An Interpretable Attention Decision Forest Model for Surface Soil Moisture Retrieval

Surface soil moisture (SSM) plays a critical role in climate change, hydrological processes, and agricultural production. Decision trees and deep learning are widely applied to SSM retrieval. The former excels in interpretability while the latter outperforms in generalization, neither, however, integrates both. To address this issue, an attention decision forest (ADF) was developed, comprising feature extractor, soft decision tree, and tree-attention modules. The feature extractor projects raw inputs into a high-dimensional space to reveal nonlinear relationships. The soft decision tree preserves the advantages of tree models in nonlinear partitioning and local feature interaction. The tree-attention module integrates outputs from the soft tree’s subtrees to enhance overall fitting and generalization. Experiments on conterminous United States (CONUS) watershed dataset demonstrate that, upon sample-based validation, ADF outperforms traditional models with an R2 of 0.868 and a ubRMSE of 0.041 m3/m3. Further spatiotemporal independent testing demonstrated the robust performance of this method, with R2 of 0.643 and0.673, and ubRMSE of 0.062 and 0.065 m3/m3. Furthermore, an evaluation of the interpretability of the ADF using the Shapley Additive Interpretative Model (SHAP) revealed that the ADF was more stable than deep learning methods (e.g., DNN) and comparable to tree-based ensemble learning methods (e.g., RF and XGBoost). Both the ADF and ensemble learning methods demonstrated that, at large scales, spatiotemporal variation had the greatest impact on the SSM, followed by environmental conditions and soil properties. Moreover, the superior spatial SSM maps produced by ADF, compared with GSSM, SMAP L4 and ERA5-Land, further demonstrate ADF’s capability for large-scale mapping. ADF thus offers a novel architecture capable of integrating prediction accuracy, generalization, and interpretability.