Enhancing multi-temporal drought forecasting accuracy for Iran: Integrating an innovative hidden pattern identifier, recursive feature elimination, and explainable ensemble learning

Abstract

Study region

The Bam and Babolar stations are located in hyperarid and humid climates in Iran, respectively.

Study focus

The primary objective of this study is to implement Artificial Intelligence (AI) to enhance multistep forecasting of the Standardized Precipitation Evapotranspiration Index (SPEI) for time horizons of one, six, and twelve months. The research employs an innovative hybrid approach that integrates a novel decomposition technique known as Hidden Pattern Feature Extraction Statistical Mode Decomposition (HPFE-SMD), along with Recursive Feature Elimination (RFE) for feature selection, and the Extra Tree Regressor (ETR) model. Additionally, the effectiveness of the suggested model (HPFE-ETR) was assessed and contrasted with two common methods, Time-Varying Filter-based Empirical Mode Decomposition (TVF-EMD) and Variational Mode Decomposition (VMD), both of which were combined with ETR.

New hydrological insights for the region

The results showed that the HPFE-ETR model consistently outperformed comparative models, and it significantly improved drought forecasting accuracy, with the largest improvements for SPEI 12 (t + 12) and more moderate gains for SPEI 12 (t + 1). In particular, the model reduced forecasting errors for SPEI 12 by about 70 % in humid climates and 43 % in hyperarid climates, demonstrating its adaptability across different climatic conditions at both study stations. Explainability results revealed that mean features had the strongest positive influence on SPEI 12 forecasts, underscoring the model’s robustness in capturing key drought drivers. These findings highlight the HPFE-ETR model’s potential to revolutionize drought early warning systems and water resource management strategies.