Deep learning-driven regional drought assessment: an optimized perspective

Climate change has become a prominent concern in recent years, with extensive research revealing a range of adverse impacts linked to ongoing global warming. The non-linear dynamics of precipitation represent a significant implication, often resulting in region-specific droughts and flooding events. Droughts, in particular, lead to numerous negative consequences. The drought index quantifies drought characteristics. Recognizing the unique attributes of each region, there is a need to adopt a suitable drought index for accurate regional drought analysis. This study compared multiple time scale drought indices-SPI, EDI, and MCZI-from 1980 to 2020, determining the 6-month SPI as the most consistent and justifiable index. To enhance drought forecasting using the identified SPI-6 index, deep learning architectures such as RNN, GRU, and LSTM were employed. Model performance was evaluated using metrics including correlation (r), root-mean-square error (RMSE), mean absolute error (MAE), and Nash-Sutcliffe efficiency (E). Notably, LSTM demonstrated superior performance, exhibiting the lowest RMSE (0.61) and MAE (0.4). Among the three models (LSTM, RNN, and GRU), LSTM achieved the highest correlation coefficient (r) of 0.85 during testing and validation phases. LSTM outperformed RNN and GRU across all datasets, exhibiting lower RMSE and MAE values, indicating higher prediction accuracy with fewer errors. These findings suggest that combining 6-month SPI with LSTM can be employed as a new, reliable integrated approach. Additionally, this study examined drought occurrences between 1980 and 2020 with a specified 6-month SPI.