Modeling of meteorological, agricultural, and hydrological droughts in semi-arid environments with various machine learning and discrete wavelet transform

Recent meteorological, hydrological, and agricultural droughts in the Mediterranean regions have raised concerns about the impact of climate change. In this study, the meteorological, agricultural, and hydrological droughts were modeled in the Wadi Ouahrane Basin using various machine learning (ML) models and standardized indices of rainfall, evapotranspiration, and runoff (SPI, SPEI, and SRI) at different times scales (1, 3, 6, 9, 12, and 24 months). The applied ML models were the linear support vector machine (SVM), quadratic SVM, cubic SVM, fine gaussian SVM, medium gaussian SVM, coarse Gaussian SVM, rational quadratic Gaussian process regression (GPR), squared exponential GPR, Matern 5/2 GPR, exponential GPR, bagged tree, and boosted tree. Moreover, the hybrid models acquired by combining these ML models with wavelet transform were evaluated. The performance of the models was analyzed using statistical criteria such as root mean square error, determination coefficient, and mean absolute error. As a result, wavelet-GPR models showed the most promising results in estimating SPI, SPEI, and SRI values. The values for SPI ( R 2 of train: 0.393; and test: 0.351), SPEI ( R 2 of train: 0.809; test: 0.746), and SRI ( R 2 of train: 0.999; test: 0.808) indicate monthly time scale. Additionally, for the time periods of 3, 6, 9, 12, and 24 months, the predictions for SPI, SPEI, and SRI generally obtain R 2 of train 0.99 and test 0.95 values. Moreover, it was determined that wavelet-based ML models, established with inputs divided into three subcomponents with Daubechies mother wavelet, showed superior results than standalone ML models. The study results could guide decision-makers and planners in developing drought risk management and mitigation strategies.