QUANTIFYING DROUGHT PERSISTENCE AND TRANSITION PROBABILITIES USING ADVANCED SPATIAL ANALYSIS OF SPI-24 DATA

ABSTRACT

Drought is still a severe worldwide threat to agriculture, water and society. Although current drought watchers combine several indices and satellite information, difficulties remain with respect to spatial location and predictive ability. Conventional methods (e.g., SVMs, NNs) have a problem of spatial autocorrelation and imbalanced data. In this study, drought prediction is augmented via DBSCAN clustering, variogram-based spatial modelling, and ROC-optimized thresholds. Our model achieves 0.955 AUC and 0.964 sensitivity and 0.943 F1-score, which yield 13% to 29% relative improvements with respect to baselines. It has strong spatiotemporal coherence and can be used for an early drought warning and for early-warning countermeasures.

Keywords: drought prediction; DBSCAN clustering; variogram modelling; ROC optimization; SPI-24; spatial autocorrelation; early warning systems; AUC; sensitivity; geo-statistics