SenFus-CHCNet: a multi-resolution fusion framework for sparse-supervised canopy height classification

Introduction Accurate forest canopy height mapping is critical for understanding ecosystem structure, monitoring biodiversity, and supporting climate change mitigation strategies. Methods In this paper, we present SenFus-CHCNet, a novel deep learning architecture designed to produce high-resolution canopy height classification maps by fusing multispectral (Sentinel-2) and synthetic aperture radar (SAR) (Sentinel-1) imagery with GEDI LiDAR data. The proposed model comprises two main components: a Multi-source and Multi-band Fusion Module that effectively integrates data of varying spatial resolutions through resolution-aware embedding and aggregation, and a Pixel-wise Classification Module based on a customized U-Net architecture optimized for sparse supervision. To discretize continuous canopy height values, we evaluate three classification schemes—coarse, medium, and fine-grained—each balancing ecological interpretability with model learning efficiency. Results Extensive experiments conducted over complex forested landscapes in northern Vietnam demonstrate that SenFus-CHCNet outperforms state-of-the-art baselines, including both convolutional and transformer-based models, achieving up to 4.5% improvement in relaxed accuracy (RA±1) and 10% gain in F1-score. Qualitative evaluations confirm that the predicted maps preserve fine-scale structural detail and ecologically meaningful spatial patterns, even in regions with sparse GEDI coverage. Discussion Our findings highlight the effectiveness of deep fusion learning for canopy height estimation, particularly in resource-limited settings. SenFus-CHCNet provides a scalable and interpretable approach for forest monitoring at regional and national scales, with promising implications for biodiversity conservation, carbon accounting, and land-use planning.