A Novel Multidimensional Statistics Denoising Algorithm Based on Gaussian Mixture Model for Photon-Counting LiDAR Data

Abstract

The micropulse multibeam photon-counting laser altimeter significantly improves the sampling density along the orbit while introducing abundant noise. Therefore, noise removal is crucial for the subsequent applications of the photon-counting laser altimeter. This study proposes a new multidimensional statistics algorithm based on the Gaussian mixture model for photon noise removal. The method can simultaneously utilize multiple point cloud statistics and avoid manually selecting denoising thresholds, dividing signals and noise adaptively. To evaluate the performance of the algorithm, it is applied to both simulated ICESat-2 data and real ICESat-2 data, and evaluation indicators including recall, precision, and the harmonic means of recall and precision are used for analysis. Experimental results indicate that compared to denoising using a single statistic, denoising with multidimensional statistics (specifically, the sum of the distance of k-nearest neighbors, the photon density, and the standard deviation of the height difference within an elliptic neighborhood of the target photon) generally yields better results. Multidimensional statistics can influence and constrain each other, leading to improved decision-making outcomes. After comparing the algorithm with mainstream DBSCAN and KNN methods, the results indicate that the GMM method achieves the best denoising results, with average F-values 5.22% and 7.26% higher than those of DBSCAN and KNN methods, respectively, across all datasets. In conclusion, this algorithm can comprehensively assess and make decisions based on different statistical measures, demonstrating robustness and adaptability in extracting signal photons, and is worth promoting in the photon denoising applications.

Keywords

• And land elevation satellite-2 (ICESat-2)
• cloud
• gaussian mixture model (GMM)
• ice
• multidimensional statistics
• photon denoising