Advanced Hybrid Machine Learning for Precise Short-Term Drought Prediction: A Comparative Study of SPI and SPEI Indices in Iran's Arid and Semi-Arid Regions

Abstract

Drought has been viewed as a climatic event of significant importance that hampers agricultural productivity, efficient management of water resources, and socio-economic development, especially in arid, semi-arid, and arid-semiarid regions. Even though improved approaches to modeling dry spells have been reported, there remains a substantial disparity in the forecasting ability of the Standardized Precipitation Index (SPI) and the Standardized Precipitation Evapotranspiration Index (SPEI) for different climatic conditions. In response to the observed disparity, the current study utilized the Tuned Q-factor Wavelet Transform (TQWT), Variational Mode Decomposition, Empirical Mode Decomposition, and Empirical Wavelet Transform (EWT), together with Gaussian Process Regression (GPR), Support Vector Machines, and Adaptive Neuro-Fuzzy Inference System (ANFIS) models. The dataset included precipitation and temperature data collected from four synoptic instrument-equipped meteorological stations from 1990 to 2022—Tabriz and Shiraz corresponding to semi-arid, and Kerman and Yazd corresponding to arid regions—and included SPI and SPEI index predictions for temporal periods of 1, 3, and 6 months. Through the use of autocorrelation diagnostics, it was possible to identify the optimal input lags (t-1, t-2, and t-3) specifically allocated for the model development process, derived from 75% of the available dataset. For the case of the 1-month temporal period, the models using the TQWT revealed the best forecasting effectiveness; most importantly, the TQWT-ANFIS model recorded the highest accuracy at the Tabriz station, while the TQWT-GPR model showed the highest accuracy values at Shiraz, Kerman, and Yazd (R2≈0.996–0.997; RMSE≈0.05–0.07). For the 3- and 6-month temporal evaluations, the EWT-ANFIS model recorded the best performance among all allocated stations, marked by the lowest error metrics (RMSE≈0.01–0.03) together with nearly perfect goodness-of-fit values (R2 and NSE≈0.999). The Shiraz and Kerman observation stations showed the best performance indices, reaching a Kling-Gupta Efficiency (KGE) of 0.99. By comparison, the report from Tabriz indicated a poorer KGE of about 0.93, while the Yazd station showed volatility in the 6-month Standardized Precipitation Index, reaching a KGE of about 0.60, suggesting a rising aridity trend. Overall, results demonstrate that while TQWT-based models dominate short-term drought prediction, EWT-ANFIS is the most robust for medium- and long-term forecasts. These findings emphasize the potential of hybrid decomposition–machine learning frameworks to improve drought monitoring and strengthen water resource management strategies in water-scarce regions. © The Author(s), under exclusive licence to Springer Nature Switzerland AG 2025.