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

Shobha Mourya Dumpati

Multispectral imagery has traditionally been used to classify benthic habitats; however, many challenges exist when using this method alone including the overlap of spectral signatures among habitat types, and the loss of signal due to water depth in coastal areas. The authors propose an innovative method that combines multispectral imagery from Sentinel-2 (Tile 17RNJ, January 30, 2024) with bathymetry data from ICESat-2 ATL24 satellite laser altimetry (January 24, 2024 and March 4, 2024) to identify five different benthic habitat types (Coral/Algae, Seagrass, Sand, Rock, and Rubble) within the Key Largo area of the Florida Keys.The authors use Random Forest classification on 15,600 fusion points (15,671 ICESat-2 bathymetric measurements co-registered with Sentinel-2 spectral information) where 12,646 samples are labeled as part of the training set and achieve a total classification accuracy of 89.25% (κ = 0.87, F1 = 0.891). GT3R ICESat-2 beam provided better-quality bathymetry data compared to the GT1L beam, resulting in 8,794 high-confidence points (using 3,736,009 total photons; 379.8% more than GT1L) L (left) and R (right). Of the 8,794 points collected on March 4, 2024, there were 7,257 points (82.5% of GT3R total). Analysis indicated that there was a significant difference in terms of depth stratification (Kruskal-Wallis H = 3149.24, p < 0.001), with seagrass limited to shallow waters at 3.52±1.74m (95% <5m), while sand occurred at the lowest depths at 5.54±2.34m. Spatial autocorrelation (Moran's I = 0.592–0.882) further demonstrated the strong clustering of habitats, with sand exhibiting the highest autocorrelation (0.882) and rock exhibiting the greatest degree of fragmentation (6.6m mean spacing). Feature importance showed the blue band (19.4%) to be the most important feature, followed by the green band (17.3%), NDWI (13.2%), NDVI (12.0%), and then depth (11.6%).Cross-validation also supported the stability of the model (86.9% ± 0.8%).Random Forest performed better than XGBoost (87.95%, 3.19s) and SVM (72.93%, 23.09s) in terms of both accuracy and processing time.Combining ICESat-2 ATL24 into the process increased accuracy by 12 – 15 % compared to processes based only on the spectral data and demonstrate the potential of the combination of multiple sensors for operational benthic habitat mapping.