Robust coral reef bathymetric mapping from ICESat-2 photons and multitemporal sentinel-2 imagery using sigma-clipping and a Stumpf-ratio-augmented random forest

Shallow-water bathymetry is essential for coral reef monitoring, but its retrieval remains challenging in complex reef environments because single-date optical imagery is susceptible to transient contamination and in situ depth measurements are limited. This study integrated ICESat-2-derived bathymetry with multitemporal Sentinel-2 imagery using Sigma-Clipping compositing and a Stumpf-ratio-augmented random forest model for Qilianyu, China. Bathymetric points were extracted from ICESat-2 ATL03 data, and Sigma-Clipping was applied to 94 Sentinel-2 images to reduce sunglint and cloud contamination. Independent pixel-wise validation against Mapper5000 airborne bathymetric LiDAR yielded an RMSE of 1.08 m, an MAE of 0.82 m, and an R2 of 0.923. Under the same compositing strategy, the random forest model reduced RMSE by 60.0% relative to the Stumpf log-ratio model. Five-fold track-disjoint validation on held-out ICESat-2 ATL03 tracks yielded a pooled out-of-fold RMSE of 1.164 m, an MAE of 0.482 m, and an R2 of 0.910. These results indicate that robust multitemporal compositing can support coral reef bathymetric mapping within Qilianyu.