<p>The Ice, Cloud and Land Elevation Satellite 2 (ICESat-2) is equipped with an advanced topographic laser altimeter system (ATLAS), which uses a photon-counting technique to achieve high-precision measurements of nearshore bathymetry. The complexity of environmental factors hinders the extraction of signal photons from ICESat-2 data. In this study, we proposed a filtering method for ICESat-2 photon-counting data combing iterative median filter, i.e., skeleton extraction and adaptive ellipse dilation (SEAED). The iterative median filter was employed to differentiate above-water and water-column photons effectively, thereby enabling the precise extraction of signal photons of the water surface. Skeleton extraction, a type of erosion operation, could greatly reduce the number of photons to be processed and maintain underwater terrain features in the raw photon data. Adaptive ellipse dilation can then extract signal photons accurately and completely based on the underwater terrain skeleton. Five regions were selected to verify the accuracy and reliability of the SEAED method. Comparison with measured water depth data and simulated data based on the measured water depth data showed that, the data estimated by SEAED had an optimal root square error of 0.25 m, and an F1 score of 0.983, indicating that the SEAED could significantly improve the filtering accuracy, particularly in complex terrain areas.</p>