The uneven quality of Airborne Laser Scanning (ALS) point clouds easily leads to feature loss during the downsampling process of deep learning. This study addresses the problem of non-uniform quality in point clouds by proposing four lightweight, tree-based enhancement strategies: jitter, nearest neighbour interpolation, internal void filling, density, and random. Based on real tree species data, the processing time was recorded and compared with Test-Time Augmentation (TTA) strategies, and systematic experiments were conducted using five deep learning models. Combined with standard testing and cross-validation approaches, the effects of different enhancement methods on distribution restoration capability, tree species classification performance, feature preservation ability, and generalization ability were quantitatively analysed. Among them, nearest neighbour interpolation enhancement performed best, achieving the best F1−score of 0.790 with a maximum relative improvement of 29.8%, while preserving tree structural features and showing the highest cross-model consistency. The distorted point clouds generated by random enhancement did not exhibit the expected performance collapse, indicating that the models possess a certain degree of tolerance to structurally heterogeneous point clouds, while further highlighting the importance of input quality for tree point cloud classification. This provides an effective improvement scheme for the lightweight processing of forestry point clouds, noise-resistant modelling of tree point clouds, and forest inventory.
Wang et al. (Thu,) studied this question.