Multi-Sensor Fusion of Sentinel-2 Imagery and ICESat-2 Satellite Laser Bathymetry for Benthic Habitat Classification in Key Largo, Florida Keys

Abstract

Accurate benthic habitat mapping is an important component of marine conservation. This study fuses freely available Sentinel-2 multispectral imagery and ICESat-2 ATL24 satellite laser bathymetry to classify five benthic habitat types — Coral/Algae, Seagrass, Sand, Rock, and Rubble — in Key Largo, Florida Keys, using a Random Forest classifier trained on 15,600 fusion points (depth + seven spectral features). Classification achieved 89.25% overall accuracy (κ = 0.87). A controlled ablation study confirmed that spectral features dominate classification performance (spectral-only OA = 89.91%), with ICESat-2 bathymetry contributing class-specific improvements for Rubble (+1.3 pp) rather than an overall accuracy gain. Statistically significant depth stratification was confirmed across all five habitats (H = 3149.24, p < 0.001). Coral/Algae fragmentation into 38 discrete patches reflects documented reef degradation in the Florida Keys National Marine Sanctuary. The results suggest that freely available satellite-based remote sensing can monitor benthic habitats in shallow, optically clear coastal waters, though further validation across diverse environmental settings is required before broader operational deployment.

Key results at a glance

1. 89.25% overall accuracy (κ = 0.87) across five habitat classes in Key Largo, Florida Keys
2. Controlled ablation study: spectral-only OA = 89.91% vs. fused OA = 89.25% — depth did not improve overall accuracy but provided class-specific benefits for Rubble (+1.3 pp)
3. ICESat-2 GT3R beam: 8,794 bathymetric points from 3,736,009 photons — 103.7% more than the next-best beam
4. Significant depth stratification across all habitats (H = 3149.24, p < 0.001)
5. Coral/Algae fragmented into 38 discrete patches — consistent with documented Florida Keys reef decline
6. Processing time: 127 seconds for 15,600 fusion points

Data and code

All data and analysis scripts are openly available on Zenodo: https://doi.org/10.5281/zenodo.20122641

The work is published in Remote Sensing Applications: Society and Environment (Elsevier). https://doi.org/10.1016/j.rsase.2026.102102

Author: Shobha Mourya Dumpati, FGS, FRGS