Abstract: ICESat-2 (Ice, Cloud, and Land Elevation Satellite-2) is one of the most advanced laser altimetry satellites currently available. Using a laser wavelength of 532 nm, it can effectively acquire bathymetric data in the shallow water areas, thus greatly promoting the development of shallow water bathymetry. However, the raw data from ICESat-2 are often clouded by noisy point clouds so that great challenges can be brought to the post-processing of the data. To improve the accuracy and efficiency of data processing, a denoising algorithm based on Multilayer Perceptron (MLP) is developed to address the feature that ICESat-2 point clouds are denser in horizontal direction than in vertical direction. This algorithm takes into account comprehensively the feature values such as point density, average distance between points, and distance between the nearest neighbors (respectively 3 and 5) within a horizontal elliptical search area, thus realizing the effective identification and removal of noise points. The developed denoising model has been validated by selecting the ICESat-2 data from an reef island region in Australia as the training set and also using the data from the Jade Bangle Reef and the East Island in the Xisha Islands of China. The experimental results show that the denoising method developed in the study has an accuracy rate of over 90%, which is significantly better than the existing OPTICS denoising algorithm and the denoising results extracted by the official ATL03 algorithm. This achievement not only provides a new solution for the noise removal from ICESat-2 data, but also provides reliable data support for the research in related fields.