Crop Yield Estimation at Gram Panchayat Scale by Integrating Field, Weather and Satellite Data with Crop Simulation Models

Abstract

Advanced technologies can improve the operational implementation of the Indian national crop insurance scheme, the Pradhan Mantri Fasal Bima Yojana (PMFBY), particularly in terms of accuracy and timeliness of the crop yield estimates that are used to determine yield losses at the Gram Panchayat (GP) level. In this study, conducted as a pilot test for PMFBY during the kharif season of 2018, technologies based on the Terrestrial Observation and Prediction System (TOPS) were tested and implemented for estimating GP-level crop yields of bajra (pearl millet, Pennisetum glaucum) in Firozabad District of Uttar Pradesh and rice (Oryza sativa) in Kendujhar District of Odisha. A combination of Synthetic Aperture Radar and optical data was used to map crop extent. Daily 2-km grids of input weather conditions were generated using a machine learning algorithm that incorporated station observations, satellite data, and reanalysis model outputs. Required crop biophysical estimates of leaf area index (LAI) and the fraction of intercepted photosynthetically active radiation (FPAR) were derived using daily cloud-screened MODIS 250-m data from the Terra and Aqua satellites and a modified MOD15 LAI/FPAR backup algorithm. A light-use-efficiency (LUE) model adapted from the MODIS (Moderate Resolution Imaging Spectroradiometer) algorithm (MOD17-GPP/NPP) was then used to spatially estimate crop yields. Crop extent maps, daily climate and gap-filled FPAR and the LUE model were used to estimate above-ground biomass, which was accumulated over the growing season and converted to crop yields using a crop-specific harvest index. The estimated yields at 250 m were aggregated within each GP and compared with crop yield data from crop cutting experiments (CCEs) conducted in 142 GPs for rice and 42 GPs for bajra. Crop extent mapping was 96% accurate in rice and 80% in bajra when validated with field surveys. A comparison of modeled yields with CCE yields showed a promising performance by the model in both crops (rice: r = 0.80, root-mean-square error (RMSE) = 411 kg/ha, mean absolute error (MAE) = 359 kg/ha, percent error (PE) = 7, Observed mean = 1500 kg/ha; Bajra: r = 0.84, RMSE = 309 kg/ha, MAE = 262 kg/ha, PE = −12.8, Observed mean = 1859 kg/ha). Although the approach showed promising results for both crops, further progress is needed to ensure consistent and reliable results. Some of the needed improvements include incorporating a dynamic crop calendar, improved maximum LUE estimates, and harvest index values that represent crop varietals grown in India. Routinely conducted CCEs in different crops and seasons around the country could provide a valuable resource for improving these parameters and, ultimately, crop yield estimates.