Intermediate Technical Report

INTERNATIONAL TROPICAL TIMBER ORGANIZATION

PD 228/03 Rev. 3 (F) Phase I

Tropical Forest Fire Monitoring and Management System Based on Satellite Remote Sensing Data in China

INTERMEDIATE TECHNICAL REPORT

By Executing Agency The Research Institute of Forest Resources Information Techniques of The Chinese Academy of Forestry (IFRIT, CAF)

Beijing, China December 20th, 2007

NAME OF THE EXECUTING AGENCY: Research Institute of Forest Resources Information Techniques of The Chinese Academy of Forestry (IFRIT, CAF)

TYPE OF REPORT: Intermediate Technical Report

TITLE OF THE PROJECT: Tropical Forest Fire Monitoring and Management System Based on Satellite Remote Sensing Data in China PD 228/03 Rev. 3 (F) Phase 1

THE PLACE AND DATE IT WAS ISSUED: Beijing, China December 20th, 2007

PROJECT TECHNICAL AND SCIENTIFIC STAFF:

NAME OF PROJECT COORDINATOR: Yi Haoruo, Project Director Ji Ping, Vice-director of the Project

NAMES OF MAIN STAFF: Yi Haoruo Ji Ping Qin Xianlin Xiao Yundan Wu Honggan Tian Yonglin Bai Lina Yang Sidao Zhong Youhong

EXECUTING AGENCY: Research Institute of Forest Resources Information Techniques of The Chinese Academy of Forestry (IFRIT, CAF)

ADRESS: Chinese Academy of Forestry, Box 33, Xiang Shan Lu Road, Hai Dian District, Beijing 100091, CHINA

TELEPHONE: 86-10-62889161

FAX: 86-10-62889161 or 86-10-62888315

E-MAIL: [email protected]

2 Table of Contents

GLOSSARY------3

ABSTRACT------4

1. INTRODUCTION------5

2. METHODS------6

2.1 Investigation and Analysis on Forest Fire and Fire Management in Tropical Region of China ------6

2.2 Design and development of TropFireMAS System------9 2.2.1 Design of TropFireMAS System ------9 2.2.2 Development of TropFireMAS System------11

2.3 Tropical Forest Fire Danger Rating Prediction Sub-System------19 2.3.1 Review of Forest Fire Danger Forecast ------19 2.3.2 Methods ------21

2.4 Tropical Forest Fire Monitoring Sub-System ------26 2.4.1 Review of Forest Fire Identification------26 2.4.2 Methods ------43 2.4.2.1 Fire Identification Method by Using AVHRR Images ------43 2.4.2.2 Fire Identification Method by Using MODIS Images------48

3. RESULTS------62

3.1 Function of the Tropical Basic Databases Sub-system ------62

3.2 Function of the Tropical Forest Fire Danger Rating Prediction Sub-system ------62

3.3 Function of the Tropical Forest Fire Monitoring Sub-system ------70

3.4 Function of Fire Management Information Sub-system------79

3.5 Function of Chinese Tropical Forest Fire Website Sub-system ------79

3.6 Function of Business Sub-systems at All Levels ------81

4. RECOMMENDATIONS------83

i 5. REFERENCES------84

6. APPENDICES------94

6.1 Published Papers by the Project------94

6.2 Forest Fire Danger Risk Predication Results ------98

6.3 Forest Fire Monitoring Results ------101

ii Glossary

CAF Chinese Academy of Forestry

EDA Experiment and Demonstration Area

GDP Guangdong Province

GIS Geographical Information System

IFRIT Research Institute of Forest Resources Information Techniques, CAF

IT Information Technology

RS Remote Sensing

SFA State Forestry Administration

TropFireMAS Tropical Forest Fire Monitoring and Management System Based on Satellite Remote Sensing Data in China

SRSIFRIS Satellite Receive Station of IFRIS

3 Abstract

“Tropical Forest Fire Monitoring and Management System Based on Satellite Remote Sensing Data in China” (TropFireMAS) has been established, based on the scientific achievements and technical accumulation of the Executing Agency on aspects of forest fire satellite monitoring, forest fire danger forecast, information system and network etc., and the combination of the scientific and technical accumulation with the study of present situation of forest fire in tropical regions of China.

TropFireMAS system is an integrated technical system based on platforms of RS, GIS, databases software. The system contains several sorts of background data loaded in Basic Databases Sub-system, and obtains satellite remote sensing data (e.g. NOAA/AVHHR, MODIS and FY data) and Meteorological Data (ground observation data) everyday, then the system issues forest fire monitoring report, forest fire danger forecast report and fire information through Chinese Tropical Forest Fire Website.

This technique report has summarized the relative methods about TropFireMAS system. The system has been structured in six subsystems, including Basic Databases Sub-system, Forest Fire Danger Forecast Sub-system, Satellite Remote Sensing Monitoring Sub-system, Fire Management Information Sub-system, Chinese Tropical Forest Fire Website and Business Sub-system. Methodology, especially that about forest fire monitoring and forest fire danger forecast which adapts to tropical region, have been introduced in detail in this report, based on the function design of all the systems. At the same time, these subsystems have been operated and verified in the EDA. Parts of results have been listed in this report. It shows that the approach of applying new-high technology for improvement of forest fire prevention is an effective way in terms of technology with less investment needed and with shorter construct period. So, the approach is the best way to improve forest fire prevention in tropical forest region of China.

Through the implementation of the Project Phase 1, the local fire prevention organizations and stuffs fully uphold and support the project, and they believe that this project will have a great potential.

1. Introduction

China’s tropical forest region is a region with high frequency of forest fires. According to statistics from 1950 to 2005, the frequency of forest fires in China amounts to more than 7 million with the total damaged forest areas of more than 300 million hectares. The forest fire loss amounts to more than several hundred billion RMB (Yuan) in China. In recent years, Chinese Government is implementing the key forestry projects of Natural Forest Protection, and Returning Farmland to Forest etc. Young forest areas have been greatly increased after afforestation in vast areas; while the implementation of closing hillside for forest cultivation makes combustible materials increase luxuriantly. All these factors have caused forest fire danger and fire calamity to be more severe in tropical regions. The forest fire problem in tropical regions of China has become a problem that urgently needs to be solved to protect and develop tropical forest of China.

Problems to be solved by the project are: First, the low level of forest fire danger forecast restricts forest fire prevention capability and effect of prevention, which leads to high frequency of forest fires caused by human activities in the forest region.

Secondly, the low level of forest fire monitoring and delayed detection of forest fires restrict timely suppression of forest fire in China’s tropical forest regions, so the fire is easy to spread and to cause heavy fire, thus leading to serious forest fire loss in the region.

Thirdly, the forest fire management level in China’s tropical region urgently needs to be enhanced due to its low level of digitization and informatization.

After investigation of forest fire, fire management and its relevant technology application in tropical forest regions of China, the project team considers that the construction of “Tropical Forest Fire Monitoring and Management System Based on Satellite Remote Sensing Data in China” (TropFireMAS) is the best way for the solution of forest fire problems in China’s tropical region with less investment and quick effectiveness.

5 2. Methods

2.1 Investigation and Analysis on Forest Fire and Fire

Management in Tropical Region of China

The project team went to EDAs of Guangdong Province to investigate the situations of forest fire and fire management in tropical region of China, to study several types of forest sub-compartment, and to survey typical fire scars in the field in May and June 2006, when the project was at the starting stage. After summarizing and analyzing the information of investigation, the project team has deeply understood the issue of tropical forest fire in China. The main conclusions of the Investigation and analysis are described as follow.

1) Regarding forest fire management

Forest fire management level has been greatly raised in tropical forest regions of China in recent years. The main indications are:

(1) Both national and local laws, stipulations and regulations regarding forest fire prevention have been carried through and put into effect. Fire management and related work are done according to laws. Forest fire prevention organizations of all levels are sound and fire fighting groups in many places possess practical fighting ability. Field fire management during fire prevention period is in order.

(2) Construction of equipments and infrastructure facilities for forest fire prevention has gained great progress. The communication condition for forest fire prevention has basically satisfied the need of forest fire prevention and suppression with the essential improvement of audio communication condition in the forest land. Large scale construction of firebreak forest belt is being developed in the region of developed economy of Guangdong Province. That will greatly ameliorate the fundamental condition of forest fire prevention there.

(3) Fire prevention work at community level has been strengthened. Multiple forms of education and publicity on forest fire prevention have been developed. Local community has been effectively organized for fire prevention action.

6 2) Regarding the infrastructure construction of forest fire prevention and capability of integrated fire prevention

The infrastructure construction of forest fire prevention in tropical forest regions of Guangdong Province is weak and the capability of integrated fire prevention is low. Its main manifestations are:

(1) The forest fire prevention facilities

The density of roads in tropical forest regions of Guangdong Province is low due to limitation of economic situation of forest regions, and the density of forest fire isolation belt and forest fire observation towers are also low in these regions. According to relevant statistics, the density of fire break belt in Guangdong Province is 95KM/100KM2, while the density of fire break forest belt is 61KM/100KM2. The density of forest fire observation towers is 1.06/100 KM2 in Guangdong. These infrastructure facilities of forest fire prevention are all by far below actual demanded quantity.

In recent years, since Chinese government invested several hundred billions of RMB in implementation of projects such as Natural Forest Preservation Project, Returning Farmland to Forest and other key forest construction projects, forest areas have been evidently increased, with an obvious augment of combustible material under forest crown. In addition, the forest in these regions is mostly conifer forests, pure forests of middle age or young forests, so the situation of forest fire prevention becomes extremely rigorous. However, it is unrealistic to invest large amount of money in a short time to strengthen construction of forest fire prevention facilities, i.e. to increase large quantities of forest roads, forest fire break belts, forest fire observation towers. So at present, what is the feasible way to enhance forest fire prevention in tropical forest regions of China? This is just the problem to be solved through this project, i.e. the realization of forest fire danger prediction with higher spatial resolution, which would help to attain the goal of greatly improving the ability to prevent forest fires in tropical regions of China with less investment of manpower and finance.

(2) The forest fire monitoring

Currently，the villagers in forest areas usually dial “109”, which is phone number for emergent fire warning, to give an alarm when forest fire is found.

7 The TropFireMAS System includes forest fire satellite monitoring sub-system for timely discovery of forest fires. The function of the sub-system is to increase the capability of timely discovering of forest fires in tropical forest regions of China. The sub-system possesses obvious advantage to find forest fire occurring in remote forest areas with sparse residents. It means that fire would be discovered without delay by using the sub-system and then there could be a possibility to put out fire promptly before it extends to large areas.

(3) The forest fire suppression

Forest fire suppression is mainly carried out at present by using simple tools with manpower as the dominant factor in tropical forest regions of China, so the fire control ability is very low. Once fire breaks out, both professional and voluntary fire fighting groups can all reach fire scenes in time and fight furiously. This is to actually remedy the defect of facilities and techniques by putting in more manpower. But it becomes more and more difficult to apply the method of putting out fire by relying on large amount of manpower, while economic system is changing in forest regions.

3) Regarding the way to enhance forest fire management in tropical region of china

The problem of forest fire calamity is a crucial problem related to local forest construction, and even to local social and economic development. Governments of various levels in tropical regions of China have attached more and more importance to prevention and suppression of forest fire. As described above, the improvement of integrated condition for forest fire prevention, such as the amplification of forest fire prevention facilities, the renewal of fire fighting equipments and techniques, the change of local farming system, the progress of local cultural tradition etc. all rely on local economic development, amelioration of local people living standard and strengthening of general national power, which is not at all one day’s merit.

What is the best way to strengthen forest fire management and prevention in tropical regions of China? The opinion of the project team is further determined after the investigation: (1) To establish TropFireMAS System with forest fire danger prediction and satellite monitoring as its main parts. The system will provide information of forest

8 fire prevention and suppression to fire management departments of all levels through communication and network platform; (2) To carry out technical training on how to use the system and its information to fire management staff of all levels. The system will improve forest fire prevention and suppression through advancing application of new-high technology in tropical forest regions of China.

2.2 Design and development of TropFireMAS System

2.2.1 Design of TropFireMAS System

TropFireMAS System is an integrated technical system based on platforms of RS, GIS, databases software. The system contains several sorts of background data loaded in Basic Databases Sub-system, and obtains satellite remote sensing data (e.g. NOAA/AVHHR, MODIS and FY data) and Meteorological Data (ground observation data) everyday, then issues forest fire monitoring report, forest fire danger forecast report and fire information through Chinese Tropical Forest Fire Website Sub-system.

The system is composed of 6 sub-systems, which include: (1) Sub-system of Basic Databases on EDAs, (2) Forest Fire Danger Forecast Sub-system for China’s Tropic Forest Region, (3) Satellite Remote Sensing Monitoring Sub-system for China’s Tropical Forest Fire, (4) Fire Management Information Sub-system for China’s Tropical Forest Region, (5) Chinese Tropical Forest Fire Website; (6) Business sub-systems at all levels (general version)

TropFireMAS System can be divided into two parts, which are Operating Center and

9 User Terminal, according to the location installed.

The Operating Center of TropFireMAS System is consisted of sub-system (1), (2), (3), (4) and (5), and is installed and operated at Executing Agency in duration of project implementation. The User Terminal of TropFireMAS System is the sub-system (6), which is installed and operated at all forest fire prevention organizations in EDAs, connected with Operating Center by internet.

The structure of TropFireMAS System is shown in figure 1.

Satellite Data Meteorological Data

Forest Fire Monitoring Forest Fire danger Forest Fire Management

Sub-system forecast Sub-system Sub-system

Basic Databases Sub-system

Forest Fire Website

Sub-system

TropFireMAS

Internet/Forestry Net

Business Business s ub - s y s te m 1 • • • • • • sub-system n

User 1 User n • • • • • •

10 Figure 1. Function Structure of TropFireMAS System

2.2.2 Development of TropFireMAS System

(1) Sub-system of Basic Databases on EDAs

This Sub-system, which stores all sorts of data for other 5 Sub-systems, is the foundation of the whole TropFireMAS. The sub-system contains all sorts of data, including: l Basic geographical elements of Guangdong Province (scale 1:50,000), l Basic geographical elements of Guangdong Province (scale 1:250,000), l Basic geographical elements of Guangdong Province (scale 1:1,000,000), l Forest distribution map of Guangdong Province (scale 1:500,000), l Forest distribution map of Guangdong Province (scale 1:2,000,000), l Vegetation map of Guangdong Province (scale 1:1,000,000), l Land usage map of Guangdong Province (scale 1:250,000), l Soil classification map of Guangdong Province (scale 1:1,000,000), l The list of the hot spots extracted from NOAA/AVHRR data (2000~2005), l The statistical table of fire events of EDAs (Yunan County, Yangchun City and Huidong County in 2000~2005), l Regulation on forest fire prevention of Guangdong Province, l Regulation on forest fire prevention of Yunan County, l Regulation on forest fire prevention of Yangchun City, l Regulation on forest fire prevention of Huidong County, l The name list of fire fighting groups of EDAs (Yunan County, Yangchun City and Huidong County), l The Time schedule on duty of EDAs (Yunan County, Yangchun City and Huidong County),

The structure of the sub-system is shown in Figure 2.

11 Meteorological RS Data Network Data Network Public Network

Geographical Data Remote Sensing Data Weather Data EDA’s Data

Forest Distribution Fire Management Information

Vegetation Distribution Training Books

Land Utility Databases

Forest Fire Monitoring Forest Fire Management Forest Fire Danger

Sub-system Sub-system Forecast Sub-system

Forest Fuel Distribution Map of Active Fires Fire Information Map of Fire Danger Website Forest Fuel Moisture

Forest Fuel Growth Trend

Public Users Users of the Forest Fire Prevention Organizations

Figure 2. The Structure of the Sub-system of Basic Databases on EDAs

12 (2) Forest Fire Danger Forecast Sub-system for China’s Tropic Forest Region

This Sub-system offers the report of forest fire danger forecast on EDAs everyday for future 24 hours during forest fire season.

This sub-system obtains needed meteorological observation data of related area as well as weather forecast data from China Meteorological Service Network. It also obtains observed data of fire factors and typical meteorological data in EDAs through Internet, as well as fuel data in DEAs extracted from satellite data. Using forecast models on forest fire danger, it thus concludes forecast reports on forest fire danger under the support of the basic databases of the System (1).

The core of the Sub-system is forecast technology on forest fire danger, including: l Fire danger rating indexes l Forecast models

The daily input data of the Sub-system include: l Sets of meteorological data of real-time observation, l Local observation data of meteorological factors and fire danger factors, l Dynamic data of vegetation cover and fuel moisture extracted from MODIS data,

The structure of the sub-system is shown in Figure 3.

Basic Databases Satellite Data Ground Observation Data Forest Distribution

Vegetation Vegetation Weather Observation Typical Administrat iv e Boundary Meteorologica Distribution Grows Forecast Data on Fire Meteorological l Observation GIS* Databases Condition Data Danger Data

(1:250,000, 1:50,000)

…… Dynimic Data of RS* Meteorological Network Internet/Forestry Net Background Data of RS of China Telephone/Fax

Prediction Models

Modify

Results

Figure 3. The Structure of Forest Fire Danger Forecast Subsystem

GIS*: Geographic Information System (GIS) RS*: Remote Sensing (RS)

14 (3) Satellite Remote Sensing Monitoring Sub-system for China’s Tropical Forest Fire

This Sub-system offers forest fire monitoring reports on EDA max. 4-6 times per day during forest fire season; and report on fire damage when forest fire with large areas occurs in EDA.

This sub-system obtains real-time remote sensing data through the satellite data receiving station, and concludes forest fire monitoring reports through digital image processing and forest fire information extraction, and by using monitoring models on forest fire under the support of the basic databases of the Sub-system (1). It also concludes reports on forest fire damage through burnt forest information extraction, and by using models on fire damage under the support of the basic databases of the Sub-system (1).

The core of the Sub-system is forest fire monitoring technology by using satellite remote sensing data, including: l Detection of small fire spot l Fire spot positioning l Area calculation of burnt forest land

The daily input data include: l MODIS data, l NOAA/AVHHR data l FY data

The structure of the Sub-system is shown in Figure 4.

Multi-Satellite Data (NOAA, MODIS, FY)

Antenna Receiver Antenna Controller

Satellite-Signals Receiving

Forest Fires Information Extracting Basic Databases Geometric and Radiometric Correction

Forest Distribution

Administrative Boundary

Ground Abnormity Hotspot Vegetation Information Burned Information GIS* Databases Information Extracting Extracting Extracting (1:250,000, 1:50,000)

……

Ground-Information

Model of Forest Model of Forest Statistical Information Fire Identification Fire Damage

Reports of Forest Fires Reports of Fire Damage

Figure 4. The Structure of Forest Fires Monitoring Sub-system

16 (4) Fire Management Information Sub-system for China’s Tropical Forest Region

This Sub-system manages various sorts of data and materials, which are stored in the Sub-system (1), regarding forest fire prevention, suppression and management, and offers these data to users with the function of searching, browsing, and so on.

The structure of the Sub-system is shown in Figure 5.

Forest Fire Danger Remote Sensing Data

Input Input Active Fires Map of Active Fires

Edit Edit GIS Weather Observation Search Search Map of Forest Fire

Fire Management Map of Weather

Database

Data User Log

Information of Forest Registrati Operation

Fire Danger Forecast Log in IP Recording

Fire Management Access User

User

Figure 5. The Structure of the Sub-system of Forest Fire Management

(5) Chinese Tropical Forest Fire Website Sub-system

17 The Sub-system of Tropical Forest Fire Web-site provides rich forest fire information, including forest fire danger forecast, forest fire monitoring and forest fire management etc., to support forest fire prevention, suppression and management.

Forest fire prevention organizations of different levels and the public could search, browse and download the data and information on the website through special forestry network and/or Internet.

(6) Business Sub-systems at All Levels (general version)

This Sub-system is User Terminal of TropFireMAS System, and is installed at forest fire prevention organizations of EDAs. The Sub-system manages various local data on forest fire prevention, and connects users of EDAs with the Operating Center, which is composed of the Sub-system (1), (2),(3),(4) and (5) of TropFireMAS System, through internet.

Persons on duty and fire commanders of EDAs and forest fire prevention offices at all levels could use the Sub-system, to get report on forest fire danger forecast, report on forest fire monitoring and various fire management information, and apply them to daily forest fire management activities.

18 2.3 Tropical Forest Fire Danger Rating Prediction Sub-System

Forest fire is a prominent global phenomenon, which not only destroys forest resource, but also poses enormous danger to wildlife as well as to human life and property. In addition, biomass burnt by fires has been identified as a significant source of aerosols, carbon fluxes, and gases, which pollute the atmosphere and contribute to radiative forcing responsible for global climate change (Maselli, 1996; Goldammer,1998). Forest fire danger rating prediction and forest fire monitoring can predict the occurrence probability and behavior of forest fire, by which damage can be eliminated as much as possible. So, it’s a very important instrument to enhance forest fire prevention abilities of tropical forest region of China.

2.3.1 Review of Forest Fire Danger Forecast

Fire danger rating system is a system that integrates the effects of existing and expected states of selected fire danger factors into one or more qualitative or numeric indices that reflect an area’s protection needs. The development and application of forest fire risk prediction technique is most 80 years. There have been many researches on forest fire risk prediction in advanced counties or organization, such as USA, Canadian and European Space Agency (ESA). A method to rate wildland fire danger has been studied as far back as 1940 by USDA. The first National Fire Danger Rating System (NFDRS) of USA has been developed in 1972. This system has since been revised many times (Burgan et al. 1988; Burgan et al. 1998). Recent improvements to fire potential assessment technology include both broad scale fire danger maps and local scale fire behavior simulations. In the context of local scale fire behavior, FARSITE and BEHAVE (Finney 2003) provide methods to simulate fire behavior for areas up to several thousand hectares. In the broad area fire danger context, spot measurements of fire danger, calculated using the NFDRS at specific weather stations, are being interpolated and mapped on a national basis through the wildland fire assessment system (Burgan et al. 1998; http://www.fs.fed.us/land/wfas/welcome.html). The U.S. maps are produced by using an inverse distance squared weighting of staffing levels. It is based on comparison of current fire danger index values with historical values. Because fire managers across the United States have not been consistent in their selection of an NFDRS index on which to base staffing levels, while staffing level itself is the only common parameter

19 with which to map fire danger. Staffing level normalizes all indexes against their historical values, so it does not matter which of the several fire danger indexes a fire manager would select. However this method neither addresses the effect of topography on fire potential, nor provides fire potential estimates for specific locations or landscape resolutions (Carlson et al. 2002; Jack et al. 2002). Canadians and ESA publish similar maps for their fire danger system on the internet (B. J. Stocks, 1987; http://www.nofc.forestry.ca/fire/cwfis; Http://ies.jrc.cec.eu.int/94.html.).

The occurrence and behavior of forest fire rely on many factors, including background constant factors (topography, fuel types) and variable factors (weather, fuel Greenness, fuel moisture content, etc.) (Xinghua Li, et al. 2001; Weiqi Zhou, et al. 2003; Haoruo Yi, et al. 2004; Xianlin Qin, 2005). In some countries, forest fire risk prediction combined measures of weather, fuel and topography. However, there are many researches that focus on forest fires risk prediction on locale scale by only using some weather variables, such as temperature, relative humidity, wind speed and precipitation to predict the potential for starting a fire. Decisions fire managers must make depend on the temporal and spatial scales involved as well as management objectives. Assessment of fire potential at any scale requires basically the same information about the fuels, topography, and weather conditions that combine to produce the potential fire environment. These factors have traditionally been measured for specific sites, with the resulting fire potential estimates produced as report, and the results applied to vaguely defined geographic areas and temporal periods, with the knowledge that the further one is displaced (in time or space) from the point where such measurements have been taken. This situation is rapidly changing because Geographic Information Systems (GIS) and space-borne observations are greatly improving the capability to assess fire potential at much finer spatial and temporal resolution. Satellite images, such as NOAA/AVHRR, TM, SPOT, Moderate Resolution Imaging Spectroradiometer (MODIS), etc., have been used in forest fire risk system (Burgan et al. 1998; Dodi Sudiana et al. 2003; etc.). In China, Remote sensing and computer techniques began to be applied in forest and grassland resources management since 1990. Achievements in fires monitoring and forest fire risk prediction have been attained by remote sensing data (Haiqing Zheng et al. 2001; Litao Wang et al. 2004; Yi Zhou et al. 2004; Guangmeng Guo et al.2005). These research achievements have supplied technology resources to the tropical

20 Forest Fire Danger Rating Prediction Sub-System.

2.3.2 Methods

The level of forest fire danger is not only related to such factors, as weather, topography, society and economy etc., but also to dynamic and static information describing the condition of forest fuels, such as forest fuel type, forest fuel greenness and forest fuel moisture content (FMC) etc. The potential of fire occurrence depends on the conjunct effect of dead and live vegetation, the moisture in live vegetation, and the moisture in dead vegetation. A key technical step for large scale forest fire danger rating prediction is to getting effectively dynamic information (e.g. greenness and FMC) of forest fuel. The development of modern information technologies, such as Remote Sensing (RS), GIS, combined with database and internet technique, have provided important advances for studying forest fire danger rating. The amount of dead and live vegetation is estimated by using data from a lower resolution sensor, such as AVHRR or MODIS. The baseline forest fuel map, given by low-spatial and high-temporal resolution satellites, such as AVHRR and MODIS, can be used to monitor near real-time changes in the vegetation vigor, which is correlated with the moisture of live vegetation. The moisture in dead vegetation is estimated by using knowledge of local weather conditions. The baseline forest fuel map and a real-time estimate data of the vegetation condition are needed for forest fire danger forecast.

Supported by RS, GIS and internet technologies, the Forest Fire Danger Index (FFDI), which combines measures of weather variables and fuel features, was developed by MODIS images, databases and meteorological data. These two basic indicators (Weather Index (WI) and Fuel Index (FI)) were selected to construct and calculate the FFDI. The resulting FFDI was classified to 5 levels: low, moderate, high, very high and extreme according to its value.

1) Construction of the FFDI

The occurrence of forest fires mainly depends on three elements: fuel, weather and topography. The fuel can be characterized by such parameters, as fuel type, biomass condition (living or dead), biomass quantity, fuel moisture content, and continuity (vertical and horizontal structure). The influence of weather on fire risk can be

21 characterized by temperature, relative humidity, wind speed, wind direction and precipitation etc. Based on statistical analysis, these parameters (relative humidity, temperature, wind speed and precipitation) are highly correlated with fire occurrence and development (X. H. Li, 2001). In this study, these four parameters are combined to calculate the WI, and other three fuel indicators (fuel greenness, fuel moisture content and fuel ignition weight) were selected to calculate the FI. We haven’t gotten enough historical topographic and fire origin data about their relationship with FFDI, so, the factors of topographic and fire origin is not considered in this model.

The FFDI, integrated by forest fuels, weather, terrain, fire origin, is an index to indicate the forest fire occurrence and to map fire potential in EDA scales. The FFDI model was developed by incorporating satellite measurements and surface observations, as well as the use of GIS to fulfill two objectives: (1) to depict fire potential at EDA scale and at 1 km2 spatial resolution, (2) to estimate forest fire potential that can be operated for Tropical Forest Fire Danger Rating Sub-System at EDA scale.

2) Calculation of the FFDI

Supported by GIS techniques, WI and FI were calculated by weather data, digital maps and MODIS images. The study area was divided into grids with a cell size of 1km2 and the calculation of each indicator was conducted in a cell.

(1) Weather Index

Weather index (WI) includes relative humidity, temperature, wind velocity and precipitation. With the comparison of three interpolation approaches (Inverse Distance Weighted (IDW) interpolate, Kriging interpolate and Spline interpolate), IDW was selected for interpolation to obtain resulting surfaces with regular grids of values. Data from irregularly spaced weather stations in Guangdong province were collected to create raster surfaces of relative humidity, temperature, wind velocity and precipitation.

(2) Fuel Index (FI)

22 The definition of FI is expressed in formula (1), in which FI is the function of the variables: Relative Greenness (RG) and Fuel Moisture Content (FMC). The value range of FI was scaled from 0-100.

FI = (1- RG)´(1- FMC) (1)

(3) Relative Greenness (RG)

The purpose of using relative greenness (RG) in the FFDI model is to define the proportion of live and dead vegetation. The RG could be derived with formula (2) in which the NDVI value used is from MODIS on Terra/AQUA satellites or AVHRR on NOAA satellites products directly.

RG = (NDi - NDmin ) (NDmax - NDmin ) (2)

Where,

NDi : highest observed NDVI value for the 8 day composite period;

NDmin : historical minimum NDVI value for a given pixel;

NDmax : historical maximum NDVI value for a given pixel.

NDmax and NDmin are maximum and minimum of the historical NDVI data from march,

2000 to march, 2004 for each pixel, and NDi is the highest observed NDVI value for the 8 day composite period. Therefore, the RG indicates how green each pixel currently is in relation to the range of historical NDVI observations for it with typical value range from 0 to 100. Low RG values indicate that the vegetation is at or near its minimum greenness (R. E. Burgan, 1998).

(4) Fuel Moisture Content (FMC)

Fuel Moisture Content (FMC) is an important parameter in determining forest fire risk and forest fire behavior (G. W. Paltridge, 1988; R. E. Burgan, 1997; E. Chuvieco, 2004, etc.). Estimation of vegetation water content is central to the understanding of biomass burning processes. The most practical, objective, and cost-effective way to monitor vegetation from a local to global scale is the use of earth observation technologies. Satellites can provide local to global coverage on a regular basis. They

23 also provide information on remote areas where ground measurements are impossible on a regular basis (E. R. Hunt, et al., 1989; S. Jacquemoud, 1990; N. R. Viney, et al., 1991; etc.).

Given an ignition source, the probability of ignition and spread of forest fire is strongly dependent on the moisture content of small dead vegetation. Many researches have demonstrated that multiple sensors currently onboard earth observation satellites are suitable for the monitoring of vegetation water content, which are usually classified into the following three categories (B. C. Gao, 1996; P. Ceccato, et al., 2001; E. Chuvieco, et al., 2001; 2002; Pei Zhang, 2004, Xianlin Qin, 2005; etc.): (1) Visible and Short Wavelength Infrared Sensors (SWIR, spectrum between 400 and 2500 nm), which provide information on vegetation biophysical parameters such as the chlorophyll content, the leaf area index, and the vegetation water content. Because water has several absorption maxima throughout the infrared region of the spectrum, quite a number of different indices and techniques have been developed for estimation of water content. (2) Long Wavelength Infrared sensors (spectrum between 6.0 and 15.0 mm), which offer information on thermal dynamics of vegetation cover hence has been used to estimate the evapotranspiration of vegetation canopies, a parameter that is closely related to water stress. (3) Radar sensors (spectrum between 0.1 and 100 cm), which give information on the dielectric constant related to vegetation water content.

The probability of forest fire igniting and spreading out strongly depends on the moisture content of small dead vegetation when a given ignition source appears. According to relevant researches, different sensors boarding on the earth observation satellites may be applicable to the monitoring of vegetation water content. The sensors with wave-band from visible light to shortwave infrared (SWIR, spectrum is between 400 and 2500 nm) provide information on vegetation biophysical parameters, such as the chlorophyll content, the leaf areas index, and the vegetation water content. There are three channels (channel 5, channel 6 and channel 7) of MODIS which belong to SWIR. The results of FMC have been compared by using SWIR and NIR channel to calculate FMC. In this system, channel 7 and channel 2 of MODIS have been selected in the FFDI model to calculate FMC. The expression is equation (3).

24 r'2 -ri Bi2 = (3) r'2 +ri

Where:

ρ’2: the correct reflectance of channel 2 of MODIS; it is calculated by formula (4);

SWIR: the reflectance of SWIR channel of MODIS;

i: channel 7 of MODIS

ρ’2 =ρ2+6.0ρ1-7.0ρ3 (4)

Where:

ρ1: the reflectance of channel 1 of MODIS (Red channel);

ρ2: the reflectance of channel 2 of MODIS (NIR channel);

ρ3: the reflectance of channel 3 of MODIS (Blue channel);

(5) Forest Fire Danger Index (FFDI)

The assumptions of FFDI model are: (1) fire danger can be assessed if proportion of weather and fuel characters is defined; (2) vegetation greenness provides a useful parameterization of the quantity of high moisture content live vegetation; (3) the FFDI in non-fuel pixel is 0.

The specific process for each pixel is to obtain the inputs from the 1-km WI, RG and FMC map. Then, equation (5) is a basic model for calculation of FFDI for each pixel with inputs 1 km WI, RG and FMC map generated by former processes. The spatial resolution of FFDI map, therefore, is 1 km and the value range scales from 0-100.

(WI +FI ) FFDI = 2 (5)

3) Categories of FFDI

The 1-km resolution FFDI map is scaled from 0-100. It was classified to 5 categories according to its value: low, moderate, high, very high and extreme. The explanations of 5 classes are shown in Table 1. At the same time, the pixels belong to water and cloud will be eliminated.

25 Table 1 Categories and Description of FFDI Fire Danger Criticality Ignitability Extend Character FFDI Grade Grade 1 Low Fuels do not ignite Fires spread slowly 0%~25% Grade 2 Moderate Fires can start from most Timber fires spread slowly to 25%~50% accidental causes moderately fast Grade 3 High All fine dead fuels ignite Fires spread rapidly and 50%~75% readily and fires start short-distance spotting easily from most causes Grade 4 Very High Fires start easily from all spread rapidly and increase 75%~90% causes and immediately quickly in intensity after ignition Grade 5 Extreme Fires start quickly spread furiously and burn >90% intensely

2.4 Tropical Forest Fire Monitoring Sub-System

2.4.1 Review of Forest Fire Identification

Timely and accurate detection of forest fires has become an important issue regarding fire management over the world. Satellite-borne sensors are available to detect fires by using the data with visible, thermal and mid-infrared bands. Active fires and fire scars can be detected day and night. Sensors with frequent overflights are needed and data from the overflights must be available in near real-time. Various international organizations, such as the International Geosphere and Biosphere Program (IGBP), have recognized the need for fire detection and monitoring. Space-borne fire detection has been a topic of intensive research since the early 1980s. Much of the work has concentrated on fire monitoring methodology or estimation of atmospheric emissions from fires. Actually, several times of satellite images covering the same location can be gotten in the duration of forest fire identification. The main satellite images which have been applied into forest fire identification of EDAs in TROPFireMAS include Advanced Very High Resolution Radiometer (AVHRR), which is on board National Oceanic and Atmospheric Administration (NOAA) Satellite, and Moderate Resolution Imaging spectrometer (MODIS), which is carried on both the Terra and Aqua satellites. So, the fire identification methodology by using NOAA/AVHRR and MODIS instrument has been reviewed in following.

1) Fire detection using NOAA/AVHRR data

Until now, most fire detection activities are based on the use of NOAA/AVHRR data.

26 The AVHRR series of sensors offer a spatial resolution of 1km and cover most of the earth’s surface every day at daytime and nighttime. AVHRR data have been widely used for fire detection because they have some unique radiometric advantages compared to other satellite data, and provide a good balance in spatial and temporal resolutions. There are many algorithms to detect fires by using AVHRR data.

(1) NOAA/AVHRR Instruments

The current sequence of NOAA satellites has been in continuous operation since Oct. 1978. The Advanced Very High Resolution Radiometer (AVHRR) is a scanning radiometer measuring reflected and emitted radiation in four channels on-board the satellites NOAA-6, - 8, and -10; in five channels on board NOAA-7, - 9, -11, -12, -13, and -14; and in six channels on board NOAA-15, -16, -17, and -18 (Table 2.1). The term “very high resolution” refers to a high radiometric resolution. AVHRR data are recorded to 10 bit precision. For the thermal infrared channels, on-board calibration exists through the regular measurement of the deep space and a blackbody of known temperature on board. No inflight calibration exists for the visible and infrared channels. However, since these channels are widely used for vegetation monitoring, efforts have been made to monitor calibration through repeated measurement of surface targets which are assumed to be stable. Daily global coverage is achieved at the expense of the wide scan angle of the sensor (which reaches a maximum of 55.4° on either side of the nadir) and the large instantaneous field of view (IFOV). This viewing geometry gives rise to a coarse 1.1 km ground resolution for pixels at nadir. The wide scan angle leads to pixels at image edges being over three times the size of pixels at nadir.

Table 2.1 Channels Characteristics of NOAA-AVHRR K, L, M, N Channel Spectral Spatial Resolution Signal to Noise (S/N) or Wavelength at nadir (km) Noise Equivalent Delta (micrometers) Temperatures (NEΔT)

1 (Visible) 0.580 - 0.68 1.1 9:1 at 0.5% Albedo

2 (Near IR) 0.725 - 1.00 1.1 9:1 at 0.5% Albedo

3A (Short Wave 1.580 - 1.64 1.1 20:1 at 0.5% Albedo Near IR)

3B (IR-Window) 3.550 - 3.93 1.1 0.12 K at 300 K

4 (IR-Window) 10.300 - 11.3 1.1 0.12 K at 300 K

27 5 (IR-Window) 11.500 - 12.5 1.1 0.12 K at 300 K

Channel 1 (0.58 μm to 0.68 μm) in the yellow/red part of the spectrum corresponds reasonably well with the spectral interval (0.6 μm to 0.7 μm) recommended for green vegetation analysis, as it corresponds with the strong chlorophyll absorption band. Channel 2 (0.725 μm to 1.1 μm) is potentially less satisfactory as it includes part of the chlorophyll absorption band as well as the more important highly reflective plateau in the near infrared (beyond 0.8 μm), which is associated with healthy green vegetation. Channel 3A (1.58 μm to 1.64 μm), which is associated with moisture content green vegetation. Channel 3B (3.55 μm to 3.93 μm), of the AVHRR is located near the spectral maximum of radiative emissions for objects radiating at temperatures around 800 K, i.e. the near-normal temperature for burning grass (Langaas and Muirhead, 1988), whilst channel 4 (10.3 μm to 11.3 μm) and channel 5 (11.5 μm to 12.5 μm) are located near the spectral maximum for normal environmental temperatures, around 300 K. As a consequence, in the case of burning grass (800 K) channel 3 will receive much more radiant energy than channel 4, and is therefore well suited to the detection of high temperature sources e.g. fire.

Full resolution AVHRR data (1.1 km) can be recorded for selected areas of the world in the LAC (Local Area Coverage) mode, or transmitted directly from the satellite in the High Resolution Picture Transmission (HRPT) mode for areas within a radius of 2500 km of a receiving station. Fire detection techniques are most reliable using these data (which can be requested on a daily basis), especially in areas where there is a good contrast between high temperature fires and cooler temperatures for the surrounding environment.

The afternoon pass of the NOAA-AVHRR (nearly 11:00am and 14:30pm equatorial crossing, such as of NOAA -14, -16 and -18), is near-optimal in terms of fire detection and monitoring in the tropical belt, when winds tend to be stronger, the ground dew has evaporated and the vegetation is dry enough for efficient ignition. However, this does not preclude the possibility of missing fire events e.g. due to cloud cover, which can be at a maximum at this time over tropical regions. Although Langaas (1992), used night-time AVHRR imagery (NOAA-10, 19:30-20:30 and NOAA-11, 01:30-02:30) to avoid problems such as channel 3 saturation and the

28 unknown contribution of daytime reflected solar radiation to channel 3, a drawback is that the satellite overpass time may be less optimal with respect to the time of maximum fire activity. In addition, even fires persisting for several days can show a strong diurnal cycle whereby they die down during the night and do not burn strongly again until the following afternoon (Belward et al. 1993). Night-time imagery therefore represents a very selective sample. Dozier (1981) introduced a theoretical approach to the study of sub-pixel temperature fields using the AVHRR. He approximated the temperature field of each pixel by two areas of uniform temperature, the background area and a target area which occupies some fraction from 0 to 1 of the pixel. Using his model, Dozier was able to show that a sub-resolution high temperature target is detectable because it has a greater effect in channel 3 than in channel 4. Also based on his model, Dozier arrived at a pair of simultaneous, nonlinear equations that could be solved for the temperature and the size of the hot target given the background temperature and the brightness temperatures in channels 3 and 4. These equations, though, are restricted to pixels that are unsaturated in both channels. Background temperature is estimated from adjacent pixels and Dozier suggested using a split-window technique to calculate a correction factor for the effect of atmospheric water vapor.

The first applications of fire detection with AVHRR were to fixed targets of known location. Matson and Dozier (1981) used NOAA-6 imagery to detect high temperature industrial sources in Detroit and waste gas flares in the Persian Gulf. Similarly, Muirhead and Cracknell (1984) detected gas flares associated with oil fields in the North Sea using NOAA-6. In neither case were the details of the detection criteria discussed but presumably manual inspection of both channels 3 and 4 was an important component. An important difference between the two studies is that Muirhead and Cracknell (1984) used daytime imagery while Matson and Dozier (1981) used nighttime imagery, citing concern over the effects of reflected solar radiation in channel 3. Following these studies of fixed targets, the AVHRR was used to detect vegetation fires. Matson et al. (1984) used the daytime image of NOAA-6, 7 over the U.S. and Brazil to detect fires based on enhanced temperatures in channel 3. They also reported on an experimental operational fire monitoring project in the western U.S that was based on manual inspection of channel 3 and channel 4 images alternating on a screen. Most importantly, these detections were largely verified by

29 forest service stations, timber companies and local police. Muirhead and Cracknell (1985) also looked at vegetation fires by applying their technique to the detection of straw burning in Great Britain using daytime NOAA-7 imagery.

The work of Flannigan and Vonder Haar (1986) and Flannigan (1985) represents the first attempt to use an automated (i.e. non-interactive) set of detection criteria. They used both daytime and nighttime data from NOAA-7 to monitor a severe forest fire outbreak in north-central Alberta, Canada. The detection criteria were based on manual inspection of a subset of the data and then applied in an automated fashion to the full dataset. Detection was based on a threshold channel 3-channel 4 brightness temperature difference of 8 K for nighttime data and 10 K for daytime data. They were also able to compare satellite detections with Alberta Forest Service daily reports which included fire location and size. These reports were based on late afternoon or early evening aerial reconnaissance. This allowed Flannigan and Vonder Haar not only to validate their ability to detect the presence of fires but also to apply the Dozier equations and compare their calculated fire areas to the reconnaissance observations. Of the 41% of the fires that were visible to the satellite (unobstructed by smoke and cloud), 80% were detected by their detection criteria. Based on comparison with the reconnaissance information, the AVHRR-based fire size estimates were 70% too large for small fires and 50% too small for large fires.

A number of case studies were then reported from diverse regions (Matson et al. 1987, Matson and Holben 1987, Stephens and Matson 1989, Langaas and Muirhead 1989). While these studies did not focus on details of detection criteria and presumably involved interactive techniques, there was an increasingly sophisticated discussion of the limitations and problems associated with fire detection with the AVHRR. This included discussion of confusion from fixed, non-fire sources (Stephens and Matson, 1989), detection of and obstruction due to smoke plumes (Matson et al. 1987, Matson and Holben 1987), and problems associated with surface reflectivity and emissivity, pixel overlap at non-nadir scan angles and incorporation of recently burnt warm areas into the two-element temperature field of the Dozier model (Langaas and Muirhead, 1989). Lee and Tag (1990) presented an alternative approach to non-interactive fire detection. Essentially they subjectively chose a threshold fire temperature and used the Dozier model to develop a look-up table specifying which combinations of satellite

30 measurements constituted a positive fire detection. Atmospheric corrections were included in the estimation of background temperatures using the method of McClain et al. (1985). They applied their technique to nighttime imagery over the San Francisco area and the Persian Gulf.

Kaufman et al. (1990a, b) applied fire detection with NOAA-9 data to the estimation of emissions from biomass burning. Fires were detected in pixels that met three detection criteria. The first criterion was for the brightness temperature of channel 3 to be elevated above a set threshold indicating a fire to be present. The second criterion specified that channel 3 - channel 4 temperature difference must be at least 10 K and the third criterion used the temperature of channel 4 to eliminate false detections from cool clouds that are highly reflective in the 3.7 μm band. Justice et al. (1996) and Scholes et al. (1996) combined AVHRR fire information in a dynamic model to generate improved tracing-gas and particulate emission estimates for Southern Africa. The approach combined satellite data on fire distribution and timing with fuel load calculated by a simplified ecosystem production model and ground-based measurements of emission ratios (Ward et al. 1996, Shea et al. 1996). Daily fires detected by the AVHRR for the entire burning season were calibrated to provide burned area estimates using Landsat MSS data. The algorithm used for fire detection is described by Justice and Dowty 1994 and is the basis for the IGBP-DIS community consensus algorithm for the AVHRR. This tracing-gas modeling was part of a larger body of fire-related research as part of the IGBP Southern African Fire Atmosphere Research Initiative (SAFARI) (Andreae et al. 1995). Setzer and Pereira (1991a) describe an operational fire monitoring program in Brazil. Using NOAA-11 data in 1989, 96% of detected fires were verified by ground crews and there were no reports of missing fires. Setzer and Pereira (1991b) used a fixed channel 3 temperature threshold to flag potential fires but also required that a smoke plume be visible in channel 1 for there to be a positive detection. In order to get a regional estimate of burned area over a burning season, they used a comparison of NOAA-9/AVHRR and Landsat/TM to determine an average fire size, 0.72 of an AVHRR pixel, and assumed that the average duration of a fire was 1.5 days. Periera et al. (1991) used AVHRR data from five consecutive days and Landsat/TM data from the sixth day to evaluate the accuracy of fire detection and burned area estimates. All AVHRR-detected fires had corresponding TM scars. When assuming that an AVHRR-detected fire burned

31 the entire pixel, the fire size was overestimated on the average by 43%. Cahoon et al. (1991) examined a large fire in China using NOAA-9 data. Using 3.7 μm saturation as their detection criteria, they found good correspondence with observed smoke plumes. In contrast, Brustet et al. (1991) reported on an AVHRR study in semi-arid West Africa. They found that a threshold in the 3.7 μm channel alone was adequate for fire detection in some cases, while for dry vegetation and a hot background surface, a second threshold applied to the 11 μm channel was necessary. Techniques developed for vegetation fires can also be used in an urban setting as shown by Doussett et al. (1993) through the detection of fires following rioting in Los Angeles in April 1992. Robinson (1991), Kennedy (1992), Justice and Malingreau (1994), and Giglio et al. (1998) all provide reviews of fire detection with AVHRR and address a number of sources of uncertainty.

2）AVHRR fire detection techniques

The primary data available consists of the measurements in all five AVHRR channels from 1) the pixel being evaluated, 2) the pixels contained within some spatial window about the central pixel and 3) the pixels in some temporal series that includes the particular scene being evaluated. Secondary data that may be available include surface reflectance (or emissivity), some measure of water vapor in the atmospheric column, aerosol optical depth and information concerning the location of permanent non-biomass burning heat sources, such as oil refineries. Most algorithms utilize a subset of the primary data and none of the secondary data (the GOES ABBA procedure is an exception). In general algorithms utilize one or more of the following: T3 (Bright temperature of channel 3), T4 (Bright temperature of channel 4), T34 (Bright temperature difference of channel 3 and channel 3, T3 - T4) and some characterization of the background in a spatial window about the pixel being evaluated. A ‘fire signal’ can be generally thought of as an enhancement in channel 3 relative to the background. There will be some level of variability, though, in the background brightness temperature as measured in channel 3. For increasingly smaller and cooler fires, the magnitude of the fire signature will decrease until it is indistinguishable from the background variability. At some point there must be a theoretical lower detection limit, applicable to all detection algorithms, that is a function of the amplitude of the background variability. Also, a practical lower detection limit exists, distinct from the theoretical limit, that reflects the ability of a

32 particular algorithm to resolve a fire signature from the background variability. Similarly, algorithms that depend on enhanced T34 for fire detection will have an upper detection limit. Fires larger and hotter than the detection limit will lead to sensor saturation and decreased values of T34. Theoretical simulations indicate that both upper and lower detection limits vary among the fire detection techniques (Dowty 1993, Giglio et al. 1998). Clearly it is important to identify non-fire factors that cause variability in the IR signal as detected by the AVHRR, particularly those that enhance the response in channel 3. Cooper and Asrar (1989) report a 6 K enhancement of channel 3 over channel 4 in the absence of fire. Simulation studies suggest that extreme cases may lead to a 15K enhancement (Dowty 1993). Factors that may be responsible for such effects include the water vapor content of the atmosphere, reflection of solar radiation in the 3.75 μm band, occurrence of sub-resolution land cover anomalies (with associated emissive and reflective properties) and occurrence of sub-resolution clouds. Some factors will act over a relatively large area and potentially can be accounted for through spatial analysis, e.g. atmospheric water vapor. Others may have isolated, sub-resolution occurrences and be more difficult to account for, e.g. isolated, sub-resolution cloud. Some algorithms based only on T34 thresholds produce false detections at cloud edges because there are a number of pixels only partially filled with cloud. The proportion of cloud may be small enough so that the pixel slips through the cloud screen but large enough so that the channel 3 response is significantly enhanced by highly reflective clouds. The channel 3 response in daytime data may be strongly enhanced by surface reflection where the satellite-surface-solar geometry results in Sun glint. This effect has been documented over oceans (Nath et al. 1993, Cracknell 1993) and can lead to false fire detections over land (Alberto Setzer, pers. comm.). Setzer noticed that for a very sharp ocean glint (low wind speed), while the reflectance in channel 1 and 2 continued to increase toward the center of the glint, above a given threshold the channel 3 stopped increasing and decreased slightly. Radiance larger than the saturation threshold in channel 3 causes a decrease in the channel 3 value rather than a constant saturated value. It appears that intense fires above saturation level will have a slightly lower value than intense fires. Pixel overlap at high scan angles is an issue that has important implications for fire detection. At the edge of a scan, over 60% of the area viewed within a pixel is also viewed by adjacent pixels (Cahoon et al. 1992a). This could lead to replicate counting of fire events and decreases the value of extreme

33 off-nadir data. At present, no techniques have been developed to account for pixel overlap effects apart from reducing the portion of the scan that is used. Finally, the effect of temporal sampling is an important issue in fire detection with AVHRR data. Presently there are four potential AVHRR sampling times; one daytime and one nighttime pass for each satellite (NOAA-16 and NOAA-18). Rarely, however, is all this data available and most fire detection studies use only one satellite pass. This pass is then used to represent the burning events for an entire day. It is important to consider the diurnal pattern of burning and the timing of the NOAA satellite overpass in order to understand the limitations of using one temporal sample. In this regard, comparison of AVHRR-derived fires with GOES-derived fires can be very useful (see Section 2). In the following text we review a number of existing AVHRR algorithm approaches for the detection of active fires. These are grouped as fixed threshold techniques, the Lee and Tag technique, and spatial analysis techniques. Although cloud screening procedures are typically integrated with fire detection, the cloud screening steps are not discussed here.

(1) Fixed Threshold Techniques

There are a number of approaches which compare AVHRR data with various fixed thresholds to determine whether a pixel is to be classified as a fire pixel. A threshold value of the channel 3, channel 4 difference (T34) is often used, sometimes in conjunction with a threshold applied to the individual value in channels 3 or 4. The method of Flannigan and Vonder Haar (1986) consisted of the following tests over a forested region:

a) T3 > T3b b) T4 > T4b c) T34 > 8 K (nighttime) T34 > 10 K (daytime)

where T3b and T4b are average values from neighboring pixels. The first two tests require that a fire pixel be hotter than the background and the third test compares T34 to a threshold. A higher threshold was used during the day because of reflective and surface heating effects. This approach includes spatial analysis in the form of T3b and T4b, but is classified here as a fixed threshold technique because of the third, more

34 critical, test. A cloud screening was performed separately based on Coakley and Bretherton (1982).

Kaufman et al. (1990a, b) used three tests to detect deforestation fires in Brazil:

a) T3 >316 K b) T34>10 K c) T4 > 250 K

The first test ensures that the pixel be hot. The second test is the same as that used by Flannigan and Vonder Haar (1986), and the third is intended to screen out clouds that may be reflective in the channel 3 band but are cool in channel 4. A number of studies in Brazil base detection on a channel 3 threshold alone, sometimes with a requirement that a smoke plume be visible in channel 1. Setzer and Pereira (1991b) detected deforestation fires using:

a) T3 > 319 K

b) manual detection of smoke plume in channel 1

Pereira et al (1991) and Pereira and Setzer (1993) also used a channel 3 threshold alone and examined raw digital counts instead of derived brightness temperatures. Pereira et al. (1991) required that channel 3 was 10 (raw digital counts) on an 8-bit (256) scale where the 8 most significant bits of the original data are used. Pereira and Setzer (1993) required that channel 3 was 8 (raw digital counts) on the same scale. These studies used daytime data.

Franca et al. (1993a, b) used the following five tests to detect fires in West African savannas in daytime imagery: a) T3 > 320 K b) T34>15 K c) T4 > 287 K d) 0

35 e) A1 < 9% reflectance

Where A1 is the reflectance in channel 1 and T45 is the difference between channels 4 and 5. The first three tests are similar to those of Kaufman (1990a, b) and the last two tests increase the ability of the algorithm to discriminate cloud, particularly subresolution cloud.

Many of the fixed threshold studies discussed above included some type of validation exercise and each of these reported good results. The fixed threshold approach seems very appropriate and performs very well within specific study areas. It must be kept in mind, though, that each of these fixed threshold algorithms was optimized for a particular region. If an algorithm is applied to a region outside what it was designed for, presumably it would require some modification.

(2) Lee and Tag Technique

The approach described here is that of Lee and Tag (1990), although the implementation is slightly modified here. They essentially compare an adjusted channel 3 brightness temperature to a threshold value that is selected on a pixel-by-pixel basis. This threshold is calculated as a function of a derived background temperature and an adjusted channel 4 brightness temperature by using Dozier's model. This method can be divided into three steps:

Step 1: Background temperatures are estimated and adjusted brightness temperatures are calculated. a) The brightness temperatures from pixels adjacent to the pixel being tested are averaged. This is actually done once with corner pixels and once with side pixels and the final results are then compared for consistency. Each channel is averaged separately resulting in three ‘background’ values: T3b, T4b and T5b. b) The atmospheric correction algorithm of McClain et al. (1985) is used with T3b T4b and T5b to calculate an estimate of background land surface temperature Tb. c) Three correction factors are calculated by subtracting T3b, T4b and T5b from Tb. d) These correction factors are added to the observed brightness temperatures of the

36 pixel to obtain adjusted values for channels 3 and 4, T3* and T4*. The idea here is that the brightness temperature of adjacent pixels with no fire should match the background temperature. Any deviation can be attributed to factors (atmosphere, surface reflectance, etc.) that would affect the central pixel also.

Step 2: The threshold channel 3 brightness temperature is calculated. Dozier's model is used here to relate surface parameters (fire size, fire temperature and background temperature) to AVHRR observations. A threshold fire temperature, Tmin, must be specified which represents the minimum fire temperature of interest. The size of a fire at the threshold temperature is found so that the predicted channel 4 brightness temperature matches the adjusted observed value, T4*. The corresponding brightness temperature predicted for channel 3 is used as a threshold.

Step 3: If the adjusted brightness temperature of channel 3 (T3*), is greater than the threshold then the pixel is flagged as a 'fire pixel'. Spatial Analysis Techniques employ variable thresholds that are derived on a pixel-by-pixel basis from the information contained in some spatial window about each pixel. These are to be distinguished from the Lee and Tag approach because here the thresholds are empirically or statistically determined whereas the Lee and Tag variable thresholds are based on theoretical calculations. The Automated Biomass Burning Algorithm (Prins and Menzel, 1993), applied to GOES VAS data, is an example of the spatial analysis approach (see Section 2). A similar approach was developed at NASA/Goddard Space Flight Center for use with AVHRR 1 km data. An important part of both of these algorithms involves a comparison of the T34 value for each pixel with a threshold that is determined on a pixel-by-pixel basis. The threshold is a function of the standard deviation of T34 in a spatial window about each pixel. This explicitly ties the detection criteria to the variability in the background.In the Goddard algorithm, pixels are flagged as ‘potential’ fire pixels if:

a) T3 >316 K

b) T4 >290 K

c) T3 > T4

37 This eliminates a large number of pixels from further consideration and reduces processing time. The T34 value of each ‘potential’ fire pixel is then compared to a threshold that is determined from the background, or neighboring pixels. This threshold is equal to the mean value of T34 of the background pixels plus twice the standard deviation of T34 of the background pixels. There is an additional stipulation that the threshold must be greater or equal to 3 K, otherwise it defaults to 3 K. If T34 of a potential fire pixel is greater than the resultant threshold, it is classified as a fire pixel. The mean and standard deviation are determined from the background pixels as follows. Background pixels that are themselves potential fire pixels are not included in the calculation of background statistics. The size of the background ‘window’ is allowed to migrate from a 3 by 3 pixel box (the eight surrounding pixels) up to a 21 by 21 pixel box as necessary until at least 25% of the background pixels are available for the calculation of background statistics. There must be at least 3 pixels used in the calculation. If the background statistics cannot be calculated then the central pixel cannot be classified as a fire pixel.

While the potential for AVHRR fire products are significant, there are also serious problems, limitations and unresolved issues associated with AVHRR data and fire detection procedures. A fundamental problem with fire detection is that it is limited to cloud-free areas. This can be a serious problem in areas of high cloudiness. A number of phenomena affect the IR signal received by the AVHRR and hence may confuse or confound fire detection (Giglio et al. 1998). These include surface reflection of solar radiation in the 3.75 μm band and the effects of atmospheric water vapor and sub-resolution cloud. The viewing geometry of the AVHRR gives rise to variable pixel size and pixel overlap problem that complicate the interpretation of detected fires. All these issues need more attention in the context of developing a MODIS operational procedure.

2) Fire detection using MODIS data

The MODIS boarding on both the Terra and Aqua satellites makes possibility of monitoring earth four times everyday. MODIS possess the ability to observe fire, smoke, and burn scar globally. The main sensors’ channels for fire detection are saturated at high brightness temperatures: 500 k for 4 µm channel and 400 k for 11 µm channel. That can only be obtained in rare circumstances at the 1 km spatial

38 resolution for fire detection. Thus, MODIS is different from other polar orbiting satellite sensors with similar thermal and spatial resolutions (such as AVHRR). In recent years, MODIS data has been broadly used for fire detection.

(1) MODIS Instruments

As part of NASA’s Earth Observing System (EOS), the Moderate Resolution Imaging spectrometer (MODIS) is carried on both the Terra and Aqua satellites. The MODIS instruments with 36 spectral channels for a wide array of land, ocean and atmospheric investigations, which began collecting data in February 2000 (Terra) and June 2002 (Aqua), are being used to generate atmospheric, oceanic, and land data products (Kaufman, et al., 1998). MODIS can provide global data every 1-2 days, its repeat cycle being 16-day. The spatial resolution of MODIS (pixel size at nadir) is 250m for channel 1 and 2 (0.6µm - 0.9µm), 500m for channel 3 to 7 (0.4µm - 2.1µm) and 1000m for channel 8 to 36 (0.4µm - 14.4µm), respectively. A detailed overview for the 36 spectral channels of MODIS is given in Table 2.2. Terra spacecraft is for global observations, its equator crossing times around 10:30am and 10:30pm local times, but the Aqua provides afternoon and night observations at 1:30pm and 1:30am. These four MODIS observations are being used to derive the daily fire and smoke distributions among other products.

Table 2.2 The parameters and purpose of MODIS bands

BD. BANDWIDTH WAVELENGT SNR REFLECTANC PURPOSE IFOV H (nm) E/EMISSIVE (m)

1 620～670 646.5 128 Land/Cloud/Aerosol 250

2 841～876 856.7 201 s boundary

3 459 ～479 465.6 243 Land/Cloud/Aerosol 500

4 545 ～565 553.7 228 s properties

5 1230 ～1250 1241.9 74

6 1628 ～1652 1629.1 275

7 2105～ 2135 2114.3 110

8 405～ 420 411.8 880 Ocean colour 1000

9 438～ 448 442.1 8380 /Phytoplankton/Bio

10 483～ 493 486.9 802 geochemistry

11 526～ 536 529.7 754

39 12 546～ 556 546.8 750

13 662～ 672 665.6 910

14 673～ 683 676.7 1087

15 743～ 753 746.4 586

16 862～ 877 866.2 516

17 890～ 920 904.1 167 Atmospheric/Water

18 931～ 941 935.3 57 Vapor

19 915～ 965 936.1 250

20 3.660～ 3.840 1382.0 0.05 0.987 Surface/Cloud

21 3.929～ 3.989 3788.2 2.00 0.985 Temperature

22 3.929～ 3.989 3992.1 0.07 0.985

23 4.020～ 4.080 3971.9 0.07 0.987

24 4.433～ 4.498 4056.7 0.25 0.985 Atmospheric

25 4.482～ 4.549 4473.2 0.25 0.985 Temperature

26 1.360～ 1.390 4545.4 1504 Cirrus Clouds

27 6.535～ 6.895 6765.4 0.25 0.992 Water Vapor

28 7.175～ 7.475 7336.7 0.25 0.997

29 8.400～ 8.700 8528.8 0.05 0.987

30 9.580～ 9.880 9734.4 0.25 0.987 Ozone

31 10.780～11.280 11018.6 0.05 0.992 Surface/Cloud

32 11.770～12.270 12032.5 0.05 0.997 Temperature

33 13.185～3.485 13365.1 0.25 0.997 Cloud Top Altitude

34 13.485～13.785 13683.5 0.25 0.997

35 13.785～14.085 13913.3 0.25 0.997

36 14.085～14.385 14195.7 0.35 0.997

(2) Relative MODIS Bands with Fire Detection

MODIS has an ability to observe fires, smoke, and burn scars globally. Its main fire detection channels saturate at high brightness temperatures: 500k at 4µm and 400k at 11µm (Kaufman, et. al.1998a), which can only be attained in rare circumstances at the 1km fire detection spatial resolution. Thus, it is unlike other polar orbiting satellite sensors with similar thermal and spatial resolutions, but much lower saturation temperatures (such as Advanced Very High Resolution Radiometer and Along Track

40 Scanning Radiometer).

Based on review of fire properties by Lobrt and Warnatz (Lobrt， Warnatz ，1993) flaming temperature can be anywhere between 800 K and 1200 K and as hot as 1800 K. Smoldering should be under 850 K and above 450 K. The actual range is probably smaller. Ward and Kaufman et al. have sensitivity studied in Brazil, measured 2-3 times as much emission of aerosol particles, CH4 and NMHC in the smoldering phase as in the flaming phase (Ward et al., 1992; Kaufman et al., 1992). According to the report of Kaufman (Kaufman, et al., 1998) on fire characteristics by using MODIS data, it shows that the sensitivity of the potential MODIS channels to the fraction of the pixel covered by flames of 1000 K was better than a fraction of the pixel covered by smoldering of 600 K. The rest of the pixel has a temperature of 300 K. At he same time, the shorter the wavelength, the stronger the sensitivity to the higher temperature region. The 1.65 µm channel is very sensitive to fraction of the pixel covered by flames and the flaming energy and not very sensitive to fraction of the pixel covered by smoldering and its energy. The 2.1 µm channel is very sensitive to flaming and somewhat sensitive to smoldering. Since the thermal energy is more concentrated in the flaming fire, the sensitivity to thermal energy is independent of smoldering or flaming. Unfortunately the MODIS 2.1 µm channel saturates at a reflectance of 0.8, which for the low solar brightness in this channel corresponds to less than 1% of the 500 m pixel being in flames. The importance of this channel is therefore limited. The 4µm channel is sensitive to both flaming and smoldering, and is 5 times more sensitive to the thermal energy emitted from flaming than from smoldering. The main results of the simulation are summarized in Table 2.3.

Table 2.3 Information on the MODIS bands that can be used for fire detection

Channels Spectral Wavelength Spatial saturation fraction of pixel that saturates Sensitivity T/f) and T/Ef ) Sensitivity (T/f) ,T/Ef ) (μm) Resolution the channel Of smoldering at 600 K of Flaming at 1000K

(m) 1000K 600 K

6 1.628～1.652 500 1 0.05 no saturation T/f=0.064 T/f=220

-6 -4 (740 K) T/Ef=9×10 T/Ef=5×10 7 2.105～2.155 500 0.8 0.007 0.65 T/f=1.2 T/f=110

-4 -4 (570 K) T/Ef=2×10 T/Ef=3×10

22 3.929～3.989 1000 500K 0.025 0.30 T/f=800 T/Ef=8300

T/Ef=0.11 T/Ef=0.02 (f=0.005) (f=0.05)

31 10.78～11.28 1000 400K 0.07 0.25 T/f=480 T/Ef=1700

T/Ef=0.07 T/Ef=0.004 (f=0.005) (f=0.05)

Energy , Ef (MWatt) Ef/f=7300 Ef/f=430000

2.4.2 Methods

Here, the fire identification methodology includes the method by using NOAA/AVHRR and MODIS. There are difference between the number and characters of MODIS and NOAA/AVHRR bands. So, the fire monitoring method has been listed separately.

2.4.2.1 Fire Identification Method by Using AVHRR Images

NOAA-16, NOAA-17 and NOAA-18 images can be received by SRSIFRIS everyday. Because NOAA-17 hasn’t median thermal bands in daytime. NOAA-16 and NOAA-18 images have been selected to fire identification in our research.

1) Pre-processing

Here, the preprocessing of NOAA/AVHRR includes Geo-reference and physical value calculation.

(1) Geo-reference

There are three kinds of methods that can georeference AVHRR images. One is the registration, which includes image to image and image to map method. The others are by using the ‘Georeference AVHRR’ and ‘Georeference from input Geometry (Built GLT) Menu’. In this work, we just used the latter method to georeference the AVHRR data by using the Tropical Forest Fire Monitoring Sub-system.

(2) Physical Value Calculation

NOAA/AVHRR sensor digital number (DN) is 10 bit. So, the digital value of AVHRR is 0 to 1023. However, the data format is un-integer 16 bit when the receive station keep the data by using a liner translation. The user should translate the data from save format to physical value. The reflectance and the radiance can be calculated by using formula (2.1).

A= a*B - b (2.1)

43 Where: A is Reflectance of visible band and NIR band or Radiance of Thermal band of AVHRR. B is the scaled data. a is slope; b is intercept;

TOA_Ref = Reflectance *1.0/cos (D2R *θs) (2.2)

Where: TOP_Ref is the TOA reflectance

θs is solar zenith angle D2R is degree to radian; the value is 0.0174533.

C BT = 2 (2.3) l C l * ln(1.0 + 1 ) 5 Ll * l

Where: BTl is the bright temperature of l central wavelength;

l is the central wavelength;(unit: µm);

Ll is the radiance of Wavelength l ; (unit: W micron^4 / m^2 sr)

C1, C2 are constants;

Here, C1 = 1.19107E+8 (unit: W micron^4 / m^2 sr)

C2 = 1.43883E+4 ( unit: micron K）

For NOAA/AVHRR thermal bands, the bright temperature correction can use formula (2.4).

Ti = BTi*Ta i +Tc i (2.4)

44 Where: Ti is the i band’s bright temperature after correction;

BTi Is the i band’s bright temperature before correction;

Ta i is the i band’s bright temperature correction scale;

Tc i is the i band’s bright temperature correction offset;

Normalized Difference Vegetation Index (NDVI) will be calculated by using formula (2.5).

rnir - rred NDVI = (2.5) rnir + rred

Where: NDVI is the Normalized Difference Vegetation Index;

rred is the reflectance of red band;

rnir is the reflectance of Near Infrared band;

2) Hotspot Detection Algorithm Description

The fire detection algorithm is based on the original AVHRR detection algorithm with modifications for the NOAA-AVHRR sensor. In addition, new tests are introduced to cope with the special environment and burning conditions of the tropical ecosystem. The algorithm consists of two major steps; marking potential fires and removing false fires. Both steps encompass threshold tests. All the tests are optimized to both detect real fires and eliminate as many false fires as possible. While the majority of tests were proposed previously, the threshold values were chosen following a trial-and-error approach based on the fire training data set. Histogram analyses of reflectance and brightness temperature corresponding to burning and non-burning pixels proved to be an effective means for obtaining optimized threshold values (Kaufman, Justice et al., 1998; Zhanqing Li, 1997; etc.).

(1) Cloud masking

Cloud detection was performed by using a technique based on that technique used in the production of the International Geosphere Biosphere Program (IGBP) AVHRR derived Global Fire Product (Stroppiana, Pinnock, et. al. 2000). The Daytime pixels are considered to be cloud if one of formula (2.6), (2.7) (2.8) conditions is satisfied.

45 Nighttime pixels are classed as cloud if the single condition formula (2.7) is satisfied.

ρ1 + ρ2 > 0.9 (2.6)

T12 < 265 K (2.7)

ìr1 + r 2 > 0.7 í (2.8) îT12 < 285k

(2) Identifying potential fire pixels

Following the pioneering work of Flannigan and Vonder Haar (1986) on automatic detection of fires using AVHRR data, numerous investigations have been conducted. Most have employed a threshold test based on the brightness temperature of AVHRR channel 3. This marks all potential fires not obscured by thick clouds by identifying the hot spots, i.e. pixels with brightness temperature (T3) higher than a certain value. According to the Planck Law, AVHRR channel 3, with a central wavelength around 3.7 m, receives maximum radiative energy from objects emitting at temperatures around 800 K. This temperature is close to the temperature for burning biomass (Kennedy et al. 1994). However, since AVHRR sensor was not designed for fire detection, it loses sensitivity at such high temperatures. For the AVHRR onboard NOAA-16 or NOAA-18, channel 3 becomes saturated at T3≈ 320 K. Nevertheless, it still proves to be the most useful AVHRR channel for fire detection as the brightness temperature for most non-fire pixels is usually significantly lower. In practice, thresholds less than the saturation temperature are often used for two reasons. First, wild fires have a large range of burning temperature ranging from lower than 500 K to higher than 1000 K and they also have a variable fraction of burning area within a pixel. To allow for the detection of all fires, a lower threshold value is needed. Secondly, Setzer and Verstraete (1994) identified an engineering design problem in the on-board processing of channel 3 output signal for NOAA-11. They found that signals greatly exceeding the saturation limit are assigned values below the saturation limit. As a result, targets that are extremely hot can have a brightness temperature lower than 320K.

A brightness temperature of T3=315K was chosen in this study. This threshold captures nearly all real fires in clear or thin-cloud covered regions, as confirmed by

46 the fire training database. On the other hand, it also leads to far too many false fire pixels. These false identifications stem from the limitations of channel 3 such as Sun glint effect, reflective soils, pixel overlap, sensor degradation, etc. Sun glint is the most serious problem encountered in our analysis due to the presence of many lakes and rivers distributed throughout the Canadian Shield. For NOAA-16 that has view geometry near the principle plane, the problem can readily be resolved by avoiding the measurements made in forward scattering directions. For NOAA-18, however, this simple solution is no longer valid since its viewing plane is far away from the principal plane and because Sun glint also occurs from cloud side in the backscattering direction. The subsequent tests are thus introduced to remove the false fire pixels.

(3) Removing false fires

Kaufman et al. (1990) introduced three tests to eliminate false fires. We adopted these tests but tuned the threshold values. The first one uses the difference between channel 3 and channel 4 brightness temperatures (T3-T4) to identify false fire pixels caused by a warm background. Some surface types (e.g. bare soil) can become warm enough to saturate channel 3. Since the spectral window of channel 4 is located in the electromagnetic spectrum that has maximum radiative emission for the ordinary earth temperatures, T3-T4 is instrumental in discriminating these false fire pixels. In the case of biomass burning, channel 3 receives much more radiant energy than channel 4 and thus the value of T3-T4 is high (Kennedy et al. 1994, Dowty 1996). For the boreal forests, the threshold value for T3-T4 was set to 14 K. All pixels with T3-T4 values lower then 14 K are considered false fire pixels caused by a warm background.

The second test employing channel 4 alone deletes false fires caused by highly reflective clouds. Radiance measured by channel 3 originates from both solar reflection and terrestrial emission. Reflection of the solar radiation by clouds can be large enough to also saturate channel 3. Since clouds usually have cold tops and thus low brightness temperatures, this test rejects marked fire pixels of T4 less then 260 K.

The third test is intended to eliminate more general bright-scene objects including both clouds and surface pixels. Note that fire hot spots have relatively low reflectance in channel 2 (R2) due to spewing ash and biomass consumption

47 (Kennedy et al. 1994). Therefore, all fire pixels with R2 > 0.22 are considered as false fires.

An additional threshold test is introduced to eliminate false fire pixels caused by thin cirrus clouds. In some cases, a combination of warm background and thin clouds can saturate channel 3 and negate the second test. Thin cirrus clouds have low T4 and high T3 because of reflection, leading to large values of T3 and T3-T4. The additional test makes use of the difference between the two thermal channels of AVHRR, channels 4 and 5. The difference of (T4-T5) has been used to identify thin cirrus clouds, which is often referred to as the split window technique (Inoue 1987). This test is implemented in combination with a relaxed test using (T3-T4). All hot spots with (T4-T5) ³ 4.1 K and (T3-T4) < 19 K are removed as false fires; these thresholds have been established by using training database.

2.4.2.2 Fire Identification Method by Using MODIS Images

The method includes the process and the hotspot detection by using MODIS data. Here, the processing methods of MODIS L1B just refer to the method of georeference and calculation of physical value.

1) Samples Test

In order to get the hotspot detection model of China, The parameters of fire have been sampled from image of SRSIFRIS for MODIS fire production, which includes bright temperature of band 21(because SRSIFRIS’s fire production hasn’t used band 22 at the beginning) ,31 and 32; Reflectance of band 1, 2 and band 7. At the same time, the relative objects parameters have been retrieved, such as cloud, water, bare soil and vegetation. Some analyses also have been developed. Some analysis results are been listed in table 2.5, Figure 2-1, Figure 2-2, Figure 2-3, Figure 2-4, Figure 2-5, Figure 2-6, Figure 2-7 and Figure 2-8.

Table 2.5 Parameters value of fire samples

R1 R2 R7 NDVI BT21 BT31 BT32 DT 23.89474 46.98303 0.039093 32.57479 304.9441 294.7382 291.966 10

48 20.61228 82.3357 0.040322 59.95593 319.5051 298.1347 294.5347 21.8 34.42226 90.47282 0.037623 44.87812 322.7301 298.9922 294.9123 25.5 30.37094 70.75482 0.044891 39.93432 376.5268 297.8852 293.207 80.1 30.16394 63.70003 0.046868 35.7284 306.6549 295.4956 292.391 12.1 44.24006 52.2569 0.038371 8.307875 344.0715 295.1971 291.4442 50.3 38.56229 69.76105 0.033267 28.80151 300.3799 285.4567 282.4492 16.6 58.78932 93.20815 0.039093 22.64434 300.3799 282.9604 279.434 19.4 69.4647 67.46849 0.042646 -1.4578 301.7879 281.1676 277.0368 22 30.28222 82.37506 0.040135 46.2401 309.1945 298.8801 296.17 10.4 40.57317 54.49043 0.042646 14.63995 307.3331 289.4587 285.9902 18.2 38.44401 38.48185 0.045773 0.049201 308.0908 289.6309 285.5433 20 36.22613 73.16546 0.043208 33.76798 326.1809 291.9408 288.6849 35.5 51.86955 31.54513 0.039841 -24.3655 333.0584 297.5312 295.6985 36.9 42.61361 88.6919 0.038371 35.09243 316.4575 301.4099 297.9679 16.4 25.25504 82.53249 0.027522 53.13922 309.4646 291.7025 288.7619 18.5 16.29482 93.0704 0.030916 70.2011 306.3598 293.4631 290.4803 14.2 44.56534 89.05596 0.039654 33.29605 303.1335 293.0562 289.922 14.5 16.44268 82.92606 0.043662 66.90574 306.6549 294.2435 290.9405 14.2 37.823 56.85186 0.042459 20.09917 314.5183 290.4682 287.4252 23.6 40.60274 83.17205 0.046013 34.39255 301.3258 290.1915 287.1135 15.1 39.53816 63.05064 0.051117 22.91915 310.4358 289.9444 285.6802 22.8 24.42703 22.34537 0.04604 -4.45062 332.9673 298.5197 294.2988 35.8 37.616 30.2365 0.040242 -10.8758 330.4686 299.258 294.6918 32.6 51.57383 43.26376 0.037463 -8.76242 327.3674 298.3602 293.623 30.4 13.10108 83.78208 0.043715 72.95489 317.2454 294.9171 291.7505 25 47.90694 81.21402 0.029366 25.79525 309.6434 296.8534 293.5043 13.1 25.19589 83.44755 0.036501 53.61728 306.8502 296.1382 292.9289 11.9 36.66971 69.6233 0.036501 31.0026 319.1052 297.2947 293.3904 23 33.26896 31.93871 0.037783 -2.04003 311.3787 297.1762 293.2517 17.3 47.25637 43.57862 0.037169 -4.04882 319.239 298.2616 293.8849 22.2 62.9885 62.41108 0.028885 -0.46046 313.1864 286.847 283.6321 28.4 98.35623 93.21799 0.040669 -2.68211 335.7796 292.5985 289.596 44.6 64.49666 78.1048 0.032599 9.542778 305.8615 293.9173 290.8698 14.1 60.50448 86.08448 0.035752 17.45016 306.6549 292.8497 289.5807 14.9 34.65883 76.66825 0.017849 37.73513 315.2775 298.7772 294.7948 17.4 40.04088 56.29102 0.033348 16.86892 314.6715 298.454 294.6918 17.4 34.18568 60.98438 0.039761 28.15874 324.5546 299.8576 295.2108 27.9 55.47729 77.91785 0.023701 16.82262 347.2421 300.177 295.6644 48.5

Figure 2.1 Character of fire samples in different bands

Figure 2.2 Physical parameters of typical Objects in Thermal Bands

Figure 2.3 Physical parameters of typical Objects in Visible Bands

Figure 2.4 Horizontal Profile of fire samples