SUPERDOVE-MODELLED BATHYMETRY USING NEURAL NETWORKS ALONG A TURBIDITY GRADIENT: BREHAT, SAINT-BARTHELEMY AND TETIAROA ISLANDS

Abstract. Despite the increasing interest in the bathymetry mapping in the context of the sea level rise and storm intensification, only 10% of the worldwide bathymetry has been mapped with reliable sonar and lidar, due to their high cost. The satellite-derived bathymetry (SDB) has therefore grown for the last decades, due to its affordability, but also to its gain in radiometric, spatial, spectral and temporal resolutions. Nevertheless, SDB products leveraging both a high spatial and a high temporal resolution are still expected by stakeholders responsible for cloudy or tidal coastal areas. This research tests the contribution of the Planetscope SuperDove CubSats (eight-band, 3 m, 1 day) four novel bands in bathymetry extraction, along a turbidity gradient (the islands of Bréhat in Channel Sea, Saint-Barthélémy in Caribbean Atlantic Ocean, and Teti’aroa in South Pacific Ocean), using neural networks calibrated, validated, and tested with recent topobathymetric lidar data. In Bréhat (turbid) and Saint-Barthélémy (clear), water depth was best modelled using the eight-band dataset with good contributions of yellow and green 1 for Bréhat (R²=0,77), and purple and green 1 for Saint-Barthélémy (R²=0,95). The very clear waters of Teti’aroa were best modelled using the combination of base (blue, green 2, red and near-infrared) + yellow (R²=0,68). The lower accuracy of bathymetry mapping in Teti’aroa revealed biases due to satellite-lidar collection time difference, lidar data specificities, and/or control quality.