Objective Typhoon-induced disturbances severely disrupt forest canopy structure, degrade ecosystem services, and diminish carbon sequestration capacity. There is an urgent need for efficient and accurate methods to assess typhoon-related forest damage. Current approaches face two major challenges: (1) the inability to characterize spatial heterogeneity and gradient-based canopy responses at the pixel scale under typhoon stress, and (2) technical limitations in inversion accuracy and spatial continuity caused by scale/noise heterogeneity in multi-source remote sensing data and detail loss in optical imagery. To address these issues, this study aims to develop a pixel-level inversion framework that integrates multi-source remote sensing data for fine-scale detection of canopy height changes before and after typhoons, and to investigate how elevation gradients and urban boundaries modulate the spatial patterns of canopy loss.

Method We propose an enhanced U-Net architecture—Hierarchical Wavelet-enhanced and Contextual Feature-integrated U-Net (HWCFU-Net). The model incorporates a hierarchical feature enhancement module based on discrete wavelet transform to strengthen both high- and low-frequency information representation, thereby mitigating multi-source data heterogeneity. It further introduces a hierarchical contextual feature integration unit that employs multi-order depthwise separable convolutions to optimize multi-scale feature transmission and selection. A pixel-wise regression strategy is adopted to independently model each pixel and directly predict continuous canopy height values, overcoming the limitations of traditional whole-image single-label or zonal averaging approaches. We integrated GEDI and ICESat-2 LiDAR data with Sentinel-1/2 optical–radar observations to construct a multi-source spatiotemporal feature set. Six pre- and post-disaster experimental scenarios were established using three representative typhoons in 2019: Lekima, Kammuri, and Hagibis. The model was systematically compared against six state-of-the-art methods: U-Net, U-Net++, AttentionRes-UNet, TSNN, Y-NET, and Random Forest.

Result HWCFU-Net consistently outperformed all benchmarks across all experimental scenarios, achieving coefficients of determination (R²) ranging from 0.62 to 0.71 and root mean square errors (RMSE) between 3.98 and 4.87 m. Compared to deep learning baselines, it improved R² by 0.01–0.10 and reduced RMSE by 0.12–1.16 m. Against Random Forest, R² increased by 0.01–0.09 and RMSE decreased by 0.13–1.03 m. The highest accuracy was achieved in the pre-typhoon Lekima scenario (R² = 0.71, RMSE = 3.98 m), demonstrating the model’s robustness and generalization capability. Spatial analysis revealed heterogeneous canopy loss patterns: low-elevation broadleaf forests suffered the greatest damage due to shallow root systems and low wind resistance, while mid- to high-elevation coniferous forests experienced relatively minor losses. Forests near urban areas endured stronger wind shear induced by surface roughness and building-induced turbulence or canyon effects, with damage intensity significantly attenuating with distance from cities. Short-term increases in local vegetation indices suggested compensatory growth triggered by post-typhoon rainfall and improved moisture conditions. Elevation gradients significantly shaped loss distribution by influencing forest composition and structural stability, while urban boundaries amplified typhoon wind fields.

Conclusion The study demonstrates that elevation gradient effects and urban boundary effects jointly govern the spatial heterogeneity of forest canopy loss. The proposed pixel-level inversion method effectively addresses data heterogeneity and detail loss, substantially improving disaster assessment accuracy. It provides a reliable theoretical foundation and technical support for forest disaster mitigation and ecological adaptation planning.