Artificial intelligence for agricultural drought monitoring based on soil moisture and crop yield under change

Water is a critical resource for food production. Climate change has shown that shifts in precipitation regimes and increases in atmospheric temperature threaten food production worldwide. Therefore, it plays an essential role as a driver in the water-environment-food nexus. Strategies to cope with impacts of climate risk require understanding the relationship between hydro-meteorological extremes and crop yield shortfalls to guide decision-making. The purpose of this paper is to propose a fully data-driven model for predicting the impacts of hydrological extremes on food production for decision-making at the municipal level. We use the Support Vector Machine (SVM) model considering a variety of kernels for predicting crop yields using reanalysis data of water storage (WS) from 0 to 28 cm of soil during the soybean growing season. We used ERA5 WS data for Paraná state in Brazil and official annual soybean crop yields (SBY) at the municipal level. We tuned a SVM radial basis function kernel with sigma and cost parameters for municipalities with significant production of soybean. The SVM model predicted crop yields accurately with R² ranging from 0.12 to 0.85. The use of soil moisture data increased model accuracy from 30 to 95% and reduced error from 5 to 58% in relation to using only SBY, except for one location. These results indicate that the model is better able to depict SBY during drier conditions, reducing prediction accuracy in years with average or above average yields. The model we proposed is useful for estimating crop losses due to water shortage at the municipal level. Our results suggest that using WS data from ERA5 as an additional input to past SBY adds relevant information for several applications such as risk transfer, irrigation planning and farm-level management with data that are made available for most countries. Our model has potential for climate impact studies coupled with projections from Global Circulation Models forced by Shared Socioeconomic Pathways (SPP).