Wavelet-Kalman-LSTM fusion: enhancing meteorological drought prediction with robust feature selection and uncertainty quantification

Accurate drought forecasting is essential for minimizing socio-economic and environmental impacts but remains challenging due to climatic variability and complexity of droughts. This study presents the Wavelet-Kalman-LSTM (WKLSTM), a new hybrid deep learning framework that combines wavelet decomposition, Kalman filtering, and Long Short-Term Memory (LSTM), improving multi-timescale drought predictions (SPI3, SPI6, SPI9, SPI12) across various regions of Australia. The WKLSTM addresses non-stationarity by decomposing signals into hierarchical subseries using discrete wavelet transforms, denoising coefficients with Kalman filtering, and utilizing LSTM’s temporal modeling. Advanced feature selection techniques such as Boruta, SHAP (SHapley Additive exPlanations), and LASSO (Least Absolute Shrinkage and Selection Operator) are integrated with eXtreme Gradient Boosting (XGBoost) to identify critical lagged predictors. The WKLSTM model is assessed against both standalone and hybrid benchmark models, including LSTM, Feedforward Neural Networks (FNN), attention-based LSTM (attLSTM), Convolutional Neural Network combined with Bidirectional Gated Recurrent Unit (CNN-BiGRU), and WLSTM. The results show that the WKLSTM outperforms by achieving R 2 values from 0.51 to 0.99 and lowered RMSE by 7% to 33% compared to benchmark models. Monte Carlo simulations quantify uncertainty with 95PPU bounds covering 50–81% of observations, indicating that although WKLSTM performs well across various regions and timescales, high predictive uncertainty persists due to the nonlinear and stochastic nature of droughts. The framework integrates noise reduction, dynamic feature selection, and probabilistic forecasting, enhancing drought prediction, aiding water resource management and climate resilience strategies. This study highlights the effectiveness of WKLSTM in addressing climatic non-linearity and offers practical insights into drought mitigation.