Accurate estimation of ground elevation on a large scale is essential and worthwhile in topography, geomorphology, and ecology. The Ice, Cloud and Land Elevation Satellite-2 (ICESat-2) mission, launched in September 2018, offers an opportunity to obtain global elevation data over the earth's surface. This paper aimed to evaluate the performance of ICESat-2 data for ground elevation retrieval. To fulfill this objective, our study first tested the availability of existing noise removal and ground photon identification algorithms on ICESat-2 data. Second, the accuracy of ground elevation data retrieved from ICESat-2 data was validated using airborne LiDAR data. Finally, we explored the influence of various factors (e.g., the signal-to-noise ratio (SNR), slope, vegetation height and vegetation cover) on the estimation accuracy of ground elevation over forest, tundra and bare land areas in interior Alaska. The results indicate that the existing noise removal and ground photon identification algorithms for simulated ICESat-2 data also work well for ICESat-2 data. The overall mean difference and RMSE values between the ground elevations retrieved from the ICESat-2 data and the airborne LiDAR-derived ground elevations are -0.61 m and 1.96 m, respectively. In forest, tundra and bare land scenarios, the mean differences are -0.64 m, -0.61 m and -0.59 m, with RMSE values of 1.89 m, 2.05 m, and 1.76 m, respectively. By analyzing the influence of four error factors on the elevation accuracy, we found that the slope is the most important factor affecting the accuracy of ICESat-2 elevation data. The elevation errors increase rapidly with increasing slope, especially when the slope is greater than 20°. The elevation errors decrease with increasing SNR, but this decrease varies little once the SNR is greater than 10. In forest and tundra areas, the errors in the ground elevation also increase with increasing vegetation height and the amount of vegetation cover.