Observed Spectral-Reflectance Zoning for Model Selection in Satellite-Derived Bathymetry of Heterogeneous Coral Reef Waters Using ICESat-2 and Sentinel-2

Accurate and spatially continuous shallow water bathymetry supports coral reef geomorphological investigation, ecosystem assessment, and marine resource management. However, spectral heterogeneity arising from substrate properties, water depth, and water-column conditions complicates model selection. This study evaluates whether bathymetric model performance differs among observed spectral-reflectance zones and whether zonal model selection improves overall retrieval. ICESat-2 bathymetric points, Sentinel-2 imagery, and Allen Coral Atlas data were analyzed at Lingyang Reef, Huaguang Reef, and East Cay Reef. Representative substrate classes identified using Allen Coral Atlas data were used to derive Sentinel-2 green-band reflectance thresholds. Water pixels were then classified into high-, medium-, and low-reflectance zones. The band-ratio, multiple linear regression (MLR), random forest (RF), and extreme gradient boosting (XGBoost) models were constructed and validated within each zone. When averaged across the three study areas, MLR achieved the best performance in the high reflectance zones, with mean R2 and RMSE values of 0.71 and 0.49 m, respectively. RF achieved the best mean performance in the medium reflectance zones, with mean R2 and RMSE values of 0.89 and 0.62 m. In the low reflectance zones, RF and XGBoost outperformed Stumpf and MLR, with XGBoost yielding mean R2 and RMSE values of 0.85 and 1.20 m, respectively. The optimal outputs from the different reflectance zones were integrated and smoothed to generate continuous bathymetric maps. The integrated zonal method achieved RMSE values of 0.76, 0.61, and 0.84 m at the three reefs, respectively, representing numerical reductions of approximately 3.8%, 4.7%, and 12.5% relative to the best unified model. These findings demonstrate that observed spectral-reflectance zoning provides a practical basis for matching bathymetric models to different reflectance conditions, supporting bathymetric mapping in heterogeneous coral reef waters.