PG-GVB: A prior constrained backscatter framework for P-band PolInSAR forest height inversion

Facilitated by the BIOMASS mission, P-band polarimetric SAR interferometry (PolInSAR) provides a key capability for large-scale forest height mapping and biomass estimation. Within this context, the random volume over ground (RVoG) model stands out for its physical consistency and broad applicability. However, at P-band, deviations of canopy backscattering from the exponential extinction assumption in RVoG lead to biased height estimates. While gaussian vertical backscatter (GVB) offers greater flexibility, its surplus parameters and unconstrained convergence raise robustness concerns, resulting in estimation biases that limit its widespread application. To address these challenges, a convolutional neural network (CNN) is used to link structure parameters to SAR observations, thereby imposing prior constraints. A parameter-guided GVB (PG-GVB) variant is developed within a three-stage framework, achieving improved performance while explicitly analyzing how different parameters influence robustness. Validation on BioSAR-2008 and AfriSAR-2016 shows that the best-constrained configuration reduces forest height errors by 20.5% (4.83 m to 3.84 m) and 26.5% (7.49 m to 5.50 m) relative to the exponential model at Krycklan and Lopé, respectively. These results reveal transferable optimization strategies, with shapley additive explanations (SHAP) analysis further quantifying feature contributions.