The Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2), owing to its sensitive photon detection system, acquires data containing a large number of background noise photons, which seriously affects the accuracy of signal extraction in complex terrain regions. To address the problem of insufficient parameter adaptability in existing denoising algorithms for slope-varying regions, we propose a denoising algorithm for slope-adaptive elliptic neighborhoods (SAEN-D), which is based on local terrain features. First, the effective signal range is intercepted by histogram statistics, and more than 88.82% of the discrete noise is preprocessed and rejected using grid statistics. Then, an adaptive elliptic neighborhood with slope angle constraints is constructed by a slope-driven segmentation strategy, and the search direction and the length of the ellipse’s long-axis are dynamically adjusted to match the signal distribution characteristics. Finally, the combination of the local distance discrepancy coefficient and OTSU’s method (OTSU) of thresholding segmentation is used to accurately distinguish signal and noise photons. Experiments are carried out in the regions of Antarctica’s flat ice cap and Greenland’s complex terrain, and the results show that in extreme terrains such as that with steep slopes and elevation faults, the value of the method described in this paper reaches 96.13–98.25%, which is 7.8% higher than that of the traditional improved local sparse coefficient (ILSC) algorithm. The results of the study confirm that SAEN-D effectively solves the problem of signal leakage and misjudgment caused by the anisotropy of photon distribution in complex terrain, and provides reliable support for high-precision elevation inversion and the dynamic monitoring of ICESat-2 data. This algorithm has broad application potential, as it can significantly improve the quality of laser altimetry satellite data and offers new insights and solutions for precise monitoring using sensor technologies in complex and variable terrains.