Estimating drought impacts on crop yield using AI and EO

Climate change has intensified droughts in many parts of the world, severely impacting different sectors. In particular, the agricultural sector is highly sensitive to precipitation deficits and the resulting soil moisture deficit, leading to a drastic reduction in crop productivity. There is an urgent need to ensure access to food for a growing population in future, making it essential to address agricultural drought induced crop yield losses. Multimodal satellite and reanalysis climate data archives, coupled with advancements in machine learning, offer a promising avenue to address this issue, but studies are often limited to the calculation of drought indices. In order to produce actionable insights and allow for time to prepare for drought-related food production deficits, specific information on crop losses is needed. Therefore, this study demonstrates the potential of the machine learning algorithm Random Forest (RF) for annual crop yield forecasting using multimodal datasets, for two agriculturally important drought-prone regions in India and Germany. Using 11 climate variables from ERA5 data and PKU GIMMS NDVI (version 1.2) from 1990 to 2021, an RF model was trained to predict crop yields for two common crops across the study sites. The model was evaluated at different spatial scales and the spatial transferability of the model was also tested, using Root Mean Square Error (RMSE; absolute error metric) and Mean Absolute Percentage Error (MAPE; relative error metric). Feature importance was also assessed across scales and across different study sites, using the mean decrease in impurity as a post-hoc explainability tool. Results show that different features are important for accurate crop yield predictions in different regions, for different crops, and across different space-time scales. Spatial transferability requires retraining the model with local data, due to the strong influence of local climatic and agricultural conditions as well as data availability. Findings pave the way for long lead time predictions of drought impacts on agricultural productivity purely open source data, contributing directly to improving global food security equitably, as the methods are equally applicable in data-rich and data-poor contexts.