Photon point clouds are widely used for forest structural parameter inversion, building height estimation, and water depth extraction. However, effective denoising remains challenging due to complex noise distributions and feature redundancy, which limit adaptability and generalisation. To address these challenges, we propose a photon-point cloud denoising algorithm based on multi-feature combination and machine learning. First, four feature systems were constructed from eigenvalue, distance, density, and height, yielding 12 features, and an improved adaptive neighbourhood curvature feature based on the minimum entropy criterion was incorporated to enhance characterisation of photon point geometry and spatial distribution under complex noise. Second, recursive feature elimination was employed to reduce feature redundancy and select six discriminative features, improving model generalisation. Third, comparative experiments on different feature combinations determine the optimal ones. Finally, the selected features were input into the classification model for denoising evaluation. The experimental results demonstrate that the adaptive neighbourhood curvature feature enhances denoising performance, and its removal leads to a 1.89% decrease in the F-score for daytime data. Moreover, multi-feature combinations significantly improve denoising accuracy. The proposed method achieved accuracies of 99.79% for night-time and 90.18% for daytime data, outperforming DBSCAN, KNN, and GMM in both scenes, indicating its robustness and practical applicability.
Hong et al. (Sat,) studied this question.