The great potential of NASA’s Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) to retrieve sea ice heights has been demonstrated. However, the presence of a significant number of noise photons in the ICESat-2 data makes accurate monitoring of sea ice changes challenging. This article proposes an adaptive clustering and kernel density estimation-based (AC-KDE) method for estimating sea ice heights in ICESat-2 photon clouds. First, the adaptive clustering method effectively detects sea ice signal photons. The method’s input parameters are determined based on the ATLAS parameters and the LiDAR transmission equation. Then, the adaptive-count signal photon aggregates are used to estimate sea ice heights, and a variable along-track resolution is obtained using the kernel density estimation method. The AC-KDE method is applied to the Multiple Altimeter Beam Experimental LiDAR (MABEL) and ICESat-2 data, and we compare it with other denoising algorithms, including the histogram-based method (HBM), density-based spatial clustering of applications with noise (DBSCAN), ordering points to identify the clustering structure (OPTICS), UMD_RDA, density-dimension method (DDM), and improved localized statistics method (ILSM) algorithms. The results indicate that the proposed method outperforms these algorithms in extracting signal photons with higher accuracy scores and F-scores, which are 0.97 and 0.97, 0.92 and 0.90, and 0.89 and 0.72 under high-medium-low signal-to-noise ratio (SNR) conditions, respectively. Additionally, the retrieved sea ice heights are compared with the ATL07 heights. The AC-KDE heights show a significant correlation with coincident ATM heights and have a lower root mean square error (RMSE) value (0.066 m) compared to ATL07 heights (0.104 m). The AC-KDE method also demonstrates a vertical height precision of 0.01 m over flat leads. The proposed method can effectively extract signal photons and accurately estimate sea ice heights in polar regions.