Integrating PolInSAR and GEDI data with machine learning for forest canopy height predicting in Pongara National Park, Gabon

Accurate monitoring of forest canopy height (FCH) is highly important to gain a proper understanding of ecosystem dynamics, biodiversity, and carbon sequestration processes. This study presents an effective approach that integrates Polarimetric Synthetic Aperture Radar Interferometry (PolInSAR) with Global Ecosystem Dynamics Investigation (GEDI) data for enabling accurate large-scale forest structure mapping. This approach aims at predicting canopy height by making use of machine learning (ML) algorithms in Pongara National Park, Gabon. For canopy height modelling, the relative height at 100% energy return (RH100) from GEDI is used as the primary reference metric, and the three regression models, namely Random Forest (RF), Classification and Regression Tree (CART), and Gradient Tree Boost (GTB), are applied using features derived from PolInSAR data as input. Furthermore, this research incorporates the Random Volume over Ground (RVoG) inversion model, which is a physics-based model using the PolInSAR data to derive forest height estimates, enabling a comparison with ML algorithms as a data-driven approach. Feature importance analysis revealed that DEM height, incidence angle, and backscattered HV power are the most influential predictors in all models. The research findings exhibited that the best training performance was for CART (RMSE = 3.011 m, R2 = 0.963), while GTB demonstrated superior generalisation during validation (RMSE = 7.713 m, R2 = 0.757), suggesting it as the most robust model. Overall, RF was very competitive, with close conformity to GTB in terms of bias, accuracy, and correlation. In comparison, the RVoG model laid out a strong correlation with the RF model (up to 0.91) and tended to estimate higher canopy heights than other ML results, reflecting its sensitivity to vertical forest structure.