DGCNN-CBAM Integration: A Novel Framework for ICESat-2 Photon Point Cloud Classification and Validated Understory Terrain Estimation
In understory terrain research, high-precision photon point cloud classification is essential for comprehensive analysis of forest ecosystems. However, ICESat-2/Advanced Topographic Laser Altimeter System (ATLAS) photon data contain substantial noise, and the raw photon data lack explicit terrain attributes, which can lead to the misclassification of noise photons as ground signals and reduce terrain-extraction accuracy. To overcome this challenge, we develop a deep learning framework that fuses a dynamic graph convolutional neural network (DGCNN) with a convolutional block attention module (CBAM), enabling robust classification of ICESat-2 photon point clouds into ground, canopy, canopy top, and noise categories. Across three contrasting forest regions—California, South Carolina, and Puerto Rico—the proposed framework consistently outperformed NASA’s official algorithm and density-based spatial clustering of applications with noise (DBSCAN). Using digital terrain model (DTM)/digital surface model (DSM) as reference, the average <inline-formula> <tex-math notation="LaTeX">$F1$ </tex-math></inline-formula>-scores were 0.85 (macro) and 0.87 (weighted); using manual reference, they were 0.77 (macro) and 0.81 (weighted). For understory terrain estimation, the extracted ground photons achieved an average R2 of 0.99 and an average root-mean-square error (RMSE) of 1.70 m, outperforming FABDEM (R<inline-formula> <tex-math notation="LaTeX">${}^{2} = 0.99$ </tex-math></inline-formula>, RMSE = 4.99 m) and Shuttle Radar Topography Mission (SRTM) (R<inline-formula> <tex-math notation="LaTeX">${}^{2} = 0.99$ </tex-math></inline-formula>, RMSE = 9.05 m). These results demonstrate the strong capability of the proposed model for photon point cloud classification and precise understory terrain recovery.