Transforming Leaf-Off LiDAR to leaf-on canopy height models using deep learning

Accurate tree canopy metrics are essential for a wide range of ecological, environmental, and climate studies. Recent advances in airborne LiDAR sensing have greatly expanded access to high-resolution three-dimensional observations of tree canopy structure. However, many widely available airborne LiDAR datasets are acquired under leaf-off conditions to support terrain mapping, and our analysis across ecologically diverse sites shows that canopy height models (CHMs) derived from these data exhibit substantial, recurring discrepancies in canopy height and extent, with an overall tendency to underrepresent canopy structure relative to leaf-on conditions. These discrepancies limit the direct use of leaf-off CHMs in applications requiring canopy information representative of leaf-on conditions. To address this limitation, we introduce a deep learning framework for reconstructing leaf-on CHMs from existing leaf-off LiDAR-derived CHMs. To our knowledge, this is the first study to formulate the direct reconstruction of leaf-on CHMs from existing leaf-off LiDAR-derived CHMs as a supervised image-to-image learning task. The framework provides two complementary models: a Pix2Pix model that uses the leaf-off CHM alone and a U-Net model that additionally incorporates multispectral imagery. Our proposed models achieve robust canopy-height accuracy across four spatially independent test sites (R²: 0.85–0.92; Mean Absolute Error: 1.3–1.6 metres), substantially reducing the discrepancies observed in the leaf-off source CHMs. They also provide sharper delineation of canopy details and higher accuracy than recent optical-imagery-based approaches. To facilitate broader use, we developed an open-access web application that deploys the pre-trained models and supports on-demand leaf-on CHM generation for user-selected areas across the U.S. where suitable leaf-off LiDAR inputs are available. Overall, the framework provides a practical means to bridge the seasonal data gap in LiDAR acquisition and improve access to reliable canopy metrics representative of leaf-on conditions.