Machine Learning for Global Drought Onset Forecasting

Extreme droughts threaten agriculture, water management, and ecosystems worldwide, and the frequency and severity of these natural disasters have been increasing with climate change. While early warning systems can improve preparedness, forecasting extreme drought 12 months ahead remains challenging due to drought's complex drivers, which span atmospheric, oceanic, and land-surface processes. This paper uses a machine-learning framework to forecast the onset of extreme drought up to one year in advance. A single global model combines three complementary data streams: ERA5-Land reanalysis climate data capturing land--atmosphere dynamics, teleconnection indices representing large-scale precipitation drivers, and local SPEI history encoding drought memory. The model predicts whether extreme drought (Standardized Precipitation Evapotranspiration Index, SPEI-3 $\le -2.0$) will occur within the next 12 months at any global land location not currently in extreme drought. Because drought is persistent in time and its risk is uneven in space, we evaluate the model per region at onset against two reference forecasts that already contain these effects: persistence and a location-dependent climatology (each cell's historical onset rate). The model outperforms both references on the pooled global score and, in the regional onset analysis, in continental Asia and a monsoon-Asia box nested inside it: $+0.074$ AUC over persistence and $+0.112$ over the climatology. We propose onset-conditional evaluation against these two references as a simple baseline for machine-learning drought forecasting.