Abstract
Abstract
The accurate forecasting of surface soil moisture (SSM) for multiple time steps is vital in the early warning systems of drying and flooding, irrigation scheduling, and hydrological modelling. The current study presents an innovative hybrid structure called ASHA–LSTM–XGB that provides 1, 3, 5, and 7-day forecasting for SSM by combining Long Short-Term Memory networks (LSTM), eXtreme Gradient Boosting (XGBoost), and the Asynchronous Successive Halving Algorithm (ASHA). The proposed model was tested and validated over Izmir Province in Western Anatolia, Türkiye, using satellite mapping of the SMAP Level 4 SSM dataset and 11 ERA5-Land environmental predictors from April 2015 to October 2025. The obtained results clearly indicate that all proposed hybrid approaches showed better performance than the traditional LSTM model for any horizon of the prediction. For the shortest term, BO-LSTM-XGB was found to be the most accurate method (NRMSE = 0.2151, R
2
= 0.8720). Meanwhile, the best combination of high prediction accuracy and stability was achieved by ASHA-LSTM-XGB, which demonstrated the best performance in the medium-term period and beyond (NRMSE = 0.2581, R
2
= 0.8153; NRMSE = 0.2666, R
2
= 0.8027 at 5- and 7-day horizons, respectively). Thus, these results demonstrate the efficiency of integrating sequential learning, non-linear ensembling, and optimal hyperparameter tuning for forecasting problems. Moreover, the results of the ablation study and analysis using SHAP values support the finding that the two most important predictor variables for SSM were the surface temperature and surface radiation variables. The integration of ASHA and LSTM-XGB has resulted in greater stability and generalization of the models. This study is considered the first application of the combined optimization-deep learning-ensemble architecture to forecast multi-step SSM. The proposed method performed better than the conventional LSTM algorithm in all forecasting scenarios and showed robust performance even at a 7-day forecasting lead time, showing promise for applications in short-range soil moisture prediction and environmental monitoring studies.