Lsr-RF: a multistage adaptive classification method for ICESat-2 Photons in intertidal wetlands by synergizing multi-scale geometrical features with Sentinel-1 and Sentinel-2 data

ICESat-2 provides near-global, high-precision elevation observations for intertidal mapping and time-series analysis. However, photon heterogeneity across ground, vegetation, water surfaces, and noise, together with tidal dynamics and dense vegetation, limits the effectiveness of conventional filtering and classification methods. To address this, we propose Lsr-RF, a multi-level adaptive framework that integrates the local sparsity rate with a random forest classifier. The method fuses ICESat-2 multi-scale geometric features with Sentinel-1 polarization and Sentinel-2 spectral features, applying RF-based feature selection and classification to distinguish noise, bare tidal-flat, vegetation, and ground-under-vegetation photons. Evaluated across strong/weak-beam and day/night conditions, Lsr-RF was compared with denoising-oriented baselines (ATL08, DBSCAN, OPTICS) and XGBoost. Lsr-RF improved overall accuracy by 0.52–28.06 percentage points and the Kappa coefficient by 0.02–0.61, and achieved a multi-class classification overall accuracy of 0.99 with a Kappa coefficient of 0.99. These results demonstrate its potential for accurate photon-level classification and broader intertidal wetland mapping applications.