Agricultural drought stress forecasting across different biomes in Brazil using machine learning

Drought events across Brazil have become more common in recent years. The northeast, marked by the dry semi-arid Caatinga biome, experiences the most extreme drought events, which can have a large impact on economic productivity including disruptions to energy and transport infrastructure as well as agricultural losses. However, droughts are an issue across Brazil and can have significant effects on vegetation across the six biomes which make up the country. Drought stress, defined by a reduction in the vegetation health index (VHI) below the 40% threshold is indicative of significant crop yield reductions. Here we evaluate monthly forecasted reductions in VHI below the 40% threshold using machine learning. We then use the results to understand how climatological features influence drought stress in each biome. This is achieved through a combination of Empirical Orthogonal Function (EOF) analysis, Shapley plots, and correlation analysis of predicted spatial trends. We observe distinct behaviour in the Caatinga biome which is rooted in its unique regional climate patterns and physical geography. For other biomes, VHI, RZSM and precipitation are less strongly linked; however high inertia in VHI allows for high forecast performance. The findings strongly imply that drought forecasting models should treat Northeast Brazil as a distinct system, training models specifically for this region to improve the prediction accuracy of vegetation health and integrated drought indices.