Airborne light detection and ranging (LiDAR) technology possesses the capability to rapidly scan wide-area surfaces, acquiring high-precision 3-D spatial data. Consequently, point cloud classification has become a core task in geographic information systems (GIS) and remote sensing applications. While deep learning methods have significantly advanced this field, processing large-scale airborne point clouds remains challenging due to data noise, uneven sampling, and complex ground structural characteristics. Traditional voxel-based methods often downsample data to improve computational efficiency, but this process inevitably leads to the loss of fine geometric details and topological information, referred to as quantization artifacts. Conversely, point-based methods preserve details but struggle with high computational costs and weak neighborhood feature abstraction capability in large scenes. Existing mainstream frameworks, such as PointNet⁺⁺ and various Transformer-based models, often face trade-offs between local neighborhood feature mining and global context capture. To address the problem of fine geometric topology information loss during voxelization and to balance computational efficiency with classification accuracy in complex urban scenarios, a novel classification architecture based on voxel and point dual-branch feature fusion is constructed in this study.

The network architecture adopted a dual-stream parallel strategy designed to leverage the complementary advantages of voxel and point representations (Fig.1). The first stream was the voxel branch, which was responsible for extracting local context information. It employed a spatial hash voxelization strategy to map unstructured point clouds into a regular grid, and utilized a local self-attention module (Fig.2) based on the point transformer. This module dynamically captured interactions between points within a local neighborhood using a fixed-radius spherical query, efficiently aggregating semantic features while suppressing noise. To expand the receptive field, a transition down module was introduced to perform sparse convolution and downsampling. The second stream was the point branch, dedicated to preserving raw geometric details. Unlike the voxel branch, it operated directly on discrete points without neighborhood aggregation, thereby maintaining high resolution. This branch employed a stacked residual multi-layer perceptron (Res-MLP) network (Fig.3) to mine deep features for each independent point, effectively preventing gradient vanishing. Finally, an adaptive fusion module was designed to integrate the features from both branches. It used nearest-neighbor interpolation to map voxel features back to the point domain and employed a channel attention mechanism to dynamically weight and fuse the coarse-grained semantic features with fine-grained point features.

Comprehensive experiments were conducted on two authoritative benchmark datasets: the ISPRS-3D dataset (Fig.4) and the WHU-Urban3D dataset (Fig.5). The experimental environment utilized an Nvidia RTX3090 GPU and the PyTorch framework.

(a) Parameter Analysis: The impact of the number of neighbors (K) and search radius (r) in the local attention module was analyzed (Table 1). The model achieved optimal performance with K=20 and r=0.2. (b) ISPRS-3D Results: Quantitative evaluation on the ISPRS-3D dataset (Table 2) showed that the proposed method achieved an overall accuracy (OA) of 82.2% and an average F₁ score of 66.1%. Comparative experiments with mainstream models (Table 6), including PointNet⁺⁺, DGCNN, RandLA-Net, and stratified transformer, demonstrated that the proposed method yielded the highest accuracy. Specifically, in categories with complex topological structures such as powerlines and fences, the F₁ score improved significantly compared to single-branch baselines (Table 4). Visual comparisons (Fig.6 and Fig.8) indicated that the dual-branch fusion strategy effectively delineated boundaries between confusing classes like impervious surfaces and low vegetation. (c) WHU-Urban3D Results: On the larger-scale WHU-Urban3D dataset, the proposed method achieved an OA of 87.0% (Table 3), outperforming the stratified transformer by 0.5% and PointNet⁺⁺ by 1.7% (Table 7). Ablation studies (Table 5) confirmed that the dual-branch architecture improved OA by 2.4% over a voxel-only baseline. Visualization results (Fig.7 and Fig.9) further verified the model's robustness in identifying street lamps, vehicles, and distinguishing between trees and buildings in dense urban environments.

The proposed voxel and point dual-branch feature fusion network successfully mitigates the issue of detail loss caused by voxelization in large-scale point cloud processing. By synergizing the efficient local context aggregation of the voxel branch with the high-resolution feature extraction of the point branch, the architecture achieves superior semantic segmentation performance. Experimental results on ISPRS-3D and WHU-Urban3D datasets demonstrate that the method outperforms current state-of-the-art networks in terms of overall accuracy and robustness, particularly for identifying small objects and refining object boundaries. This study provides an effective solution for the refined classification of airborne LiDAR data in complex urban scenes.