Structural Observability Determines When UAV-Based Structure-from-Motion Approximates LiDAR in Forests

Abstract. Accurate three-dimensional characterization of forest structure is essential for ecological monitoring and operational forest inventories, not least because vegetation structure fundamentally shapes remotely sensed observations. While UAV-based laser scanning (ULS) enables detailed reconstruction of canopy and terrain structure, Structure-from-Motion (SfM) photogrammetry offers a substantially lower-cost alternative using consumer-grade imaging systems. However, the performance of SfM in forests is strongly influenced by canopy occlusion, observation geometry, and forest structural complexity. Here, we evaluate under which conditions optimized UAV-based SfM produces structural representations comparable to ULS and investigate the role of structural observability in explaining similarities and differences between both sensing approaches. We compared ULS with two image-based SfM acquisition strategies: (i) a standard nadir-oriented mapping workflow and (ii) a multi-angle acquisition employing oblique viewing geometries designed to improve canopy observability. Data were acquired during leaf-off and leaf-on conditions across three contrasting temperate forest plots in Switzerland representing deciduous-dominated, mixed, and conifer-dominated stand structures. Datasets were compared visually and quantitatively using metrics characterizing canopy height, cover, and vertical structural complexity, together with assessments of ground visibility and individual-tree detection performance. The multi-angle SfM configuration consistently outperformed the standard nadir-oriented workflow and substantially improved the representation of canopy structural detail, the completeness of canopy surfaces, and vertical structure. Several structural metrics showed improvements approaching a factor of two relative to the standard SfM acquisition. Under favourable conditions, optimized SfM closely reproduced LiDAR-derived canopy height and plot-scale vertical canopy profiles and occasionally matched or exceeded LiDAR point frequencies within externally visible canopy layers. Nevertheless, important limitations remained. SfM exhibited strongly reduced ground visibility beneath closed canopies and incompletely represented lower canopy layers, sub-canopy structures, and fine woody elements, particularly in structurally complex and fully foliated stands. Individual-tree detection performance also decreased markedly in dense and conifer-dominated forests, whereas ULS remained substantially more robust across conditions. Overall, our results indicate that the principal limitation of UAV-based SfM in forests is not the geometric accuracy of reconstructed surfaces but the observability of structural elements that remain partially or completely occluded from the camera network. Observation geometry emerged as a key factor controlling this observability and therefore the degree of agreement between SfM and ULS. Optimized UAV-based SfM can provide a cost-effective alternative for plot-scale characterization of canopy structure when high-quality terrain information is available and the structural elements of interest are sufficiently observable from the camera network.