Tree species classification serves as a core task in forest resource management and ecological monitoring, playing a crucial role in biodiversity conservation and carbon cycle research. Compared to optical or Synthetic Aperture Radar (SAR) data, three-dimensional (3D) point clouds generated by Light Detection and Ranging (LiDAR) can more accurately characterize the geometric structure of trees, such as trunk topology and leaf cluster distribution, leading to their widespread application in tree species classification. However, the generalization capability of existing classification methods in real-world complex forest environments remains a significant challenge, primarily due to the combined effects of scene complexity, data distribution shift, and class imbalance. To address these challenges, this paper proposes a novel tree point cloud classification network, KPCTree (Kernel Point Convolution for Tree). The model employs Kernel Point Convolution (KPConv) as its backbone, integrates a pointwise channel attention mechanism to enhance feature discrimination capability, and introduces specialized data augmentation strategies tailored for tree point cloud characteristics. Furthermore, a dynamic class weighted loss function is adopted to mitigate the data imbalance problem. Experimental results demonstrate that KPCTree significantly outperforms existing methods across multiple datasets, exhibiting excellent generalization and universality.