A Bidirectional Analysis Method for Extracting Glacier Crevasses from Airborne LiDAR Point Clouds

Abstract

A crevasse is an important surface feature of a glacier. This study aims to detect crevasses from high-density airborne LiDAR point clouds. However, existing methods continue to suffer from the data holes within the crevasse region and the influence of the undulating non-crevasse glacier surfaces. Therefore, a bidirectional analysis method is proposed to robustly extract the crevasses from the point clouds, which utilizes their vertical and horizontal characteristics. First, crevasse points are separated from non-crevasse points using a hybrid-entity method, where the height difference and the nearly vertical characteristic of a crevasse sidewall are considered, to better distinguish the crevasses from the undulating non-crevasse glacier surfaces. Second, the crevasse regions/edges are adaptively delineated by a local statistical analysis method that is based on a novel feature of the Delaunay triangulation mesh of non-crevasse points in the horizontal plane. Last, the pseudo-crevasse points and regions are removed by a cross-analysis method. To test the performance of the proposed method, this study selected airborne LiDAR point clouds from two sites of Alaskan glaciers (i.e., Tyndall Glacier and Seward Glacier) as the experimental datasets. The results were verified by a comparison with the ground truth that was manually delineated. The proposed method achieved acceptable results: the recall, precision, and $F_1$ scores of both sites exceeded 94.00%. Moreover, a comparative experiment was carried out and the results confirmed that the proposed method achieved superior performance.