2020, Arumugam, Ponraj, Chemura, Abel, Schauberger, Bernhard, Gornott, Christoph

Immediate yield loss information is required to trigger crop insurance payouts, which are important to secure agricultural income stability for millions of smallholder farmers. Techniques for monitoring crop growth in real-time and at 5 km spatial resolution may also aid in designing price interventions or storage strategies for domestic production. In India, the current government-backed PMFBY (Pradhan Mantri Fasal Bima Yojana) insurance scheme is seeking such technologies to enable cost-efficient insurance premiums for Indian farmers. In this study, we used the Decision Support System for Agrotechnology Transfer (DSSAT) to estimate yield and yield anomalies at 5 km spatial resolution for Kharif rice (Oryza sativa L.) over India between 2001 and 2017. We calibrated the model using publicly available data: namely, gridded weather data, nutrient applications, sowing dates, crop mask, irrigation information, and genetic coefficients of staple varieties. The model performance over the model calibration years (2001–2015) was exceptionally good, with 13 of 15 years achieving more than 0.7 correlation coefficient (r), and more than half of the years with above 0.75 correlation with observed yields. Around 52% (67%) of the districts obtained a relative Root Mean Square Error (rRMSE) of less than 20% (25%) after calibration in the major rice-growing districts (>25% area under cultivation). An out-of-sample validation of the calibrated model in Kharif seasons 2016 and 2017 resulted in differences between state-wise observed and simulated yield anomalies from –16% to 20%. Overall, the good ability of the model in the simulations of rice yield indicates that the model is applicable in selected states of India, and its outputs are useful as a yield loss assessment index for the crop insurance scheme PMFBY.

2021, Arumugam, Ponraj, Chemura, Abel, Schauberger, Bernhard, Gornott, Christoph

Accurate and spatially explicit yield information is required to ensure farmers’ income and food security at local and national levels. Current approaches based on crop cutting experiments are expensive and usually too late for timely income stabilization measures like crop insurances. We, therefore, utilized a Gradient Boosted Regression (GBR), a machine learning technique, to estimate rice yields at ~500 m spatial resolution for rice-producing areas in India with potential application for near real-time estimates. We used resampled intermediate resolution (~5 km) images of the Moderate Resolution Imaging Spectroradiometer (MODIS) Leaf Area Index (LAI) and observed yields at the district level in India for calibrating GBR models. These GBRs were then used to downscale district yields to 500 m resolution. Downscaled yields were re-aggregated for validation against out-of-sample district yields not used for model training and an additional independent data set of block-level (below district-level) yields. Our downscaled and re-aggregated yields agree well with reported district-level observations from 2003 to 2015 (r = 0.85 & MAE = 0.15 t/ha). The model performance improved further when estimating separate models for different rice cropping densities (up to r = 0.93). An additional out-of-sample validation for the years 2016 and 2017, proved successful with r = 0.84 and r = 0.77, respectively. Simulated yield accuracy was higher in water-limited, rainfed agricultural systems. We conclude that this downscaling approach of rice yield estimation using GBR is feasible across India and may complement current approaches for timely rice yield estimation required by insurance companies and government agencies.