Accurate multi-temporal terrain change detection in mountainous regions is crucial for disaster monitoring and environmental management; however, it remains challenging due to the complex topography, dense vegetation, and occlusion. This study proposes an integrated, automated framework for terrain change detection that combines 2D semantic segmentation and 3D geometric analysis, utilizing UAV imagery and point cloud data. This method uses the DeepLabV3 neural network to perform multi-class semantic segmentation of orthophotos, combining Fast Point Feature Histograms (FPFH) and Random Sample Consensus (RANSAC) algorithms, followed by Iterative Closest Point (ICP) based local refinement. Geometric changes are subsequently quantified using the Multiscale Model-to-Model Cloud Comparison (M3C2) distance metric. When applied to 0.06 km2 of mountainous terrain in the Guanziling area, Tainan City, Taiwan, between January 2024 and January 2025, the framework achieved a mean Intersection over Union (mIoU) of 87.05% in semantic segmentation and a root-mean-square error (RMSE) of 4.2 cm in geometric registration for rigid structures. The model's capacity to generalize was evaluated by validation against 100 independent, manually annotated tiles, which showed a strong coefficient of determination (R2 = 0.9251) between the predicted and actual change proportions. Comparisons in complex "Medium Change" zones showed that while 2D-only methods identified 9% of changes due to shadows and occlusion, the integrated 3D analysis successfully detected physical displacements averaging 2.12 m. Overall, the proposed 2D-3D fusion framework enables reliable terrain change detection and supports post-disaster assessment, landslide monitoring, and infrastructure management in complex mountainous environments.
Hou et al. (Mon,) studied this question.