Assimilating Remote Sensing Data In Crop Insurance Models: Multiple Approaches to Estimating Joint Distribution

Abstract

Accurate measurement and assessment of crop risks has eluded the insurance industry across the world, particularly in the global South. The insurance industry suffers from ‘incomplete markets’, despite support from national governments, for example, in India, the European Union (EU) and the United States (US). Remote sensing data could provide objective data that could aid the risk assessment; however, it requires novel approaches and their validation. Specifically, it remains to be seen how best to obtain the joint distribution of crop yields and the environmental variables, such as temperature and rainfall. Further, updating the risk estimates with evolving information and conditionalities during a season is also necessary from the dual perspectives of operational efficiency and risk management. We explore a solution in the context of the north Indian wheat crop grown in the winter (Rabi) season. For estimation of distribution, we consider three approaches a) Independently fitting distribution on each explanatory variable and then determining the crop yield distribution using the linear relationship of crop yield with explanatory variables, b) Jointly fitting standard conditional predictive distribution on the crop yield along with explanatory variables, and c) Fitting Bayesian conditional predictive distribution on the crop yield given explanatory variables. In each approach, various parametric and nonparametric distributions can be included. Our early results point to an improved estimation of risks when compared to the United States’ Risk Management Agency (RMA) approach. Further, the versatility of joint conditional distribution is impressive, and it explains the nonlinear interactions with vegetation and weather variables quite well. We see promise in these methods to address the food security challenges in the wake of climate change. Apart from immense practical significance, our work furthers a methodological understanding of conditional distribution through standard statistical conditional inference and Bayesian inference.