A geometry-aligned and elevation-guided U-Net for canopy mapping of Populus euphratica trees using UAV LiDAR and multispectral imagery

Abstract

ABSTRACT Accurate plot-scale canopy mapping of Populus euphratica trees in desert riparian forests remains challenging because highly reflective bare soil, mixed shrubs and fragmented crowns often cause severe spectral confusion. In addition, residual local inconsistencies between uncrewed aerial vehicle (UAV) multispectral imagery and the light detection and ranging (LiDAR)-derived canopy height model (CHM) may persist after preprocessing, which may introduce boundary artefacts and unstable multimodal predictions. To address these issues, this study constructed a five-channel dataset by combining four-band UAV multispectral imagery with CHM data from representative plots in a P. euphratica riparian forest in the middle reaches of the Tarim River. We developed a geometry-aligned and elevation-guided U-Net (GA-EUNet) for pixel-level canopy mapping. The model incorporates a geometric alignment module (GAM) to reduce the influence of residual local cross-modal inconsistency at the feature level, and an elevation-guided attention module (EGA) to reweight spectral features using CHM-derived structural cues. Under a fixed plot-level data partition and training protocol, GA-EUNet achieved the highest validation-plot Intersection over Union (IoU) and F1-score among the compared models, with an IoU of 0.911 and an F1-score of 0.953. In the independent test plot, GA-EUNet also maintained higher F1-scores than the compared models in the 0–20% canopy-closure intervals, where fragmented crowns and background interference were prominent. These results support the potential of multimodal spectral – structural fusion for UAV plot-scale canopy mapping of P. euphratica in heterogeneous desert riparian scenes, while broader cross-site and cross-season validation remains necessary before operational regional application.