High-precision airborne LiDAR remains essential for urban forestry: Revealing the limitations of recent large-scale canopy height products in urban contexts

Abstract

Urban trees play a crucial role in delivering ecosystem services, yet their structural complexity poses significant challenges for remote sensing. Although recent large-scale, open-access canopy height models (CHMs) offer potential alternatives to airborne LiDAR, their suitability for urban environments remains uncertain. This study systematically evaluated four prominent CHMs (Potapov, Lang, Tolan, and Malambo) using high-resolution airborne LiDAR data from Washington, D.C., with complementary analyses from four additional cities. We assessed their performance in both canopy classification and height prediction using classification and regression metrics, spatial autocorrelation, consistency tests, and explainable machine learning. Results revealed consistent limitations across products, including widespread misclassification of canopy, systematic tree height prediction biases — characterized by overestimation of low and underestimation of high canopies (the OLUH effect) — and pronounced spatial clustering of errors along urban–forest edges. Among the models, only the Lang CHM passed the Bland–Altman consistency test, showing marginal statistical agreement with reference data. Tree characteristic variables, especially canopy height itself, emerged as dominant drivers of height errors, while topography and built-up context also contributed. Consistent patterns observed across four additional cities indicated that these limitations are systemic rather than location-specific. We conclude that high-precision airborne LiDAR remains essential for urban forestry and recommend enhancing canopy height mapping techniques to better capture the structure of urban trees. A promising direction is the development of urban-specific CHMs with finer spatial and temporal resolution, improved temporal consistency, and integration with high-resolution imagery, contextual deep learning models, and local calibration strategies.