A Full Perception Layered Convolution Network for UAV Point Clouds Data Towards Landslide Crack Detection

Abstract

Landslide cracks are essential signs of landslide creeping and sliding, that are also very useful to landslide prediction and risk prevention. In order to improve the ability to extract local structural features of landslide crack, this article proposed a full perception layered convolution network (FPLC-Net) model for landslide crack detection using UAV point clouds data. The proposed network model consists of three steps: first, sub-neighborhoods are divided by down-sampling and neighborhood full perception grouping, and local fine features of subneighborhoods are aggregated to sub-neighborhood description points. Second, the features of the sub-neighborhood description points are updated based on the distance from the sub-neighborhood description points to the center point. Finally, the feature extraction and synthesis of neighborhood points are realized by aggregating the features of sub-neighborhood description points to the central point. The model was validated in the eastern part of the Qinghai-Tibet Plateau. In addition, we explored the best combination of sub-neighborhood description points and sub-neighborhood points through several contrast experiments. The results show that the FPLC-Net achieves the highest mean intersection over union of 0.784 when the number of sub-neighborhood description points and sub-neighborhood points is four and six, which is higher than PointNet++ (0.708), DGCNN (0.637) and PointConv (0.678).

Keywords

• Convolution
• crack detection
• feature extraction
• remote sensing