Abstract
The processing of LiDAR point cloud data is of critical importance in the context of forest resource surveys, as well as representing a pivotal element in the realm of forest physiological and ecological studies.Nonetheless, conventional denoising algorithms frequently exhibit deficiencies with regard to adaptability and denoising efficacy, particularly when employed in relation to disparate datasets.To address these issues, this study introduces DEN4, an unsupervised, deep learning-based point cloud denoising algorithm designed to improve the accuracy of single tree segmentation in LiDAR point clouds.DEN4 introduces a multilevel noise separation module that effectively distinguishes between signal and noise, thereby improving the signal-to-noise ratio (SNR) and reducing the error.The experimental results demonstrate that DEN4 significantly outperforms traditional denoising methods in several key metrics, including mean square error (MSE), SNR, Hausdorff distance, and structural similarity index (SSIM).In the 60 sample dataset, DEN4 achieved the best mean and standard deviation on all metrics: Specifically, the MSE mean was found to be 0.0094, with a standard deviation of 0.0008, the SNR mean was 149.1570, with a standard deviation of 0.5628, the Hausdorff mean was 0.8503, with a standard deviation of 0.0947, and the SSIM mean was 0.8399, with a standard deviation of 0.0054. For instance, in the S10 dataset, DEN4 attained a 70.2% diminution in MSE and a 37.8% augmentation in SNR in comparison with PTD.The findings demonstrate the efficacy of DEN4 in multiple forest datasets, its ability to maintain geometric integrity, and its enhanced stability without the necessity for pre-labelled data. The algorithm's superior performance and robustness in diverse forest environments underscores its potential application in single tree segmentation and forest resource management.