in the territory in the last 71 years. However, the 9-month SPI results for this year did not show consistent results throughout the six analyzed stations as in the two previously selected years.

The results derived from Aguablanca, El Zembo, and Huasca de Ocampo showed dry conditions for 2011 (see Fig. 5). Values ranging from mild drought (0 to -0.99) to extreme drought values of -4.5 in the month of July were recorded in El Zembo Meteorological Station.

Conversely, Alchoyola, Atotonilco, and Sta Maria Amajac Meteorological Stations showed wet conditions consistently peaking in the month of July, with values above 1.5 showing very wet conditions. Since it was of interest to study the year 2011 beforehand, it was decided to analyze it and acquire conclusions derived from satellite imagery.

Note in Fig. 5 that throughout the different stations, the year 2000 and the year 2005 followed the same trend. The year 2000 CLASS NAME Area in Km2 % of the Territory Forest 211.29 19.47% consistently showed positive values of Irrigated Agriculture 23.81 2.19% Pastureland 224.65 20.71% SPI (normally wet conditions) and Rainfed Agriculture 413.26 38.09% Shrubland 183.29 17.71% Water Bodies 26.76 0.35% 2005 continually showed negative SPI TOTAL AREA 1085 100.00% values (corresponding to abnormally dry conditions). The 9-month SPI values for 2011 shows diverse trends throughout the three stations (Aguablanca, El Zembo, and Huasca), which show negative SPI values. The meteorological station located in El Zembo recorded values of very severe drought with SPI values of -3.7 and -4.54 (below 3 to 4 times the standard deviation) for the months of June and July, respectively. Contrarily, the other three stations inside the study area

(Alchoyola, Atotonilco, and Sta Maria Amajac) showed normally wet conditions and the trends seem similar with peaking values in July and August, and then decreasing for the followed 84 months.

3.2 LULC Maps

Upon the completion of the supervised classification, 18 LULC maps were created from

Landsat ETM+ images from the years 2000, 2005, and 2011. Given that paved areas (to be considered urban settlements) represented around 1% of the study area, they were not taken into account for the inclusion of urban areas in the LULC maps.

Based on ground truthing the accuracy of the finalized LULC maps derived from the

Landsat ETM+ images acquired for different periods of the year was calculated. LULC maps consisted of 6 classes (see Fig. 6) that mainly define this rural landscape in Central Mexico:

Forests (F), Irrigated Agriculture (IA), Pastureland (P), Rainfed Agriculture (RA), Shrubland

(SH) and Water Bodies (W). Due to the SLC-off, approximately 10% of the pixels in the study area was lost for all images corresponding to years 2005 and 2011.

Table 3: The average area for each class for all LULC maps.For each of these images, an error matrix was generated. User and Producer’s accuracy were calculated (Appendix 1 and 2). Images from summer and late summer had the best overall accuracy and kappa accuracy (Appendix 3).

The lowest scores of overall accuracy resulted from the images corresponding to spring and fall.

The class that was better classified was the water bodies, as their spectral signature is quite unique and was not confused with other classes.

The user and producer accuracy results showed how individual class was misclassified with each other. Pastureland and Rainfed agriculture had mixed results and the lowest percentages because of their rapid response due to precipitation. However, during the summer months, the distinction was clearer (Appendix 1 and 2). Saadat et al.’s simplified methodology 85 for classifying with the accuracy of LULC maps for semidry areas proved to be successful in classifying many images in a short time and with an overall accuracy above 75%. Given the resolution of the images and the data gaps presented, these results were considered to have sufficient accuracy for the purposes established in this study.

3.3 NDVI and Vegetation Cover Change

NDVI was processed for all 18 Landsat ETM+ images covering years 2000, 2005, and

2011 throughout different stages of the growing season. Using the LULC classification, the

Figure 6 A, B and C: One sample of Land Use Land Cover (LULC) map for each of the three selected years 2000, 2005 and 2011.

Figure 9: Results for Vegetation Condition in Rainfed Agriculture

A 86

B

Figure 7: Mean NDVI for all classes in the 18 images. (F) Forest, (IA) Irrigated Agriculture, (P) Pastureland, (Sh) Shrubland, (RA) Rainfed Agriculture and (W) Water Bodies. 87

C

NDVI statistics were extracted for each of the LULC classes. In this way, it was possible

to verify the ‘greenness’ trends of each class in different stages of the growing season (Fig. 7).

Each class presented consistent patterns of mean NDVI throughout the 18 scenes. 88

Forests consistently presented the highest value trend of NDVI. Forests showed a response of seasonal high NDVI values that peaked during the wet months (June, July, and

August). Shrubs were shown to be seasonal and very responsive to wetness conditions, with the lowest values during the dry months, and rapidly increasing NDVI values in wet months. The agricultural cycle in irrigated agriculture was evident in the NDVI trends for the three analyzed years. This can be determined by how the moderate NDVI values (between 0.4 and 0.5) during the planting months of April and May evolved by June and picked up in August, fading by

November after the harvest.

The mean NDVI pattern in Rainfed Agriculture and Pasture lands was very similar due to their response to rain; however, Rainfed Agriculture consistently showed higher values (Fig.7).

In the year 2000, Rainfed Agriculture followed a similar pattern to Irrigated Agriculture but with lower values that peaked in August and faded by November for the post-harvest. In the year

2005, the deficit in precipitation can be presumed by the very low values in June (the mean

NDVI of the class was 0.21) and then by July, even though the NDVI values increased by almost

15%, by August NDVI declined instead of peaking. June 2005 (Fig. 8) was characterized to be the month where most of the vegetation classes seemed to be stressed, especially Rainfed

Agriculture and Pastureland, with NDVI values ranging between 0.01 and 0.26.

The year 2011 was different (Fig. 7). Even though June presented low NDVI values for most classes, by August NDVI values peaked higher than any other month, especially for those classes that were more responsive to wetness conditions (Rainfed Agriculture, Pastureland, and 89

Shrubs). The values gradually decreased by the following months due to harvesting in September and October. By November, the post-harvest values were the lowest for this year.

3.4 NDVI Change Detection

Coinciding months of the years analyzed were compared (e.g. the image of April 2000 vs.

April 2005 and 2011). Change was detected for the different classes with the purpose of detecting alteration in the vegetative cover that was influenced by drought and not only to normal phenomenology or different plantation conditions. Ten change maps were processed.

Table 4 and 5 show positive and negative change in the mean values of NDVI for each class. To interpret the results, phenomenological patterns and agricultural practices were taken into account.

Planting season in central Mexico starts around the months of May and June, when the rainy season starts. Crops are harvested by the months of September and October. By November, harvest is finished in most areas. Therefore, as it has been described above, in a normal year, seasonal vegetative covers increase their NDVI values during the wet months.

When the year 2005 was compared with the year 2000, a number of changes in NDVI values were detected. The months of April and May resulted in negative changes in most classes

(except Irrigated Agriculture); however, in the t-test performed of 100 random NDVI pixel samples inside the Rainfed Agriculture no statistical significant difference was found. For the months of June, July, and August, significant difference was found and values of negative change up to values of -34% were found (Table 4). Therefore, these negative changes were directly attributed to drought. Vegetation was expected to peak in August; however, it decreased the values of NDVI, resulting in stressed crops of 0.36. 90

For the year 2011, changes in vegetation were diverse (Table 5). During the months of

May and June, the results were similar to that of the pattern in 2005. May did not present a significant difference with the year 2000, and June presented a significant decrease in NDVI values, especially for those classes more dependent on rain for their development. However, the

NDVI values increased significantly by the month of August for all classes, especially Rainfed

Agriculture and Shrubland. The negative changes in NDVI values for June can therefore hardly be attributed to the effects of drought, given that for the following months NDVI values peaked in August, corresponding to a normal or even a wet year where vegetation seemed to be healthy enough by the time of harvest when then the values of NDVI were reduced corresponding to normal harvest and post-harvest conditions.

Table 4: NDVI change 2000-2005 for each of the vegetated LULC classes.

Apr May Jun Jul Aug Nov F -0.34% -3.13% -11.51% 5.21% 0.53% 5.56% IA 8.76% 5.82% -19.05% 11.11% -7.84% -9.85% P -5.52% -0.73% -24.28% -9.61% -13.89% 2.79% SH -11.56% -11.31% -34.16% -8.39% -10.87% 15.66% RA -2.26% 1.63% -17.14% -6.72% -16.18% 1.93%

Table 5: NDVI change 2000-2011 for each of the vegetated LULC classes.

Class May Jun Aug Nov F -4.58% 3.27% 4.52% 6.54% IA 8.68% -7.84% 6.18% 5.01% P 1.26% -19.06% 9.67% 2.58% SH -5.83% -20.55% 17.26% 13.50% RA 0.52% -13.10% 9.59% 1.38% 91

3.5 VCI and the Vegetation Cover Change

The Vegetation Condition Index (VCI) is derived from the values of NDV so the results were used to reinforce the results discussed above and furthermore to have an understanding of how the weather condition of the NDVI acts in response.

The VCI was calculated to understand the effects of weather impact on vegetation. VCI for Rainfed Agriculture (RA) exhibited persistent low values compared to other classes. The evolution of RA (Fig 9) in the year 2000 showed low values during the beginning of the season and peaked in July (at 70%) to then gradually decrease. In the year 2005, VCI peaked at a low value of 38.4% in the early summer, and then kept decreasing until it reached the lowest value of the summer in August (at 27%).

On the other hand, VCI values for 2011 evolved in a similar manner to that of 2000, peaking in late summer at 49% and then again decreasing until the post-harvest month in

November. Even though the results of VCI show that RA persistently had relatively low values, indicating low output, it is noticeable how weather conditions affected directly and most severely

Rainfed Agricultural Crops during the year 2005.

4. Conclusion

An innovative combination of methods was applied in this study with the final objective of detecting crop development and failure during abnormally dry years in the Tortugas-Tepezata subwatershed in central Mexico. The delineation accuracy of agricultural lands and specifically where rainfed farming is practiced was of great importance. For this reason, Saadat et al’s (2011) innovative protocol for LULC classification was used. This protocol was designed for classifying 92 large areas using single date Landsat ETM+ images and incorporating ancillary topographic and climatic data (Saadat et al. 2011). Such a methodology for LULC classification had not been tested for drought monitoring. This methodology was effective in the classification of rainfed agricultural lands for the selected study area in Central Mexico. The LULC maps were then used to extract NDVI values with the purpose of detecting crop failure and development during dry and normal years. Even though there was a limitation in the imagery availability for certain months in each year, a clear pattern of crop failure was found in the three years analyzed.

The NDVI time series was used to differentiate the evolution of each of the vegetation classes and it enabled each one to be distinguished from one another successfully. Pastureland and Rainfed Agriculture classes appeared to have similar patterns of NDVI throughout all the images, and very similar values for summer images (June, July, and August). It is concluded that this result is produced due to their dependency on water for their successful development. The changes perceived in the values of NDVI and VCI supported the conclusion that vegetation condition was affected by the weather patterns in 2005. This was noticed in the patterns and the very low values of NDVI and VCI where, according to the normal year selected, it peaked in

August. This was positively correlated with the results obtained from the precipitation analysis that matched the SPI results indicating clear deficits in precipitation for the year 2005.

On the other hand, an unexpected result occurred out for 2011. Even though the Mexican

Government announced that half of the country was affected by drought in January 2012, it seems that the study area explored in this paper was not affected by this drought. The SPI results showed inconclusive results on the presence of drought because three of the six analyzed meteorological stations presented precipitation results that were contradictory. Aguablanca, El 93

Zembo, and Huasca de Ocampo Meteorological Stations showed dry conditions for the year

2011, even with values of extreme drought (SPI of -4.5). Conversely, Alchoyola, Atotonilco, and

Sta Maria Amajac Meteorological Stations showed wet conditions, peaking in the month of July with values above 1.5, showing very wet conditions. When the NDVI and VCI analysis was conducted, it was concluded that even though for the month of June 2011, the conditions in vegetation seemed to be indicating the presence of a drought, by the end of the summer, the VCI peaked in values even higher than the normal year analyzed. These results could be interpreted in the two following ways:

1. The dry conditions did not occur and the plantation conditions were very different

than those recorded for the year 2000; however, that would not explain the low values

in the other classes.

2. Most likely, dry conditions did appear for the month of June as for the rest of the

country that the government recorded at a national level. However, by the end of June

or sometime in July (where no satellite imagery was available), there was enough

precipitation that allowed vegetation to recuperate rapidly and resulted in peaking

values of NDVI and VCI by August; this slowly decreased as the harvest took place

in the following months.

The methodology presented here resembles that of Lenney (1996), Anyamba (2005),

Fadhil (2011), and Vanderpost (2011) in the use of Landsat products to detect the influences of drought on vegetation using different indices in semi-arid regions. However, the methodology proposed in this paper made improvements to those approaches by detecting the different response of different vegetation classes. Furthermore, in the time series used, it showed more 94 continuity in the years selected and therefore was able to account for vegetation phenomenology, as well as detect when vegetation responded in stress. Moreover, a study of this nature has, to date, not been developed for Central Mexico.

The methodology developed in this paper presented several advantages but also some disadvantages. Using medium resolution images, such as Landsat, allowed for the production of accurate classification of the vegetative cover. Therefore, when analyzing drought, these maps can provide good insights of how vegetation responds to the effects of dry conditions with a high confidence level. However, developing this method for larger regions would be time consuming.

Furthermore, the temporal resolution of Landsat (repeat coverage of 16 days) limits the time series analysis of vegetation status. The method proved to be useful to verify the development of rainfed agriculture for small regions to provide information about the vegetation status and response towards droughts. It could be used for other developing regions where isolated communities depend on rainfed agriculture and other sources of data for detecting are not available.

Vegetation indices were used to detect drought effects by representing the negative impact of adverse meteorological and hydrological conditions. However, some authors have failed to find linear correlations between vegetation indices, hydrological, and meteorological data (Bhuiyan et al. 2006; Bayarjargal 2006; Owrangi 2011). Owrangi concluded that when the scales of the parameters are different, there is the possibility of weak correlation, even when in reality one parameter is entirely dependent on the other. Bhuiyan et al. (2006) argued that negative SPI anomalies do not always correspond to drought. On the contrary, drought may appear in the hydrologic and vegetative spheres in spite of positive SPI, and vice versa. 95

Drought is considered by many to be the most complex but least understood of all natural hazards, affecting more people than any other hazard. It is, however, a normal feature of climate ant its recurrence is inevitable (Wilhite 2003). In nature, droughts are slow processes that take time to creep up and appear and therefore it is difficult to detect, since many variables come into play. Such variables are precipitation patterns, plantation conditions, management, and available irrigation infrastructure.

Because of the complexity of droughts, there is much to be done in the future for drought detection. The integration of data from different sources and spatial and temporal resolutions is essential. Remotely sensed data should be integrated with other sources of data such as static sensing, and mobile sensing systems to aid with the ground truthing of data. Similarly, with the wide availability of handheld sensing systems (e.g., smartphones) that have the potential to take photographic evidence of droughts, georeferencing, and data validation to enhance in situ data quality could give local stakeholders the opportunity to automatically upload relevant data. The aforementioned has been implemented through projects like the Ushahidi Platform , where crowdsourcing has been used by people in developing regions to provide an alert system of disaster or emergency by sensing SMS messages, videos, and images from smartphones or online reports. The long term future of environmental research and emergency relief should focus on increasing computer power to the point where all relevant research and reports can be accessed in an integral way to be displayed at different scales in space and time. Only then can the uncertainty of the diverse influences of variables be better understood. These tools will enable decision makers to take pertinent and well informed actions to manage resources to prevent serious and unavoidable droughts in the future. 96 97

Appendix 1: Confusion Matrices

April 22-2000 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 23 4 0 0 2 29 Irrigated Agriculture 0 1 10 0 0 0 11 Rainfed Agriculture 0 0 0 34 12 0 46

Pastureland 0 0 0 12 18 0 30 Shrubland 0 3 0 5 0 21 29 Column Total 8 27 14 51 30 23 153

May 13-2000 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 23 3 0 0 3 29 Irrigated Agriculture 0 0 9 2 0 0 11 Rainfed Agriculture 0 0 5 35 6 0 46

Pastureland 0 0 0 11 19 0 30 Shrubland 0 3 1 0 0 25 29 Column Total 8 26 18 48 25 28 153

June 25-2000 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 25 4 0 0 0 29 Irrigated Agriculture 0 1 10 0 0 0 11 Rainfed Agriculture 0 0 0 40 6 0 46

Pastureland 0 0 0 10 20 0 30 Shrubland 0 5 0 2 0 22 29 Column Total 8 31 14 52 26 22 153 98 99

July 03-2000 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 23 4 0 0 2 29 Irrigated Agriculture 0 0 9 2 0 0 11 Rainfed Agriculture 0 0 4 36 6 0 46

Pastureland 0 0 0 10 20 0 30 Shrubland 0 2 1 0 0 26 29 Column Total 8 25 18 48 26 28 153

August 28-2000 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 26 3 0 0 0 29 Irrigated Agriculture 0 1 10 0 0 0 11 Rainfed Agriculture 0 0 0 38 8 0 46

Pastureland 0 0 0 7 23 0 30 Shrubland 0 5 0 2 0 22 29 Column Total 8 32 13 47 31 22 153

November 16-2000 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 24 3 0 0 2 29 Irrigated Agriculture 0 0 9 2 0 0 11 Rainfed Agriculture 0 0 3 34 9 0 46

Pastureland 0 0 0 10 20 0 30 Shrubland 0 5 0 3 0 21 29

Column Total 8 29 15 49 29 23 153 100

April 04, 2005

Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 24 4 0 0 1 29 Irrigated Agriculture 0 1 10 0 0 0 11 Rainfed Agriculture 0 0 0 36 10 0 46

Pastureland 0 0 0 12 18 0 30 Shrubland 0 3 0 4 0 22 29 Column Total 8 28 14 52 28 23 153

May 22, 2005 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 22 3 0 0 4 29 Irrigated Agriculture 0 0 9 2 0 0 11 Rainfed Agriculture 0 0 5 35 6 0 46

Pastureland 0 0 1 11 18 0 30 Shrubland 0 3 1 0 0 25 29 Column Total 8 25 19 48 24 29 153

Jun 07, 2005 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 25 4 0 0 0 29 Irrigated Agriculture 0 0 9 2 0 0 11 Rainfed Agriculture 0 0 0 40 6 0 46

Pastureland 0 0 0 12 18 0 30

Shrubland 0 3 0 4 0 22 29 Column Total 8 28 13 58 24 22 153 101

102

July 09, 2005 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 25 4 0 0 0 29 Irrigated Agriculture 0 0 10 0 0 1 11 Rainfed Agriculture 0 0 0 39 7 0 46

Pastureland 0 0 0 10 20 0 30 Shrubland 0 6 0 2 0 21 29 Column Total 8 31 14 51 27 22 153

August 10, 2005 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 23 4 0 0 2 29 Irrigated Agriculture 0 0 9 2 0 0 11 Rainfed Agriculture 0 0 4 36 6 0 46

Pastureland 0 0 0 10 20 0 30 Shrubland 0 2 1 0 0 26 29 Column Total 8 25 18 48 26 28 153

November 14, 2005 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 22 3 0 0 4 29 Irrigated Agriculture 0 0 9 2 0 0 11 Rainfed Agriculture 0 0 5 35 6 0 46

Pastureland 0 0 1 11 18 0 30 Shrubland 0 3 1 0 0 25 29 Column Total 8 25 19 48 24 29 153 103

May 23, 2011 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 25 4 0 0 0 29 Irrigated Agriculture 0 1 10 0 0 0 11 Rainfed Agriculture 0 0 0 37 9 0 46

Pastureland 0 0 0 12 18 0 30 Shrubland 0 3 0 4 0 22 29 Column Total 8 29 14 53 27 22 153

June 08, 2011 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 0 0 0 0 0 0 Irrigated Agriculture 0 25 3 1 0 0 29 Rainfed Agriculture 0 1 10 0 0 0 11

Pastureland 0 0 0 38 8 0 46 Shrubland 0 0 0 7 23 0 30 Column Total 0 5 0 2 0 22 29

August 11, 2011 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 26 3 0 0 0 29 Irrigated Agriculture 0 1 10 0 0 0 11 Rainfed Agriculture 0 0 0 37 9 0 46

Pastureland 0 0 0 12 18 0 30 Shrubland 0 3 0 2 0 24 29 Column Total 8 30 13 51 27 24 153 104

September 28, 2011 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 25 4 0 0 0 29 Irrigated Agriculture 0 1 10 0 0 0 11 Rainfed Agriculture 0 0 0 37 9 0 46

Pastureland 0 0 0 12 18 0 30 Shrubland 0 3 0 4 0 22 29 Column Total 8 29 14 53 27 22 153

October 30, 2011 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 25 4 0 0 0 29 Irrigated Agriculture 0 0 9 2 0 0 11 Rainfed Agriculture 0 0 0 40 6 0 46

Pastureland 0 0 0 12 18 0 30 Shrubland 0 3 0 4 0 22 29 Column Total 8 28 13 58 24 22 153

November 15, 2011 Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total ------Water Body 8 0 0 0 0 0 8 Forest 0 22 3 0 0 4 29 Irrigated Agriculture 0 0 9 2 0 0 11 Rainfed Agriculture 0 0 5 35 6 0 46

Pastureland 0 0 0 11 19 0 30 105

Shrubland 0 3 1 0 0 25 29 Column Total 8 25 18 48 25 29 153 106

Appendix 2: Accuracy Totals

April 22-2000

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 27 29 23 85.19% 79.31% Irrigated Agriculture 14 11 10 71.43% 90.91% Rainfed Agriculture 51 46 34 66.67% 73.91% Pastureland 30 30 18 60.00% 60.00% Shrubland 23 29 21 91.30% 72.41% Totals 153 153 114 Overall Classification Accuracy = 74.51%

May 13-2000

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 0 0 0 ------Irrigated Agriculture 26 29 23 88.46% 79.31% Rainfed Agriculture 18 11 9 50.00% 81.82% Pastureland 48 46 35 72.92% 76.09% Shrubland 28 29 25 89.29% 86.21% Totals 153 153 119 Overall Classification Accuracy = 77.78%

June 25-2000

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 31 29 25 80.65% 86.21% Irrigated Agriculture 14 11 10 71.43% 90.91% Rainfed Agriculture 52 46 40 76.92% 86.96% Pastureland 26 30 20 76.92% 66.67% Shrubland 22 29 22 100.00% 75.86% Totals 153 153 125 Overall Classification Accuracy =81.70 % 107 108

July 03-2000

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 0 0 0 ------Irrigated Agriculture 25 29 23 92.00% 79.31% Rainfed Agriculture 18 11 9 50.00% 81.82% Pastureland 48 46 36 75.00% 78.26% Shrubland 28 29 26 92.86% 89.66% Totals 153 153 122 Overall Classification Accuracy = 80.06%

August 28-2000

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Clouds Water Body 8 8 8 100.00% 100.00% Forest 32 29 26 81.25% 89.66% Irrigated Agriculture 13 11 10 76.92% 90.91% Rainfed Agriculture 47 46 38 80.85% 82.61% Pastureland 31 30 23 74.19% 76.67% Shrubland 22 29 22 100.00% 75.86% Totals 153 153 127 Overall Classification Accuracy = 83.01%

November 16-2000

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 0 0 0 ------Irrigated Agriculture 29 29 24 82.76% 82.76% Rainfed Agriculture 15 11 9 60.00% 81.82% Pastureland 49 46 34 69.39% 73.91% Shrubland 23 29 21 91.30% 72.41% Totals 153 153 116 Overall Classification Accuracy = 77.82% 109

April 04-2005

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 28 29 24 85.71% 82.76% Irrigated Agriculture 14 11 10 71.43% 90.91% Rainfed Agriculture 52 46 36 69.23% 78.26% Pastureland 28 30 18 64.29% 60.00% Shrubland 23 29 22 95.65% 75.86% Totals 153 153 118 Overall Classification Accuracy = 78.77%

May 22-2005

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 25 29 22 88.00% 75.86% Irrigated Agriculture 19 11 9 47.37% 81.82% Rainfed Agriculture 48 46 35 72.92% 76.09% Pastureland 24 30 18 75.00% 60.00% Shrubland 29 29 25 86.21% 86.21% Totals 153 153 117 Overall Classification Accuracy = 76.47%

June 07-2005

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 28 29 25 89.29% 86.21% Irrigated Agriculture 13 11 9 69.23% 81.82% Rainfed Agriculture 58 46 40 68.97% 86.96% Pastureland 24 30 18 75.00% 60.00% Shrubland 22 29 22 100.00% 75.86% Totals 153 153 122 Overall Classification Accuracy = 79.74% 110

July 09-2005

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 31 29 25 80.65% 86.21% Irrigated Agriculture 14 11 10 71.43% 90.91% Rainfed Agriculture 51 46 39 76.47% 84.78% Pastureland 27 30 20 74.07% 66.67% Shrubland 22 29 21 95.45% 72.41% Totals 153 153 123 Overall Classification Accuracy = 80.39%

August 10-2005

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 25 29 23 92.00% 79.31% Irrigated Agriculture 18 11 9 50.00% 81.82% Rainfed Agriculture 48 46 36 75.00% 78.26% Pastureland 26 30 20 76.92% 66.67% Shrubland 28 29 26 92.86% 89.66% Totals 153 153 122 Overall Classification Accuracy = 81.49%

November 14-2005

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 0 0 0 ------Irrigated Agriculture 25 29 22 88.00% 75.86% Rainfed Agriculture 19 11 9 47.37% 81.82% Pastureland 48 46 35 72.92% 76.09% Shrubland 29 29 25 86.21% 86.21% Totals 153 153 117 111

Overall Classification Accuracy =76.47% 112

May 23, 2011

May 13-2000 Overall Kappa Statistics 0.7207 Class Name Kappa ------Water 1 Forest 0 Irrigated Agriculture 0.7507 Rainfed Agriculture 0.7939 Pastureland 0.6516 Shrubland 0.8312

July 03-2000 Overall Kappa Statistics 0.7453 Class Name Kappa ------Water 1 Forest 0 Irrigated Agriculture 0.7527 Rainfed Agriculture 0.7939 Pastureland 0.6832 Shrubland 0.8734

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 29 29 25 86.21% 86.21% Irrigated Agriculture 14 11 10 71.43% 90.91% Rainfed Agriculture 53 46 37 69.81% 80.43% Pastureland 27 30 18 66.67% 60.00% Shrubland 22 29 22 100.00% 75.86% 113

June 25-2000 Overall Kappa Statistics 0.7682 Class Name Kappa ------Water 1 Forest 0 Irrigated Agriculture 0.827 Rainfed Agriculture 0.8999 Pastureland 0.8024 Shrubland 0.7181

Totals 153 153 120 Overall Classification Accuracy = 78.86%

June 08, 2011 August 28-2000 November 16,-2000 May 22, 2005 Overall Kappa Statistics 0.7855 Overall Kappa Statistics 0.6948 Overall Kappa Statistics Class Name Kappa Class Name Kappa Class Name Kappa ------Water 1 Water 1 Water 1 Forest 0.8692 Forest 0 Forest 0.7115 Irrigated Agriculture 0.9006 Irrigated Agriculture 0.7873 Irrigated Agriculture 0.7924 Rainfed Agriculture 0.749 Rainfed Agriculture 0.7984 Rainfed Agriculture 0.6516 Pastureland 0.7074 Pastureland 0.6162 PasturelandShrubland 0.71810.5256 Shrubland 0.6753 Shrubland 0.8298

April 04, 2005 Overall Kappa Statistics 0.7102 Class Name Kappa ------Water 1 Forest 0.789 Irrigated Agriculture 0.8999

Rainfed Agriculture 0.6707 Pastureland 0.5104 Shrubland 0.7159 June07, 2005 Overall Kappa Statistics 0.7417 Class Name Kappa ------Water 1 Forest 0.8312 Irrigated Agriculture 0.8013 Rainfed Agriculture 0.7899 Pastureland 0.5256 Shrubland 0.7181 114

July 09, 2005 Overall Kappa Statistics0.7519 Class Name Kappa ------Water 1 Forest 0.827 Irrigated Agriculture 0.8999 Rainfed Agriculture 0.7717 Pastureland 0.5952 Shrubland 0.6778 November 14,2005 Overall Kappa Statistics0.7045 Class Name Kappa August 10, 2005 ------Overall Kappa Statistics 0.7453 Water 1 Class Name Kappa Forest 0.7115 ------Irrigated Agriculture 0.7924 Water 1 Rainfed Agriculture 0.6516 Forest 0.7527 Pastureland 0.5256 Irrigated Agriculture 0.7939 Shrubland 0.8298 Rainfed Agriculture 0.6832 Pastureland 0.5984 0.8734 Shrubland

June 08, 2011 Overall Kappa Statistics 0.777 Class Name Kappa ------Water 1 Forest 0.827 Irrigated Agriculture 0.9006 Rainfed Agriculture 0.7466 Pastureland 0.7074 Shrubland 0.7181

October 30, 2011 0.7530 August 11, 2011 Overall Kappa Statistics Overall Kappa Statistics 0.7516 Class Name Kappa Class Name Kappa ------Water 1 Water 1 Forest 0.8312 Forest 0.8713 Irrigated Agriculture 0.8013 Irrigated Agriculture 0.9006 Rainfed Agriculture 0.7899 Rainfed Agriculture 0.7065 Pastureland 0.5256 Pastureland 0.5143 Shrubland 0.7181 Shrubland 0.7955 115

November 15, 2011 September 28, 2011 Overall Kappa Statistics Overall Kappa Statistics 0.7265 Class Name Kappa Class Name Kappa ------Water 1 Water 1 Forest 0.7115 Forest 0.8298 Irrigated Agriculture 0.7939 Irrigated Agriculture 0.8999 Rainfed Agriculture 0.6516 Rainfed Agriculture 0.7007 Pastureland 0.5617 Pastureland 0.5143 Shrubland 0.8298 Shrubland 0.7181

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 31 29 25 80.65% 86.21% Irrigated Agriculture 13 11 10 76.92% 90.91% Rainfed Agriculture 48 46 38 79.17% 82.61% Pastureland 31 30 23 74.19% 76.67% Shrubland 22 29 22 100.00% 75.86% Totals 153 153 126 Overall Classification Accuracy =82.35 %

August 11, 2011

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 30 29 26 86.67% 89.66% Irrigated Agriculture 13 11 10 76.92% 90.91% Rainfed Agriculture 51 46 37 72.55% 80.43% Pastureland 27 30 18 66.67% 60.00% Shrubland 24 29 24 100.00% 82.76% 116

Totals 153 153 123 Overall Classification Accuracy = 80.39% 117

September 28, 2011

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 29 29 25 86.21% 86.21% Irrigated Agriculture 14 11 10 71.43% 90.91% Rainfed Agriculture 53 46 37 69.81% 80.43% Pastureland 27 30 18 66.67% 60.00% Shrubland 22 29 22 100.00% 75.86% Totals 153 153 120 Overall Classification Accuracy = 78.77%

October 30, 2011

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 28 29 25 89.29% 86.21% Irrigated Agriculture 13 11 9 69.23% 81.82% Rainfed Agriculture 58 46 40 68.97% 86.96% Pastureland 24 30 18 75.00% 60.00% Shrubland 22 29 22 100.00% 75.86% Totals 153 153 122 Overall Classification Accuracy = 79.77%

November 15, 2011

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy

Water Body 8 8 8 100.00% 100.00% Forest 0 0 0 ------Irrigated Agriculture 25 29 22 88.00% 75.86% Rainfed Agriculture 18 11 9 50.00% 81.82% Pastureland 48 46 35 72.92% 76.09% Shrubland 25 30 19 76.00% 63.33% Totals 153 153 118 Overall Classification Accuracy = 77.12% 118

Appendix 3: Kappa Index

April 22-2000 Overall Kappa Statistics 0.6774 Class Name Kappa ------Water 1 Forest 0 Irrigated Agriculture 0.7488 Rainfed Agriculture 0.8999 Pastureland 0.6087 Shrubland 0.6753 119 120 121

References

Appendini, Kirsten, & Diana Liverman. (1994). “Agricultural policy, climate change and food

security in Mexico.” Food Policy 19.2 (1994): 149-164. Web. 14 Oct 2013.

Barbosa, H. A., A.R. Huete, and W.E. Baethgen. “A 20-Year Study of NDVI Variability over the

Northeast Region of Brazil.” Journal of Arid Environments 67.2 (2006): 288-307. Web.

17 Oct 2013

Boken, Vijendra K., et. al. “Monitoring and Predicting Agricultural Drought: A Global Study.”

Ebrary. Ebrary, 2005. Web. 16 Oct 2013.

Cihlar, J. “Land Cover Mapping of Large Areas from Satellites: Status and Research Priorities.”

International Journal of Remote Sensing 21.6-7 (2000): 1093-1114. Print.

Cohen, Y., and M. Shoshany. “A National Knowledge-Based Crop Recognition in Mediterranean

Environment.” ITC Journal 2002.4 (2002): 75-87. Print.

CONANP. (2003). Programa de Manejo Reserva de La Biosfera Barranca de Metztitlán. Mexico

DF.

Congalton, Russell G. “A Review of Assessing the Accuracy Of Classifications of Remotely

Sensed Data.” Remote Sensing of Environment 37.1 (1991): 35-46. Print. de Asis, Alejandro M., and Kenji Omasa. “Estimation of Vegetation Parameter for Modeling Soil

Erosion Using Linear Spectral Mixture Analysis of Landsat ETM Data. ISPRS Journal of

Photogrammetry and Remote Sensing 62.4 (2007): 309-324. Web. 15 Oct 2013.

Eakin, Hallie. (2000). “Smallholder Maize Production and Climatic Risk: A Case Study from

Mexico.” Climatic Change 45.1 (2000): 19-36. Web. 16 Oct 2013. 122

Eakin, Hallie. “Institutional Change, Climate Risk, and Rural Vulnerability: Cases from Central

Mexico.” World Development 33.11 (2005): 1923-1938. Web. 14 Oct 2013.

Eakin, Hallie, and Amy Lynd Luers. “Assessing the Vulnerability of Social-Environmental

Systems.” Annual Review of Environment and Resources 31.1 (2006): 365-394. Web. 13

Oct 2013.

Fadhil, Ayad Mohammed. “Drought Mapping Using Geoinformation Technology for Some Sites

in the Iraqi Kurdistan Region.” International Journal of Digital Earth 4.3 (2011):

239-257. Web. 15 Oct 2013.

Focardi, Silvia, et. al. “Satellite-Based Indices in the Analysis of Land Cover for Municipalities

in the Province of Siena, Italy.” Journal of Environmental Management 86.2 (2008):

383-389. Web. 09 Oct 2013.

Im, J., J.R. Jensen, and J.A. Tullis. “Object-Based Change Detection Using Correlation Image

Analysis and Image Segmentation.” International Journal of Remote Sensing 29.2

(2008): 399-423. Print.

Joshi, P. K., et. al. “Vegetation Cover Mapping in India Using Multi-Temporal IRS Wide Field

Sensor (WiFS) Data.” Remote Sensing of Environment 103.2 (2006): 190-202. Web. 09

Oct 2013.

Kogan, F. N. (1990). “Remote sensing of weather impacts on vegetation in non-homogeneous

areas.” International Journal of Remote Sensing 11.8 (1990): 1405-1419. Web. 13 Oct

2013.

Kogan, F. N. “Application of Vegetation Index and Brightness Temperature for Drought

Detection.” Advances in Space Research 15.11 (1995): 91-100. Web. 12 Oct 2013. 123

Kogan, Felix N. “Droughts of the Late 1980s in the United States as Derived from NOAA Polar-

Orbiting Satellite Data.” Bulletin of the American Meteorological Society 76.5 (1995):

655-668. Web. 12 Oct 2013.

Labrecque, S., et. al. “A Comparison of Four Methods to Map Biomass from Landsat-TM and

Inventory Data in Western Newfoundland.” Forest Ecology and Management 226.1-3

(2006): 129-144. Print.

Lenney, Mary Pax, et. al. “The Status of Agricultural Lands in Egypt: The Use of Multitemporal

NDVI Features Derived from Landsat TM.” Remote Sensing of Environment 56.1 (1996):

8-20. Web. 08 Oct 2013.

Liverman, Diana M. “Drought Impacts in Mexico: Climate, Agriculture, Technology, and Land

Tenure in Sonora and Puebla.” Annals of the Association of American Geographers 80.1

(1990): 49-72. Print.

McKee, T.B.N., J. Doesken, and J. Kleist. “Drought Monitoring with Multiple Time Scales.”

Ninth Conference of Applied Climatology. American Meteorological Society (1993):

233-236. Print.

Mishra, Ashok K., and Vijay P. Singh. “A Review of Drought Concepts.” Journal of Hydrology

391.1–2 (2010): 202-216. Web. 15 Oct 2013.

Mishra, Ashok K., and Vijay P. Singh. “Drought Modeling – A Review.” Journal of Hydrology

403.1–2 (2011): 157-175. Web. 15 Oct 2013.

Peters, Albert J., et. al. “Drought Monitoring with NDVI-Based Standardized Vegetation Index.”

Photogrammetric Engineering and Remote Sensing, 68.1 (2002): 71-75. 124

Peters, Albert J., et. al. “Satellite Assessment of Drought Impact on Native Plant Communities of

Southeastern New Mexico, U.S.A.” Journal of Arid Environments 24.3 (1993): 305-319.

Web. 08 Oct 2013.

29.11 (1995): 2449-2454. Print.

Rahimzadeh Bajgiran, Parinaz, et. al. “Using AVHRR-Based Vegetation Indices for Drought

Monitoring in the Northwest of Iran.” Journal of Arid Environments 72.6 (2008):

1086-1096. Web. 12 Oct 2013.

Rahimzadeh-Bajgiran, Parinaz, Kenji Omasa, and Yo Shimizu. “Comparative Evaluation of the

Vegetation Dryness Index (VDI), the Temperature Vegetation Dryness Index (TVDI) and

the Improved TVDI (iTVDI) for Water Stress Detection In Semi-Arid Regions of Iran.”

ISPRS Journal of Photogrammetry and Remote Sensing 68.0 (2012): 1-12. Web. 12 Oct

2013.

Renschler, Chris S., and Jon Harbor. “Soil Erosion Assessment Tools from Point to Regional

Scales—The Role of Geomorphologists in Land Management Research and

Implementation.” Geomorphology 47.2–4 (2002): 189-209. Print.

Saadat, H., et. al. “Land Use and Land Cover Classification over a Large Area in Iran Based on

Single Date Analysis of Satellite Imagery.” ISPRS Journal of Photogrammetry and

Remote Sensing 66.5 (2011): 608-619. Print.

Sivanpillai, Ramesh, et. al. “Estimating Regional Forest Cover in East Texas Using Advanced

Very High Resolution Radiometer (AVHRR) Data.” International Journal of Applied

Earth Observation and Geoinformation 9.1 (2007): 41-49. Web. 15 Oct 2013. 125

Story, Michael, and Russell G. Congalton. “Accuracy Assessment-A User's Perspective.”

Photogrammetric Engineering and Remote Sensing 52.3 (1986): 397-399. Print.

Wardlow, Brian D., Stephen L. Egbert, and Jude H. Kastens. “Analysis of Time-Series MODIS

250 m Vegetation Index Data for Crop Classification in the U.S. Central Great Plains.”

Remote Sensing of Environment 108.3 (2007): 290-310. Web. 08 Oct 2013.

Wilhite, Donald A., Mark D. Svoboda, and Michael J. Hayes. “Understanding the Complex

Impacts of Drought: A Key to Enhancing Drought Mitigation and Preparedness.” Water

Resources Management 21.5 (2007): 763-774. Web. 12 Oct 2013.

Xiaofan, Liu, et. al. “Assessing Vegetation Response to Drought in the Laohahe Catchment,

North China.” Hydrology Research 43.1/2 (2012): 91-101. Web. 15 Oct 2013. 126

Preface to Chapter 4

Chapter 3 presented the analysis of a time series analysis of Standardized Precipitation Index (SPI) in order to detect periods of time with normal and abnormally dry conditions. Then a new GIS based simplified version of the algorithm to perform land use land cover (LULC) classification presented by Saadat et al (2011) and then vegetation indices and change detection was performed with the objective of detecting drought. Chapter 4 describes a more detailed analysis of LULC classification with the aim of performing the highest overall classification accuracy possible. This was performed using early summer and late summer stages to identify the effects of drought in agriculture. A special attention was made to rainfed agricultural systems.

Chapter 4 has been submitted to the Journal of Agriculture and Environment for International

Development. The manuscript has been co-authored by Andres Sierra-Soler, Jan Adamowski,

Hossein Saadat, Zhiming Qi and Santosh Pingale. 127

Chapter 4

High Accuracy Land Use Land Cover (LULC) Maps for Detecting Agricultural Drought

Effects in Rainfed Agro-Ecosystems in Central Mexico.

A. Sierra-Soler, J. Adamowski, H. Saadat, Z. Qi, S. Pingale Department of Bioresource Engineering, Macdonald Campus, McGill University, 21111 Lakeshore, Ste-Anne-de-Bellevue, QC, Canada, H9X3V9 [email protected], [email protected], [email protected], [email protected] Abstract

Satellite remote sensing provides a synoptic view of the land and a spatial context for measuring drought impacts, which have proved to be a valuable source of spatially continuous data with improved information on monitoring vegetation dynamics. Many studies have focused on detecting drought effects over large areas, given the wide availability of low-resolution images. In this study, however, the objective was to focus on a smaller area (1085 km2) using

Landsat ETM+ images (multispectral resolution of 30m and 15 m panchromatic), and to process very accurate Land Use Land Cover (LULC) classification to determine with great precision the effects of drought in specific classes. The LULC classification was processed using a new method based on available ancillary information plus analysis of three single date satellite images. The newly developed LULC methodology developed produced overall accuracies ranging from 87.88% to 92.42%. Spectral indices for vegetation and soil/vegetation moisture were used to detect anomalies in vegetation development caused by drought and furthermore, the 128 area of water bodies was measured and compared to detect changes in water availability for irrigated crops.

Key Words: remote sensing, drought, LULC maps, Spectral Indices

1. Introduction

Droughts occur in all climatic zones and are mostly related to the reduction of precipitation received over an extended period, such as a season or a year. The impacts of a drought increase slowly and often accumulate over the months and may linger years after the end of the drought. Unlike other natural hazards, the impacts of droughts are non-structural and spread over large geographical areas. Droughts rank first among all natural hazards when

measured in terms of the number of people Figure 1: Location of the Study Area affected (Obasi 1994; Hewitt 1997; Wilhite

2000; Mishra et al. 2010), which emphasizes

the importance of investigating how this

natural phenomenon disturbs ecological and

socio-economic systems.

Droughts affect virtually all regions; however,

it does so in such a variety of ways that each

system is apt to have its own concept of

drought, and its own criteria for gauging the

severity of drought (Whitmore 2005).

Defining drought is therefore difficult (Redmond 2002), since it is relative to the system of interest. Differences in hydro-meteorological variables and socio-economic factors as well as the 129 stochastic nature of water demands in different regions around the world have become an obstacle to having a universal definition of drought (Mishra et al. 2010).

This study focused on developing an original methodology for investigating agricultural drought in semiarid developing regions where drought poses a great threat to rainfed smallholder farmers. The methodology was tested in central Mexico but has the potential to be used in other regions with similar conditions. Rainfed agricultural production in these areas is focused mainly on producing self-sufficiency crops for the poorest sectors in society that depend on this economic activity. Agricultural drought occurs when there is not enough available plant soil moisture in the root zone (Johan 2003). Agricultural drought is generally characterized by two key factors: the estimated water demand and expected water supply. The formulation of these key factors for a region largely depends on the agro-climatic conditions (Yurekli and Kurunc

2006). The capacity of soils to retain and release water depends on factors such as soil texture, depth, structure, organic matter content, and biological activity (Bot 2005).

Remotely sensed approaches have been used to detect agricultural drought, and are often associated with image reflectance using spectral indices to indicate vegetation condition such as the normalized difference vegetation index (NDVI) and soil/vegetation wetness such as the normalized difference wetness index (NDWI) and the tasseled cap wetness (TCW), described below.

The normalized difference vegetation index (NDVI) is a spectral index that has been shown to be highly correlated with parameters associated with plant health and productivity, and has been widely used for drought detection (Lenney 1995; Kogan 1995; Rahimzadeh 2002;

Peters et al. 2002; Barbosa et al. 2006). The NDVI has been used successfully to identify stressed 130 and damaged crops and pastures but interpretive problems arise when results are extrapolated over non-homogeneous areas. It can be said that NDVI has two components: ecology and weather. The need of multi temporal NDVI analysis for detecting drought is that, in any single

NDVI image in a given growing season, barren fields may be indistinguishable from temporarily fallow healthy fields (Wallace et al. 1993), and immature crops with low-density cover could be confused with poor crops (Lenney et al. 1996). Therefore, when NDVI has been used for the analysis of weather impact on vegetation, the weather component must be separated from the ecosystem component (Rahimzadeh et al. 2008). It has also been found that NDVI correlates with net primary production, biomass, vegetation fraction, and yield (Goward et al. 1987;

Maselli et al. 1992; Rasmussen 1992; Quarmby et al. 1993; Hayes and Decler 1996; Korgan

1997; Unganai and Korgan 1998; Lenney et al. 1996).

The NDWI proposed by Gao (1996) can be obtained using the near infrared (NIR) and shortwave infrared (SWIR) channels. He found that NDWI is a measure of liquid water molecules in vegetation canopies that interacted with the incoming solar radiation and is less sensitive to atmospheric scattering effects than NDVI. Guet al. (2007 and 2008) evaluated the relationship between NDVI and NDWI using MODIS 500-m satellite imagery in Oklahoma and

Kansas, USA. They found good correlation between the two indices and furthermore NDWI was more sensitive to drought than NDVI. Rhee et al. (2010) sought to identify a drought index that had the possibility to be used for agricultural drought monitoring in arid/semiarid regions, as well as humid/sub-humid regions, while also using MODIS data. They found that NDWI has a better response in arid/semiarid regions than in humid regions when detecting drought because of good correlations with other precipitation and temperature indices. It is convenient to detect 131 drought at larger scales using coarse resolution imagery such as MODIS; however, this approach cannot provide detailed land cover response to precipitation anomalies.

In semi-arid developing regions, medium resolution remotely sensed data has been used to assess the impacts of drought. Vanderpost et al. (2011) used Landsat imagery to assess the long-time condition of rangeland in semiarid areas of Botswana. By calculating vegetation indices, they found significant degradation in vegetation corresponding to the droughts between

1984 and 2000. The technique used was limited by the many gaps in the long-time coverage, and hence there was a lack of continuity in terms of change.

The tasseled cap transformation (TCT) has been used widely for vegetation mapping and monitoring land cover change (Oetter et al. 2001; Jin and Sader 2005; Fadhil 2011). The TCT of

Landsat thematic mapper consists of six multispectral features, all of which could be potentially differentiated in terms of stability and change in a multitemporal data set (Jin 2004). The first three features have been labeled brightness, greenness, and wetness (band 1, 2, and 3, respectively). The third feature, wetness, has been shown to be sensitive to soil plant moisture

(Jin and Sader 2004). Tasseled cap wetness (TCW) contrasts the sum of the visible and the near- infrared bands with the sum of the shortwave bands ( Jin and Sader 2004).

The NDWI and TCW have both been used in studies to detect drought or disturbance in ecosystems. Jin and Sader (2004) used a time series of both NDWI and TCW derived from

Landsat to compare forest disturbances in Maine, USA. They found high correlations (>0.95) between the two indices. Fadhil (2011) used NDWI and TCW to detect drought effects on vegetation in the Iraqi Kurdistan region. He derived both spectral indices from two Landsat images from consecutive years to calculate five vegetation and soil/vegetation moisture indices 132 and performed change detection. This study did not include a classification of the vegetation, so it is unclear how the different kinds of vegetation were affected by the drought.

The response to drought effects is differentiated by the vegetation cover, but none of the above-mentioned spectral indices has the potential to classify on their own crops from other classes. The knowledge of Land Use Land Cover (LULC) plays an important role in identifying areas where the effects of drought can cause damage in ecosystems and in agricultural lands. The identification of rainfed crops and dynamics over different precipitation conditions has not been Figure 2: Image Preprocessing thoroughly and explicitly explored in the past. In this study, the delineation accuracy of Landsat ETM+ Roads Map 1:50 agricultural lands and specifically where dry-land farming000 is practiced was of great importance.

There are many approaches that have been used to associate image reflection data with FLAASH module vegetation characteristics. Over the lastprocessing decades, various studies have shown the efficacy of satellite imagery in characterizing vegetation cover (Joshi et al. 2006; de Asis Omasa 2007; Geo-referencing Focadi et al. 2008) forests (Labrecque et al. 2006; Sivanpillai et al. 2007) and crops (Cohen and

Study Area Boundary Shoshany 2002; Wardlow et al. 2007).Geo-referenced Additionally, images several authors have used LULC maps in satellite based agricultural drought monitoring (Diouf and Lambin 2001; Wilhelmi and Wilhite

Clip 2002; Rhee et al. 2010).

In this study, the objective was to developClipped and an Geo-Referenced original methodology Images to develop very accurate Land Use Land Cover (LULC) maps to study in depth the effects of the drought in rainfed agricultural lands in Central Mexico. Four medium resolution Landsat ETM+ images were used, corresponding to early and late summer where vegetation in a normal year is in development and then reaches a mature state and before the harvest commences. Such months in

Central Mexico correspond to June and August and the years analyzed were 2000 and 2005 133 because according to Sierra-Soler et al. (submitted), precipitation analysis was found to have normal conditions. After the classification of the four images was completed, spectral indices indicating vegetation condition (NDVI) and soil/vegetation moisture (NDWI and TCW) were calculated. The classification helped to interpret how different vegetative covers were affected by the low precipitation registered in 2005. This methodology has the potential to be used as a tool to identify in detail the effects of drought in rainfed agricultural lands in developing regions, and it can also be used as a mechanism to prevent and provide relief in the event of droughts.

2. Study Area

The Tortugas-Tepezata sub watershed (Fig.1) is one of the sub-watersheds of the

Moctezuma River and is located in the state of Hidalgo in central Mexico. The Moctezuma River is inside the Panuco hydrological region that flows from the Valley of Mexico into the Gulf of

Mexico. The Panuco River Hydrologic Region covers most of the state of Hidalgo, with runoff coefficients in the range of 20 to 30% (CONANP 2003). This hydrologic region has been divided into two: High Panuco and Lower Panuco. The study area is in the first, the High Panuco.

The interest of studying this semiarid area of central Mexico is due to the importance of rainfed agriculture in the region. Even though there are other economic activities such as tourism, the majority of the rural towns in the area depend directly or indirectly on local agriculture for food security (CONAPO 2010). Furthermore, social exclusion, poverty, illiteracy, and access to services are the variables that make this particular region vulnerable in the event of droughts.

The Metztitlán basin, along with Tortugas Tepezata basin, originates in the Amajac River.

The main feature of the sub watershed is the river canyon that flows 100 km and is named 134 according to the region. There are three identified three major sections, the first entry to South

Canyon with the Rio Grande Tulancingo, the second at the junction with the San Sebastian River and Venados River, and the third starts at the District 08 Metztitlán named Metztitlán River, north of the sub-basin. This river runs SE to NW and N flows into the Metztitlán Lagoon north of Tortugas-Tepezata watershed. In the geological past, the river ran without forming the lagoon, but during the Holocene the Cerro El Tajo suffered a massive collapse forming the reservoir

(CONANP 2003). The climate is generally warm, dry, and semidry in different parts, which is determined by the rain shadow effect the Sierra Madre Oriental has on this region. The mean

o annualFigure temperature 3: Extraction is around of Sampling 20-24 PointC. In Location the rainy Map season, in summer, trade winds release their moisture on the windward side and the elevated parts of the mountain range, where forests are Climatic Zones and Clipped and Geo-Referenced Images Topography pine and oak, among other cold temperate vegetation (CONANP 2003).

Unsupervised Classification In Mexico, drought monitoring at the national scale is performed by North America Combine

Drought Monitor (NADM), which is a cooperative effort betweenRaw Classified drought Map experts in the United

Initial Homogeneous Map States, Mexico, and Canada. The data produced by NADM focuses only on three data driven indexes: the SPI, the PDSI, and the Percent of Average Precipitation. Other studies in Mexico Combine have reported qualitative drought impacts, documenting smallholder rainfed maize production

Potential LU map and climatic risk (Eakin 2000), climate change impacts on food security (Appendini and

Liverman 1994), estimates of the effectsSampling of El Point Niño Location Southern Map Oscilation (ENSO) on crop yield

(Adams et al. 2003) and vulnerability and adaptation to drought (Liverman 1990). However, very little has been done in the region using remotely sensed data in combination with meteorological analysis to estimate the effects of drought on rainfed agricultural production, making this study both original and significant. 135

3. Materials and Methods

3.1 Materials

1. Four Landsat ETM+ Satellite Images: Corresponding to the available images for

initial and final stages of summer (June 25 and August 28, 2000 and June 07 and

August 10, 2005). These images were chosen because they were all the available

images available that were clear from clouds for the desired period. Each scene

selected included the entire surface of the study area. The scenes were selected for

different months of different years, with the purpose of comparing vegetation

development in seasons with different precipitation conditions.

2. Two digital ancillary layers were also collected to assist in the interpretation and

classification of the remotely sensed data. These were a 1:25 000 topographic map,

and a 1:50 000 scale climatic zone map both made available by CONABIO (2005).

Figure 4: SupervisedThese maps Classification were used. to follow Saadat et al.’s (2010) methodology for the

Clipped and Geo- Sampling Pont referenced Imagesdevelopment of the LULC maps.

3. Ground Truth Data: 132 samples were collected Identifyingvia fieldwork across all the study Field Data Collection additional areas where roads permitted access. Each samplesampling comprised points cataloguing different

Visual land use and land cover classes that were to be used to create ground truth maps for Ground Truth Classification

assessing the accuracy of the supervised classification performed by the remote

Remaining Unknown sensingSupervised software (explained below). The visits wereAreas (Class performed 0) accompanied by a Classification YES hired native guide and driver familiar with the area and with a good understanding of

the local agricultural practices. Farmers, residents, and local authorities were engaged Primary Classified LU Map Unknown with 90% confidence level. areas NO in conversationPrimary with the intent of investigating agricultural practices, past natural Classified

Threshold with 90% confidence level. Secondary Classified

Class selection in only 5 LU classes 136

disasters (such as droughts, fires and floods) and land use changes in the past 10-15

years.

4. ERDAS Imagine (version 8.7) and Arc Info (version 9) software were used for image

classification, drought indices, processing, and data analyses.

3.2 LULC Classification

3.2.1 General description

Classification of Landsat ETM+ images following Saadat et al’s (2010) methodology proceeded in five general steps: (i) preprocessing of the images, (ii) random extraction of a training sampling location: an unsupervised classification and two digital ancillary layers served in identifying potential LULC areas to aid in identifying sampling points, (iii) supervised classification of the image into LULC classes, (iv) enhancement of the LU classification via image segmentation and zonal statistics, and (v) enhancement of the LC classification via NDVI.

Each step is described below.

3.2.2 Preprocessing images

The Landsat ETM+ satellite products have 8 individual bands, each representing different portions of the electromagnetic spectrum. The four Landsat ETM+ were subject to preprocessing, which consisted of several steps (Fig. 2). Bands 1-5 and 7 range in the visible spectrum blue, green, red, near-infrared (NIR), and mid infrared (MIR), with 30m of spatial resolution, a panchromatic band with 15m band 8, and a thermal band with 60 m resolution band

6. Given the low resolution of the thermal band, 6.1 and 6.2 were not used. Bands 1-5 were combined into a multilayer image and the study area was clipped. Pan-sharpening (or image fusion) was performed with the objective of providing better image resolution. This was performed by fusing the 30m resolution multilayer with the 15m resolution panchromatic image. 137

The PCA method was used because a major goal of this technique is to reduce data file size yet retain the spectral information of the six ETM+ bands (Saadat et al 2011).

Atmospheric effects such as the amount of water vapor, distribution of aerosols, and scene visibility affect the raw imagery. Such effects were eliminated so that images taken in different times could be compared accurately.

The Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) atmospheric correction module in ENVI software was used for atmospheric correction to retrieve spectral reflectance data from the multispectral Landsat ETM+ images. Images were subjected to geo-referencing and image-to-image registration was performed. This process was verified by 138 importing the tracks recorded from the GPS that were obtained during fieldwork as ground controlFigure points. 5: Image After Segmentation preprocessing and Zonal of all Statistics images was completed, the study area was clipped from each scene. Clipped and Geo-referenced Secondary Classified Map 3.2.3 ExtractionImages of a training sampling location

TrainingImage samples were chosen to encompass a full variety of potential LULC classes such Segmentation Zonal as forest, irrigated agriculture, pastureland,Statistics rainfed agriculture, shrubs, and water bodies across the study area. This process (Fig. 3) was performed by processing an unsupervised classification Extracted Objects (spatial distribution) with 20 different classes with 95% of convergence threshold and maximum iterations of 10. This Polygon Map Classified Map map was usedPolygoniz as support to identify sampling locations across the study area, which was used to support the accuracy of the supervised classification. The total number of training sampling Class Selection (6 locations used for the study was 150; however, due to the inaccessibilityclasses) of some of these areas, only 132 was recorded during the performed fieldwork during the months of January-March

2013. Rainfed Forest (F) Pastureland (P) Shrubland (SH) Water Bodies (W) Agriculture (RA) 3.2.4 Supervised classification of the images into LULC classes.

Six classes were identified for describing the rural landscape in Central Mexico for the supervised classification (Fig. 4). Land use land cover (LULC) was classified as follows:

Rainfed Agriculture (RA), Irrigated Agriculture (IA), Forest (F), Shrubland (Sh), Pastureland (P), and Water Bodies (W). Saadat et al(2011) further classified pasturelands and forests using differences in densities; however, for this purposes of this study, the process was not needed since the focus was to understand the response of rainfed agriculture to drought. To increase the accuracy of these maps, a χ2 threshold at 90% confidence level was applied to the results. In this process, the identification pixels had a 10% chance of being misclassified. These pixels were put 139 into class “0” and defined as “unknown” areas. These unknown areas were reviewed observing their spectral signatures and resampling adjacent areas. Again, a supervised classification was performed with the samples plus the spectral signature sampling points until only a small pixel group (about 1% of the total area) was left unclassified.

3.2.5 Step 4: Image segmentation and zonal statistics

In an effort to further increase classification accuracy, an image segmentation algorithm was applied to the Landsat ETM+ images (Fig. 5) using the Bonnie Ruefenacht algorithm

(Ruefenacht et al. 2002). Image segmentation is a process used as a way of partitioning raster 140 images into segments based on pixel values and locations. Pixels that are spatially connected and have similar values are grouped in a single segment to isolate objects of varying size, shape, and homogeneity. This algorithm merges groups of pixels into polygon objects (raster to vector format) but it is unable to classify them (Saadat et al., 2011), thus the need for combining the images with zonal statistical analysis.

The zonal statistics present a distribution of each LULC within each segmented polygon.

The reason for image segmentation is to use the resultant polygon vector map in combination with the supervised classification raster map and zonal statistics to generate a new classified polygon map, the idea being to eliminate mixed pixels. It was found that for the six LU classes, the majority distribution was always more than 90% within any one segmented polygon. Thus, each segmented polygon was fully classified to the 90% majority, creating one layer for each of the six LU classes. 141 142

3.2.6 Assessing the accuracy of remotely sensed data

To evaluate the accuracy of the LULC maps that were processed, reference sampling locations were chosen to encompass the complete variety of the classes throughout the study area. Also, climate and topographic maps were taken into account to perform sampling. In total,

132 sites were extracted based on the stratified random sampling procedure as described by

Stehman (1999). Each sample comprised cataloguing different land use and land cover classes that were to be used to create ground truth maps for assessing the accuracy of the supervised classification performed by the remote sensing software to be explained below. The field work was performed accompanied by a hired local guide and driver who had significant knowledge of the area and of the local agricultural practices. Farmers, residents, and local authorities were engaged in conversation with the intent of investigating agricultural practices, past natural disasters (such as droughts, fires and floods), and land use changes in the past 10-15 years. This process was recorded by importing the tracks and control points recorded from the GPS and complemented with photographic evidence. 143

3.2.7 Spectral analysis for detecting drought effects on vegetation

A drought index is a prime variable for assessing the effect of a drought and defining different drought Year 2000 parameters, which CLASSNAME Area Km2 % of the Territory include intensity, Clouds 11.0287 1.02% duration, severity, Forest 211.298 19.47% Irrigated Agriculture 23.81 2.19% and spatial extent Pastureland 224.655 20.71% (Mishra and Singh Rainfed Agriculture 413.264 38.09% 2010). It should be River 15.8838 1.46% noted that a drought Shrubland 181.298 16.71% Water 3.76173 0.35% variable should be able TOTAL AREA 1084.99 100.00% to quantify the 144 drought for different time scales for which a long time series is essential. The most commonly used time scale for drought analysis is a year, followed by a month (Mishra and Singh 2010).

The yearly time scale can be used to abstract information on the regional behavior of droughts, while the monthly time scale is more appropriate for monitoring the effects of drought in situations related to agriculture, water supply, and groundwater abstractions (Panu and Sharma

2002). A time series of drought indices provides a framework for evaluating drought parameters of interest (Mishra and Singh 2010). To encompass a drought’s effects on the different classes of vegetation cover, two vegetation conditions and vegetation/soil moistures were chosen.

3.2.8 Normalized difference vegetation index (NDVI)

NDVI was designed on the premise that healthy vegetation has a low reflectance in the visible portion of the electromagnetic spectrum due to the absorption by chlorophyll and other pigments and high reflectance in the Near Infrared (NIR) because of the internal reflectance by the mesophyll spongy tissue of a green leaf (Campbell 1987). NDVI is calculated as the ratio of the red (RED) and the Near Infrared (NIR) bands of a sensor system and is represented by the following equation (Kogan 1995):

[1]

NDVI values range from -1 to +1. Because of high reflectance in the NIR portion of the electromagnetic spectrum, healthy vegetation is represented by high NDVI values between 0.05 and 1. Higher NDVI indicates a greater level of photosynthetic activity (Sellers 1985; Tucker et al. 1991). Conversely, non-vegetated surfaces such as water bodies yield negative values of

NDVI. Bare soil areas represent NDVI values close to 0 due to higher reflectance in both the 145 visible and the NIR portions of the electromagnetic spectrum (Lillesand and Kiefer 1994;

Rahimzadeh et al. 2008).

3.2.9 The normalized difference water index (NDWI).

The normalized difference water index (NDWI) is a more recent satellite-derived index from the NIR and short wave infrared (SWIR) channels that reflect changes in both the water content and spongy mesophyll in vegetation canopies. NDWI is calculated as follows (Gao

1996):

[2]

Where NIR is the Near Infrared band and MidIR is the Mid Infrared Band.

Because NDWI is influenced by both desiccation and wilting of vegetative canopy, it may be a more sensitive indicator than NDVI for drought monitoring (Mishra and Singh 2010).

NDVI and NDWI sense similar depths through vegetation canopies. However, NDWI is less sensitive to atmospheric effects than NDVI. NDWI does not completely remove the background soil reflectance effects, similar to NDVI because the information about vegetation canopies contained in the 1.24µm channel is very different from that contained in the red channel (Gao

1996). For this reason, NDWI should be considered complementary but not a substitute for

NDVI.

NDWI was calculated for the 18 Landsat ETM+ images. In the same way as previously done for NDVI, the NDWI output layers were divided by the LULC classes and in this way each class could be analyzed separately and a NDWI time series per class was plotted with the purpose of discerning the different NDWI values per class in different stages of the growing seasons in the two dry and normal years. 146 Figure 6: LULC maps of June 2000 (overall accuracy 90%) and August 2005 (overall accuracy 97%). Note the effects of SLC-off in the 2005 map 3.2.10 Tasseled cap transformationwhere almost wetness 10% of the (TCW) image was lost.

Numerous methods have been developed for transforming available information from multispectral sensors to derive features to interpret characteristics of the land surface. Such methods include the three indices based on ratios and differences of bands. The Tasseled Cap

Transformation of Landsat Multispectral Scanner and Thematic mapper (Kauth and Thomas

1976; Crist and Cicone 1984) offers a way to optimize data viewing for vegetation studies.

The different bands in a multispectral image can be visualized as defining an N- dimensional space, where N is the number of bands. Each pixel, positioned according to its data file value in each band, lies within the N-dimensional space. This pixel distribution is determined by the absorption/reflection spectra of the imaged material. For viewing purposes, it is advantageous to rotate the N-dimensional space such that one or two of the data structure axes are aligned with the View X and Y axes. In particular, the axes that are largest for the data structure produced by the absorption peaks are of special interest for the application (Crist and

Kauth 1986). Research has produced three data structure axes that define the vegetation information content. This option can show these three axes (or layers) as a degree of brightness, greenness, and wetness, as calculated by the Tasseled Cap coefficients used. Layer 1 (red) outputs the brightness component and indicates areas of low vegetation and high reflectors, layer

2 (green) is the greenness component and indicates vegetation status, and layer 3 (blue) is the wetness component that indicates water and moisture in the scene.

The Tasseled Cap Transformation Wetness (TCW) was used to determine the amount of moisture being held by the vegetation or soil, thus termed wetness

(Fadhil, 2009). TCW images were derived from ETM images of the study area 147

using tasseled cap transformation algorithm with ER Mapper according to the

following equation (Jin and Sader 2005):

[3]

Where B, G, R, NIR, SWIR, and SWIR2 are the Landsat ETM+ 6 bands, excluding the thermal

bands and the panchromatic band.

3.2.11 Change detection

Change detection is the process through which changes in the state a phenomenon are

identified by observing it over repeated time intervals. Given the availability of repetitive

coverage and a constant image quality, it is one of the main applications of remote sensing

(Beland et al. 2005).

Figure 7: NDVIThe change change detectionmaps of June approach and August. adopted The comparisonin this study was was made a post-classification between the normal yearcomparison (2000) and the detected dry year (2005). followed by a change threshold algorithm. It consists of using the spectral indices calculated

(NDVI, NDWI and TCW) from the four Landsat ETM+ images and then the data classification

to produce an image difference algorithm threshold calculation. The products from this

procedure were the two change maps corresponding to the difference in values for the beginning

and the end of the summer (June and August). Furthermore, to investigate and validate the

changes in specific areas in rainfed agriculture, the 44 samples taken during the field work were

used to extract the values of the spectral indices for the different images. The spectral index

values for the same locations were compared and tested for significant differences for all pairs of

images corresponding with an α of 0.01. Statistical significance was tested for the same months

in different years. 148

4. Results

4.1 Land Use Land Cover (LULC) Maps

. Area in Km2 Some Some Decreased Decrease Increase Increase >10% (1%-9.9) (1%-9.9%) >10% Total Forest 133.49 38.80 3.47 0.88 176.64 Irrigated Agriculture 17.31 1.50 0.47 0.40 19.68 Pastureland 127.78 37.64 2.87 2.72 171.01 Rainfed Agriculture 318.27 21.87 3.60 2.69 346.42 Shrubland 153.13 5.56 1.38 1.32 161.40 Total 749.99 105.38 11.78 8.00 875.15 Four Land Use Land Cover Maps were processed derived from Landsat ETM+ images.

Area in Km2 Some Some Decreased Decrease Increase Increase >10% (1%-9%) (1%-9%) >10% Total Forest 37.85 66.16 71.12 0.00 175.13 Irrigated Agriculture 8.56 4.91 5.98 0.00 19.45 Pastureland 117.96 43.10 0.00 19.51 180.57 Rainfed Agriculture 146.65 98.70 97.21 0.00 342.57 Shrubland 86.39 41.46 30.82 0.00 158.66 Total 397.41 254.33 205.13 19.51 876.38 Based on ground-truth, accuracy of the finalized LULC maps derived from the Landsat ETM+ images acquired for different periods of the year was calculated. LULC maps consisted of 6 classes (see Fig. 6) that mainly define this rural landscape in Central Mexico: Forests (F),

Irrigated Agriculture (IA), Pastureland (P), Rainfed Agriculture (RA), Shrubland (SH), and Water 149

Bodies (W). Due to the SLC-off, 15.95% of the pixels in the study area is lost for all images corresponding to 173.19 km2 for year 2005. The use of Saadat et al’s methodology to process the

LULC maps resulted in very high accuracy, ranging from 87.88% to 92.42% accuracy

(Appendices 1, 2, and 3). The landscape in central Mexico is very diverse and heterogeneous

(see images in Appendix 4). Farmers have shaped the land to satisfy their needs and have altered the landscape to overcome abnormally dry conditions by building small dams and water reservoirs. Moreover, farmers have adapted management practices such as irrigating Figure 8: Comparison 44 samples in Rainfed Agriculture NDVI values for the initial and final stages pronominally the most profitable crop in the region,of summer. which is alfalfa because of the importance of self-subsistence livestock in the region. Also, fertilizers and pesticides are used to increase yields.

Table 1: Average area for each class for June and August 2000.

Table 2: Average area for each class for June and August 2005

Year 2005 CLASSNAME Area Km^2 % of the Territory Forest 176.64 19.94% Irrigated Agriculture 19.68 2.22% Pastureland 171.01 19.30% Rainfed Agriculture 346.42 39.10% River 9.16 1.03% Shrubland 161.40 18.22% Water 1.763 0.20% TOTAL AREA 886.073 100.00%

The importance of having very accurate maps is that when detecting change of spectral indices and then evaluating it for each class, the results provide a very precise insight on how the 150 specific class behaved in the year, with an anomalous deficit of precipitation. Such evidence has the potential to be very useful for decision makers to provide solutions for people depending on such activities. Note in table 1 and 2 the difference of areas for all classes between those of 2000 and 2005. All areas in 2000 are bigger due to the data gaps of the Landsat imagery SLC-off of

2005. However, the percentage of areas is very similar.

4.2 NDVI Results

Sierra-Soler et al. (submitted) performed a precipitation analysis of monthly precipitation date since January 1980 to December 2011. The SPI was processed using this monthly data. The

9-month SPI compares the precipitation for that period with the same 9-month period over the historical record and it is a good indicator of seasonal conditions affecting agriculture. The year

2000 was observed to have normal wetness conditions because it presented consistent positive results within approximately one standard deviation in the 9-month SPI values that increased in

April, peaking in the month of June and then diminishing over the consequent months. The year

2005 persistently showed consistently dry conditions throughout the 6 meteorological stations inside the study area. This year presented values ranging from mild drought during the first

Figure 9: NDWI trend for all classes in the 18 images (F) Forest, (IA) Irrigated Agriculture, (P) Pastureland, (Sh) Shrubland, (RA) Rainfed Agriculture and (W) Water Bodies. 151

months of the year to extreme drought in May, June, and July, according to the SPI results

recorded from meteorological stations.

During summer, rainfed crops in central Mexico can be expected to have developed and

should be entering mature stages by the end of summer to be harvested by September

(SAGARPA 2005). NDVI change detection of the images corresponding to the initial final stages

of summer was performed (Fig. 7). The identified negative change in NDVI is significant in the

two periods analyzed. For the month of June, most classes showed a significant decrease in

NDVI values in most regions (Table 3). It was found that for NDVI in June, 85.69% of the

vegetated areas (representing 749.99 km2) decreased by 10% or more. Even though this decrease

is relevant, it should be noted that there is an 18 day gap between the two images; however, there

were no images available for closer dates to be compared, and the study compelled such a

Figure 10 NDWI change of June and August. The comparison was made between the normal year (2000) and the limitation. detected dry year (2005). 152

Area in Km2 Some Some Decrease Decrease Increase Increase >10% (1%-9.9) (1%-9.9%) >10% Total Forest 120.08 41.78 12.93 1.36 176.14 Irrigated Agriculture 17.37 1.29 0.75 0.24 19.65 Pastureland 116.49 56.88 6.01 2.13 181.52 Rainfed Agriculture 319.38 19.76 4.86 2.60 346.60 Shrubland 151.75 7.80 1.35 0.56 161.46 Total 725.08 127.50 25.91 6.89 885.38

Area in Km2 Some Some Decrease Decrease Increase Increase >10% (1%-9.9) (1%-9.9%) >10% Total Forest 11.51 106.25 51.63 5.12 174.51 Irrigated Agriculture 5.22 9.67 3.38 1.21 19.48 Pastureland 36.13 126.04 13.61 4.16 179.93 Rainfed Agriculture 78.20 210.38 46.45 7.72 342.75 Shrubland 39.32 108.36 9.66 1.47 158.82 Total 170.39 560.70 124.73 19.67 875.48 153

Figure 11: TCW change of June and August. The comparison was made between the normal year (2000) and the detected dry year (2005).

Low NDVI values mean there is little difference between the red and NIR signals. This manifestation of a negative change condition is therefore interpreted and associated with a lower photosynthetic activity of an important part of the study area that was analyzed. The August

NDVI change map presented a less severe scenario of negative change; however, 74% of the

Table 7: Area in km2 of each LULC class categorized for the change in TCW for June 2000 vs. June 2005

Some Some Decrease Decrease Increase Increase >10% (1%-9.9) (1%-9.9%) >10% Total Forest 68.09 71.01 0.01 35.51 174.61 Irrigated 15.97 3.42 0.00 3.40 22.80 Agriculture Pastureland 157.08 19.49 0.01 3.23 179.81 Rainfed 288.81 48.44 0.02 3.33 340.60 Agriculture Shrubland 120.59 26.37 0.00 12.01 158.97 Total 650.53 168.74 0.04 57.48 876.80 154

Area in Km2 Some Some Decrease Decrease Increase Increase >10% (1%-9.9) (1%-9.9%) >10% Total Forest 68.09 75.01 25.51 5.67 174.27 Irrigated Agriculture 13.97 2.42 2.40 0.95 19.75 Pastureland 127.08 19.49 3.23 29.35 179.15 Rainfed Agriculture 248.81 38.44 31.69 23.47 342.41 Shrubland 110.59 24.37 19.01 4.83 158.80 Total 568.53 159.74 81.84 64.26 874.38 total territory did experience at least some decrease in photosynthetic activity compared to the normal year. According to the LULC maps, Rainfed Agriculture represents approximately 58% of the total region, and for both periods analyzed these were the greatest extensions of land to have significant decreases in NDVI values (see Table 3 and 4). It can be noted that 28% of the area classified as Rainfed Agriculture presented a slight positive change in NDVI.

Forests also presented a slight increase in NDVI with an increase ranging between 1-9% of 71.12 km2 of forested area. Rainfed Agriculture seemed to return significant negative of change representing more than 10% or more in negative change. 155

Table 3: Area in km2 of each LULC class categorized for the change in NDVI for June 2000 vs. June 2005

Table 4: Area in km2 of each LULC class categorized for the change in NDVI for August 2000 vs. August 2005.

During fieldwork, 44 out of the 132 samples corresponding to rainfed agriculture were documented as ground truth data. Such samples were used to extract and compare NDVI values for the particular sampled points representing Rainfed Agricultural lands (Fig. 8). The negative change in the mean values of NDVI for both dates is significant. During the month of June, there is no overlap of NDVI values, representing an evident decrease in vegetation health and a negative mean value of 19.34% for the sampled points. By August, even though there is an overlap of some of the samples taken, the mean value of all points was reduced by 16%. This indicates that the overall health of vegetation for the sampled locations just before the harvest was affected. 156

4.3 NDWI and Soil/Vegetation Moisture

The Normalized Difference Water Index (NDWI) is a measure of liquid water molecules in vegetation canopies that interact with the incoming solar radiation. The NDWI algorithm was processed for 18 Landsat ETM+ images used in the study performed by Sierra-Soler et al.

(submitted) with the purpose of analyzing the trends in vegetation moisture in the study area in

2000, 2005, and 2011. Additionally, the months of June and August 2000 and 2005 were studied in depth since during summer, rainfed crops in Central Mexico are expected to have developed and entered mature stages by the end of summer, to be harvested by September (SAGARPA

2005).

Using the LULC classification, the NDWI statistics were extracted for each of the LULC classes. In this way, it was possible to verify the moisture trends of each class in different stages of the growing season (Fig. 9). Each class presented consistent patterns of mean NDWI throughout the 18 scenes, with values that slowly increased before the summer and continued to increase until peaking in the month of August, and then decreased by November when the rain season finished. Similarly to the NDVI trend studied by Sierra-Soler et al. (submitted), forests consistently presented the highest values of NDVI, although there was an apparent response to 157 seasonal high NDWI peaking during the wet months (June, July, and August). All classes peaked during the summer months and peaked in August. Rainfed Agriculture and Pasturelands consistently received the lowest values in NDWI, but followed the same trend as other classes.

The values for all classes in the year 2005 showed lower than normal values presented in the year 2000. The year 2011 presented a peculiar behavior, with low values and decreasing for

June and peaking very high in August. The year 2000 and 2005 were further investigated to compare the behaviors.

Change maps of NDWI were created (Fig. 10). Similar to the NDVI, the identified negative change in NDWI was substantial in the two periods analyzed, and especially for June

2005. For the month of June, most classes showed a significant decrease in NDVI values in most regions (Table 5). It was found that for NDWI in June, 81.9% of the vegetated areas

(representing 725.08 km2) decreased by 10% or more. This manifestation of negative change condition is interpreted and associated with lower moisture in the vegetation of great parts of the study area analyzed.

The case of August change was different. In this month, 20% of the study area presented a significant decrease in NDWI values while 64% presented some decrease. For the area classified as Rainfed Agriculture, 84% presented at least some decrease in NDWI; however, 20% of this area presented a 10% or more decrease. 158

Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total

Water Body 2 0 0 0 0 0 2 Forest 0 22 1 0 0 2 25 Irrigated Agriculture 0 0 4 0 0 0 4 Rainfed Agriculture 0 0 0 42 5 0 47

Pastureland 0 0 0 2 26 0 28 Shrubland 0 2 0 0 0 24 26 Column Total 2 24 5 44 31 26 132

Table 5: Area in km2 of each LULC class categorized for the change in NDWI for June 2000 vs. June

2005Table 6: Area in km2 of each LULC class categorized for the change in NDWI for June 2000 vs. June

2005

The NDWI did not show severe negative change for the month of August for most classes. The decrease of (1%-9%) could be dismissed; however, it is important to note that

Rainfed Agriculture and Pasturelands continuously showed the greatest negative response for the study period.

4.4 TCW and Soil/Vegetation Moisture

The results of soil/vegetation moisture indicator TCW demonstrated a clear decrease in the soil/vegetation during the period studied (Fig. 11). This particular spectral index showed a very similar negative response for both months analyzed. Significant decrease in TCW was found for June (72.4%) and August (64.9%) of the whole study area (Tables 7 and 8). This represented 84% of the Rainfed Agriculture, which showed a decrease of more than 10% in

TCW values. The areas classified as forest presented the most resilient characteristics by showing the least of the negative decrease and some increase for both periods. 159 160

Table 8: Area in km2 of each LULC class categorized for the change in TCW for June 2000 vs. June 2005 5. Conclusion

Climatic variability has the potential of producing extremes in weather. In many semiarid developing regions such as Central Mexico, droughts have the potential to cause crop failure and threaten the food security of poor farmers. Rainfed agriculture is practiced widely in these regions, and consequently, poor smallholder producers are exposed to losing their self- sufficiency production.

The original methodology presented in this study had the objective of investigating rainfed agriculture response to dry conditions using remotely sensed data. Four Land Use Land

Cover (LULC) maps were processed from single imagery using a new methodology obtaining overall accuracies ranging from the lowest at 87.88% to the highest at 92.42%. The importance of developing high accuracy LULC maps is that when detecting change of spectral indices and then evaluating it for each class, the results provide a detailed insight on how specific classes develop in years with dry conditions. The period of study selected corresponded to the months of

June and August (initial and final stages of summer) with the objective of understanding vegetation response to dry conditions at a crucial state of development. This period is where rainfed crops in Central Mexico are expected to have developed and entered mature stages by the end of the summer in order to be harvested by September (CONANP 2003). Three spectral indices (NDVI, NDWI, and TCW) were processed and analyzed to detect change of a normal and a dry year (2000 and 2005, respectively).

All spectral indices processed indicated a clear decrease in photosynthetic activity

(NDVI) and soil/vegetation moisture (NDWI and TCW) for the year 2005 compared to 2000, 161 particularly in the month of June. However, the negative change for the final stages of summer

(month of August) should not be underestimated. Persistently, Rainfed Agriculture and

Pastureland were the classes that presented the greatest percentages of decrease in all the obtained indices.

It was determined that these results were produced due to their dependency on water for their successful development. The changes perceived in the values of NDVI, NDWI, and TCW supported the conclusion that vegetation condition was negatively affected by the low rainfall presented in the year 2005. In June 2005 compared to June 2000, 85.7% of the total study area presented a decrease of vegetation condition (NDVI) and 81.8% vegetation/soil moisture

(NDWI). In both cases, the decrease was of 10% or more. For August 2005 compared to August

2000, 74.3% of the total study area presented at least some decrease of vegetation condition

(NDVI) and 83.4% of vegetation/soil moisture (NDWI). There is an 18 day gap between the two dates compared; however, there were no images available for closer dates to be compared.

Droughts are difficult to define because of their intricate nature, the variables that influence the phenomenon, and the way it is perceived by the systems in the region of interest.

The results presented in this paper could be compared to other semi arid areas; however, drought perception is different for every region because of the different water demand, ecosystems, and infrastructure. In summary NDVI, NDWI, and TCW post classification comparisons and image differencing techniques have proven to be useful methods for tracing environmental changes in the study area.

The presented methodology has strengths and weaknesses. High accuracy maps enabled an understanding of how different classes of vegetation changed in condition and soil/vegetation 162 moisture for specific key months of development. This detailed information has the potential to be used as background information for prevention and relief action. Vegetation and soil/ vegetation moisture spectral indices used in this June 2005 study Overall Kappa Statistics have been very well documented in the literature Class Name Kappa and are ------pertinent to detect drought effects on vegetative Water 1 cover. Forest 0.9013 The first disadvantage is the above- Irrigated Agriculture 1 mentioned Rainfed Agriculture 0.7222 18-day gap between the images analyzed. Pastureland 0.8557 Shrubland 0.9042 Furthermore, to quantify the severity of the drought studied in this paper, it should be compared with other periods where dry conditions were present. Landsat 8 was launched in February 11, 2013 and it will provide new imagery of the earth at medium resolution, which could have great potential to be used for drought monitoring in the future. Future research in the area should focus on comparing other dry events with the results obtained in this study, where significant environmental negative change was found related to dry conditions.

Unlike other natural hazards, the impacts of drought August 2005 are non- Overall Kappa Statistics 0.8599 structural and spread over large geographical Class Name Kappa areas, which ------results in difficulty in the quantification of Water 1 impact and for Forest 0.852 the provision of relief (Mishra and Singh Irrigated Agriculture 0.9734 2010). The Rainfed Agriculture 0.8107 present study found a negative response of Pastureland 0.9057 vegetation and Shrubland 0.8563 soil/vegetation moisture for the year 2005; however, further efforts must be made to understand longer time series of imagery to compare the response of this drought to others. 163

The results have the potential to be used as background for decision makers such as the

Water Council of the State of Hidalgo (CEAAH in Spanish) or the Water National Commission

(CONAGUA in Spanish) to identify areas that are more prone to be vulnerable to droughts because of the meteorological and environmental context in which they are situated. 164

Appendix 1: Confusion Matrices

June2000

Figure 11: A photo of pasturelands. (February 2013)

August 2000

Figure 12: A photo of a water body. (February 2013)

Figure 13: A photo of the river at the bottom of the canyon. (February 2013) 165

Figure 14: A photo of Rainfed Agriculture prepared for plantation before the rain season (surrounded by pastures). (January 2013)

Figure 15: Continuously productive Irrigated Agriculture plantations of alfalfa. Some Rainfed Agriculture and Pastures can be seen. (January 2013)

Figure 16: A photo of forested areas. (February 2013) 166

Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total

Water Body 2 0 0 0 0 0 2 Forest 0 22 0 0 0 3 25 Irrigated Agriculture 0 0 4 0 0 0 4 Rainfed Agriculture 0 0 0 43 4 0 47

Pastureland 0 0 0 2 26 0 28 Shrubland 0 1 0 0 0 25 26 Column Total 2 23 4 45 30 28 132

June 2005

Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total

Water Body 2 0 0 0 0 0 2 Forest 0 23 0 0 0 2 25 Irrigated Agriculture 0 0 4 0 0 0 4 Rainfed Agriculture 0 0 0 38 9 0 47

Pastureland 0 0 0 3 25 0 28 Shrubland 0 2 0 0 0 24 26 Column Total 2 25 4 41 34 26 132

August 2005

Water Irrigated Rainfed Classified Data Bodies Forest Agriculture Agriculture Pastureland Shrubland Row Total

Water Body 2 0 0 0 0 0 2 Forest 0 22 0 0 0 3 25 Irrigated Agriculture 0 0 4 0 0 0 4 Rainfed Agriculture 0 0 0 41 6 0 47

Pastureland 0 0 0 2 26 0 28 Shrubland 0 3 0 0 0 23 26 167

Column Total 2 25 4 43 32 26 132 168

Appendix 2: Accuracy Totals

June 2000

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy Water Body 2 2 2 100.00% 100.00% Forest 24 25 22 91.67% 88.00% Irrigated Agriculture 5 4 4 80.00% 100.00% Rainfed Agriculture 44 47 42 95.45% 89.36% Pastureland 31 28 26 83.87% 92.86% Shrubland 26 26 24 92.31% 92.31% Totals 132 132 120

Overall Classification Accuracy = 90.91%

August 2000

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy Water Body 2 2 2 100.00% 100.00% Forest 23 25 22 95.65% 88.00% Irrigated Agriculture 4 4 4 100.00% 100.00% Rainfed Agriculture 45 47 43 95.56% 91.49% Pastureland 30 28 26 86.67% 92.86% Shrubland 28 26 25 89.29% 96.15% Totals 132 132 122 Overall Classification Accuracy = 92.42%

June 2005

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy Water Body 2 2 2 100.00% 100.00% Forest 25 25 23 92.00% 92.00% Irrigated Agriculture 4 4 4 100.00% 100.00% Rainfed Agriculture 41 47 38 92.68% 80.85% Pastureland 34 28 25 73.53% 89.29% Shrubland 26 26 24 92.31% 92.31% Totals 132 132 116 Overall Classification Accuracy = 87.88% 169

August 2005

Class Reference Classified Number Producers Users Name Totals Totals Correct Accuracy Accuracy Water Body 2 2 2 100.00% 100.00% Forest 25 25 22 88.00% 88.00% Irrigated Agriculture 4 4 4 100.00% 100.00% Rainfed Agriculture 43 47 41 95.35% 87.23% Pastureland 32 28 26 81.25% 92.86% Shrubland 26 26 23 88.46% 88.46% Totals 132 132 118 Overall Classification Accuracy = 89.39 % 170

Appendix 3: Kappa Index of Agreement

June 2000 Overall Kappa Statistics 0.8799 Class Name Kappa ------Water 1 Forest 0.8533 Irrigated Agriculture 0.9734 Rainfed Agriculture 0.8404 Pastureland 0.9066 Shrubland 0.9042 171

Appendix 4: Photographic samples of each LULC class in the study area 172 173

References

Anyamba, A., and C.J. Tucker. “Analysis of Sahelian vegetation dynamics using NOAA-

AVHRR NDVI data from 1981-2003.” Journal of Arid Environments 63 (2005): 596–

614. Print.

Appendini, Kirsten, & Diana Liverman. (1994). “Agricultural policy, climate change and food

security in Mexico.” Food Policy 19.2 (1994): 149-164. Web. 14 Oct 2013.

Bot, A. “The Importance of Soil Organic Matter: Key to Drought-Resistant Soil and Sustained

Food Production.” FAO. FAO.org, 2005. Web. 16 Oct 2013.

Cohen, Y., and M. Shoshany. “A National Knowledge-Based Crop Recognition in Mediterranean

Environment.” ITC Journal 2002.4 (2002): 75-87. Print.

CONANP. (2003). Programa de Manejo Reserva de La Biosfera Barranca de Metztitlán. Mexico

DF.

Congalton, Russell G. “A Review of Assessing the Accuracy Of Classifications of Remotely

Sensed Data.” Remote Sensing of Environment 37.1 (1991): 35-46. Print.

Crist, Eric P, and Richard C. Cicone. “A Physically-Based Transformation of Thematic Mapper

Data---The TM Tasseled Cap.” Geoscience and Remote Sensing, IEEE Transactions 3

(1984): 256-263. Print. de Asis, Alejandro M., and Kenji Omasa. “Estimation of Vegetation Parameter for Modeling Soil

Erosion Using Linear Spectral Mixture Analysis of Landsat ETM Data. ISPRS Journal of

Photogrammetry and Remote Sensing 62.4 (2007): 309-324. Web. 15 Oct 2013. 174

Diouf, A., and E.F. Lambin. “Monitoring Land-Cover Changes in Semi-Arid Regions: Remote

Sensing Data and Field Observations In The Ferlo, Senegal.” Journal of Arid

Environments 48.2 (2001): 129-148. Web. 15 Oct 2013.

Eakin, Hallie. (2000). “Smallholder Maize Production and Climatic Risk: A Case Study from

Mexico.” Climatic Change 45.1 (2000): 19-36. Web. 16 Oct 2013.

Eakin, Hallie. “Institutional Change, Climate Risk, and Rural Vulnerability: Cases from Central

Mexico.” World Development 33.11 (2005): 1923-1938. Web. 14 Oct 2013.

Eakin, Hallie, and Amy Lynd Luers. “Assessing the Vulnerability of Social-Environmental

Systems.” Annual Review of Environment and Resources 31.1 (2006): 365-394. Web. 13

Oct 2013.

Fadhil, Ayad Mohammed. “Drought Mapping Using Geoinformation Technology for Some Sites

in the Iraqi Kurdistan Region.” International Journal of Digital Earth 4.3 (2011):

239-257. Web. 15 Oct 2013.

Focardi, Silvia, et. al. “Satellite-Based Indices in the Analysis of Land Cover for Municipalities

in the Province of Siena, Italy.” Journal of Environmental Management 86.2 (2008):

383-389. Web. 09 Oct 2013.

Goward, Samuel N, Compton J. Tucker, and Dennis G. Dye. “North American Vegetation

Patterns Observed with the NOAA-7 Advanced Very High Resolution Radiometer.”

Vegetatio 64.1 (1985): 3-14.

Gu, Yingxin, et. al. “A Five-Year Analysis of MODIS NDVI and NDWI for Grassland Drought

Assessment over the Central Great Plains of the United States.” Geophysical Research

Letters, 34.6 (2007). Print. 175

Gu, Yingxin, et. al. “Evaluation of MODIS NDVI and NDWI for Vegetation Drought Monitoring

Using Oklahoma Mesonet Soil Moisture Data.” Geophysical Research Letters 35.22

(2008). Print.

Hewitt, Kenneth. Regions of Risk: A Geographical Introduction to Disasters. London: Longman.

1997. Print.

Im, J., J.R. Jensen, and J.A. Tullis. “Object-Based Change Detection Using Correlation Image

Analysis and Image Segmentation.” International Journal of Remote Sensing 29.2

(2008): 399-423. Print.

INEGI. “Anuario Estadistico del Estado de Hidalgo.” Inegi. Inegi.org, 2007. Web. 12 Oct 2013.

Jin, Suming, and Steven A. Sader. “Comparison of Time Series Tasseled Cap Wetness and the

Normalized Difference Moisture Index in Detecting Forest Disturbances.” Remote

Sensing of Environment 94.3 (2005): 364-372. Web. 13 Oct 2013.

Johan, Rockström. “Resilience Building and Water Demand Management for Drought

Mitigation.” Physics and Chemistry of the Earth, Parts A/B/C 28.20–27 (2003): 869-877.

Web. 14 Oct 2013.

Jin, Suming, and Steven A. Sader. “Comparison of Time Series Tasseled Cap Wetness and the

Normalized Difference Moisture Index in Detecting Forest Disturbances.” Remote

Sensing of Environment 94.3 (2005): 364-372. Web. 13 Oct 2013.

Johan, Rockström. “Resilience Building and Water Demand Management for Drought

Mitigation.” Physics and Chemistry of the Earth, Parts A/B/C 28.20–27 (2003): 869-877.

Web. 14 Oct 2013. 176

Joshi, P. K., et. al. “Vegetation Cover Mapping in India Using Multi-Temporal IRS Wide Field

Sensor (WiFS) Data.” Remote Sensing of Environment 103.2 (2006): 190-202. Web. 09

Oct 2013.

Kogan, F. N. (1990). “Remote sensing of weather impacts on vegetation in non-homogeneous

areas.” International Journal of Remote Sensing 11.8 (1990): 1405-1419. Web. 13 Oct

2013.

Kogan, F. N. “Application of Vegetation Index and Brightness Temperature for Drought

Detection.” Advances in Space Research 15.11 (1995): 91-100. Web. 12 Oct 2013.

Kogan, Felix N. “Droughts of the Late 1980s in the United States as Derived from NOAA Polar-

Orbiting Satellite Data.” Bulletin of the American Meteorological Society 76.5 (1995):

655-668. Web. 12 Oct 2013.

Labrecque, S., et. al. “A Comparison of Four Methods to Map Biomass from Landsat-TM and

Inventory Data in Western Newfoundland.” Forest Ecology and Management 226.1-3

(2006): 129-144. Print.

Lenney, Mary Pax, et. al. “The Status of Agricultural Lands in Egypt: The Use of Multitemporal

NDVI Features Derived from Landsat TM.” Remote Sensing of Environment 56.1 (1996):

8-20. Web. 08 Oct 2013.

Liverman, Diana M. “Drought Impacts in Mexico: Climate, Agriculture, Technology, and Land

Tenure in Sonora and Puebla.” Annals of the Association of American Geographers 80.1

(1990): 49-72. Print.

Liverman, Diana M. “Vulnerability and Adaptation to Drought in Mexico.” Natural Resources

Journal 39 (1999): 99-116. Print. 177

Liverman, Diana M., and Karen L. O'Brien “Global Warming and Climate Change in Mexico.”

Global Environmental Change 1.5 (1991): 351-364. Web. 12 Oct 2013.

Maselli, F., et. al. “Inclusion of Prior Probabilities Derived from a Nonparametric Process into

the Maximum-Likelihood Classifier.” Photogrammetric Engineering & Remote Sensing

58.2 (1992): 201-207. Print.

McKee, T.B.N., J. Doesken, and J. Kleist. “Drought Monitoring with Multiple Time Scales.”

Ninth Conference of Applied Climatology. American Meteorological Society (1993):

233-236. Print.

Mishra, Ashok K., and Vijay P. Singh. “A Review of Drought Concepts.” Journal of Hydrology

391.1–2 (2010): 202-216. Web. 15 Oct 2013.

Mishra, Ashok K., and Vijay P. Singh. “Drought Modeling – A Review.” Journal of Hydrology

403.1–2 (2011): 157-175. Web. 15 Oct 2013.

Obasi, GOP. “WMO's Role in the International Decade for Natural Disaster Reduction.” Bulletin

of the American Meteorological Society 75.9 (1994): 1655-1661. Print.

Oetter, Doug R., et. al. “Land Cover Mapping in an Agricultural Setting Using Multiseasonal

Thematic Mapper Data.” Remote Sensing of Environment 76.2 (2001), 139-155. Web. 12

Oct 2013.

Panu, U.S., and T.C. Sharma. “Challenges in Drought Research: Some Perspectives and Future

Directions.” Hydrological Sciences Journal 47.S1 (2002): S19-S30. Print.

Peters, Albert J., et. al. “Drought Monitoring with NDVI-Based Standardized Vegetation Index.”

Photogrammetric Engineering and Remote Sensing, 68.1 (2002): 71-75. 178

Peters, Albert J., et. al. “Satellite Assessment of Drought Impact on Native Plant Communities of

Southeastern New Mexico, U.S.A.” Journal of Arid Environments 24.3 (1993): 305-319.

Web. 08 Oct 2013.

Quarmby, J., and C.F. Forster. “An Examination of the Structure of UASB Granules.” Water

Research 29.11 (1995): 2449-2454. Print.

Rahimzadeh Bajgiran, Parinaz, et. al. “Using AVHRR-Based Vegetation Indices for Drought

Monitoring in the Northwest of Iran.” Journal of Arid Environments 72.6 (2008):

1086-1096. Web. 12 Oct 2013.

Rahimzadeh-Bajgiran, Parinaz, Kenji Omasa, and Yo Shimizu. “Comparative Evaluation of the

Vegetation Dryness Index (VDI), the Temperature Vegetation Dryness Index (TVDI) and

the Improved TVDI (iTVDI) for Water Stress Detection In Semi-Arid Regions of Iran.”

ISPRS Journal of Photogrammetry and Remote Sensing 68.0 (2012): 1-12. Web. 12 Oct

2013.

Rasmussen, M. S. “Assessment of Millet Yields and Production in Northern Burkina Faso Using

Integrated NDVI from the AVHRR.” International Journal of Remote Sensing 13.18

(1992): 3431-3442. Print.

Redmond, Kelly T. “The Depiction of Drought: A Commentary.” Bulletin of the American

Meteorological Society 83.8 (2002), 1143-1147. Print.

Renschler, Chris S., and Jon Harbor. “Soil Erosion Assessment Tools from Point to Regional

Scales—The Role of Geomorphologists in Land Management Research and

Implementation.” Geomorphology 47.2–4 (2002): 189-209. Print. 179

Rhee, Jinyoung, Jungho Im, and & Gregory J. Carbone. “Monitoring Agricultural Drought for

Arid and Humid Regions Using Multi-Sensor Remote Sensing Data.” Remote Sensing of

Environment 114.12 (2010): 2875-2887. Web. 10 Oct 2013.

Ruefenacht, Bonnie, et. al. New Technique for Segmenting Images. Salt Lake City: USDA Forest

Service Remote Sensing Applications Center. 2002. Print.

Saadat, H., et. al. “Land Use and Land Cover Classification over a Large Area in Iran Based on

Single Date Analysis of Satellite Imagery.” ISPRS Journal of Photogrammetry and

Remote Sensing 66.5 (2011): 608-619. Print.

Sivanpillai, Ramesh, et. al. “Estimating Regional Forest Cover in East Texas Using Advanced

Very High Resolution Radiometer (AVHRR) Data.” International Journal of Applied

Earth Observation and Geoinformation 9.1 (2007): 41-49. Web. 15 Oct 2013.

Unganai, Leonard S., and Felix N. Kogan. “Drought Monitoring and Corn Yield Estimation in

Southern Africa from AVHRR Data.” Remote Sensing of Environment 63.3 (1998):

219-232. Web. 12 Oct 2013.

Wardlow, Brian D., Stephen L. Egbert, and Jude H. Kastens. “Analysis of Time-Series MODIS

250 m Vegetation Index Data for Crop Classification in the U.S. Central Great Plains.”

Remote Sensing of Environment 108.3 (2007): 290-310. Web. 08 Oct 2013.

Wilhelmi, Olga V., and Donald A. Wilhite. “Assessing Vulnerability to Agricultural Drought: A

Nebraska Case Study.” Natural Hazards 25.1 (2002): 37-58. Web. 09 Oct 2013.

Wilhite, Donald A., Mark D. Svoboda, and Michael J. Hayes. “Understanding the Complex

Impacts of Drought: A Key to Enhancing Drought Mitigation and Preparedness.” Water

Resources Management 21.5 (2007): 763-774. Web. 12 Oct 2013. 180

Xiaofan, Liu, et. al. “Assessing Vegetation Response to Drought in the Laohahe Catchment,

North China.” Hydrology Research 43.1/2 (2012): 91-101. Web. 15 Oct 2013.

Yurekli, K., and A. Kurunc. “Simulating Agricultural Drought Periods Based on Daily Rainfall

and Crop Water Consumption.” Journal of Arid Environments 67.4 (2006): 629-640.

Web. 14 Oct 2013. 181

Chapter 5

5.1 Summary and General Conclusions

This study investigated the use of geographic information systems for analyzing agricultural drought effects on rainfed lands in semiarid developing regions. The study was tested in a relatively small watershed (1095 km2) in Central Mexico. Nonetheless, the method could be applied in other developing regions with similar environmental conditions and agricultural practices.

The methodology presented in this study monitored the behavior of different vegetative covers in the event of abnormally dry conditions detected, using a time series analysis of precipitation data for the period between January 1980 to December 2011. To attain the aim of this research, the study proceeded in three major steps. The first step was the use of the standardized precipitation index (SPI) to make a selection of three years to study the spring and summer rainfed agricultural cycle, one year with normal conditions and two with abnormally dry conditions. The second step was the accurate delineation of agricultural lands, and specifically means where rainfed farming is practiced. The third step was to extract different spectral indices with the purpose of crop vigor, development, and moisture during dry and normal years.

A unique aspect of this thesis is the inclusion of high accuracy land use land cover

(LULC) maps for detecting drought effects on vegetation cover. The methodology presented in this thesis produced a total of 22 LULC maps processed from Landsat ETM+. In chapter 3, 18

LULC maps of the growing season corresponding to the years 2000, 2005, and 2011 yielded 182 overall accuracies of between 77.5% and 83.2%. In chapter 4, four LULC maps of the early and late summer of 2000 and 2005 yielded accuracies ranging between 87.88% and 92.42%.

The LULC maps, in combination with vegetation status indices and vegetation/soil moisture, permitted the analysis of the different vegetative classes that compose the landscape.

The results demonstrated persistently that Rainfed Agriculture and Pastureland classes presented the greatest negative change during the dry periods. The method proved to be useful to verify the development of rainfed agriculture for small regions to provide information about the vegetation status and response towards droughts. For that reason, accurate LULC maps can be effective tools in aid of agricultural drought detection effects. This study also corroborates other studies that found that late-summer images present the best information for the LULC classification .

Vegetation and soil/vegetation moisture spectral indices used in this study have been very well documented in the literature and are pertinent to detect drought effects on vegetative cover.

The combination of such indices overlap the effects of droughts experienced in the years selected.

The vegetation health indices (NDVI and VCI) analysis through the growing season produced the following results. Forests consistently presented the highest value trend of vegetation condition. Forests showed a response of seasonal high NDVI and VCI values that peaked during the wet months (June, July, and August). Shrubs were seasonal and very responsive to wetness conditions, with the lowest values during the dry months, and with rapidly increasing health in wet months. The agricultural cycle in irrigated agriculture was evident in the

NDVI trends for the three analyzed years. This can be determined by how the moderate 183 vegetation condition during the planting months of April and May evolved by June, peaked in

August, and faded by November after the harvest.

The mean NDVI pattern in Rainfed Agriculture and Pasturelands was very similar due to their response to rain; however, Rainfed Agriculture consistently showed higher values. In the year 2000, Rainfed Agriculture followed a similar pattern to irrigated agriculture but with lower values, which peaked in August and faded by November for the post-harvest. In 2005, the deficit in precipitation can be presumed by the very low values in June and then by July, even though

June 2005 was characterized to be the month where most of the vegetation classes seemed to be stressed, especially Rainfed Agriculture and Pastureland, with values indicating the lowest detected vegetation condition.

The vegetation/soil moisture (NDWI and TCW) analysis through the growing season yielded the following results. Each class presented consistent patterns throughout the 18 scenes analyzed, with values that slowly increased before the summer and continued to increase until peaking in the month of August, and then decreased by November when the rain season finished.

Similar to the vegetation spectral trends, forests consistently presented the best moisture condition, although there was an apparent response to seasonal high precipitation peaking during the wet months (June, July, and August). All classes peaked during the summer months and peaked in August. Rainfed Agriculture and Pasturelands consistently presented dry conditions, but followed the same trend as other classes. The values for all classes in 2005 showed lower than normal values compared to the year 2000. The year 2011 was a peculiar with low values and decreasing for June and then peaking very high in August. The years 2000 and 2005 were further 184 investigated to compare the behaviors. It was found that for the year 2005 clear decrease in vegetation activity and soil/vegetation moisture for the initial and the final stages of summer.

5.2 Contributions to Knowledge

Given that the research presented in this thesis was tested in central Mexico, which represents an example of a specific site with specific climatic, and physiological characteristics, the contributions to knowledge for this study are:

1. To the author’s knowledge, no research of this nature has been carried out at such a

scale in Mexico resulting as background information for decision makers and future

researchers in the field.

2. The thesis presented an original protocol for the creation of land use land cover

(LULC) classification that yields high overall accuracies that ranged between 87.88%

and 92.42%.

3. The results have enhanced the understanding of relationships between vegetative

cover, with its response to anomalous dry conditions. Furthermore, it also raises

questions on the different methods for detecting agricultural drought in semiarid

developing regions.

4. The methodology here presented has the potential to be used by decision makers in

water management in semi-arid regions of Mexico. Preparedness to take action in

early stages of drought can mitigate the great economic and social costs of droughts.

5.3 Suggestions for Future Research

It is expected that in time, further analysis of existing data-bases in the global environmental research will lead to more and better relationships between different sources of 185 data. With the growing availability of improved temporal and spatial scale sensing systems, there will be a need to incorporate different sorts of data. Remotely sensed data has the potential to be integrated with other sources of data such as static sensing and mobile sensing systems to aid in the ground truthing of data. Similarly, the wide availability of handheld sensing systems (e.g., smartphones) that have the potential to take photographic evidence of droughts, georeferencing, and data validation to enhance in situ data quality, could give local stakeholders the opportunity to upload relevant data automatically. Projects like the Ushahidi Platform, where crowdsourcing has been used by people in developing regions to alert of a disaster or emergency by sensing

SMS messages, videos, and images from smartphones or online reports, have been shown to be effective.The long term future of environmental research and emergency relief should focus on increasing computer power to the point where all relevant research and reports can be accessed in an integral way to be displayed at different scales in space and time. 186

Bibliography

Adler-Golden, Steven M. et. al. “Atmospheric Correction for Short-Wave Spectral Imagery

based on MODTRAN 4.” Paper presented at the PROC SPIE INT SOC OPT ENG, 1999.

Anyamba, A., and C.J. Tucker. “Analysis of Sahelian vegetation dynamics using NOAA-

AVHRR NDVI data from 1981-2003.” Journal of Arid Environments 63 (2005): 596–

614. Print.

Appendini, Kirsten, & Diana Liverman. (1994). “Agricultural policy, climate change and food

security in Mexico.” Food Policy 19.2 (1994): 149-164. Web. 14 Oct 2013.

Barbosa, H. A., A.R. Huete, and W.E. Baethgen. “A 20-Year Study of NDVI Variability over the

Northeast Region of Brazil.” Journal of Arid Environments 67.2 (2006): 288-307. Web.

17 Oct 2013.

Barkin, David. (2002). “The Reconstruction of a Modern Mexican Peasantry.” The Journal of

Peasant Studies 30.1 (2002): 73-90. Print.

Bayarjargal, Y., et. al. “A Comparative Study of NOAA–AVHRR Derived Drought Indices Using

Change Vector Analysis.” Remote Sensing of Environment 105.1 (2006): 9-22. Web. 14

Oct 2013.

Beland, M, et. al. “Assessment of Land-Cover Changes Related to Shrimp Aquaculture Using

Remote Sensing Data: A Case Study in the Giao Thuy District, Vietnam.” International

Journal of Remote Sensing 27.8 (2006): 1491-1510. Print.

Bhuiyan, C., R.P. Singh, and F.N. Kogan. “Monitoring Drought Dynamics in the Aravalli Region

(India) Using Different Indices Based on Ground and Remote Sensing Data.” 187

International Journal of Applied Earth Observation and Geoinformation 8.4 (2006):

289-302. Web. 17 Oct 2013.

Boken, Vijendra K., et. al. “Monitoring and Predicting Agricultural Drought: A Global Study.”

Ebrary. Ebrary, 2005. Web. 16 Oct 2013.

Bot, A. “The Importance of Soil Organic Matter: Key to Drought-Resistant Soil and Sustained

Food Production.” FAO. FAO.org, 2005. Web. 16 Oct 2013.

Brown, Pete. “Institutions, Inequalities, and the Impact Of Agrarian Reform On Rural Mexican

Communities.” Human Organization 56.1 (1997): 102-110. Print.

Cihlar, J. “Land Cover Mapping of Large Areas from Satellites: Status and Research Priorities.”

International Journal of Remote Sensing 21.6-7 (2000): 1093-1114. Print.

Cohen, Y., and M. Shoshany. “A National Knowledge-Based Crop Recognition in Mediterranean

Environment.” ITC Journal 2002.4 (2002): 75-87. Print.

CONANP. (2003). Programa de Manejo Reserva de La Biosfera Barranca de Metztitlán. Mexico

DF.

Congalton, Russell G. “A Review of Assessing the Accuracy Of Classifications of Remotely

Sensed Data.” Remote Sensing of Environment 37.1 (1991): 35-46. Print.

Crist, Eric P, and Richard C. Cicone. “A Physically-Based Transformation of Thematic Mapper

Data---The TM Tasseled Cap.” Geoscience and Remote Sensing, IEEE Transactions 3

(1984): 256-263. Print. de Asis, Alejandro M., and Kenji Omasa. “Estimation of Vegetation Parameter for Modeling Soil

Erosion Using Linear Spectral Mixture Analysis of Landsat ETM Data. ISPRS Journal of

Photogrammetry and Remote Sensing 62.4 (2007): 309-324. Web. 15 Oct 2013. 188

Diouf, A., and E.F. Lambin. “Monitoring Land-Cover Changes in Semi-Arid Regions: Remote

Sensing Data and Field Observations In The Ferlo, Senegal.” Journal of Arid

Environments 48.2 (2001): 129-148. Web. 15 Oct 2013.

Eakin, Hallie. (2000). “Smallholder Maize Production and Climatic Risk: A Case Study from

Mexico.” Climatic Change 45.1 (2000): 19-36. Web. 16 Oct 2013.

Eakin, Hallie. “Institutional Change, Climate Risk, and Rural Vulnerability: Cases from Central

Mexico.” World Development 33.11 (2005): 1923-1938. Web. 14 Oct 2013.

Eakin, Hallie, and Amy Lynd Luers. “Assessing the Vulnerability of Social-Environmental

Systems.” Annual Review of Environment and Resources 31.1 (2006): 365-394. Web. 13

Oct 2013.

Fadhil, Ayad Mohammed. “Drought Mapping Using Geoinformation Technology for Some Sites

in the Iraqi Kurdistan Region.” International Journal of Digital Earth 4.3 (2011):

239-257. Web. 15 Oct 2013.

Focardi, Silvia, et. al. “Satellite-Based Indices in the Analysis of Land Cover for Municipalities

in the Province of Siena, Italy.” Journal of Environmental Management 86.2 (2008):

383-389. Web. 09 Oct 2013.

Foley, Jonathan A., et. al. “Global Consequences of Land Use.” Science 309.5734 (2005):

570-574. Print.

Gao, Bo-cai.. “NDWI—A Normalized Difference Water Index for Remote Sensing of Vegetation

Liquid Water from Space.” Remote Sensing of Environment 58.3 (1996): 257-266. Web.

12 Oct 2013. 189

Glantz, Michael H. Drought Follows the Plow: Cultivating Marginal Areas. Cambridge:

Cambridge University Press. 1994. Print.

Goward, Samuel N, Compton J. Tucker, and Dennis G. Dye. “North American Vegetation

Patterns Observed with the NOAA-7 Advanced Very High Resolution Radiometer.”

Vegetatio 64.1 (1985): 3-14.

Gu, Yingxin, et. al. “A Five-Year Analysis of MODIS NDVI and NDWI for Grassland Drought

Assessment over the Central Great Plains of the United States.” Geophysical Research

Letters, 34.6 (2007). Print.

Gu, Yingxin, et. al. “Evaluation of MODIS NDVI and NDWI for Vegetation Drought Monitoring

Using Oklahoma Mesonet Soil Moisture Data.” Geophysical Research Letters 35.22

(2008). Print.

Guerschman, J. P., et. al. (2003). “Land Cover Classification in the Argentine Pampas Using

Multi-Temporal Landsat TM Data.” International Journal of Remote Sensing 24.17

(2003): 3381-3402. Web. 13 Oct 2013.

Hartmann, Tomas, Carlos Di Bella, and Patricio Oricchio. “Assessment of the Possible Drought

Impact on Farm Production in the SE of the Province Of Buenos Aires, Argentina.”

ISPRS Journal of Photogrammetry and Remote Sensing 57.4 (2003): 281-288. Web. 10

Oct 2013.

Hayes, M.J., and W.L. Decker. “Using NOAA AVHRR Data to Estimate Maize Production in the

United States Corn Belt.” Remote Sensing 17.16 (1996): 3189-3200. Print.

Herrera-Rodriguez, Mauricio. (2012). “Social Change and Land Tenure Regimes in Mexico.”

GeoJournal 77.5 (2012): 633-649. Web. 12 Oct 2013. 190

Hewitt, Kenneth. Regions of Risk: A Geographical Introduction to Disasters. London: Longman.

1997. Print.

Im, J., J.R. Jensen, and J.A. Tullis. “Object-Based Change Detection Using Correlation Image

Analysis and Image Segmentation.” International Journal of Remote Sensing 29.2

(2008): 399-423. Print.

INEGI. “Anuario Estadistico del Estado de Hidalgo.” Inegi. Inegi.org, 2007. Web. 12 Oct 2013.

Jin, Suming, and Steven A. Sader. “Comparison of Time Series Tasseled Cap Wetness and the

Normalized Difference Moisture Index in Detecting Forest Disturbances.” Remote

Sensing of Environment 94.3 (2005): 364-372. Web. 13 Oct 2013.

Johan, Rockström. “Resilience Building and Water Demand Management for Drought

Mitigation.” Physics and Chemistry of the Earth, Parts A/B/C 28.20–27 (2003): 869-877.

Web. 14 Oct 2013.

Jornada, La. “Heladas y Sequía Causan Estragos al Campo en Hidalgo y Nayarit. La Jornada.

Jornada.anum.mx, 2012. Web. 10 Oct 2013.

Joshi, P. K., et. al. “Vegetation Cover Mapping in India Using Multi-Temporal IRS Wide Field

Sensor (WiFS) Data.” Remote Sensing of Environment 103.2 (2006): 190-202. Web. 09

Oct 2013.

Kauth, Richard J., and G.S. Thomas. (1976). “The Tasselled Cap--A Graphic Description of the

Spectral-Temporal Development of Agricultural Crops as seen by Landsat.” Paper

presented at the LARS Symposia, 1976. 191

Kogan, F. N. (1990). “Remote sensing of weather impacts on vegetation in non-homogeneous

areas.” International Journal of Remote Sensing 11.8 (1990): 1405-1419. Web. 13 Oct

2013.

Kogan, F. N. “Application of Vegetation Index and Brightness Temperature for Drought

Detection.” Advances in Space Research 15.11 (1995): 91-100. Web. 12 Oct 2013.

Kogan, Felix N. “Droughts of the Late 1980s in the United States as Derived from NOAA Polar-

Orbiting Satellite Data.” Bulletin of the American Meteorological Society 76.5 (1995):

655-668. Web. 12 Oct 2013.

Labrecque, S., et. al. “A Comparison of Four Methods to Map Biomass from Landsat-TM and

Inventory Data in Western Newfoundland.” Forest Ecology and Management 226.1-3

(2006): 129-144. Print.

Lenney, Mary Pax, et. al. “The Status of Agricultural Lands in Egypt: The Use of Multitemporal

NDVI Features Derived from Landsat TM.” Remote Sensing of Environment 56.1 (1996):

8-20. Web. 08 Oct 2013.

Li, Ainong, et. al. “Eco-Environmental Vulnerability Evaluation in Mountainous Region Using

Remote Sensing and GIS—A Case Study in the Upper Reaches Of Minjiang River,

China.” Ecological Modelling, 192.1–2 (2006): 175-187. Web. 10 Oct 2013.

Liverman, Diana M. “Drought Impacts in Mexico: Climate, Agriculture, Technology, and Land

Tenure in Sonora and Puebla.” Annals of the Association of American Geographers 80.1

(1990): 49-72. Print.

Liverman, Diana M. “Vulnerability and Adaptation to Drought in Mexico.” Natural Resources

Journal 39 (1999): 99-116. Print. 192

Liverman, Diana M., and Karen L. O'Brien “Global Warming and Climate Change in Mexico.”

Global Environmental Change 1.5 (1991): 351-364. Web. 12 Oct 2013.

Lucas, Richard, et. al. “Rule-Based Classification of Multi-Temporal Satellite Imagery for

Habitat and Agricultural Land Cover Mapping.” ISPRS Journal of Photogrammetry and

Remote Sensing 62.3 (2007): 165-185. Web. 11 Oct 2013.

Luers, Amy L., et. al. “A Method for Quantifying Vulnerability, Applied to the Agricultural

System of the Yaqui Valley, Mexico.” Global Environmental Change 13.4 (2003):

255-267. Web. 16 Oct 2013.

Maselli, F., et. al. “Inclusion of Prior Probabilities Derived from a Nonparametric Process into

the Maximum-Likelihood Classifier.” Photogrammetric Engineering & Remote Sensing

58.2 (1992): 201-207. Print.

Maxwell, S. K., et. al. “An Automated Approach to Mapping Corn from Landsat Imagery.”

Computers and Electronics in Agriculture 43.1 (2004): 43-54. Web. 10 Oct 2013.

McKee, T.B.N., J. Doesken, and J. Kleist. “Drought Monitoring with Multiple Time Scales.”

Ninth Conference of Applied Climatology. American Meteorological Society (1993):

233-236. Print.

Mishra, Ashok K., and Vijay P. Singh. “A Review of Drought Concepts.” Journal of Hydrology

391.1–2 (2010): 202-216. Web. 15 Oct 2013.

Mishra, Ashok K., and Vijay P. Singh. “Drought Modeling – A Review.” Journal of Hydrology

403.1–2 (2011): 157-175. Web. 15 Oct 2013.

Moreira, Elsa E., et. al. “SPI-Based Drought Category Prediction Using Loglinear Models.”

Journal of Hydrology 354.1–4 (2008): 116-130. Web. 08 Oct 2013. 193

Moreira, Elsa E., et. al. “Analysis of SPI Drought Class Transitions Using Loglinear Models.”

Journal of Hydrology 331.1–2 (2006): 349-359. Web. 08 Oct 2013.

The National Drought Mitigation Center. Interpretation of 6-Month, 9-Month and 12-Month

Standardized Precipitation Index Map. Washington: NDMC. 2012. Print.

Obasi, GOP. “WMO's Role in the International Decade for Natural Disaster Reduction.” Bulletin

of the American Meteorological Society 75.9 (1994): 1655-1661. Print.

Oetter, Doug R., et. al. “Land Cover Mapping in an Agricultural Setting Using Multiseasonal

Thematic Mapper Data.” Remote Sensing of Environment 76.2 (2001), 139-155. Web. 12

Oct 2013.

Owrangi, Amin, et. al. “Drought Monitoring Methodology Based on AVHRR Images and SPOT

Vegetation Maps.” Journal of Water Resource and Protection (2006): 325-334. Print.

Panu, U.S., and T.C. Sharma. “Challenges in Drought Research: Some Perspectives and Future

Directions.” Hydrological Sciences Journal 47.S1 (2002): S19-S30. Print.

Peters, Albert J., et. al. “Drought Monitoring with NDVI-Based Standardized Vegetation Index.”

Photogrammetric Engineering and Remote Sensing, 68.1 (2002): 71-75.

Peters, Albert J., et. al. “Satellite Assessment of Drought Impact on Native Plant Communities of

Southeastern New Mexico, U.S.A.” Journal of Arid Environments 24.3 (1993): 305-319.

Web. 08 Oct 2013.

Pongracz, R., I. Bogardi, and L. Duckstein. “Application of Fuzzy Rule-Based Modeling

Technique to Regional Drought.” Journal of Hydrology 224.3–4 (1999): 100-114. Web.

13 Oct 2013. 194

Prado, Henia. “Denuncian Suicidios Por Hambre.” Reforma. Reforma.com, 2012. Web. 16 Oct

2013.

Quarmby, J., and C.F. Forster. “An Examination of the Structure of UASB Granules.” Water

Research 29.11 (1995): 2449-2454. Print.

Rahimzadeh Bajgiran, Parinaz, et. al. “Using AVHRR-Based Vegetation Indices for Drought

Monitoring in the Northwest of Iran.” Journal of Arid Environments 72.6 (2008):

1086-1096. Web. 12 Oct 2013.

Rahimzadeh-Bajgiran, Parinaz, Kenji Omasa, and Yo Shimizu. “Comparative Evaluation of the

Vegetation Dryness Index (VDI), the Temperature Vegetation Dryness Index (TVDI) and

the Improved TVDI (iTVDI) for Water Stress Detection In Semi-Arid Regions of Iran.”

ISPRS Journal of Photogrammetry and Remote Sensing 68.0 (2012): 1-12. Web. 12 Oct

2013.

Rasmussen, M. S. “Assessment of Millet Yields and Production in Northern Burkina Faso Using

Integrated NDVI from the AVHRR.” International Journal of Remote Sensing 13.18

(1992): 3431-3442. Print.

Redmond, Kelly T. “The Depiction of Drought: A Commentary.” Bulletin of the American

Meteorological Society 83.8 (2002), 1143-1147. Print.

Renschler, Chris S., and Jon Harbor. “Soil Erosion Assessment Tools from Point to Regional

Scales—The Role of Geomorphologists in Land Management Research and

Implementation.” Geomorphology 47.2–4 (2002): 189-209. Print. 195

Rhee, Jinyoung, Jungho Im, and & Gregory J. Carbone. “Monitoring Agricultural Drought for

Arid and Humid Regions Using Multi-Sensor Remote Sensing Data.” Remote Sensing of

Environment 114.12 (2010): 2875-2887. Web. 10 Oct 2013.

Rojas, O., A. Vrieling, and F. Rembold. “Assessing Drought Probability for Agricultural Areas in

Africa with Coarse Resolution Remote Sensing Imagery.” Remote Sensing of

Environment 115.2 (2011): 343-352. Web. 15 Oct 2013.

Romero-Polanco, Emilio. (2012). “Por Efectos de la Sequía, Más de dos Millones de Mexicanos

Están en Riesgo de Hambruna.” Boletín (2012): 53. Print.

Ruefenacht, Bonnie, et. al. New Technique for Segmenting Images. Salt Lake City: USDA Forest

Service Remote Sensing Applications Center. 2002. Print.

Saadat, H., et. al. “Land Use and Land Cover Classification over a Large Area in Iran Based on

Single Date Analysis of Satellite Imagery.” ISPRS Journal of Photogrammetry and

Remote Sensing 66.5 (2011): 608-619. Print.

Sellers, P. J. “Canopy Reflectance, Photosynthesis and Transpiration.” International Journal of

Remote Sensing 6.8 (1985): 1335-1372. Print.

SEMARNAT. Atlas Geográfico del Medio Ambiente y Recursos Naturales. Mexico City: La

Nueva. 2010. Print.

Shahid, Shamsuddin, and Houshang Behrawan. Drought Risk Assessment in the Western Part of

Bangladesh. Natural Hazards 46.3 (2008): 391-413. Web. 13 Oct 2013.

Sheffield, Justin, and Eric F. Wood. Drought: Past Problems and Future Scenarios. London:

Earthscan. 2011. Print. 196

Sims, Aaron P, and Sethu Raman. “Adopting Drought Indices for Estimating Soil Moisture: A

North Carolina Case Study.” Geophysical Research Letters 29.8 (2002): 21-24. Print.

Sivanpillai, Ramesh, et. al. “Estimating Regional Forest Cover in East Texas Using Advanced

Very High Resolution Radiometer (AVHRR) Data.” International Journal of Applied

Earth Observation and Geoinformation 9.1 (2007): 41-49. Web. 15 Oct 2013.

Sonmez, Kemal F. et. al. “An Analysis of Spatial and Temporal Dimension of Drought

Vulnerability in Turkey Using the Standardized Precipitation Index.” Natural Hazards

35.2 (2005): 243-264. Web. 14 Oct 2013.

Stehman, Stephen V. “Basic Probability Sampling Designs for Thematic Map Accuracy

Assessment.” International Journal of Remote Sensing 20.12 (1999): 2423-2441. Web.

08 Oct 2013.

Story, Michael, and Russell G. Congalton. “Accuracy Assessment-A User's Perspective.”

Photogrammetric Engineering and Remote Sensing 52.3 (1986): 397-399. Print.

Tang, Chunling, and Thomas C. Piechota. “Spatial and Temporal Soil Moisture and Drought

Variability in the Upper Colorado River Basin.” Journal of Hydrology 379.1–2 (2009):

122-135. Web. 12 Oct 2013.

Tucker, Compton J. “Red and Photographic Infrared Linear Combinations for Monitoring

Vegetation.” Remote Sensing of Environment 8.2 (1979): 127-150. Web. 07 Oct 2013.

Tucker, Compton J., and Bhaskar J. Choudhury. “Satellite Remote Sensing of Drought

Conditions.” Remote Sensing of Environment 23.2 (1987): 243-251. Web. 11 Oct 2013. 197

Unganai, Leonard S., and Felix N. Kogan. “Drought Monitoring and Corn Yield Estimation in

Southern Africa from AVHRR Data.” Remote Sensing of Environment 63.3 (1998):

219-232. Web. 12 Oct 2013.

Villarini, Gabriele, et. al. “On the Stationarity of Annual Flood Peaks in the Continental United

States during the 20th Century.” Water Resources Research 45.8 (2009): W08417. Print.

Wardlow, Brian D., Stephen L. Egbert, and Jude H. Kastens. “Analysis of Time-Series MODIS

250 m Vegetation Index Data for Crop Classification in the U.S. Central Great Plains.”

Remote Sensing of Environment 108.3 (2007): 290-310. Web. 08 Oct 2013.

Wilhelmi, Olga V., and Donald A. Wilhite. “Assessing Vulnerability to Agricultural Drought: A

Nebraska Case Study.” Natural Hazards 25.1 (2002): 37-58. Web. 09 Oct 2013.

Wilhite, Donald A., Mark D. Svoboda, and Michael J. Hayes. “Understanding the Complex

Impacts of Drought: A Key to Enhancing Drought Mitigation and Preparedness.” Water

Resources Management 21.5 (2007): 763-774. Web. 12 Oct 2013.

Xiaofan, Liu, et. al. “Assessing Vegetation Response to Drought in the Laohahe Catchment,

North China.” Hydrology Research 43.1/2 (2012): 91-101. Web. 15 Oct 2013.

Yurekli, K., and A. Kurunc. “Simulating Agricultural Drought Periods Based on Daily Rainfall

and Crop Water Consumption.” Journal of Arid Environments 67.4 (2006): 629-640.

Web. 14 Oct 2013.

Zabludovsy, Karla. “Food Crisis as Drought and Cold Hit Mexico” New York Times 31 January

2012. A7. Print. 198

Zierl, B. “A Water Balance Model to Simulate Drought in Forested Ecosystems and Its

Application to the Entire Forested Area in Switzerland.” Journal of Hydrology 242.1–2

(2001): 115-136. Web. 13 Oct 2013.