Detection and automatic identification of loess sinkholes from the perspective of LiDAR point clouds and deep learning algorithm

Abstract

Nowadays, the detection and automatic identification of the three-dimensional structure of sinkholes is extremely lacking, which has resulted in significant gaps in sinkholes mapping, soil erosion estimation and morphological studies. In this study, we discovered 249 sinkholes on a river terrace (about 2050 m long and 100 m wide) in a small watershed of Chinese Loess Plateau. Subsequently, we used the unmanned aircraft systems (UAS) and handheld laser scanner (HLS) to investigate these loess sinkholes in detail. We introduced the PointNet ++ deep learning model to train the point cloud dataset for 50 epochs and then selected the best model. In order to evaluate the identification accuracy and transferability of the model, we input point clouds of the unknown prediction area into the trained model to predict the sinkhole point clouds. The trained model exhibits excellent transferability and can effectively identify the sinkhole point clouds in the predicted area (OA = 0.935, IoU (Sinkhole) = 0.662, mIoU = 0.794, AUC = 0.966, Recognition rate = 82.46 %), and even sinkholes with complex connected structures can be accurately identified. This study provides a new perspective for future large-area LiDAR surveys, mapping, and assessment of sinkholes.