A Point Cloud Filtering Method for Reservoir Bank Slopes Integrating Scene Classification and Ridge Detection

Reservoir bank slopes are characterized by pronounced terrain relief and dense vegetation, which lead to severe mixing of ground and non-ground points in LiDAR point clouds and pose significant challenges to accurate ground filtering and terrain reconstruction. Traditional filtering methods based on uniform thresholds often fail to balance filtering accuracy and terrain structure preservation under complex and heterogeneous terrain conditions. This study proposes a LiDAR point cloud filtering method that integrates scene classification and ridge detection. A depth image interpolated from the point cloud is used to derive directional gradients and multi-dimensional terrain descriptors, enabling automatic classification of the study area into gentle and mountainous regions. A feature-triangle-based region growing approach is developed to extract ridge lines in mountainous areas, which are incorporated as structural constraints to preserve critical terrain features during filtering. Within a progressive TIN densification framework, scene-adaptive parameter strategies are applied to iteratively extract ground points and update the TIN. Experiments conducted on UAV-borne LiDAR data from the Cheyiping reservoir bank slope along the Lancang River show that the proposed method achieves ground point misclassification rates of 0.64% and 3.42% in gentle and mountainous regions, respectively, demonstrating high consistency with manually interpreted reference data. Compared with conventional methods, the proposed approach effectively suppresses top-clipping and excessive slope smoothing, while significantly improving terrain structure preservation in complex environments.