Deep learning-based vegetation canopy height mapping with polarimetric SAR: Application of a Polarization Fusion U-Net in Gabon’s tropical forests

Forests provide essential ecosystem services, including carbon storage and the conservation of biodiversity. This highlights the need for accurate and scalable methods to assess forest structure. Active remote sensing, particularly Synthetic Aperture Radar (SAR), has significant potential for estimating forest structure owing to its ability to penetrate vegetation layers and interact with different forest elements. In particular, L- and P-band SAR signals are able to penetrate canopy layers, making them suitable for retrieving forest structure information.
We present a novel approach utilizing full-polarimetric SAR backscatter data to estimate canopy height as relevant forest structure variable. To capture the non-linear complex relationship between the SAR data and the vegetation canopy height, we propose a Polarization Fusion U-Net (PF-Unet) designed to enhance canopy height estimation from SAR backscatter data by effectively utilizing multi-polarization channels (e.g., HH, HV, and VV). Specifically, the proposed model is tailored to the physical properties of the SAR backscatter, incorporating: (a) a polarization fusion layer, (b) attention gates applied to each layer in the decoder blocks, and (c) Exponential Linear Unit (ELU) activation and Huber loss functions. To assess the potential of the PF-Unet model, L- and P-band SAR data, collected over the tropical forests of Gabon, were separately used. The model was evaluated in a complex tropical forest environment and compared against traditional Machine Learning (ML) approaches (Random Forest (RF) and Light Gradient Boosting Machine (LGBM)) as well as the standard U-Net model. The PF-Unet consistently outperformed all the baselines for both SAR datasets. The PF-Unet model achieved an RMSE of 4.35 m (15.73%) for L-band and 4.43 m (15.95%) for P-band, which was an improvement over the U-Net model’s RMSE of 5.02 m (18.15%) and 4.59 m (16.53%), respectively. This suggests promising potential for enhanced canopy height estimation, which is particularly valuable for upcoming spaceborne missions like NISAR and BIOMASS.