Drought forecasting across multiple temporal scales using CMIP6 projections and hybrid deep learning models for Central Anatolia, Türkiye

Abstract

This study proposed a multi-temporal drought forecasting framework that integrates CMIP6 climate projections with hybrid deep learning models to enhance forecasting accuracy under climate change. For the Kayseri, Nevs & cedil; ehir, and K & imath;rs & cedil;ehir stations located in the Central Anatolia Region of T & uuml;rkiye, climate uncertainty was reduced by adopting a multi-model ensemble mean derived from 19 CMIP6 global climate models. Subsequently, monthly precipitation outputs under three socio-economic pathways (SSP1-2.6, SSP2-4.5, and SSP5-8.5) were used to compute the Standardized Precipitation Index at 1-, 3-, and 6-month time scales (SPI-1, SPI-3, and SPI-6). The SPI time series were decomposed using the Tunable Q-factor Wavelet Transform (TQWT), Maximal Overlap Discrete Wavelet Transform (MODWT), Variational Mode Decomposition (VMD), and Intrinsic Time-Scale Decomposition (ITD). The resulting components were combined with Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Gated Recurrent Units (GRU), and Long Short-Term Memory (LSTM) architectures to develop sixteen hybrid models. Model training was conducted using a chronological data partitioning strategy, while future projections were exclusively employed for performance evaluation. The results indicate that the TQWT-GRU model consistently outperforms other approaches at short-term (1-month) and long-term (6-month) time scales, whereas the MODWT-RNN model exhibits superior performance at the medium-term (3-month) scale across all stations and scenarios. For these optimal models, the coefficients of determination (R2) and Nash-Sutcliffe efficiency (NSE) values predominantly exceed 0.99, while the mean absolute error (MAE) and root mean square error (RMSE) remain within a low-error range of approximately 0.01-0.05. As a result, the findings underscore the necessity of scale- and region-specific hybrid model selection in drought forecasting and provide a robust framework for climate risk assessment and sustainable water resources management.