Traditional 3D point semantic segmentation networks based on PointNet++ tend to sacrifice the accuracy of minority classes to maintain the overall accuracy dominated by majority classes. A new CNN is proposed to improve the segmentation accuracy of PointNet+ when processing airborne LiDAR point clouds with long-tailed distribution,which mainly consists of two aspects. The first is cluster-based farthest point sampling(FPS).Through intra-class FPS under proportional constraints, meanshift clustering based on confidence and zoning FPS combined with neighborhood compensation, the samplesof minority classes in airborne LiDAR point clouds can be retained to the maximum extent, and can be well learned by the network through re-weighting. The second is local feature learning under the spatial self-attention mechanism. By using different spatial encoding methods, a new spatial self-attention mechanism is constructed to facilitate learning the complete structure of the target from sparse sample data. Therefore, the learning ability of the network model for minority classes is improved while ensuring the good learning ability of the majority classes. Experiments on public data set show that the overall accuracy(OA)and F,score in this article have a significant improvement,which is 6.3% and 6.6% higher than those of PointNet+Compared with other 6 networks based on PointNet+ and the top10 network model in recent publications, the proposed algorithm has the best performance,good generalization ability and application value.