Transforming Leaf-Off Lidar to Leaf-On Canopy Height Models Using Deep Learning

Abstract

Accurate tree canopy metrics are essential for various ecological, environmental, and climate studies. Recent advancements in airborne LiDAR sensing technologies have revolutionized the acquisition of three-dimensional Earth observations. However, extensive validation across diverse ecological sites reveals that conventional methods significantly underestimate canopy height and coverage, underscoring the need for more accurate leaf-on canopy height model generation approaches. To overcome this limitation, we introduce a deep learning framework leveraging Pix2Pix GAN and U-Net architectures to transform leaf-off LiDAR data into accurate leaf-on canopy height models. Our proposed models substantially enhance canopy height estimation, achieving robust performance (R²: 0.85–0.92; Mean Absolute Error: 1.3–1.6 meters), clearly outperforming conventional approaches. Additionally, our method provides superior delineation of canopy details and consistently higher accuracy compared to recent optical imagery-based approaches. To facilitate widespread adoption, we developed an open-access web application that deploys our pre-trained models, enabling users to generate leaf-on canopy height models directly from leaf-off LiDAR datasets across the U.S. Our approach effectively bridges the seasonal data gap in LiDAR data collection, enriching the ecological and environmental research community by providing accurate, reliable canopy metrics during critical vegetative periods.

Keywords: Canopy Height Model, Airborne LiDAR, Leaf-off to Leaf-on Transformation, Deep Learning, Forest Structure Mapping