Integrating Synthetic Aperture Radar (SAR) Sentinel 1A and CROPGRO Peanut Simulation Model for Spatial Yield Gap Analysis

1. Introduction

To bridging the yield gap, especially in developing nations, is a strategic step for alleviating the problem of food security. The difference between the actual yield that farmers are currently producing and the potential yield that could be obtained by utilizing the best agronomic methods is known as the yield gap. This can aid in predicting future crop results for specific places and in identifying the contributing elements to the disparity.

In order to discover production barriers, examine measures to enhance agricultural productivity, and identify intervention regions, it is vital to quantify crop yield gaps and assessing food security of nation [1,2,3]. Furthermore, accurately assessing productivity, land use pattern and food security requires a thorough investigation of existing yield gaps as well as production limits [4,5].

Production capacity changes of farmland, changes in crop productivity, and current yield differences must all to be analysed holistically to potentially offer for farmers policy decision makers so that they optimize their farm land management systems and boost production of food by limiting additional area expansion [6,7,8].

Mostly agricultural yield disparities are on two purposes [5], first is to determine the level of yield gap, which is defined as the difference between potential yield (Yp) and average attainable yields, to be able to determine how much average yields may be raised by management adjustments. The second is to figure out the fundamental factors that lead to yield variations so that policy, research, and extension endeavors can prioritise strengthening land and labour productivity.

Measuring potential yield (Yp) from the highest farmer yield, field trials, and crop simulation models are the three most prominent approaches for assessing potential yield of crops whilst using yield gap (Yg) analysis [9,10].

A comprehension of farming practices at various scales and the assessment of production gaps can both be greatly improved by remote sensing approaches. Technological breakthroughs in remote sensing technologies, like multispectral and hyperspectral sensors on drones and geosynchronous orbits (Synthetic Aperture Radar), have expanded the possibility to perceive small and marginal landholding systems at greater spatiotemporal resolutions, allowing a more accurate assessment of crop yields while integrating with simulation models in those areas. Crop simulation modelling and satellite image-based earth observation technologies have advanced to quickly allow for better regional crop yield monitoring and forecasting. Remote sensing has been an effective method for coping with heterogeneity during the past two decades, and it can be used in conjunction with more conventional strategies like field tests and simulation models. In fact, remote sensing using sensors mounted on aircraft or satellites has the capacity to offer observations for a single field throughout each growing season. In order to investigate yield gap variability, remote sensing might augment surveys or field experiments by overcoming geospatial data limitations.

Through the use of data with high spatial and temporal resolutions, the most recently launched sophisticated remote sensing satellites aids us in making precise estimates of crop area and yield. Additionally, enhanced spatial, spectral, and temporal resolution make it easier for us to distinguish between crops [11]. Due to the availability of these high-resolution satellite sensors, software with automated processing chains, and advanced agricultural yield models, reliable information on crop acreage, crop conditions, and yields have now been provided. The accuracy of crop classification is improved by time series satellite data's repeated covering of the cultivated area. [12]. SAR sensors have an especially impressive capability to detect structural changes in crop canopies. Additionally, they can accurately identify the crop type due to their sensitivity to a variety of crop biophysical variables, such as LAI, yield, and canopy height [13].

Crop simulation models are used to simulate potential crop yields which are then compared to actual yields in many yield gap analysis [5]. When potential yield (Yp) is taken from maximal farmer yield and field studies then there shall be many constraints [14]. In order to figure out Yp from field experiments, it is quite challenging to establish ideal growth circumstances across the crop growth period for a number of seasons to account for climate change. Corresponding to this, the maximum farmers' yield is determined by individual farmers' estimates, which may incorporate flawed information. On the other hand, crop simulation models are very important because they allow for interconnections between crops, soil, weather, water regimes, and management to generate a more precise estimate of Yp [5]. Models also account for the unpredictability of weather conditions across many years. Thus, notwithstanding difficulties relating to parameter estimates or structure functioning, crop simulation models provide the most accurate Yp for a certain area [4,15,16].

In order to address a wide range of complex agronomic problems at both the local and provincial scale, two distinct advancements remote sensing and crop growth simulation models have been created [17,18]. Despite the fact that these two technologies have frequently been explored independently, there is rising interest in using data from remote sensing to update or enhance agricultural ecosystem model simulations. This is because these two technologies are inherently complementary [19,20,21].

Understanding the causes of yield gap variability can help to improve site specific land management and yields [22]. However, on a micro level, there is a lack of understanding of reasons for yield gap. This is mainly due to understanding factors limiting agricultural production at a local level have used approaches like, surveys and field experiments [23]. These techniques have spatial data limitations since they use a limited number of randomly selected units and fall short of giving a thorough knowledge of yield gaps at the micro level when considering the variety that can be found even within individual fields and plots [14]. The purpose of this research paper is to work for the potential value of satellite-based microwave remote sensing and crop simulation model to measure and explain crop yield gaps.

2. Material and Methods

2.1. Study area

Thiruvannamalai district comes under Northeastern agro climatic zones of Tamil Nadu with the spread over of 6191 square kilometers and 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 is presented as Figure 1. The prevailed climate is hot sub humid to semi-arid. The mean annual maximum and minimum temperature are 32.6^o C and 22.2^o C, respectively. The annual rainfall of the Northeastern zone excluding hills varied from 800-1400 mm and more rainfall was received during Rabi 2020-21. Predominant soil type were red sandy loam and clay loam soil which encourages cultivation of Peanut, rice, blackgram and sesame.

2.2. Data used

The benefit of Synthetic Aperture Radar (SAR) is that it can collect data day or night and operates at wavelengths that are not hindered by clouds or poor lighting. With its C-SAR equipment, Sentinel 1A can provide dependable, ongoing wide-area monitoring. Sentinel 1's C-band imaging system features four imaging modes with various resolutions and coverage areas. It is capable of dual polarisation, has a brief revisit time, and delivers products quickly. For crop identification and mapping in the study area during Rabi 2020–21, Ground Range Detected (GRD) information from Sentinel-1A Synthetic Aperture Radar image acquisition with VH polarization in Interferometric Wide (IW) swath mode are gathered at 12-day intervals. The data were downloaded from https://scihub.copernicus.eu/dhus/ based on the usual Peanut growing season in the research location to have complete coverage during the crop growing season, beginning October 4, 2020, to January 8, 2021.

2.2.1. SAR data processing

The multi-temporal space-borne Sentinel 1A SAR IW-GRD data were converted into terrain-geocoded ° values using an entirely automated processing pipeline created by [24]. The processing string was a module of the MAPscape-RICE programme, created by the Swiss company sarmap. The SAR time series data performed a number of important processing steps, as indicated below and illustrated in Figure 2. to produce terrain geocoded degree values.

2.3. Crop classification and accuracy assessment

2.3.1. Maximum likelihood classification (MLC)

To classify imaging pixels into a land use / cover category based on pixel value is the purpose of the image classification process. In this work, crop area identification was done using the maximum likelihood classification (MLC) algorithm. The spectral response pattern is used by the MLC to statistically analyze the individual categories variance and covariance while identifying an unidentified pixel. It is assumed that the training set's distribution is gaussian. According to this premise, the average vector and covariance matrix can completely represent the distribution of a training set for a class. The statistical possibility that a given pixel belongs to a specific class can be calculated assuming these variables.

To determine Peanut, the image was first categorized using MLC with multi-temporal superimposed SAR images. Based on training pixels acquired at various locations, the training spectral signature for Peanut crop was created. The categorization for VH polarization is then completed in order to distinguish the Peanut field from other categories.

2.3.2. Accuracy Assessment

The accuracy of the Peanut area map is validated using the error matrix and kappa statistics. To measure the classification accuracy, each pixel's class assignment in the classified image is contrasted with the matching class assignment on the reference data. An error matrix is created using the agreement and disagreement pixels. The elements of the matrix indicate the number of pixels in the testing dataset, while the columns and rows represent the total number of classes [25]. All of the classified output's classification accuracy was examined.

A farmer survey was used to confirm that the observed post-harvest situation accurately reflected the presence of a Peanut crop during the surveilled season. Generally, the accuracy assessment in fields was performed during the pod filling or maturity phase before harvesting, but in some cases, the field assessment was conducted during post-season and Peanut haulms. The error matrix was used to determine the accuracy metrics, including overall accuracy, producer accuracy, and user accuracy [26]. The proportion of correctly identified cases located along the diagonal, which represent the overall accuracy, were determined as follows:

By dividing the total number of correctly categorized samples by the total number of reference

samples, the producer's accuracy (errors of omission) for each class was calculated as follows:

By dividing the total number of samples in a class that were verified to belong to that class by the number of samples in that class that were correctly classified, the user's accuracy (errors of commission) of each class was calculated as follows:

Kappa Coefficient

The kappa coefficient, which estimates the proportionality (or percentage) increase by the classifier over a simply random assignment to classes, is a different metric for classification accuracy [27]. The formula shown below was used to estimate the kappa coefficient.

In the case of a confusion matrix containing r rows and correspondingly r columns, Where,

A = the sum of r diagonal elements, which is the numerator in the computation of overall accuracy

B = sum of the r products (row total x column total)

N = the number of pixels in the error matrix (the sum of all r individual cell values)

2.4. Crop Cutting Experiments (CCEs) for Actual Yield estimation

Actual yield of Peanut or farmers attainable yield during Rabi 2020-21 was estimated with the help of Department of Agriculture Economics and Statistics by organizing CCEs at 31 monitoring sites of Thiruvannamalai district. In each monitoring sites four CCEs were organized at randomly selected area at a size of 5x5 m. The average of Four CCEs from each monitoring site were converted to yield per ha and use it as actual yield for that monitoring site. The estimated yield data used as actual attainable yield for the measure of yield gap.

2.5. Crop simulation using CROPGRO Peanut model for Potential yield estimation

DSSAT is a software package for microcomputers that combines crop, soil, and weather information in common formats for crop modelling and application programme evaluation. The DSSAT was employed in the current analysis because the user may simulate the results of cultivation techniques over a number of years for various crops in any area in the world. The elements of the crop simulation models for DSSAT shown in Figure 3.

In DSSAT version 4.5, the CROPGRO-Peanut model can be utilized to simulate the daily growth and development of the Peanut crop. DSSAT is one-dimensional model that calculates daily variations in soil water content in a single layer through soil evaporation, infiltration, unsaturated flow, irrigation, vertical drainage, root water uptake and plant transpiration. In addition to crop coefficients that were thought to be less variable or even more conservative among crop cultivars, it also required cultivar coefficients (cultivar-specific characteristics) as an input to the model. The model could represent how high temperatures affect peanut development and growth. The maximum potential yield was computed using the input datasets for 31 monitoring locations by DSSAT CROPGRO peanut model. It needed just the mentioned little datasets for its execution.

2.5.1. Weather file

Automatic Weather Stations (AWS) and regular observatories located at the study districts provided the daily meteorological information on maximum temperature (°C), minimum temperature (°C), solar radiation (MJ m^-2 day^-1), and rainfall (mm) for Rabi 2020-21, which were used to create a weather file for running the CROPGRO-Peanut model. To anticipate the yield, the DSSAT model needed weather information for the whole crop's growing season.

2.5.2. Soil data file

The Department of Remote Sensing and Geographical Information System at the Tamil Nadu Agricultural University in Coimbatore provided the soil data used to create the soil files. This was accomplished by using a digital soil information system that had data on the following soil characteristics (1:50000) viz., depth, texture, pH ,EC, BD, drainage, organic carbon, soluble P and K, Ca and Mg, and CEC. The soil profile information’s as required in DSSAT were fed into the S-Build tool in DSSAT to create a soil file. The Thiruvannamalai district comprises of 12 soil series viz., Perunthurai, Devikapuram, Pedakallupalli, Meyyur, Vetavalam, Vilukam, Kolathur, Ayyalur, Perumalthangal, Mattathari, Parapperi, Kalugachalapuram and details were used for S build tool presented in Appendix 1.

2.5.3. Crop management file

The inputs to the model for each of the 31 monitoring locations that will be stimulated for Rabi 2020–21 were documented in the crop management (X built) file. The actual figures of the experimental conditions and field characteristics, including the name of the weather station, the soil, the initial soil moisture, the field's description, planting geometries, the conditions of inorganic nitrogen, irrigation and water management, fertilizer management, tillage operations, the application of organic residue and chemicals, environmental modifications, harvest management, simulation control mechanisms (specification of simulation options, such as starting dates, on/off options for water and nitrogen balances, symbiosis) and output options were included in the X built tool as crop management file.

2.5.4. Estimation of genetic co-efficient Peanut

The tweaking of genetic parameters, also known as model calibration or parameterization, ensuring that simulated and observed values are closely comparable. The genetic variables of cultivars were estimated using the data collected from the experiment. In order to achieve the most accurate fit between the observed and simulated number of days to the plant developmental events and grain yield at harvest, the genetic coefficients that affect the occurrence of developmental stages in the CROPGRO-Peanut model entrenched in DSSAT model were derived successively by manipulating the relevant coefficients. For developing varietal genetic coefficient, the field experiment with randomized block design was conducted for popular Peanut varieties at Oilseeds Research Station, Tamil Nadu Agricultural Research Station, Tindivanam during Rabi 2019-20 against the ruling varieties viz., TVM 7, TMV 13, VRI 2 and G 7 were used for generating varietal coefficient.

2.6. Yield estimation by integrating Remote Sensing and DSSAT Crop growth model

SAR Sentinel-1A data integrated with DSSAT CROPGRO Peanut for estimating Peanut yield were presented in Figure 4. LAI values retrieved from dB images of the SAR data was used to integrate the DSSAT simulated yield with the remote sensing data. A linear regression equation was generated between DSSAT simulated yield and spatially simulated LAI values for estimating Peanut yield.

2.6.1. Generation of LAI for Peanut yield estimation

Five LAI readings were taken from each monitoring field, and the average was calculated. In order to provide information to the CROPGRO-Peanut system for simulation and validation, yield data was collected at the completion of the season from each monitoring field and the entire dataset was finalized over the study area with measurements of Peanut pod yield and LAI. Regression analysis was carried out between the observed and expected yield using the data from the study region pooled into a single dataset.

2.6.2. Retrieving LAI from dB images of SAR data

Using the point sample tool in QGIS 2.18.4, the dB (back scattering) readings of Peanut crops were gathered for monitoring fields. In the monitoring Peanut fields of the study region, the equation for linear regression was created between the dB readings and the simulated LAI values at pod growth stage (Figure 5). The raster calculator tool in ArcMap gives users the option to create selection queries, enter Map Algebra syntax, or do basic arithmetic using operators and functions. Therefore, in this investigation, the dB values of dB pictures from mature stages were substituted for the calculated regression values using the raster calculator in ArcMap. To construct yield maps and statistics spatially, an empirical method was used to extrapolate the field observed and DSSAT calculated LAI values (Figure 6).

Yield gap is defined as the difference between potential yield (Yp) and average yield (Y_a). The potential yield was derived from crop simulation model CROPGRO peanut by a trail conducted at Oilseeds Research Station, Tindivanam. Basic information essential for DSSAT model calibrate and validate were used from this trial and estimated the potential yield of Peanut for the study area. Actual yield data of Peanut were collected by organizing the crop cutting experiments in the study area.

3. Results and Discussion

3.1. Peanut area and accuracy assessment

Peanut area map for the Thiruvannamalai district were derived from multi temporal C-band SAR imagery of Sentinel-1A. Using the shape files of administrative boundaries, district wise and block wise maps and statistics of Peanut area were extracted for 18 blocks of Thiruvannamalai district. The classified Peanut area for each block are displayed in Table 1 and Figure 7 and the total peanut growing area during Rabi 2020-21 was estimated to be 32298 ha. Thandrampet block in Thiruvannamalai district had the maximum area, with 6970 ha followed by Thurinjapuram, Thiruvannamalai, Keelpennathur and Chengam blocks, which had 5576, 5197, 4989 and 4111 ha, respectively. The Jawadhu hills block of Thiruvannamalai district recorded a minimum area of 29 ha.

The Peanut area maps were assessed for accuracy on a Peanut or non-Peanut basis, where all other land cover types were grouped into non-Peanut class. In total 103 validation points covering 69 Peanut and 34 non-Peanut points were considered for validation and confusion matrix were presented in Table 2. 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 per cent) and kappa score (>0.50) indicated that the assessment was of high quality. The results were in accordance with [28,29,30,31] has delineated the crop area estimation with high classification accuracy using multi temporal sentinel 1A data had indicated the suitability of remote sensing-based products for crop area estimation and monitoring for policy decisions, including crop insurances.

3.2. Genetic Coefficient development and calibration of DSSAT model

Crop simulation model genetic coefficient development and calibration are vital for establishing their credibility and ability in simulating growth, development, yield attributes and yield of crops [32]. Among the four varietal genetic coefficients (Table 3) generated during the Rabi season of 2019-20, the time between emergence to flower appearance (EM-FL) and time between the first flower to first pod (FL-SH) ranged from 17.50 to 19.00 photothermal days and 7.00 to 8.00 photothermal days, respectively. In contrast, the time between the first flower to first seed (FL-SD) and the time between first seed to physiological maturity stage (SD-PM) was found to be a range of 17.30 to 18.50 and 58.00 to 62.00 photothermal days, respectively. The maximum leaf photosynthesis rate ranged from 1.09 to 1.25 mg CO₂ m^-2 s^-1. The Seed filling duration (SFDUR) and seeds per pod ranged from 31 to 33 photothermal days and 1.65, respectively, for six varieties. Similarly, [33,34,35,36] developed genetic coefficient for groundnut cultivars to various agro climatic zones of India for yield estimation using simulation models.

The capability of the DSSAT CROPGRO peanut model to predict growth and development of the four predominant Peanut cultivars viz. TMV7, TMV13, VRI 2 and G7 (Table 4) were assessed to predict crop response as influenced by season, weather, management and soil distribution by field experiment conducted at Oilseeds Research Station, Tindivanam.

Further, the DSSAT model resulted in a simulated canopy height of 56 to 69 cm, and the maximum LAI for these six varieties ranged from 4.93 to 5.60. These simulation results were in line with the findings of [37], who opined that when independent crop and soil datasets were used to evaluate, the PNUTGRO model was influenced by seasons and location on simulating growth and yield. Apart from that, [38,39,40] tried to simulate the varietal influence on the growth and physiological process at different regions using DSSAT crop simulation model. The simulated Peanut pod yield of different Peanut varieties ranged from 3935 to 4657 kg ha^-1. Accordance with [28], simulated the potential yield of rainfed Peanut across major production zones of Tamil Nadu, which correlated highly with the results obtained in the present study among the varieties. Similar finding on pod yield simulation by DSSAT in a different location was reported by [36,37,41,42,43] and simulation model overestimated the pod yield then actual attainable yield because simulation of crops has been simulated the yield without limiting factors. The CROPGRO peanut model was used by [34] to simulate the phenological events, yield and yield attributing characters of Peanut cultivars of GG 2 and GG 20 in Gujarat precisely. In the past years, many research efforts were taken to simulate the growth and productivity of Peanut and other crops.

3.3. Spatial estimation of Potential, Actual yield and Yield gap in Peanut

3.3.1. Potential Yield Estimation

DSSAT based potential yield for monitoring sites and block wise spatial Peanut potential yield of Thiruvannamalai district were furnished in Table 5 and Table 6. According to [44] regression analysis is a useful method for creating yield prediction models. Among the monitoring sites, the maximum potential production of 4843 kg/ha was simulated using CROPGRO peanut at Mel palur and Kil palur villages, while the least potential yield of 3194 kg/ha was simulated at Tandarai village of Thiruvannamalai district. Similarly, from the block wise spatial potential yield of 18 blocks of Thiruvannamalai district, minimum spatially estimated potential yield varied from 3711 to 3805 kg/ha recorded at Pernamallur and Polur blocks, respectively with mean minimum yield of 3750 kg/ha. The maximum spatially estimated potential yield varied from 3936 to 4306 kg/ha recorded at Thandarampet and Vembakkam blocks, respectively with mean maximum of 4029 kg/ha. The overall mean spatially estimated potential yield for Thiruvannamalai district was 3874 kg/ha and spatially estimated potential yield map was presented in Figure 8.

As the technique simulates based on given soil, water, climate and management to their full capacity, DSSAT based potential yield estimation always provides higher yield than farmers attainable yield similar to the findings [45]. The DSSAT model simulates daily leaf area index (LAI), vegetation status parameters, biomass production and potential yield using soil, crop management and daily weather data as input. As a result of taking into consideration significant interactions between crops, weather, soil, water regimes, and management, as well as variations in weather across years and areas, simulation models also offer the most precise estimation of potential yield. Within the climatic, soil, and management context of a certain cropping system, they enable assessment of potential or water-limited productivity. Crop models will therefore perform better if provided a technique to incorporate both temporal and spatial variation to the level that data are place specific than any empirical method can do [5].

3.3.2. Actual Yield Estimation

Crop Cutting Experiments (CCE) based actual yield of monitoring sites and block wise spatially estimated actual yields were presented in Table 5 and Table 6. Actual Peanut yields recorded at monitoring locations of Thiruvannamalai district during Rabi 2020-21 ranged from 1228 to 3106 kg/ha in Agaram and Mel palur villages, respectively. Among the blocks of Thiruvannamalai district, minimum spatially estimated actual yields were 1394 to 1644 kg/ha in the Tandarampet and Jawadhu hills blocks, while maximum yields ranged from 1858 to 2526 kg /ha in the Vembakkam and Tandarampet blocks, respectively. The overall mean actual spatial yield was 1740 kg/ha and the actual spatial yield map was presented in Figure 9.

During Rabi 2020-21, the actual crop actual decreased in comparison to previous years. However, such a decrease on the actual yield was accompanied by a significant year to year fluctuation, which may be attributed to rainfall variability. The existing relationship between crop growth period rainfall and overall actual production is supported by [46]. Although it is still in its early stages, the decline in real yield may be due to a change in the region's adoption of better agronomic measures during the course of the indicated period.

3.3.3. Peanut Yield Gap Estimation

The yield gap of monitoring sites and block wise spatial yield gap are presented in Table 5 and Table 6. Among the monitoring sites, Ponnaiyur recorded the more yield gap of 2346 kg/ha whereas Mangalam village recorded the least yield gap of 1217 kg/ha compared to other monitoring sites. For the Thiruvannamalai district, the minimum yield gap ranged from 1456 to 2072 kg/ha at Thiruvannamalai and Vembakkam blocks while the highest yield gap of 2231 to 2508 kg/ha prevailed at Thellar and Thiruvannamalai blocks respectively. The overall mean spatial yield gap of Thiruvannamalai district was estimated at 2134 kg/ha. By combining remote sensing (SAR Sentinel-1A) data with crop simulation models (DSSAT CROPGRO peanut) the spatial yield gap of Peanut was investigated in this paper for Rabi 2020-21 and yield gap map was presented in Figure 10. The yield gap recorded was 1217 to 2346 kg/ha and the major production constraints of Peanut for yield gap in the district were seasonal rainfall variability, poor soil fertility, lack of improved varieties and inadequate weed managements. The current yield gap could be managed by combining traditional agronomic approaches and latest technologies with involvement of different stakeholders. A substantial amount of work has been devoted to mapping and interpreting yield variations within a single field in order to inform precision agriculture applications [4].

4. Conclusion

Integration of SAR Sentinel-1A with CROPGRO Peanut for Spatial yield gap assessment of Peanut during Rabi 2020-21 at Thiruvannamlai district had shown that 2134 kg/ha of average spatial yield gap with mean minimum of 1890 kg/ha and maximum of 2324 kg/ha. The key production constraints of this Peanut growing areas in Tamil Nadu was due to seasonal rainfall unpredictability, poor soil fertility, lack of improved varieties, and inadequate weed management, which created a wide gap between potential and attainable yield. The current yield gap could be bridged by combining traditional agronomic approaches and latest technologies with the involvement of different stakeholders. As a result, remote sensing to examine the magnitude and causes of yield gaps on a regular basis with improved yield gap awareness will be critical in satisfying future crop demand at reasonable prices while minimizing the environmental effect.