ICESat-2 data classification and estimation of terrain height and canopy height

Abstract

ICESat-2 (Ice, Cloud, and Land Elevation Satellite-2) was launched in 2018 with a photon-counting LiDAR (Light Detection and Ranging) system, ATLAS (Advanced Topographic Laser Altimeter System). It is collecting massive earth elevation data all over the world, which has shown the potential of large-scale forest monitoring. However, the energy emitted by the LiDAR system is low, and the received signals are easily affected by noise. Accurate classification of photons is an important step for forest parameter retrieval. Given the limitations of existing photon classification algorithms in areas with complex terrain, we proposed an improved local outlier factor algorithm with a rotating search area (LOFR). First, photons are transformed to the along-track direction, and noise photons are preliminarily filtered out by using the elevation histogram and elevation statistical methods. Next, ground photons are extracted by using the LOF (Local Outlier Factor) with the horizontal ellipse search area algorithm (LOFE) during an initial classification stage to filter photons that are far away from the ground. During the refined classification stage, which is the core of the algorithm, the terrain slope is calculated according to the ground photons extracted during initial classification. The elliptic search area is then rotated to align its long axis with the terrain slope. Finally, the LOFR scores of the photons are calculated to remove noise photons and signal photons are classified into top-of-canopy photons, canopy photons, and ground photons. The results show that the algorithm can effectively classify photons. Both estimated terrain height and canopy height derived using the classified photons are in good agreement with airborne LiDAR data. The mean absolute error (MAE) of the estimated terrain height relative to airborne data was 1.45 m and the root mean square error (RMSE) was 2.82 m. For canopy height validation, the correlation coefficient (R2), MAE, and RMSE at the best study scale (80 m) were 0.86, 1.82 m, and 2.72 m, respectively. These results demonstrated that the proposed LOFR algorithm can improve photon classification over complex terrain without prior knowledge of the terrain. Therefore, it could provide a robust approach for large-scale ATLAS data processing.