ICESat-2 bathymetric

Given the limitations of existing unsupervised machine learning denoising algorithms applied to ICESat-2 satellite data, this study introduces a deep learning based denoising method. This method aims to improve the accuracy of extracting shallow water depth from ICESat-2 data, thereby reducing on-site measurement costs and accelerating data acquisition speed. It is expected that these developments will make significant contributions to advancing marine scientific research and economic activities. This study optimized the performance of a multi-layer perceptron (MLP) neural network model for denoising ICESat-2 data by integrating recursive feature elimination, data standardization, and grid search with hierarchical cross validation for hyperparameter adjustment. Special emphasis is placed on independently training neural networks for underwater point clouds to enhance classification results. The MLP model performs well in the entire data denoising task. When applied to datasets from the Great Barrier Reef in Australia, Oahu Island in the Hawaiian Islands, Ganquan Island, and Puerto Rico, the accuracy of the model is all above 0.9. Although the accuracy of underwater data denoising is slightly lower, with a denoising rate of over 0.6 at the same location, the results are still very valuable for practical applications. To verify the effectiveness of the MLP method, a comparative analysis was conducted between the DBSCAN algorithm and the confidence based classification method. The results highlight the superiority of deep learning based methods. After refractive correction, the correlation coefficient exceeds 0. 80 observations were made between the extracted seabed depth data and actual depth measurements. In addition, the average absolute error is less than 1 m, and the root mean square error is less than 1. 0 meters, 2