A Semantic Segmentation Network for ALS Point Clouds Considering Geographic Location and Feature Interaction

The airborne LiDAR scanning technology is widely used in various fields due to its ability to quickly acquire geospatial features. As a crucial step in data applications, the automatic acquisition of point-by-point labels has become a hot topic in current research on point cloud data processing. Recent deep learning algorithms for feature extraction in the form of kernel point convolution (KPConv) are currently considered mainstream due to their outstanding performance. Among these, KPConv is the most representative. Although it has demonstrated good performance in various point-cloud classification tasks, the local window convolution approach overlooks global semantic information and does not focus on the relationship between point distribution and features. These shortcomings are particularly prominent in large-scale point-cloud classification. This paper proposes a novel network structure, LI-Net, that considers geographic location and feature fusion. First, location-enhanced kernel point convolution (LFKPConv) was designed. Subsequently, a new mask attention module was introduced to integrate the global features. Compared to traditional attention mechanisms, this module focuses only on a small number of prominent feature points, allowing for effective feature interactions while reducing memory consumption. Finally, to enable cross-layer feature fusion, a feature-weighting unit was designed in the encoding phase to enhance the significance of the semantic features. The proposed method achieved competitive results on the ISPRS, LASDU, and DFC2019 datasets, as well as a new state-of-the-art result on the GML dataset, with an average F1 score of 72.0% and an accuracy of 97.3%.