Investigating the contribution of decomposition techniques to machine learning accuracy in SPEI-based drought forecasting for multiple Köppen-Geiger climates

Drought is a disaster that affects everything related to humans, particularly the economy. Therefore, predicting its effects before they occur is crucial. However, due to its nature, droughts are more challenging to detect than other natural disasters. This study aims to investigate the effect of decomposition techniques (VMD, DWT, EMD, and EEMD) on the drought forecasting performance of machine learning methods (network-based: MLP, KAN, RNN, BiLSTM, and BiGRU, as well as tree-based methods: RF, GB, XGB, AB, and M5P) in different climate types. To this end, the Standardised Precipitation Evapotranspiration Index (SPEI), which was calculated using 52 years of precipitation and temperature values from 1969 to 2020 for three meteorological stations in Türkiye with different Köppen-Geiger climate classifications, was employed. Drought predictions were made for three SPEI time scales: 3, 6, and 12 months. The results of the analysis revealed that decomposition increased the power of prediction compared to raw drought data, and VMD was the most effective decomposition technique. For instance, the NSE values, which was approximately 0.5 in SPEI-3, 0.7 in SPEI-6, and 0.9 in SPEI-12, increased to above 0.95 across all time scales following the implementation of the VMD method to different climate types. Besides, MLP, KAN, and M5P proved to be the most effective machine learning methods with this value above 0.98 in all data sets. Performance improved as the time scale increased in recurrent neural network-based methods (RNN, BiLSTM, and BiGRU). Consequently, irrespective of the climate region, models employing the decomposition method (VMD and DWT) exhibited considerably enhanced performance.