Rabi Groundnut Area Estimation using Synthetic Aperture Radar (SAR) in Thiruvannamalai District of Tamil Nadu

Groundnut (Arachis hypogaea L.) is predominant oilseed crop grown in tropical and subtropical climates around the world. It is grown in different rainfall and temperature regimes on a variety of soils. India occupied the second position in acreage and production of groundnut with an area of 3.35 million hectares and a production of 9.11 million tonnes with an average productivity of 2720 2720 kg ha^-1 during 2018-19 (Directorate of Economics and Statistics, 2020).
       
In Tamil Nadu, it is the major oilseed crop grown under rainfed and irrigated condition accounting for 5.7% of the total cropped area. Groundnut is grown on an area of 3.35 lakh hectares with a production of 9.11 lakh tonnes during 2018-19 in the state (Directorate of Economics and Statistics, 2020). The districts viz., Tiruvannamalai, Vellore, Villupuram, Namakkal, Erode and Salem constituted 54.9% of the area under groundnut in Tamil Nadu.
       
In general, crop cultivated area is estimated using agricultural statistics obtained from field visits and interview with the farmers, which is extremely tedious, time-consuming, labour intensive, less precise, expensive, inconsistent and too generalized (Prasad et al., 2006). It is, however unable to provide timely information on regional spatial distribution of crop growing areas. One of the most common uses of remote sensing in agriculture is crop acreage and production estimation (CAPE). The prime objective of CAPE is the estimation of crop acreage and production of major crops to assist advanced planning for food production, distribution and supply. The remote sensing approach has the ability to provide information on agricultural crops quantitatively, instantaneously and above all, non-destructively over broad areas. Crop phenological information data can also derived through remote sensing (Karnieli, 2003 and Xin et al., 2002).
       
In crop production, crop acreage is the most important component. Crop acreage must be monitored and estimated on a national scale in order to evaluate the national, regional food demand and supply balance, as well as to assess social security. Veloso et al., (2017) reported that crop mapping and estimation of biophysical parameter using SAR backscatter and NDVI temporal behaviors would help better understanding of agricultural activities and environmental conditions.
       
Crop identification, discrimination and early estimation are essential for planning delivery estimates, harvest and storage needs. However, operations that rely entirely on optical images for productivity and yield monitoring are subjected to data gaps during critical crop growth stages due to clouds and haze have little effect on SAR. SAR data also give complementary and unique characterizations of vegetation due to its sensitivity to canopy geometry as compared to the information provided by optical image, in addition to this of tested advantage (Jia et al., 2012). According to Haldar et al., (2017) due to volume scattering that as the plant develops the RADAR backscatter increases as plant grows. Cotton and groundnut crops are detected and discriminated from the other major land covers using a decision rule-based model and the use of multi temporal SAR data appears to be promising, crop backscattering is dynamic during the growth stage. During cloudy and rainy conditions, image collection is limited due to optical sensor relies on solar light.  As a result the remote sensing system that can work in all weather conditions.
       
Radar has a great potential to provide information in such a manner. Previous studies have shown that SAR images have a lot of potential of crop acreage and monitoring (McNairn et al., 2013). Furthermore, the accuracy of crop classification is not always as good as required for better decision making given the current state of the art of SAR imagery interpretation methodologies. To improve crop identification and area estimation, there is need to have a better understanding of the crop and the underlying soil characteristics that influence radar backscatter throughout the growing season, to find a suiTable methodology to extracting crop information from SAR imagery and to evaluate the multi temporal SAR data for crop identification.

Study area
 
Thiruvannamalai district comes under North Eastern Agro Climatic Zones of Tamil Nadu with the spread over of 6191 square kilometers. It lies between 11° 55' to 13° 15' North latitude and 78° 20' to 79° 50' East longitude. This district has eighteen blocks consisting of 1067 revenue villages and study area map were presented in Fig 1. The climate is hot sub humid to semi-arid. The mean annual maximum and minimum temperature are 32.6 and 22.2°C, respectively. The annual rainfall of the North Eastern Zone excluding hills varied from 800-1400 mm. Predominant soil type was red sandy loam and clay loam soil, which encourages the cultivation of groundnut, rice, blackgram and sessame. The experiment was conducted at Department of Remote Sensing and GIS, Tamil Nadu Agriculture University, Coimbatore, Tamil Nadu during rabi season 2020-21.
 
 

 
Data used
 
Synthetic aperture radar (SAR) has the advantage of operating at wavelengths not impeded by cloud cover or lack of illumination and can acquire data over a site during day or night time. Sentinel 1A, with its C-SAR instrument can offer reliable, repeated wide area monitoring. The Sentinel 1 mission includes C-band imaging system operating in four imaging modes with different resolutions and area coverage. It provides dual polarization capability, short revisit period and rapid product delivery. Ground range detected (GRD) and single look complex (SLC) datasets from Sentinel 1A Synthetic Aperture Radar imaging system with VV and VH polarizations in Interferometric Wide (IW) swath mode are obtained at 12 days interval and used for crop identification and mapping in the study area. Based on the normal growing season of groundnut in the study area, the data were downloaded from https://scihub.copernicus.eu/ dhus/ for the period of 4^th October 2020 to 8^th January 2021 to have the full coverage during the crop growth period in the study area.
 
Ground truth collection
 
Ground truth points were collected during Rabi 2020-21 in the groundnut growing areas of the study area at different crop growth stages and the crop calendar were presented in Fig 3, 4. Totally 103 groundnut points and 34 other crop points were collected during the ground survey for training and validation using random stratified sampling method.
 

 
SAR data processing
 
A fully automated processing chain developed by Holecz et al., (2013) was used to convert the multi-temporal space-borne Sentinel 1A SAR IW-GRD data into terrain-geocoded σ° values. The processing chain was a module within the MAPscape-RICE software, developed by sarmap, Switzerland. The SAR time-series data underwent a series of basic processing steps to generate terrain-geocoded σ° values as detailed below and presented in Fig 2.
 

 
(i) Strip mosaicking
 
To facilitate the overall data processing and data handling.
 
ii) Co-registration
 
Images acquired with the same observation geometry were co registered in slant range geometry.
 
iii) Time-series speckle filtering
 
To balance differences in reflectivity between images.
 
iv) Terrain geocoding
 
Radiometric calibration and normalization.
 
v) ANLD filtering
 
To get smoothened homogeneous targets.
 
vi) Removal of atmospheric attenuation
 
σ° values were corrected by means of an interpolator.
 
vii) Subsetting
 
The rectangular extent of the study was extracted from the base map and the raster images were subsetted to an extent from 11° 55' to 13° 15' North latitude and 78° 20' to 79° 50' East longitude. Subsetting of raster data reduces the time in further processing.
 
Multi-temporal feature extraction
 
Multi-temporal features viz., minimum, maximum, mean, minimum date, maximum date and span ratio of VV, VH polarizations and minimum, maximum and mean features of coherence (cc) data were extracted using feature extraction tool in MAPscape-RICE software. These multi-temporal features are having certain range regarding groundnut crop, which were extracted using point sampling tool of QGIS 3.20.0.
 
Crop classification
 
The aim of the image classification is to categorize the image pixels into a land cover category based on the pixel value. This section explains the classification methodology used in this study for groundnut crop area identification and classification.
 
Maximum likelihood classification (MLC)
 
Maximum likelihood classification (MLC) algorithm was used in this study for crop area identification. The MLC quantitatively evaluates both the variance and covariance of the category by spectral response pattern when classifying an unknown pixel. Class mask (containing different classes) generated using multi temporal features of groundnut fields for both VH and VV polarization images in study area were used in the classification. The class mask file was used to precisely segregate the groundnut pixels from other class pixels.
 
Accuracy assessment
 
The error matrix and Kappa statistics are used for evaluating the classification accuracy. The class allocation of each pixel in classified image is compared with the corresponding class allocation on reference data to determine the classification accuracy. The ground reference data are used for validation and the pixels of agreement and disagreement are compiled in the form of an error matrix. The accuracy measures such as overall accuracy, producer’s accuracy and user’s accuracy are estimated. Kappa coefficient is another measure of classification accuracy. It is a measure of the proportional improvement by the classifier over a purely random assignment to classes.

Radar backscattering signature
 
SAR data have a proven ability to detect groundnut through the unique temporal signature of the backscatter coefficient (also termed sigma naught - σ°) exhibited by the crop. The radar backscattering coefficient (σ°) is a measure of crop biomass, plant height, water content, underlying soil, crop phenology etc. The SAR data collected during the cropping period was processed and analyzed using training pixels from ground truth points to derive the temporal backscattering coefficient (σ°) for groundnut from the study area. The temporal backscattering signatures of groundnut during rabi 2020-21 were generated by stacking nine SAR acquisitions from 4^th October 2020 to 8^th January 2021. The signature curves of mean backscattering (dB) values for groundnut showed a marginal increase in backscattering at seedling to vegetative stage (D₁ to D₃) and a steep increase from flowering to pod development stage (D₄ to D₆) followed by a decline thereafter at maturity (D₇ to D₉) and it will present in Table 1 and Fig 4a, 4b. Temporal backscatter values were recorded in 25 test sites across Thiruvannamalai districts and the details are presented in Table 2 and Fig 4c. In that district, backscattering values were found to be ranging from -19.22 to -15.57 dB and -19.37 to -15.50 dB at D₁ (16^th October 2020) and D₂ (28^th October 2020) in VH polarization. At D₄ corresponding to flowering and peg penetration stage, the values were -18.79 to -15.06 dB. The backscattering values increased further and reached a maximum of -18.44 to -14.62 dB and -19.59 to 15.10 at D₆ and D₇ corresponding to pod development to maturity stages and declined thereafter. The increased backscattering values at initial growth stages to flowering stages primarily influenced by LAI and biomass of the crop (Deiveegan et al., 2016) and decrease in backscatter values at the later stages might have been probably caused by maturity of the crop, which lowered the water content of the vegetation or related to the vegetation biomass and or related to the reduced volumetric scattering due to maturity (Panigrahy and Mishra, 2003).
 

 

 

 



 
Multi temporal features extraction
 
Temporal signatures were extracted for each monitoring site and used to generate the dB curves for groundnut fields shows the temporal signature for selected representative pixels to visualize the resulting maximum likelihood classification using multi temporal features (MTF) from C-band SAR imagery of Sentinel-1A. The multi temporal features (in dB value) viz. max, min, mean, max date, min date and span ratio were generated using nine acquisitions during rabi 2020-21 and presented in Table 2. Among the features, the max feature i.e. maximum value for different monitoring fields of groundnut ranged from -19.23 to -15.57 at VH Polarization. Min feature i.e. minimum value ranged from -19.86 to -16.77 at VH polarization. Similarly, mean i.e. mean value and span ratio for groundnut fields are ranged from -18.29 to -15.39 and 2.15 to 5.10, respectively at VH polarization. In groundnut fields of Tiruvannamalai districts, the max date features for VH polarization was recorded at D₉ (8^th January 2021). The min date feature for VH polarization was recorded between D₁ (4^th October 2020) and D₃ (28^th October 2020) with majority of the fields recording minimum date as D₂ (16^th October 2020). SAR based C band imageries of Sentinel 1A were used to generate information of groundnut, which was in line with the findings of Nelson et al., (2014), they were used SAR sensors at short wavelengths in the bands of X, C, K_a and K_u at greater incident angles are suitably sensitive to detect even tiny seedlings immediately after emergence. Whereas, Venkatesan et al., (2019) used microwave SAR Sentinel data for multi temporal features extraction on maize crop at Pereambalur and Ariyalur district of Tamil Nadu.

Groundnut area and accuracy
 
Groundnut area map for the Thiruvannamalai district were derived from multi temporal C-band SAR imagery of Chengam blocks with 5576, 5197, 4989 and 4111 ha, respectively. The minimum area of 29 ha was recorded at Jawadhu hills block of Thiruvannamalai district (Fig 5).
 

       
The accuracy assessment for the groundnut area maps was conducted on a groundnut or non-groundnut basis, where all other land cover types were grouped into single non-groundnut class. In total, 103 validation points covering 69 groundnut and 34 non-groundnut points were considered for validation and confusion matrix were presented in Table 4. The overall classification accuracy was 87.4 per cent with a kappa score of 0.75 indicating the accuracy of classification. The overall accuracy (>80) and kappa index (>0.50) indicated that good level of assessment. The high accuracy of classification in the study area demonstrated that the methodology was appropriate for groundnut area estimation using multi-temporal Sentinel 1A data and indicated the suitability of these remote sensing-based products for policy decisions including crop insurances as quoted by Deiveegan et al., (2016). The findings were in line with Venkatesan et al., (2019) estimated maize area with an accuracy of 91 per cent and Sudarmanian et al., (2017)  estimated rice  area of Thiruchirapalli district of Tamil Nadu with an high accuracy and quantify the methane emission from paddy field using remote sensing datasets.


 

The rabi groundnut area estimated using SAR Sentinel-1A data by supervised maximum likelihood classification and thorough analysis of temporal backscattering during crop growth stages. The estimated groundnut area was 32298 ha with a higher accuracy percentage of 87.4 and kappa coefficient of 0.75.

The authors thank the Agriculture and Land Eco System Division, Space Application Centre, ISRO, Ahmedabad for providing the fund to carry out the research work in a project mode (Space Technology Utilization for Food Security, Agricultural Assessment and Monitoring-SUFALAM Project) and also thank the Professor and Head and staff of the Department of Remote Sensing and GIS, TNAU for their valuable comments and constructive suggestions on the manuscript.