A Deep Learning Perspective on Meteorological Droughts in the Mun River Basin

Accurate drought prediction is crucial for enhancing resilience and managing water resources. Developing robust forecasting models and understanding the variables influencing their outcomes are essential. This study developed models that integrate wavelet transformation (WT) with advanced AI models, increasing prediction accuracy. This study investigates the prediction of meteorological droughts using standalone bootstrapped random forest (BRF) and bi-directional long short-term memory (Bi-LSTM) models, compared to wavelet-decomposed hybrid models (WBRF, WBi-LSTM). These models were evaluated in the Mun River Basin, Thailand, utilizing monthly meteorological data (1993-2022) from the Thai Meteorological Department. The predictions were assessed using statistical metrics (R², MAE, RMSE, and MAPE). For the Standardized Precipitation Index (SPI), the hybrid WBRF model consistently outperformed the standalone BRF across various metrics and timescales, demonstrating higher R² (0.89 to 0.97 for SPI-3) and lower error metrics (MAE: 0.144 to 0.21 for SPI-6, RMSE: 0.2 to 0.3 for SPI-12). Similarly, the hybrid WBi-LSTM model outperformed the standalone Bi-LSTM in SPI predictions, exhibiting higher R² (0.87 to 0.91 for SPI-3) and lower error metrics (MAE: 0.19 to 0.23 for SPI-6, RMSE: 0.27 to 0.81 for SPI-12) across all timescales. This trend was also observed for the China Z-index (CZI), Modified China Z-index (MCZI), Hutchinson Drought Severity Index (HDSI), and Rainfall Anomaly Index (RAI), where hybrid models achieved superior performance compared to standalone models. The WBiLSTM model emerged as the preferred choice across different timespans. The integration of WT enhanced the predictive accuracy of hybrid models, making them effective tools for drought prediction.