Chen et al. (2025) An Interpretable Attention Decision Forest Model for Surface Soil Moisture Retrieval

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-10-17
• Authors: Jianhui Chen, Zuo Wang, Ziran Wei, Chang Huang, Yongtao Yang, Ping Wei, Hu Li, Yuanhong You, Shuoqi Zhang, Zhijie Dong, Hao Liu
• DOI: 10.3390/rs17203468

Research Groups

The specific research groups, labs, or departments involved are not explicitly stated in the provided text.

Short Summary

This study developed an Attention Decision Forest (ADF) model to integrate interpretability and generalization for surface soil moisture (SSM) retrieval. ADF demonstrated superior performance compared to traditional models and produced high-quality large-scale SSM maps while maintaining interpretability comparable to tree-based ensemble methods.

Objective

• To develop a novel model architecture, the Attention Decision Forest (ADF), that effectively integrates prediction accuracy, generalization, and interpretability for surface soil moisture (SSM) retrieval.

Study Configuration

• Spatial Scale: Conterminous United States (CONUS) watershed dataset; large scales.
• Temporal Scale: Not explicitly defined, but the study involved spatiotemporal independent testing and considered spatiotemporal variation as a key factor influencing SSM.

Methodology and Data

• Models used: Attention Decision Forest (ADF) comprising a feature extractor, soft decision tree, and tree-attention module. Performance was compared against traditional models, Deep Neural Networks (DNN), Random Forest (RF), and XGBoost. Interpretability was evaluated using the Shapley Additive Interpretative Model (SHAP). Spatial SSM maps were compared with GSSM, SMAP L4, and ERA5-Land.
• Data sources: Conterminous United States (CONUS) watershed dataset.

Main Results

• ADF achieved an R² of 0.868 and an unbiased Root Mean Square Error (ubRMSE) of 0.041 m³/m³ on sample-based validation, outperforming traditional models.
• Robust performance was demonstrated in spatiotemporal independent testing, with R² values of 0.643 and 0.673, and ubRMSE values of 0.062 m³/m³ and 0.065 m³/m³.
• ADF's interpretability was more stable than deep learning methods (e.g., DNN) and comparable to tree-based ensemble learning methods (e.g., RF, XGBoost).
• Both ADF and ensemble learning methods indicated that spatiotemporal variation had the greatest impact on SSM at large scales, followed by environmental conditions and soil properties.
• ADF produced superior spatial SSM maps compared to GSSM, SMAP L4, and ERA5-Land.

Contributions

• Development of the Attention Decision Forest (ADF), a novel model architecture that successfully integrates high prediction accuracy, strong generalization capability, and robust interpretability for surface soil moisture retrieval.
• Demonstration of ADF's superior performance in large-scale surface soil moisture mapping and its ability to provide stable interpretability, addressing a critical gap in existing retrieval methods.

Funding

No funding information was provided in the paper text.

Citation

@article{Chen2025Interpretable,
  author = {Chen, Jianhui and Wang, Zuo and Wei, Ziran and Huang, Chang and Yang, Yongtao and Wei, Ping and Li, Hu and You, Yuanhong and Zhang, Shuoqi and Dong, Zhijie and Liu, Hao},
  title = {An Interpretable Attention Decision Forest Model for Surface Soil Moisture Retrieval},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17203468},
  url = {https://doi.org/10.3390/rs17203468}
}