Extracting accurate terrain in vegetated areas from ICESat-2 data

The uncertainty of ICESat-2 terrain accuracy, especially in vegetated areas, limits its scientific application, and there is barely any comprehensive modeling evaluation for this uncertainty. In this study, we propose a terrain quality classification model with uncertainty measurement for extracting accurate vegetated terrain from ICESat-2 altimetry products, which includes two main parts: 1) training samples are used to construct a terrain elevation quality classification model; 2) the relationship between the vote entropy and elevation accuracy is analyzed to measure the uncertainty of the predicted results of the model. Compared with airborne LiDAR data from multiple areas of the world, it is confirmed that the extracted results can meet the different quality requirements for terrain elevation data in vegetated areas (95th percentile of the absolute error: 1 m, 2 m, and 3 m) with a higher than 90% purity (proportion of accurate terrain). The accuracy of the extracted results is ∼0.5–0.9 m, ∼0.9–1.3 m and ∼1.1–2.9 m, respectively, and that of the eliminated results is ∼1.2–3.8 m, ∼2.0–4.4 m and ∼3.4–5.8 m, respectively. The results also show that the method can extract high-accuracy terrain in high vegetation cover areas (where the tree cover index is higher than 80%), and has the potential of applying to large-scale and even global-scale vegetated terrain. Moreover, the method is also suitable for non-vegetated areas. The extracted results can meet the corresponding quality requirements for non-vegetated areas (root-mean-square error: 0.333 m, 0.667 m, and 1 m), and their purity is also high than 90%.