Drought prediction using advanced hybrid machine learning for arid and semi-arid environments
Mohsen Rezaei, Mehdi Azhdary Moghaddam, Jamshid Piri, Gholamreza Azizyan, Aliakbar Shamsipour
Abstract
Drought monitoring and forecasting are crucial for efficient water resources management, particularly in arid and semi-arid regions like South Baluchestan, a sub-basin in southeastern Iran known for its tropical fruit cultivation. This study aims to provide reliable predictions of the agricultural Standardized Precipitation Index (ASPI) at various time scales (1, 3, 6, and 12 months) using advanced hybrid machine learning models (ANN-POA, ANFIS-POA, and SVM-POA). Data from ten rain gauge and evaporation stations were utilized for this purpose. The results indicate that the SVM-POA model outperformed both ANN-POA and ANFIS-POA in predicting ASPI drought events. Interestingly, increasing the time scale led to a decrease in the frequency of drought events, while simultaneously causing them to last longer. Additionally, the accuracy of all forecasting methods improved with a longer time scale. A comprehensive evaluation of the models was conducted using six statistical indices (RMSE, MARE, BIAS, NSE, WI, CI), along with visualizations such as clustered column charts, Taylor diagrams, and time-series scatter plots. These findings highlight the potential of the SVM-POA model for short-term drought forecasting, which can significantly contribute to sustainable water resource management and support the cultivation of tropical fruits in the arid and semi-arid South Baluchestan sub-basin.