Forest Canopy Height Retrieval from WorldView-3 Stereo Imagery Using Learned Feature Matching and Local Geometric Rectification

Forest canopy height is a critical parameter for monitoring forest structure, biomass, and ecosystem condition. High-resolution satellite stereo imagery provides a practical source for canopy-height retrieval over areas where repeated field or airborne LiDAR surveys are difficult, but dense forest scenes remain challenging because tree crowns often contain repeated, weak, and view-dependent textures. This study proposes a canopy-height extraction workflow for WorldView-3 stereo imagery that combines DISK feature extraction, LightGlue feature matching, local epipolar rectification using an affine camera approximation, semi-global matching, and sub-pixel refinement. The resulting disparity maps are triangulated to generate digital surface models and are evaluated against airborne LiDAR-derived canopy heights over a dense pine forest plot in Jacksonville, Florida, USA. Across ten stereo-pair combinations, the proposed DISK+LightGlue workflow achieved a mean RMSE of 2.28 m, MAE of 1.46 m, and Pearson correlation coefficient of 0.78, compared with 3.06 m, 2.16 m, and 0.68 for the SIFT baseline. The average RMSE and MAE reductions were 25.5% and 32.3%, respectively. Additional baseline and ablation experiments indicate that learned feature matching, local geometric rectification, and sub-pixel refinement each contribute to the final height accuracy. The results also show that stereo-pair convergence angle strongly affects canopy-height reconstruction, with larger improvements over SIFT for pairs with convergence angles of at least 20 degrees. These findings suggest that learned correspondence and geometry-aware local rectification can improve canopy-height retrieval from high-resolution satellite stereo imagery when stereo-pair geometry is explicitly considered.