RGB-Derived Canopy Height Models for Riparian Woody Vegetation Monitoring Using Depth Anything V2

Abstract

Rivers and riparian zones support a variety of woody plant species and play an important role in riverine ecosystems and fluvial processes. Rapid establishment of herbaceous and woody vegetation occurs in unvegetated river channels, mainly because of dam construction and hydrological alterations. However, repeated monitoring of the three-dimensional structure over broad river corridors remains challenging. We aimed to assess the applicability of RGB-derived canopy-height models (CHMs) inferred using Depth Anything V2 for monitoring riparian woody vegetation. A monocular depth-estimation framework was trained using the National Agriculture Imagery Program-CHM dataset and applied to three Korean riverine environments. The inferred CHMs were evaluated against the LiDAR-derived CHMs and further tested using two application-oriented assessments: individual tree detection (ITD) and woody vegetation area classification. The RGB-derived CHMs reproduced the overall spatial patterns of the riparian canopy height, although the accuracy varied with vegetation structure and image acquisition conditions. The mean absolute error between RGB- and LiDAR-derived CHMs was approximately 0.96 m across the study sites. In the ITD assessment, the RGB-derived CHM achieved an F1 score of 0.73, compared with 0.80 for the LiDAR-derived CHM. For woody vegetation area classification, the RGB-derived CHM showed high pixel-level accuracy, while Dice coefficient and IoU varied depending on the canopy-height threshold and study site. These results suggest that the RGB-derived CHMs can serve as supplementary data for monitoring riparian woody vegetation when LiDAR acquisition is limited.