Predicting below-average NDVI anomalies for agricultural drought impact forecasting

Abstract

Agricultural droughts, driven by deficits in root-zone soil moisture, pose challenges to food security and economic stability in Africa, which is simultaneously vulnerable to frequent droughts and strongly relies on rainfed agriculture. Current Earth observation (EO)-based monitoring systems rely on a near-real-time assessment of vegetation conditions — often through monitoring the Normalized Difference Vegetation Index (NDVI)- and are thereby allowing for reactive rather than proactive drought management. This study presents a machine learning-based forecasting system to predict below-average NDVI anomalies as a proxy for agricultural drought impact, focusing on recently drought-affected and crises-prone countries. By integrating EO data, meteorological forecasts, soil moisture, and static environmental descriptors, we developed a system that forecasts below-average NDVI anomalies up to three months in advance and explicitly considers ensemble uncertainty. The forecast shows an improved accuracy over using near-real-time NDVI anomalies and similar temporal patterns during the 2021–2022 growing seasons, which was used for independent validation. Our forecasted results are comparable to existing NDVI-based monitoring products such as the Agricultural Stress Index System developed by FAO. Despite these advancements, the modelling system struggles during transitions between rainy and dry seasons, often coinciding with the start and end of the growing season. Uncertainties in meteorological forecasts burden effective estimates of important phenological dates such as emergence or harvest up to three months in advance. This study complements existing soil moisture forecasting tools with impact on vegetation and presents a benchmark for the potential of integrating predictive models into anticipatory strategies in existing drought management frameworks.