CART and IDC – Based Classification of Irrigated Agricultural Fields Using Multi-Source Satellite Data

Geocarto International, 2016 http://dx.doi.org/10.1080/10106049.2016.1232312

CART and IDC – based classification of irrigated agricultural fields using multi-source satellite data

Virupakshagouda C. Patila,b, Khalid A. Al-Gaadia,c, Rangaswamy Madugundua , ElKamil Tolaa, Ahmed M. Zeyadaa, Samy Mareya and Chandrashekhar M. Biradard aPrecision Agriculture Research Chair, King Saud University, Riyadh, Saudi Arabia; bElectron Science Research Institute, Edith Cowan University, Joondalup, Australia; cDepartment of Agricultural Engineering, College of Food and Agriculture Sciences, King Saud University, Riyadh, Saudi Arabia; dGeoinformatics Unit, International Center for Agricultural Research in the Dry Areas, Amman, Jordan

ABSTRACT ARTICLE HISTORY To understand water productivity of crops cultivated in the Eastern Province Received 30 August 2015 of Saudi Arabia, this study was conducted to generate a reliable crop type Accepted 7 August 2016 map using a multi-temporal satellite data (ASTER, Landsat-8 and MODIS) and KEYWORDS crop phenology. Classification And Regression Tree (CART) and ISO-DATA Crop phenology; decision Cluster (IDC) classification techniques were utilized for the identification tree; spectral separability; of crops. The Ideal Crop Spectral Curves were generated and utilized for vegetation indices the formulation of CART decision rules. For IDC, the stacked images of the phenology-integrated Normalized Difference Vegetation Index were utilized for the classification. The overall accuracy of the classified maps of CART was 76, 77 and 81% for ASTER, MODIS and Landsat-8, respectively. For IDC, the accuracy was determined at 67, 63 and 60% for ASTER, MODIS and Landsat-8, respectively. The developed decision rules can be efficiently used for mapping of crop types for the same agro-climatic region of the study area.

1. Introduction Crop type mapping is a key factor for the efficient management of land and water resources (Biradar et al. 2009; Heller et al. 2012). Several researchers used crop type maps in various agricultural studies, such as cropping patterns based on crop water needs (Alzahrani et al. 2012), quantification of water use efficiency (Patil et al. 2015), irrigation management (Uddin et al. 2004), decisions on crop rotation (Biradar et al. 2008), nutrient management (Patil et al. 2014), yield forecasting (Ferencz et al. 2004) and economic policies and price optimization (Thornton et al. 1997; Wang et al. 2010). The use of satellite-based data-sets for studying agricultural fields and addressing resource management strategies started in the 1990s (de Leeuw et al. 2010; Liaghat & Balasundram 2010). The remote sensing methods used to identify crop types mainly rely on the spectral signatures of crops (Sakamoto et al. 2005; Wardlow et al. 2007; Vincikova et al. 2010) and their temporal profiles of vegetation indices (Xiao et al. 2005; Biradar & Xiao 2011). Due to the dynamic nature of the agricultural crops, the spectral reflectance of a crop may vary with respect to its phenology. On the other hand, the use of crop phenology-integrated spectral profiles improved the classification accuracy (Blaes et al.

2005; Zafar & Waqar 2014). For example, Pena-Barragan et al. (2011) achieved an overall accuracy of 79% in the classification of crops by incorporating phenology. Hence, the Ideal Crop Spectral Curves (ICSCs), which represent the phenology-integrated multi-spectral and multi-temporal profiles of a specific crop, are essential for the generation of an accurate crop type map. For the incorporation of the phenological changes in crop type mapping, the use of multi-temporal image analysis was found to be superior over single image analysis (Wardlow et al. 2007; Ozdogan 2010; Ozdogan et al. 2010; Foerster et al. 2012). However, during multi-temporal image analysis, the spectral reflectance of forage crops, such as alfalfa and Rhodes grass, can be influenced by the cutting schedule, which needs to be considered in discriminating agricultural crops (Yang et al. 2013). Classification of irrigated crops requires not only the detection of significant spectral differences, but also an algorithm that can successfully identify crops. In general, supervised (for example, the Classification And Regression Tree (CART)) and unsupervised (for example, ISO-DATA Cluster (IDC)) classification methods have been widely used to classify agricultural crops (Ahmad & Sufahani 2012). The CART method works on a sequence of binary decisions formulated in the classification strategy (Safavian & Landgrebe 1991). Depending on the decision rule, the first conditional statement leads to the second, the second to the third and so on (Friedl & Brodley 1997). However, a CART decision tree constructed to classify one data-set (e.g. Landsat-8) may not be able to classify another data-set (e.g. ASTER) due to the variation in the spectral profile or band width, where both data- sets do not cover exactly the same regions of the electromagnetic spectrum. On the other hand, the unsupervised IDC algorithm works via an iterative process through which it re-clusters the pixels to achieve relatively homogeneous groups separable in the spectral space (Ball & Hall 1965; Tou & Gonzalez 1977; Shen et al. 2009). The IDC requires a number of clusters and a number of additional user-supplied parameters as inputs to control the clustering process. Numerous studies used CART and IDC classification methods for the discrimination of land use

and land cover classes (Hansen et al. 2000; Sesnie et al. 2008; Xie et al. 2008; Tooke et al. 2009; Punia et al. 2011). Most of the researchers used vegetation indices (Normalized Difference Vegetation Index (NDVI), EVI and SAVI) and crop phenology as a base for formulating the decision rules for crop type mapping (Sakamoto et al. 2005; Wardlow et al. 2007; Liu et al. 2014). In addition to vegetation indices, individual bands such as Red, NIR and SWIR were also utilized for crop separability and crop classification studies (Sharma et al. 1995; Dadhwal et al. 1996; Manjunath et al. 1998; Panigrahy et al. 2009; Mondal et al. 2014). A reliable crop type map provides vital information on cropping patterns for the efficient management of agricultural inputs and available water resources. In view of the determination of crop water requirements, this study was carried out to generate a reliable crop type map by employing the CART and IDC classification techniques. The specific key objectives of the study were (i) to generate crop-specific ICSCs using a multi-temporal satellite data (ASTER, Landsat-8 and MODIS) and crop phenology, (ii) to classify and generate crop type maps utilizing the obtained ICSCs and (iii) to com- pare the accuracy of the CART and IDC classified maps.

2. Study area The study was carried out in Todhia Arable Farm (TAF), which spread across an area of 6967 ha with 47 agricultural fields (2400 ha) under centre pivot irrigation systems. Each field was about 50 ha. The TAF was located between Al-Kharj and Haradh cities in the Eastern Province of Saudi Arabia, within the latitudes of 24°10′22.77″ and 24°12′37.25″ N and longitudes of 47°56′14.60″ and 48°05′08.56″ E (Figure 1). The study area was under a dry continental climate with hot summers (40 ± 1.7 °C) and cold to moderate winters (15 ± 1.3 °C) with an average annual temperature of 35 °C. Tube wells located in the TAF were used to supply irrigation water to the cultivated fields. The major crops cultivated in the TAF were wheat, alfalfa, Rhodes grass, corn and barley. Geocarto International 3

Figure 1. Location map of Todhia Arable Farm in the eastern region of Saudi Arabia.

3. Data collection 3.1. Field data A reconnaissance survey was conducted to understand the cropping pattern of the TAF and to determine the sampling approach for the development of classification strategies. Wheat and barley were cultivated during the winter season (November–April), while corn was grown twice a year (March–June and July–November). Rhodes grass and alfalfa were cultivated as biennial multi cut crops. In some instances, Rhodes grass was cultivated as a catch crop after the harvest of wheat or barley. Out of the 47 fields of the TAF, 11 fields (23%) were randomly selected and considered as sample plots. For the convenience of the study, the selected 11 fields were earmarked for ground truth data collection based on the area coverage of each crop (one field for wheat, one field for barley, three fields for corn, three fields for alfalfa and three fields for Rhodes grass). The sample plots were visited at frequent intervals (once in 16 ± 2 days) corresponding to the date of satellite over-pass (ASTER/ MODIS or Landsat-8) during the study period (February 2012–May 2014). From each sample plot, four to five homogeneous patches (>3 × 3 pixels) were identified and used to monitor the changes in the spectral reflectance with respect to crop phenology. The geo-location of each homogeneous patch was recorded, along with the field data, which included crop type, phenology and growth stage (Table 1).

3.2. Satellite imagery A total of 43 cloud-free images were acquired for the study, 15 (ASTER), 15 (MODIS) and 13 (Landsat-8). ASTER data were procured from the Japanese Space Centre (http://ims.aster.ersdac. jspacesystems.or.jp), while MODIS (MOD09A1) and Landsat-8 data were downloaded from the por- tal of the USGS Earth Explorer (http://earthexplorer.usgs.gov). The details of satellite data used in this study are provided in Table 2. The acquired images covered the entire growth period of wheat, barley and corn. However, for alfalfa and Rhodes grass, at least a complete growth cycle between two harvests was covered.

4. Methods In order to generate a reliable crop type map, agricultural crops were classified based on the response of phenology-integrated multi-spectral and multi-temporal profiles (i.e. ICSCs) of satellite data. Initially, the obtained ASTER (15 m), MODIS (250 m) and Landsat-8 (30 m) images were pre-processed and utilized for the generation of ICSCs and phenology-integrated NDVI profiles. Thereafter, the resultant data-sets were used for the identification of crops by applying CART and IDC classification techniques. 4 V. C. Patil et al.

Table 1. Cropping pattern – Todhia Arable Farm.

2012 2013 2014 Pivot Code Crop Sowing Harvest Crop Sowing Harvest Crop Sowing Harvest Palace W Jan Mar R June Dec R Jan Apr C Mar/Aug July/Dec – – – – – – TE-10 A Jan Nov W Jan Apr C Feb Apr W Dec Dec C Sep Dec – – – TE-11 W Jan Apr A Jan Dec A Jan Apr R May Nov – – – – – – TE-9 R Feb Nov B Jan Mar C Feb Apr B Dec Dec R Apr Dec – – – TE-8 A Jan Dec C Mar/Aug June/Dec C Feb Apr TE-1 A Jan Dec C Mar/Aug June/Dec C Jan Apr LN-1 A Jan Dec C Mar/Aug June/Dec C Jan Apr LN-2 W Jan Mar C Mar/Aug June/Dec – – – A Apr Dec – – – – – – P2 R Feb Nov A Jan Dec A Jan Apr P4 W Jan Mar – – – – – – R Mar Nov A Jan Dec A Jan Apr P17 R Feb Nov B Jan Mar C Jan Apr B Dec Dec R Apr Dec – – – 3–5 R Feb Nov W Jan Mar A Jan Apr W Dec Dec R Apr Nov – – – P5 R Feb Feb B Jan Feb A Jan Apr C Mar/Aug Jul/Nov C Mar/Aug July/Dec – – – 12C A Jan Dec R Mar Dec R Jan Apr P14 W Jan Mar A Jan Dec – – – R Mar Nov – – – – – – TE-4 A Jan Dec C Mar/Aug Jun/Dec C Feb Apr TE-5 A Jan Dec C Mar/Aug Jun/Dec C Jan Apr Note: W – wheat; C – corn; R – Rhodes grass; B – barley; A – alfalfa.

Table 2. Details of satellite images used in this study.

Sl. No. Sensor Path/row Date of satellite pass 1 ASTER (15 m) 164/11 1 June, 17 June, 3 July, 19 July, 4 Aug, 7 Oct 2012 2 165/10 17 Feb, 4 Mar, 20 Mar, 21 Apr, 12 Sep, 15 Nov, 2012, 12 Feb, 16 Mar and 19 May 2013 3 Landsat-8 (30 m) 165/43 3 June, 19 June, 5 July, 21 July, 22 Aug, 25Oct, 10 Nov 2013, 29 Jan, 14 Feb, 18 Mar, 3 Apr, 19 Apr and 5 May 2014 4 MODIS (MOD09A1) (500 m) 22/6 (Tile no.) 17 Feb, 4 Mar, 20 Mar, 21 Apr,1 June, 17 June, 3 July, 19 July, 4 Aug, 12 Sep, 7 Oct, 15 Nov 2012, 13 Feb, 16 Mar and 19 May 2013

4.1. Pre-processing of satellite images The ASTER (Level-1B) and Landsat-8 (L1T) data used in this study were geo-referenced to the Universal Transverse Mercator (UTM) map projection with World Geodetic System 84 (WGS84) data. For ASTER (Level-1B), each acquired scene was geo-registered with the use of ‘embedded hdf’ metadata (Abrams et al. 2004) utilizing the ‘Georeferenced ASTER tool’ of ENVI (ver. 5.0) software program. Subsequently, the geo-rectified data-sets of ASTER and Landsat-8 (digital numbers) were converted into spectral radiance (Abrams et al. 2004; Chander et al. 2009; Ghulam 2009; Mishra et al. 2014) and transformed into Top-Of-Atmosphere (TOA) reflectance. However, in case of MODIS, atmospherically corrected surface reflectance data (MOD09A1) were used. The obtained TOA reflectance images were used to generate field boundaries, crop-specific ICSCs and temporal profiles of NDVI and GNDVI. Geocarto International 5

4.2. Extraction of field boundaries The Feature Extraction module of ENVI (ver. 5.0) software program was used to extract the field boundaries by employing the Object-Based Image Classification (OBIC) technique (Kettig & Landgrebe 1976; Mason et al. 1988). Since the field boundaries did not change for the entire study period, a single image of ASTER (4 March 2012) was used to extract the field boundaries. The OBIC was achieved in two steps: (i) image segmentation to create image objects, and (ii) classification of the obtained image objects as agricultural or non-agricultural areas. In the segmentation process, the selected ASTER image was divided into segments based on the spectral, spatial and textural characteristics of bands 1, 2 and 3 N. Thereafter, classification was performed to fuse all adjacent objects that were assigned the same land cover category (i.e. agricultural and non-agricultural areas). Subsequently, the classified output of OBIC was converted into ArcGIS shape file using ‘Raster to vector’ tool of ENVI (ver. 5.0) software program. The polygons pertaining to agricultural fields were then labelled with field identification numbers (IDs) for further analysis. Finally, the accuracy assessment of the extracted field boundaries was performed by comparing the location and shape of polygons (i.e. extracted field boundaries) against the actual field’s spatial location determined through a GPS survey.

4.3. Integration of the crop phenology to the multi-temporal satellite data Different crops can be identified on the basis of their phenological and temporal profiles. Vegetation indices, such as NDVI which is a good indicator of vegetation health, were successfully utilized for efficient analyses of spectral and temporal profiles. In this study, NDVI threshold method (Jeong et al. 2011) was adopted for the extraction of crop phenological profile. Subsequently, the phenology-integrated NDVI images were stacked according to their crop calendar (Table 1). The maximum and minimum slopes of NDVI curve, as described in Equation (1) by Jeong et al. (2011), were estab- lished and the corresponding NDVI values (NDVI(t)) were used as thresholds for the determination of crop stages.

t [NDVI ( + 1)−NDVI(t)] NDVIslope (t) = (1) NDVI(t)

In the phenology-integrated NDVI profile, the minimum NDVI value indicated the sowing period (i.e. at the start). At the mid of the phenological profile, NDVI values reached the peak and again decreased at the start of maturity stage. After validating the phenological profile of each crop against the temporal profiles of NDVI, the Phenology Day (PD) was determined for the construction of ICSCs.

4.4. Generation of the ICSCs The ICSCs represent the multi-temporal behaviour of a crop with respect to its phenology and are capable of providing information on crop spectral behaviour at a given growth stage (Liu et al. 2014). Therefore, once the data on crop phenology are available, crop-specific ICSCs can serve as a tool for accurate identification of crop types (Wardlow et al. 2007; Siebert & Ewert 2012). To generate the crop-specific ICSCs, the spectral profiles of the sampled plots (pivot fields) were examined and the corresponding information (i.e. band-wise spectral reflectance) was exported to ASCII files for further analysis. Due to a relatively large size of the fields (~50 ha), 4–5 homogenous patches from each sampled field (i.e. 3 by 3 pure pixels of ASTER and Landsat-8 images) were identified for the study. For MODIS, a single pixel (250 m) corresponding to the geo-location of pure homogeneous patches of ASTER was considered because ASTER and MODIS were onboard of the same satellite platform (NASA EOS Terra). Hence, they were complementary in spatial and temporal resolutions, which offered a unique opportunity for scale-related studies (Liu et al. 2007). 6 V. C. Patil et al.

Figure 2. ASTER band-wise crop reflectance curves.

To deal with coarser resolution of MODIS data, equal-area projection technique was used. Initially, 15-m grids (corresponding to ASTER pixels) were generated for the entire agricultural field and the

purity of MODIS pixels was analysed. When the majority area (>60%) of MODIS pixel matched with the ASTER homogeneous patch, the MODIS pixels were labelled as pure crop pixels and considered for the study. Otherwise, the pixel was labelled as unclassified and eliminated from the study. Sensor- and band-wise spectral reflectance of each sampled crop was assessed across the growing period. The spectral reflectance pertaining to the same crop phenological stage of all sampled fields was averaged out to obtain a crop specific ICSC (Figures 2–4).

4.5. Crop type mapping The sensor-wise (ASTER, MODIS and Landsat-8) generated ICSC images were stacked according to the crop phenology across the growth period. For example, ASTER-generated ICSCs of corn crop for the year 2012 were stacked into two files, one for each season (February–June 2012 and July–November 2012). The ICSC stacked files were used as inputs for CART and IDC algorithms.

4.5.1. Classification and regression tree In this study, CART analysis was accomplished using ERDAS Imagine (Ver. 2010) software program. It involved segmentation (object creation), training data preparation, decision rules preparation (decision tree), classification, post-classification check and pruning/manual editing. A decision tree served as a layered classification procedure to split a data-set into recursively smaller groups on the basis of a set of rules or thresholds. Since the ICSCs were generated based on the crop-wise dynamics of spectral reflectance and crop phenology, they were directly related to the phenological characteristics and physical processes of crops. Hence, the ICSCs related to green, red, NIR and SWIR bands and their ratios (NDVI and GNDVI) were used to formulate the decision rules for the discrimination of agricultural crops (Tables 3–5). Consequently, the same set of classification parameters was consistently applied to multiple seasons/ years. A separate set of decision rules was prepared with 5, 7 and 10 nodes for ASTER, Landsat-8 Geocarto International 7

Figure 3. MODIS band-wise crop reflectance curves. and MODIS, respectively (Figures 5–7). Care was taken to ensure that there was no spectral overlap between the classes when fixing the decision boundaries.

4.5.2. ISO-DATA clustering (IDC) The IDC is an advanced algorithm that can automatically adjust the number of clusters (a group of pixels with similar spectral characteristics) for a given data-set and find the best cluster centres through an iterative approach. It also uses a number of different heuristics to determine whether to merge or split clusters (Shen et al. 2009). The IDC classification was performed on the phenology-integrated NDVI layers derived from the ICSC stacked images using the Spatial Modeller of ENVI (ver. 5.0) software 8 V. C. Patil et al.

Figure 4. LANDSAT-8 band-wise crop reflectance curves.

Table 3. The lower and upper decision boundaries used to separate the crops on ASTER image.

Spectral class Level Crop season Band Lower boundary Upper boundary Harvested fields/crop residues 2 Jul & Nov–Dec Red 32 36 Wheat, barley 2 Jan & Feb NDVI 0.41 0.63 Wheat 3 Feb & Mar Red 54 62 Corn, alfalfa 4 Sep & Oct NIR 56 64 Corn-1 4 Aug–Nov NIR 49 58 Corn-2 4 Mar–May NDVI 0.57 0.68 Alfalfa 5 Mar & Oct GNDVI 0.55 0.75 Rhodes grass-2 5 Apr & July Green 15 18

program. For this study, the model was assigned to classify up to 60 clusters based on 30 iterations and a 0.95 convergence threshold. In the post-classification process, clusters were merged in two cases: (i) when the number of pixels in a cluster was less than the convergence threshold, or (ii) when the distance between the centres of two clusters was less than the threshold distance. However, a cluster was split into two different clusters when the cluster’s standard deviation exceeded a predefined value and the number of pixels was twice the threshold set for the minimum number of threshold members. Geocarto International 9

Table 4. The lower and upper decision boundaries used to separate the crops on Landsat-8 image.

Spectral class Level Crop season Band Lower boundary Upper boundary Harvested fields/crop residues 2 Jul & Nov–Dec SWIR-1 11 16 Corn, alfalfa 3 May & Oct Red 25 30 Alfalfa-1 4 Apr–July SWIR-1 39 50 Alfalfa-2 5 Jan–Feb NDVI 0.57 0.68 Corn, Rhodes grass 6 Mar–May SWIR-2 40 54 Corn-1 6 May–Sep NIR 40 60 Corn-2 6 Nov GNDVI 0.38 0.58 Rhodes grass-1 7 Mar–May SWIR-2 22 33 Rhodes grass-2 7 Nov GNDVI 0.55 0.69

Table 5. The lower and upper decision boundaries used to separate the crops on MODIS image.

Spectral class level Crop Season Band Lower boundary Upper boundary Harvested fields/crop residues 2 Jul & Nov–Dec SWIR-2 25 45 Rhodes grass, corn 3 Sep–Nov Green 17 25 Rhodes grass-1 4 Apr–Jun Blue 11 15 Rhodes grass-2 6 Sep–Nov Green 18 24 Corn-1 5 Sep–Nov NIR-2 42 49 Corn-2 4 Apr–Jun GNDVI 0.38 0.46 Barley 7 Feb NIR-1 45 62 Wheat, alfalfa 8 Jan–Nov Red 11 18 Wheat-1 9 Jan & Feb SWIR-1 30 35 Wheat-2 10 Mar SWIR-2 22 28 Alfalfa-1 9 Mar &Nov NIR-1 42 46 Alfalfa-2 10 Feb GNDVI 0.42 0.57

Figure 5. Decision tree designed for ASTER image. 10 V. C. Patil et al.

Figure 6. Decision tree designed for Landsat-8 image.

Figure 7. Decision tree designed for MODIS image. Geocarto International 11

4.5.3. Accuracy assessment The crop classification map derived from remote sensing was validated using the ground truth data. The classified maps of CART and IDC were subjected to accuracy assessment adopting pixel-based validation process. In case of ASTER and Landsat-8 data-sets, the classified outputs were assessed against 3–4 homogeneous patches (3 × 3 pixels) in each sample plot which were utilized for the construction of ICSCs. For MODIS data, a single (250 m) pixel was used to extract the ICSCs. Out of the 47 fields of TAF, 11 fields were considered as sample plots and the remaining 36 fields were utilized for accuracy assessment. Overall accuracy and Kappa statistics were used for the assessment of the classification accuracy.

5. Results 5.1. Crop growth stage and phenology The collected ground data of agricultural crops (crop type, sowing date, growth stage and harvesting date) were analysed for the determination of the PD of each crop. As presented in Table 6, the PD was tagged to the corresponding NDVI and ICSCs. For wheat, NDVI values exhibited a distinct seasonal pattern (Figure 8). In 2012, wheat crop NDVI upper limits of 0.38–0.60 were recorded for PD 48 (stem elongation) and 80 (head emergence), respectively. In 2013, NDVI upper limits of 0.32–0.52 were observed for the PD 43–75, respectively, i.e. for stem elongation and head emergence stages in 2012. For corn at initial stages (PD 22), the upper level of NDVI values for both of summer and winter corn were 0.14–0.16, respectively. The peak NDVI (0.74 upper limit) was observed at the head emergence

Table 6. Phenology tagged with acquired satellite images.

Crop growth stage Year DOY Sensor Wheat Barely Corn Rhodes grass Alfalfa 2012 48 A, M SE – – GE LD 64 A, M HF – – TI EB & FL 80 A, M MI – SL LD EB & FL 112 A, M RI – TI HF LD 153 A, M – – HE LD LD 169 A, M – – FL LD EB 185 A, M – – GD LD FL & H 201 A, M – – GD HF LD 217 A, M – – – LD EB 256 A, M – – TI LD FL & H 281 A, M – – SE LD EB 320 A, M – – HE HF FL 2013 43 A, M SE SE – LD LD 75 A, M HF HF – LD EB 139 A, M – – – HF EB 154 L8 – – – LD LD 170 L8 – – SL LD EB & FL 186 L8 – – LD HF EB & FL 202 L8 – – SE LD LD 234 L8 – – HE LD EB & FL 298 L8 – – FL HF EB & FL 314 L8 – – GD LD & H EB & FL 2014 29 L8 – – – LD LD 45 L8 – – – HF EB 77 L8 – – SL LD FL 93 L8 – – LD LD LD 109 L8 – – SE HF EB 125 L8 – – HE LD FL Notes: DOY = Day of year; A = ASTER; M = MODIS; L8 = Landsat-8; GE = Germination; SE = Seedling; TI = Tillering; SE = Stem elongation; LD = Leaf development; EB = Early bud stage; HF = Heading and flowering; HE = Head emergence; FL = Flowering; MI = Milking; RI = Ripening; GD = Grain development; H = Harvest/cut. 12 V. C. Patil et al. stage (PD 82) for summer corn crop; however for winter corn, the peak NDVI (0.82 upper limit) was reached at PD 110. Rhodes and alfalfa crops showed less distinct seasonal NDVI behaviour. The growth rate for alfalfa was reduced greatly, especially in winter season where the peak NDVI value was 0.72 at the start of the heading and before the flowering stage (PD 40). Higher growth rate was observed in early summer months as the peak NDVI reached the same value of 0.72 at PD 20. The NDVI values varied throughout the year for Rhodes grass, with the largest variability during leaf development (PD 16–38) and after the first harvest, from PD 12 until initiation of blooming or flowering (PD 42), as illustrated by Figure 8. The average spectral (ICSCs) variability among the crops was 76% for Landsat-8 followed by 72% for MODIS and 63% for ASTER data. There was an overlap among the crops for ASTER sensor as the NIR region, which was the only available portion for vegetation studies, could provide clear differen- tiation between two crops. Hence, concise phenology adjustments assisted in spectral separability of ASTER data, along with NDVI and GNDVI.

5.2. Comparative assessment of IDC and CART in crop type mapping The results of the accuracy assessment of the CART and IDC classified crop type maps are given in Tables 7 and 8. These results indicated that the overall accuracy of identification of crop signatures using ASTER images was of 79.71% (2012) and 72.87% (2013) for CART analysis. For IDC, the accuracy was 66.23% (2012) and 68.09% (2013). Crop-wise accuracy assessment for ASTER data-sets indicated that both methods resulted in similar accuracy levels for alfalfa (77–81% for CART and 78–88% for IDC). However, about 61 and 59% of barley areas were accurately classified by CART and IDC, respectively. The classification accuracy of corn for CART ranged between 53 and 69%. For IDC, the accuracy ranged between 56 and 74%. More than 75% of Rhodes grass cultivated area

was classified accurately by both CART (80%) and IDC (76%) methods for the year 2012. However, in 2013, CART model provided a very low accuracy of 36% compared 75% for IDC. For CART, the classification accuracy of wheat crop varied between 50% (2013) and 73% (2012). While for IDC, it

Figure 8. Crop-wise NDVI response against crop phenology across growing season: C is the cut numbers. Note: For growth stages of each see Table 6. Geocarto International 13

Table 7. Accuracy assessment of CART and IDC classified crop type maps.

CART IDC Classification accuracy Classification accuracy Sensor Year (%) Kappa coefficient (%) Kappa coefficient ASTER 2012 79.71 0.69 66.23 0.59 2013 72.87 0.74 68.09 0.52 Average 76.29 0.72 67.16 0.56 MODIS 2012 83.06 0.70 62.95 0.67 2013 71.51 0.68 62.91 0.64 Average 77.29 0.69 62.93 0.66 Landsat-8 2013 86.40 0.71 59.67 0.64 2014 74.93 0.75 60.95 0.61 Average 80.67 0.73 60.31 0.63 Overall mean 78.08 0.71 63.47 0.61

ranged between 30% (2013) and 60% (2012). Harvested/non-cropped area was classified with accuracies of 63% (2012) to 75% (2013) for CART and 56% (2012) to 64% (2013) for IDC. Overall, the results indicated that the best accuracies of crop signature identification using ASTER data-sets for alfalfa, barley, corn, Rhodes grass and wheat crops were 88% (IDC), 61% (CART), 74% (IDC), 80% (CART) and 73% (CART), respectively. Accuracy assessment results of MODIS derived crop type maps with both CART and IDC methods are presented in Table 8. Overall accuracies of 83% (2012) and 72% (2013) were achieved with CART. While for IDC, a classification accuracy of about 63% was achieved for both 2012 and 2013. Crop- wise accuracy assessment results using MODIS indicated that the highest accuracy was achieved for Rhodes grass (81%; CART), followed by alfalfa (78%; IDC), corn (76%; CART), barley (61%; CART) and wheat (73%; CART). Generally, MODIS-based classification resulted in a higher accuracy com-

pared to ASTER. As presented in Table 8, Landsat-8 classified crop type maps were achieved with overall accuracies of 75–86% and 60–61% for CART and IDC methods, respectively. During the period of the study when Landsat-8 data were collected, wheat and barley crops were not cultivated. Hence, these crops were not included in the preparation of decision rules. Crop-wise accuracy assessment results using Landsat-8 revealed that Rhodes grass was identified with an accuracy of 78–82% (CART) and 58–69% (IDC). For alfalfa crop, the classification accuracy was 69–72% (CART) and 62–67% (IDC). However, for corn crop, the results showed a classification accuracy of 70–73% (CART) and 52–55% (IDC).

6. Discussion The crop types investigated in the study were divided into three categories, each category was treated individually by the classification rules. The first category included wheat, barley and corn (i.e. winter crops). While, the second category comprised summer corn, where the NDVI variations were caused by phenological phase transitions, such as sowing and harvesting in the mid of summer (i.e. July/ August–November/December). These two categories were considered to have a ‘normal’ NDVI profiles within the growing season. Other crop types that did not have normal NDVI profiles were character- ized as the third category, which included alfalfa and Rhodes grass. Due to the flexible planting dates and multiple short growing seasons (~30–45 days) of these crops, they cannot be properly addressed by the sigmoid nature of the NDVI profile. As illustrated in the sensor-wise crop type maps provided in Figures 9–11, the CART and IDC classification techniques performed well on phenology-integrated ICSC profiles. However, there was a great variability in the overall accuracy of the classified images, where the overall accuracy of CART was 80, 83 and 86% for ASTER, MODIS and Landsat-8 images, respectively. These results are in good agreement with the overall accuracy of CART of 92% reported by Biswal et al. (2013). For IDC, the 14 V. C. Patil et al. – – – – – – – – – 5.8 62.2 17.2 11.4 52.3 20.6 69.2 23.7 64.3 60.9 60.8 63.2 2014 IDC – – – – – – 7.1 7.2 67.2 10.3 15.5 54.7 15.4 46.5 23.2 57.6 22.1 56.1 59.7 61.8 66.4 2013 Landsat-8 – – – – – 5.2 5.8 72.0 15.5 10.3 69.7 23.3 16.7 11.1 78.2 17.8 19.2 78.2 74.9 74.9 77.9 2014 CART – – – – – – – – – – 9.2 7.9 68.6 11.4 22.8 73.2 82.1 63.7 86.4 78.7 73.7 2013 – – 9.4 6.52 69.7 18.8 16.6 59.2 32.9 62.8 19.3 26.9 57.6 20.5 17.9 30.2 38.6 64.4 62.9 63.4 66.4 2013 IDC – – – – 7.8 6.6 77.6 20.8 64.6 17.3 32.6 23.3 66.5 12.5 16.8 60.1 14.7 11.6 62.9 64.9 69.6 2012 MODIS – 8.4 70.7 15.9 11.9 61.2 18.8 75.6 22.6 19.8 25.0 76.9 18.2 24.2 50.0 37.5 25.6 44.7 71.5 69.7 70.6 2013

CART – – – – 68.9 11.9 18.2 69.2 32.5 27.9 25.0 80.9 11.7 10.6 10.3 72.7 23.1 62.5 83.1 76.5 72.7 2012 – 7.7 7.2 77.5 15.5 22.2 22.8 59.2 32.9 55.8 15.4 37.2 75.8 23.1 16.7 30.2 38.6 64.4 68.1 60.1 54.0 2013 IDC – – – – 5.2 6.2 87.8 26.2 74.4 19.2 18.6 41.7 75.2 17.8 11.5 60.0 14.7 55.9 66.2 62.6 61.3 2012 ASTER 6.9 8.35 81.3 14.6 12.5 61.3 18.6 68.8 22.7 18.8 25.0 36.1 12.5 24.2 50.1 37.5 12.7 75.1 72.9 73.4 76.9 2013 CART – – – 9.8 77.3 12.4 13.6 53.3 32.5 32.9 25.0 79.8 15.4 10.6 10.3 72.7 31.6 62.5 79.7 74.4 71.7 2012 verall) verall) O hodes grass ccuracy assessment of crop type identification (%). typeccuracy identification of crop assessment hodes G RA SS hodes grass verall) O fa/ R accuracy ( verall classification classification verall orn orn as alfalfa orn as no-crop orn as R hodes grass hodes grass hodes grass as alfalfa hodes grass hodes grass as corn as corn hodes grass hodes grass as no-crophodes grass Table 8. A Table Classes Crop A lfalfa lfalfa as corn A lfalfa as corn A lfalfa as no-crop A lfalfa as R Barley Barley as wheat Barley as wheat C C C C R R R R Wheat Wheat Wheat mixed with barley mixed Wheat Wheat mixed with alfal - mixed Wheat N o-crop/harvested O Producers accuracy Producers User accuracy ( User Geocarto International 15

Figure 9. Crop type map developed using ASTER data: (A),(B) and (C) are the CART classified maps; (D), (E) and (F) are Iso-data cluster analysed maps: (A), (B) corresponding to the season-1 and season-2 of cropped area of the year 2012; (C) is corresponding to the season-1 of year 2013.

attained overall accuracy was 68, 63 and 61% for ASTER, MODIS and Landsat-8 images, respectively. The better performance of CART approach over IDC can be attributed to the intuitive classification structure and its ability to handle noisy and missing data. In addition, CART has the ability to work without prior assumptions regarding the distribution of input data. Compared to previous studies, the overall accuracy of crop signature identification obtained in this study (80–86% for CART and 61–68% for IDC) was found to be within the acceptable range. For example, a study that utilized MODIS-NDVI imagery for mapping crop type resulted in an overall accuracy of 87% (Peterson et al. 2011). Another study by Sencan (2004) reported an accuracy of the CART model ranging between 66 and 91%. MODIS classified images showed, in this study, a higher accuracy compared to ASTER images. This was attributed to the use of SWIR, NDVI and GNDVI data at a specified phenology triggering period. For example, MODIS 8-day composite data of NDVI and GNDVI enhanced the accuracy in spectral separation. Given the 16-day revisit cycle, a higher temporal frequency could not be obtained with Landsat-8 and ASTER, which influenced their ability to properly capture phenological shifts, such as variations in crop type, condition and vigour. As depicted in Figure 8, the results of this study 16 V. C. Patil et al.

Figure 10. Crop type map developed using MODIS data: (A), (B) and (C) are the CART classified maps; (D), (E) and (F) are Iso-data cluster analysed maps: (A), (B) corresponding to the season-1 and season-2 of cropped area of the year 2012; (C) is corresponding to the season-1 of year 2013.

indicated that the accuracy of MODIS-NDVI data in crop type mapping was consistent with the results obtained by Shao et al. (2010) with user’s accuracies of 87, 82, 81 and 70%, for corn, soybean, wheat and alfalfa, respectively. The low accuracy attained with the IDC compared to CART was attributed to the fact that the computer algorithm used for the classification process in IDC identified groups of pixels based on statistical results without taking crop signatures into consideration. Moreover, the image classification is often a per-pixel operation, which is based only on the spectral component of the data; hence, the spatial component is ignored in IDC. For this reason, the mismatch in spatial distribution between the resultant spectral classes and the actual crop types were substantial leading to lower map accuracy with IDC compared to CART.

7. Conclusion To determine crop water needs, a satellite-based reliable crop type map was successfully generated by the integration of crop phenology data with the multi-temporal data-sets (i.e. ICSCs) of ASTER, Geocarto International 17

Figure 11. Crop type map developed using Landsat-8 data: (A) and (B) are the CART classified maps; C( ) and (D) are Iso-data cluster analysed maps: (A) and (C) are corresponding to the season-2 of year 2013 and (B) and (D) are of season-1 of year 2014.

Landsat-8 and MODIS satellite images. The study compared the ability of CART and IDC methods in identifying the crops cultivated in the study area and found that the CART model performed better with an overall accuracy of 80–86%, compared to IDC, which provided an overall accuracy ranging between 61 and 68%. The incorporation of SWIR band in the formulation of decision rules enhanced the crop separability for Landsat-8 data compared to ASTER, where the SWIR bands did not work properly. The ICSCs developed in this study can be used to map crop types for any given year without the need for additional ground truth data. However, the agricultural fields to be mapped have to be within the same agro-climatic region of the study area.

Acknowledgements The assistance provided by the graduate students, namely, Eng. Mohammed Elsiddig Ali Abass and Eng. Ahmed Galal Kayad in the field was quite valuable. The unstinted cooperation and support extended by Mr. Jack King and Mr. Alan King is gratefully acknowledged. Authors also thank Dr. R. Houborg, Water Desalination and Reuse Center, King Abdullah University of Science and Technology, Saudi Arabia, for the important suggestions in the manuscript preparation.

Disclosure statement No potential conflict of interest was reported by the authors.

Funding This project was financially supported by King Saud University, Vice Deanship of Research Chairs.

ORCiD Rangaswamy Madugundu http://orcid.org/0000-0002-5326-4785 18 V. C. Patil et al.

References Abrams M, Hook S, Ramachandran B. 2004. ASTER User handbook, Version 2 Jet Propulsion Laboratory, California Institute of Technology. Available from: http://asterwebjplnasagov/content/03_data/04_Documents/aster_user_ guide_v2pdf Ahmad A, Sufahani SF. 2012. Analysis of Landsat-5 TM data of Malaysian covers using isodata clustering technique. In: Proceedings of IEEE Asia-Pacific Conference on Applied Electromagnetism (APACE 2012); p. 92–97, Melaka, Malaysia, December 2012. http://eprints.utem.edu.my/6472/1/apace2012_asmala_published.pdf. Alzahrani KH, Muneer SE, Taha AS, Baig MB. 2012. Appropriate cropping pattern as an approach to enhancing irrigation water efficiency in the Kingdom of Saudi Arabia. J Anim Plant Sci. 22:224–232. Ball GH, Hall DJ. 1965. ISODATA, a novel method of data analysis and pattern classification; Technical report NTIS AD 699616. Stanford, CA: Stanford Research Institute. Biradar CM, Gowda PH, Hernandez JE, Howell TA, Marek TH, Xiao X. 2009. Irrigated area mapping in the Northern high plains of Texas using Landsat thematic mapper data. In: AGU Fall Meeting; 2009 December 14–18; San Fransisco, CA. Biradar CM, Thenkabail PS, Platonov A, Xiao X, Geerken R, Noojipady P, Turral H, Vithanage J. 2008. Water productivity mapping methods using remote sensing. J Appl Remote Sens. 2:023522–023544. Biradar CM, Xiao X. 2011. Quantifying the area and spatial distribution of double- and triple-cropping croplands in India with multi-temporal MODIS imagery in 2005. Int J Remote Sens. 32:367–386. Biswal S, Ghosh A, Sharma R, Joshi PK. 2013. Satellite data classification using open source support. J Indian Soc Remote Sens. 41:523–530. Blaes X, Vanhalle L, Defourny P. 2005. Efficiency of crop identification based on optical and SAR image time series. Remote Sens Environ. 96:352–365. Chander G, Markham BL, Helder DL. 2009. Summary of current radiometric calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote Sens Environ. 113:893–903. doi:http://dx.doi.org/10.1016/jrse200901007. Dadhwal VK, Parihar JS, Medhavy TT, Ruhal DS, Jarwal SD, Khera AP. 1996. Comparative performance of thematic mapper middle-infrared bands in crop discrimination. Int J Remote Sens. 17:1727–1734. De Leeuw J, Georgiadou Y, Kerle N, Gier A, Inoue Y, Ferwerda J, Smies M, Narantuya D. 2010. The function of remote sensing in support of environmental policy. Remote Sens. 2:1731–1750. doi:http://dx.doi.org/10.3390/rs2071731 Ferencz C, Bognar P, Lichtenberger J. 2004. Crop yield estimation by satellite remote sensing. Int J Remote Sens. 25:4113–4149. doi:http://dx.doi.org/10.1080/01431160410001698870 Foerster S, Kaden K, Foerster M, Itzerott S. 2012. Crop type mapping using spectral – temporal profiles and phenological information. Comput Electron Agric. 89:30–40. Friedl MA, Brodley CE. 1997. Decision tree classification of land cover from remotely sensed data. Remote Sens Environ. 61:399–409. Ghulam A. 2009. How to calculate reflectance and temperature using ASTER data. Center for Environmental Sciences, Saint Louis University. Available from: http://wwwgissluedu/RS/ASTER_Reflectance_Temperature_Calculationphp Hansen MC, DeFries RS, Townshend JRG, Sohlberg R. 2000. Global land cover classification at 1 km spatial resolution using a classification tree approach. Int J Remote Sens. 21:1331–1364. doi:http://dx.doi.org/10.1080/014311600210209 Heller E, Rhemtulla JM, Lele S, Kalacska M, Badiger S, Sengupta R, Ramankutty N. 2012. Mapping crop types, irrigated areas, and cropping intensities in heterogeneous landscapes of southern India using multi-temporal medium- resolution imagery: implications for assessing water use in agriculture. Photogramm Eng Remote Sens. 78:815–827. Jeong SJ, Ho CH, Gim HJ, Brown ME. 2011. Phenology shifts at starts vs end of growing season in temperate vegetation over the Northern Hemisphere for the period 1982–2008. Global Change Biol. 17:2385–2399. doi:http://dx.doi. org/10.1111/j1365-2486201102397x Kettig RL, Landgrebe DA. 1976. Classification of multispectral data by extraction and classification of homogeneous objects. IEEE Trans Geosci Electron. 14:19–26. Liaghat S, Balasundram SK. 2010. A review: the role of remote sensing in precision agriculture. Am J Agric Biol Sci. 5:50–55. Liu MW, Ozdogan M, Zhu X. 2014. Crop type classification by simultaneous use of satellite images of different resolutions. IEEE Trans Geosci Remote Sens. 52:3637–3649. doi:http://dx.doi.org/10.1109/TGRS20132274431 Liu Y, Yamaguchi Y, Ke C. 2007. Reducing the discrepancy between ASTER and MODIS land surface temperature products. Sensors (Basel). 7:3043–3057. Manjunath KR, Kundu N, Panigrahy S. 1998. Evaluation of spectral bands and spatial resolution of LISS II and LISS III sensors onboard IRS satellites for crop identification. J Indian Soc Remote Sens. 26:197–208. Mason DC, Corr DG, Cross A, Hogg DC, Lawrence DH, Petrou M, Tailor AM. 1988. The use of digital map data in the segmentation and classification of remotely-sensed images. Int J Geog Inf Syst. 2:195–215. Mishra N, Haque MO, Leigh L, Aaron D, Helder D, Markham. B. 2014. Radiometric cross calibration of Landsat 8 operational land imager (OLI) and Landsat 7 enhanced thematic mapper plus (ETM+). Remote Sens. 6:12619–12638. doi:http://dx.doi.org/10.3390/rs61212619. Mondal S, Jeganathan C, Sinha NK, Rajan H, Roy T, Kumar P. 2014. Extracting seasonal cropping patterns using multi- temporal vegetation indices from IRS LISS-III data in Muzaffarpur District of Bihar, India. Egypt J Remote Sens Space Sci. 17:123–134. Geocarto International 19

Ozdogan M. 2010. The spatial distribution of crop types from MODIS data: temporal unmixing using independent component analysis. Remote Sens Environ. 114:1190–1204. doi:http://dx.doi.org/10.1016/jrse201001006. Ozdogan M, Yang Y, Allez G, Cervantes C. 2010. Remote sensing of irrigated agriculture: opportunities and challenges. Remote Sens. 2:2274–2304. doi:http://dx.doi.org/10.3390/rs2092274. Panigrahy RK, Ray SS, Panigrahy S. 2009. Study on the utility of IRS-P6 AWiFS SWIR band for crop discrimination and classification. J Indian Soc Remote Sens. 37:325–333. Patil VC, Al-Gaadi KA, Madugundu R, Tola E, Marey S, Aldosari A, Biradar CM, Gowda PH. 2015. Assessing agricultural water productivity in desert farming system of Saudi Arabia. IEEE J Sel Top Appl Earth Obs Remote Sens. 8:284–297. Patil VC, Al-Gaadi KA, Madugundu R, Tola E, Marey S, Al-Omran AM, Khosla R, Upadhyaya SK, Mulla DJ, Al-Dosari A. 2014. Delineation of management zones and response of spring wheat (Triticum aestivum) to irrigation and nutrient levels in Saudi Arabia. Int J Agric Biol. 16:104–110. Pena-Barragan JM, Moffatt KN, Richard EP, Johan S.2011 . Object-based crop identification using multiple vegetation indices, textural features and crop phenology. Remote Sens Environ. 115:1301–1316. doi:http://dx.doi.org/10.1016/ jrse201101009. Peterson D, Whistler J, Egbert S. 2011. Mapping irrigated lands by crop type in Kansas Pecora 18 – Forty Years of Earth Observation. In: the Conference of Understanding a Changing World; 2011 November 14–17; Herndon, Virginia. Punia M, Joshi PK, Porwal MC. 2011. Decision tree classification of land use land cover for Delhi, India using IRS-P6 AWiFS data. Expert Syst Appl. 38:5577–5583. Safavian SR, Landgrebe D. 1991. A survey of decision tree classifier methodology. IEEE Trans Syst Man Cybern. 21:660– 674. doi:http://dx.doi.org/10.1109/2197458. Sakamoto T, Yokozawa M, Toritani H, Shibayama M, Ishitsuka N, Ohno H. 2005. A crop phenology detection method using time-series MODIS data. Remote Sens Environ. 96:366–374. doi:http://dx.doi.org/10.1016/jrse200503008. Sencan S. 2004. Decision tree classification of multi-temporal images for field-based crop mapping [thesis]. Turkey: The Graduate School of Natural and Applied Sciences, Middle East Technical University; p. 139. Sesnie SE, Gessler PE, Finegan B, Thessler S. 2008. Integrating Landsat TM and SRTM – DEM derived variables with decision trees for habitat classification and change detection in complex neo-tropical environments. Remote Sens Environ. 112:2145–2159. Shao Y, Lunetta RS, Ediriwickrema J, Iiames JS. 2010. Mapping cropland and major crop types across the great lakes basin using MODIS-NDVI data. Photogramm Eng Remote Sens. 76:73–84.

Sharma SA, Bhatt HP, Ajai. 1995. Oilseed crop discrimination: selection of optimum bands and role of middle infrared. ISPRS J Photogramm Remote Sens. 50:25–30. Shen Y, Li J, Mantena V, Jakkula S. 2009. Seasonal adaptation of vegetation color in satellite images for flight simulations. J Intell Learn Sys Appl. 1:42–51. Siebert S, Ewert F. 2012. Spatio-temporal patterns of phenological development in Germany in relation to temperature and day length. Agric For Meteorol. 152:44–57. Thornton PK, Bowen WT, Ravelo AC, Wilkens PW, Farmer G, Brock J, Brink JE. 1997. Estimating millet production for famine early warning: an application of crop simulation modeling using satellite and ground-based data in Burkina Faso. Agric For Meteorol. 83:95–112. doi:http://dx.doi.org/10.1016/S0168-1923(96)02348-9. Tooke TR, Coops NC, Goodwin NR, Voogt JA. 2009. Extracting urban vegetation characteristics using spectral mixture analysis and decision tree classifications. Remote Sens Environ. 113:398–407. doi:http://dx.doi.org/10.1016/ jrse200810005. Tou JT, Gonzalez RC. 1977. Pattern recognition principles. Reading (MA): Addison-Wesley; p. 377. Uddin S, Al-Rumikhani YA, Latif MS. 2004. Use of remote sensing and agro meteorology for Irrigation management in arid lands: a case study from northwestern Saudi Arabia. J Environ Hydrol. 12:paper-9. 1–14. http://hydroweb. com/jeh/jeh2004/din.pdf Vincikova H, Hais M, Brom J, Prochazka J, Pecharova E. 2010. Use of remote sensing methods in studying agricultural landscapes – a review. J Landscape Stud. 3:53–63. Wang B, Kocaoglu DF, Damin TU, Yang J. 2010. A decision model for energy resource selection in China. Energy Policy. 38:7130–7141. doi:http://dx.doi.org/10.1016/jenpol201007031. Wardlow BD, Egbert SL, Kastens JH. 2007. Analysis of time-series MODIS 250 m vegetation index data for crop classification in the US central great plains. Remote Sens Environ. 108:290–310. Xiao X, Zhang Q, Hollinger D, Aber J, Moore-III B. 2005. Modeling gross primary production of an evergreen needleleaf forest using MODIS and climate data. Ecol Appl. 15:954–969. doi:http://dx.doi.org/10.1890/04-0470. Xie Y, Sha Z, Yu M. 2008. Remote sensing imagery in vegetation mapping: a review. J Plant Ecol. 1:9–23. doi:http:// dx.doi.org/10.1093/jpe/rtm005. Yang C, Everitt JH, Du Q, Luo B, Chanussot J. 2013. Using high-resolution airborne and satellite imagery to assess crop growth and yield variability for precision agriculture. Proc IEEE. 101:582–592. doi:http://dx.doi.org/10.1109/ JPROC20122196249. Zafar S, Waqar MM. 2014. Crop type mapping by integrating satellite data and crop calendar over Okara District, Punjab (Pakistan). J Space Technol. 4:21–25.