Integrating CNN With PolInSAR Multidimensional Backscattering for Enhanced Forest Height Retrieval

The operational deployment of P-band Synthetic Aperture Radar (SAR) constellations (e.g., ESA’s BIOMASS) has established low-frequency Polarimetric Interferometric SAR (PolInSAR) as a pivotal technology for retrieving vertically resolved forest parameters such as forest height and biomass, enabling precise quantification of terrestrial carbon stocks. Nevertheless, the conventional Random Volume over Ground (RVoG) model, which postulates an exponential decay of backscattering profiles, introduces systematic biases in low-frequency penetration scenarios. Under the assumption of two-layer vegetation structure, an adaptive inversion method based on multi-dimensional backscattering models is proposed to mitigate the backscattering error caused by diverse vertical structure. The proposed method establishes a scalable vertical structure modeling framework through parametric characterization of stabilized backscatter profiles. Employing a pixel-wise Convolutional Neural Network (CNN) classification architecture, it dynamically optimizes backscattering profile selection while integrating Shapley Additive Explanation (SHAP) interpretability modules to decipher feature contribution mechanisms across multi-dimensional parameter spaces. The validation based on BioSAR-2008 and AfriSAR-2016 airborne P-band single baseline data, demonstrates that the multi-model joint strategy significantly enhances the inversion accuracy for distinct forest types compared to the best-performing single model. Among these results, the Root Mean Square Error (RMSE) in coniferous forest areas is reduced from 3.57 m (for the new model) to 3.07 m (a relative improvement of 14%), while in tropical rainforest areas it is reduced from 6.08 m (for the RVoG model) to 3.70 m (a relative improvement of 39%). Subsequent SHAP feature attribution analysis further revealed that the 33 PolInSAR features at the tropical rainforest site were more closely associated with backscatter, and the complex nonlinear relationship between the two was better captured. The proposed method has the capacity to effectively compensate for estimation biases caused by model mismatch, realize the complementary advantages of multiple vertical structure profiles, and provide a new technical path for forest parameter inversion.