SSFE-PFE-DGCNN: A Segmented Statistical Filtering Enhanced Physics-Feature Dynamic Graph Convolutional Network for ICESat-2 Photon Cloud Denoising

Abstract

Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) is currently one of the most advanced spaceborne lidar systems, and its data have been widely applied in ocean bathymetry research. However, due to atmospheric scattering, water refraction, and other factors, its data contain substantial noise. Consequently, acquiring clean and pure spaceborne lidar data remains a key research focus. This study proposes SSFE-PFE-DGCNN, a three-stage physics-guided framework for denoising ICESat-2 underwater photon point clouds: 1) physics-constrained preprocessing: sea surface photons are extracted via dual Gaussian fitting, with underwater points separated based on the sensor's ≤50 m bathymetric range. Segmented Gaussian filtering subsequently removes statistical outliers; 2) depth-density fused denoising: a dynamic graph convolutional neural network leverages two physics-guided features—depth and Gaussian kernel density estimation probability—to discriminate seafloor signals from water-column noise; and 3) residual refinement: fine-grained Gaussian filtering eliminates residual noise in initially denoised points. A total of 30 ICESat-2 ATL03 datasets spanning diverse seabed topographies, water depths (0–5 m to >20 m), and photon density regimes are partitioned into 18 training, 6 validation, and 6 independent test sets. On the test sets, the proposed method achieves classification accuracies ranging from 97.63% to 98.96% (average F1-score 98.86%), outperforming DBSCAN, OPTICS, and PointNet++ baselines. Compared to the NASA ATL24 standard product, the method demonstrates superior signal retention in deep-water regions, sparse point clouds, and complex terrain. Validation against in situ bathymetry achieves R2 ≥ 0.97 (MAE ≤ 0.32 m) in gently sloping areas and maintains coherent topographic mapping in rugged terrain with R2 = 0.91 (MAE = 0.40 m and RMSE = 0.46 m).