Hybrid machine learning for drought prediction at multiple time scales: a case study of Ağrı station, Türkiye

Abstract

Drought is a prolonged period of significantly reduced precipitation, resulting in water scarcity and environmental stress. In this study, Ağrı province, situated in the eastern region of Türkiye, where most of the land cannot be irrigated and the livelihood is based on agriculture, was selected as the study area. Meteorological droughts in Ağrı province were forecasted using hybrid machine-learning models, leveraging monthly precipitation and temperature series from 1965 to 2022. The study employed the standardized precipitation index (SPI), relying solely on precipitation data, and the standardized precipitation evapotranspiration index (SPEI), which also considers both temperature and precipitation data. Various timescales, including 1M (1 month), 3M, 6M, 9M, and 12M, were taken into consideration. The best model for each hybrid model was determined using data at time points t, t-₁, t_-2, t_-3, and t_-4 for the relevant time series. The study combined ensemble least squares boosting algorithms (LSBoosting), adaptive network-fuzzy inference system (ANFIS), support vector machines (SVM), Gaussian process regression (GPR), and M5 model tree (M5Tree) approaches with the variational mode decomposition (VMD) technique to create hybrid models. The results indicate that certain models perform better at different timescales, with M5Tree and GPR generally providing higher accuracy. For instance, the M5Tree model achieved the lowest MAE (0.0714 and 0.0555) and RMSE (0.0909 and 0.0732) values for the _9MSPI and _12MSPI timescales, respectively, making it the best-performing model at these scales. Similarly, the GPR model stood out for the _1MSPI and _6MSPI scales with the lowest MAE values (0.1336 and 0.0736, respectively). Based on the performance criteria, the best hybrid model for the _1MSPI was the GPR approach. For the SPEI, except for _3MSPEI, the M5Tree approach showed the best performance at other timescales. However, since the prediction outcomes gave similar results according to classical performance criteria, a one-sided Wilcoxon sign rank test was applied to the outcomes of ANFIS, GPR, and M5Tree models for _3MSPEI, _6MSPI, _9MSPI, and _12MSPI. It has been determined that these three models are not superior to each other. Additionally, the one-sided Wilcoxon signed-rank test found no statistically significant difference between ANFIS, GPR, SVM, and M5Tree models for the _3MSPI. This research concluded that the performance of hybrid machine-learning methods applied to different timescales of SPI and SPEI varies.