Evaluating and Calibrating ICESat-2 Canopy Height: Airborne Validation and Machine Learning Enhancement Across Boreal and Tropical Forests

Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) represents a major advancement in remote sensing for terrestrial observation, substantially improving the capability to map vegetation structural parameters. However, spatial heterogeneity poses significant challenges to data accuracy. To evaluate the performance of ICESat-2 and improve its inversion accuracy, this study used airborne LiDAR data to validate ICESat-2 terrain and canopy height measurements in boreal forests of Alberta, Canada, and in three tropical rainforest regions—Costa Rica, French Guiana, and Gabon. Machine-learning approaches were further applied to calibrate ICESat-2 canopy height estimates. Our results show that the uncalibrated ICESat-2 data exhibit strong consistency in boreal forests, with higher accuracy under snow-covered nighttime conditions (terrain error < 1 m, canopy height error of 3.19 m). In contrast, the uncertainties in tropical rainforests are considerably larger, with terrain errors of 3–7 m and canopy height errors of 5–7 m. After calibration, XGBoost reduced canopy height error by 0.84 m in boreal forests, whereas Random Forest calibration improved canopy height accuracy by 1.09 m in tropical regions. Overall, our findings provide additional scientific evidence supporting the reliability of ICESat-2 measurements and substantially enhance the accuracy of satellite-based canopy height estimation.