Abstract: Median filtering, as a classical image processing technique, is widely used for image denoising and noise reduction in point cloud data. However, it has not been applied to the denoising processing of underwater point cloud data from ICESat-2. The noise points in ICESat-2 data exhibit characteristics of salt-and-pepper noise, and median filtering has shown significant effectiveness in eliminating such noise. Therefore, this study proposes a new denoising method that is based on the traditional median filtering algorithm and combines the specific noise patterns and spatial distribution characteristics of ICESat-2 point cloud data to separate noisy point clouds from signal point clouds. By using frequency statistics to distinguish between sea surface and underwater data, the underwater data are transformed into a grid format and processed using median filtering technology. Based on the number of grids, pixel sets are classified as underwater signals or discrete points, and clustering analysis is further used to divide discrete points into signal and noise points. Finally, the selected signal pixel sets are augmented with data to improve the accuracy and completeness of the data. And by comparing with the commonly used DBSCAN denoising results and classification results based on confidence level for point cloud denoising, the superiority of the proposed method in preserving seafloor terrain data was confirmed. The experimental results show that the water depth data extracted by this method has a correlation coefficient with the true water depth of more than 0.95, an average absolute error of less than 0.5, a root mean square error of less than 1.0, and an average relative error of less than 0.5%, demonstrating high accuracy and reliability. This provides a new and effective method for denoising processing of ICESat-2 marine survey and mapping data.