Research on drought prediction models by multiple machine learning methods in a semi-arid region

Study region The Ziya River Basin (ZRB) is located in northern China, characterized by increasing drought frequency due to climate change. Study focus This study aims to develop and compare three machine learning-based drought prediction models - Random Forest (RF), Support Vector Machine (SVM), and Extreme Learning Machine (ELM) - for the ZRB. Unlike traditional single-factor drought indices, the proposed models comprehensively integrate multiple meteorological and hydrological variables, including precipitation, temperature, potential evapotranspiration, vegetation characteristics, and soil moisture. The Moisture Index ( MI ) was identified as the optimal target variable through correlation analysis. Model performances were evaluated using multiple metrics ( R², MAE, RMSE, NSE, KGE, and PBIAS ) and validated against in-situ soil moisture measurements at various depths. The optimal model was then applied to simulate the spatiotemporal distribution of drought and to statistically analyze drought-affected areas under different land use types, soil textures, elevations, and slopes. New hydrological insights for the region The ELM-based drought prediction model (ELM-DI) significantly outperformed RF and SVM, achieving the highest fitting accuracy ( R² : 0.948–0.989; MAE: 0.062–0.158; RMSE: 0.086–0.309; NSE > 0.90; KGE > 0.90; PBIAS < 5%). Validation against measured soil moisture data showed that ELM-DI exhibited the strongest correlation with deep-layer soil moisture. Spatiotemporal analysis revealed that cropland, loam soils, elevations of 1000–2000 m, and slopes of 2°–6° experienced the largest drought-affected areas across all seasonal scales.