Geometry-based point cloud fusion of dual-layer UAV photogrammetry and a modified unsupervised generative adversarial network for 3D tree reconstruction in semi-arid forests

Abstract

We present the first application of geometry-based relationship constraints for point-cloud registration and unsupervised 3D reconstruction of tree structure in semi-arid forest using unmanned aerial vehicle (UAV) photogrammetry. Accurate three-dimensional (3D) reconstruction of tree structure is essential for a plethora of subsequent tasks like assessing ecosystem health and informing sustainable forest management strategies, in particular over ecologically sensitive arid and semi-arid ecosystems that increasingly face decline due to prevalence of environmental stressors. This highlights the need for high-resolution geospatial monitoring approaches. While UAV-based photogrammetry offers a flexible and cost-effective means of capturing forest structure, conventional top-of-canopy imaging fails to sufficiently represent critical under-canopy features, including stem morphology and lower crown structure. Here, we suggest an integrated 3D reconstruction framework that combines dual-layer UAV photogrammetry, acquiring data from both above and below the canopy, with an innovative geometry-based point cloud registration method. Unlike conventional approaches like Iterative Closest Point (ICP) and Random Sample Consensus (RANSAC), this method leverages spatial relationships among individual trees to robustly align multi-view point clouds acquired under occluded and variable conditions. To further refine the reconstructed tree models, we suggest an updated unsupervised Generative Adversarial Network (Denoise-GAN), enabling both noise reduction and structural completion without reliance on labeled training data. The resulting models were used to extract key phenotypic features with high accuracy compared to reference data (root collar diameter (DRC) R² = 0.93, height R² = 0.97,Crown area R² = 0.99, number of stems R² = 1), providing vital indicators for quantifying forest structure and health. The presented methodology not only enhances the completeness and accuracy of 3D tree reconstruction in semi-arid forest, but also represents a significant advancement toward a scalable, data-driven semi-arid forest monitoring system. This workflow offers substantial potential for ecological applications, particularly in degraded and topographically complex ecosystems.