Li et al. (2026) Physical process-based attention encoder-decoder LSTM model to improve global soil moisture prediction

Identification

• Journal: Agricultural and Forest Meteorology
• Year: 2026
• Date: 2026-04-02
• Authors: Qingliang Li, Jian Hong, Cheng Zhang, Wei Shangguan, Zhongwang Wei, Lu Li, Wenzong Dong, Jinlong Zhu, Xiao Chen, Yuguang Yan, Fanhua Yu, Yongjiu Dai
• DOI: 10.1016/j.agrformet.2026.111161

Research Groups

• College of Computer Science and Technology, Changchun Normal University, Changchun, China
• Research Institute for Scientific and Technological Innovation, Changchun Normal University, Changchun, China
• Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Guangdong Province Key Laboratory for Climate Change and Natural Disaster Studies, School of Atmospheric Sciences, Sun Yat-sen University, Guangzhou, China

Short Summary

This study introduces the AEDLSTM-HBV model, which integrates physical features from the Hydrologiska Byråns Vattenbalansavdelning (HBV) model into an Attention-Enhanced Encoder-Decoder Long Short-Term Memory (AEDLSTM) network to improve global soil moisture prediction. The model significantly outperforms state-of-the-art methods, particularly in permafrost and desert regions, by effectively leveraging the fusion of physical and deep learning features.

Objective

• To address the limitations of deep learning models in integrating relevant physical features for soil moisture prediction by proposing the AEDLSTM-HBV model, which fuses physical features from the HBV model with an attention-enhanced encoder-decoder LSTM network to achieve more accurate and reliable global soil moisture predictions.

Study Configuration

• Spatial Scale: Global
• Temporal Scale: Prediction of soil moisture (specific horizon not detailed, but implies time-series forecasting).

Methodology and Data

• Models used: AEDLSTM-HBV (Attention-Enhanced Encoder-Decoder Long Short-Term Memory with Hydrologiska Byråns Vattenbalansavdelning features), HBV (Hydrologiska Byråns Vattenbalansavdelning) model features.
• Data sources: LandBench1.0 dataset.

Main Results

• The proposed AEDLSTM-HBV model surpasses state-of-the-art models in global soil moisture prediction.
• The model demonstrates robust performance in predicting soil moisture in permafrost and desert regions.
• Achieved an average R² improvement of up to 20% over the baseline AEDLSTM model.

Contributions

• Introduction of a novel deep learning architecture, AEDLSTM-HBV, that effectively integrates physical process features from the HBV model into an attention-enhanced encoder-decoder LSTM network.
• Demonstration of the significant potential of fusing physical model features with deep learning to enhance the predictive capabilities for global soil moisture.
• Improved accuracy and reliability of soil moisture predictions, particularly in challenging environments such as permafrost and desert regions.

Funding

• Not specified in the provided text.

Citation

@article{Li2026Physical,
  author = {Li, Qingliang and Hong, Jian and Zhang, Cheng and Shangguan, Wei and Wei, Zhongwang and Li, Lu and Dong, Wenzong and Zhu, Jinlong and Chen, Xiao and Yan, Yuguang and Yu, Fanhua and Dai, Yongjiu},
  title = {Physical process-based attention encoder-decoder LSTM model to improve global soil moisture prediction},
  journal = {Agricultural and Forest Meteorology},
  year = {2026},
  doi = {10.1016/j.agrformet.2026.111161},
  url = {https://doi.org/10.1016/j.agrformet.2026.111161}
}