Machine learning integration for robust real-time drought monitoring and early warning systems

This study proposes a KNN–LSTM fusion framework for long-term meteorological drought forecasting using SPI at 6- and 12-month scales. Monthly precipitation data (1971–2022) from six stations in Ankara Province were used to compute SPI, with lagged features selected via ACF/PACF. A KNN module retrieved historical analogs, while the LSTM modeled nonlinear temporal dependencies. Model outputs were combined using a Gradient Boosting Regressor as a meta-learner to produce the final hybrid forecasts. The performance of the model was assessed using the root mean square error, the mean absolute error, the symmetric mean absolute percentage error and the Nash-Sutcliffe efficiency. Across all stations and both SPI timescales, the hybrid framework consistently outperformed standalone models, reducing RMSE by over 60% and achieving NSE values near or above 0.95, indicating excellent agreement with observations. Furthermore, the Diebold-Mariano test was employed to formally compare forecast accuracy. The hybrid model significantly outperformed KNN and LSTM in prediction accuracy (Formula: see text), providing strong evidence of the robustness and effectiveness of the proposed ensemble approach. These results show that combining analog-based retrieval with deep sequential learning provides a reliable and generalizable framework for operational drought prediction, effectively capturing long-term persistence in semi-arid regions.