Transformer-based Reconstruction of Canopy Profiles from Large-Footprint Waveform LiDAR

Tahrir Siddiqui , Keith Krause, Jan van Aardt

Spaceborne laser scanning (SLS) presents a cost-effective means for frequent, global-scale monitoring of forest ecosystem parameters. Compared to airborne laser scanning (ALS), SLS offers substantially greater spatial coverage and revisit frequency, but at the cost of larger footprints, sparser sampling, and attenuated return signals. These constraints typically result in a loss of fine-scale vertical canopy structure in large-footprint waveform LiDAR, thereby limiting the retrieval of ecologically meaningful forest structural metrics. To address this challenge, we developed an encoder–decoder Transformer architecture to reconstruct high-resolution vertical canopy profiles from large-footprint waveform LiDAR observations. Using waveform data acquired by NASA’s Land, Vegetation, and Ice Sensor (LVIS) – a high-altitude ALS instrument commonly used as a proxy for spaceborne missions – we trained the model to recover fine-scale canopy structure by leveraging overlapping ALS point clouds as reference data. The proposed Transformer leverages long-range vertical dependencies within waveform signals to infer canopy structural details that are degraded or unresolved in large-footprint, high-altitude observations. Results show that the proposed approach substantially improves the agreement between LVIS-derived and ALS-derived canopy structural complexity metrics, increasing correlations from R = 0.62 to 0.84 and from R = 0.76 to 0.90 for two representative metrics. This framework is readily transferable to current and future SLS missions, enabling the retrieval of super-resolved vertical canopy profiles and supporting large-area assessment of ecologically meaningful canopy structural metrics.