Document Type : Original Article
Authors
1 University of Hormozgan-Banda abbas
2 University of Hormozgan
Abstract
This study presents a comprehensive evaluation of Support Vector Machine (SVM) and Gaussian Process Regression (GPR) models for drought prediction across Iran's diverse climate zones using the Standardized Precipitation-Evapotranspiration Index (SPEI). The research integrates teleconnection indices, satellite data, and machine learning to address limitations of traditional drought forecasting methods. Results demonstrate the superior performance of GPR with Laplace kernel, achieving higher accuracy (R²: 0.91-0.75 in training, 0.85-0.37 in testing) and better uncertainty quantification (UA: 1.12-2.33, PICP: 1.0) compared to SVM-RBF. This practical improvement translates to a 10-15% increase in the explained variance of drought intensity, a critical distinction for activating different levels of emergency response. The Laplace kernel's flexibility in modeling abrupt climatic variations and GPR's probabilistic framework provide more reliable drought forecasts, particularly in extreme climates. Random Forest analysis revealed distinct climatic drivers, with temperature and evapotranspiration dominating arid regions, while oceanic oscillations (ENSO, WHWP) controlled humid zones. The UNEEC method provided robust uncertainty assessment, showing GPR's consistent performance across different climate classifications. While SVM-RBF remained competitive in moderate climates, its accuracy declined in complex conditions. The findings highlight GPR's advantages for precision drought forecasting in operational early warning systems, where reliable probabilistic forecasts can optimize reservoir management and agricultural advisory services, while acknowledging SVM's computational efficiency for large-scale monitoring applications.
Keywords
Main Subjects
)