Improved multi-class drought classification through climate data analysis and advanced imbalanced data handling

Study region The Urmia Lake Basin, a terminal endorheic system in semi-arid northwestern Iran, faces severe hydrological decline. 51 years of monthly precipitation records (1971–2021) from four upper sub-basin rain gauges are analyzed alongside 11 climate teleconnections. Study focus Detecting rare, extreme drought events in terminal basins is frequently hindered by class imbalance in historical hydroclimatic records, biasing machine learning classifiers toward normal conditions. We benchmark a complete ladder of nine data-augmentation methods, including SMOTE, ADASYN, stand-alone CTGAN and VAEs, and two novel hybrids, to improve multi-class Standardized Precipitation Index nowcasting (3–12-month scales) under an identical, strictly leakage-free chronological evaluation protocol. New hydrological insights Under chronological validation, three-class nowcasting yields a moderate operational baseline of approximately 0.42 macro-F1. This exposes a massive temporal leakage gap from the artificial 0.90 F1-score obtained via conventional shuffled splitting, which fits the strong temporal autocorrelation of cumulative SPI. While simpler adaptive oversamplers maximize macro-F1, the hybrids achieve the highest minority-class recall and G-mean on the rare dry states. This trade-off is mathematically explained by CTGAN learning the non-convex probability manifold, whereas SMOTE is geometrically restricted to the minority class’s convex hull. Longer scales function as low-pass filters, smoothing synoptic noise and isolating persistent planetary signals. This nowcasting tool can be directly translated into regional water governance through an operational Traffic-Light Policy Roadmap at Bukan Dam.