GEDI-Guided Deep Learning for Forest Canopy Height Completion Using Sentinel-2 Imagery, Sparse LiDAR Observations, and Gradient-Based Reconstruction

Abstract

Accurate Forest canopy height estimation is essential for biomass assessment, carbon stock estimation, biodiversity monitoring, and forest management. Spaceborne LiDAR missions such as the Global Ecosystem Dynamics Investigation (GEDI) provide valuable measurements of forest vertical structure; however, their sparse spatial sampling limits direct wall-to-wall canopy height mapping. Most existing deep learning approaches use GEDI-derived height metrics mainly as supervisory labels during training, without fully exploiting available GEDI observations as structural guidance during inference. To address this limitation, this paper proposes GEDI-Guided Sentinel Network (GSen-Net), a deep learning framework for dense canopy height completion from Sentinel-2 multispectral imagery and sparse GEDI RH98 observations. The proposed framework reformulates canopy height estimation as a gradient-based completion problem, in which spatial height gradients are predicted and integrated with sparse LiDAR observations through a multi-resolution differentiable reconstruction module. GSen-Net adopts an attention-enhanced U-Net architecture to fuse spectral and structural features and capture both local canopy patterns and long-range spatial dependencies. Experiments over ecologically diverse regions in Vietnam and India demonstrate the effectiveness of the proposed approach. Incorporating GEDI observations as an input modality reduces mean absolute error from 7.09 m to 5.71 m in Vietnam and from 6.83 m to 4.77 m in India. Ablation and comparative analyses further show that multi-modal fusion, gradient-based reconstruction, and attention-enhanced feature extraction jointly improve canopy height estimation.