Assessing the contribution of super-resolution in satellite derived bathymetry in the Antarctic

Abstract

The difficulty of defining the depth of near-shore seas (bathymetry) arises from the limits imposed by traditional ship-based approaches during data collection. Although LiDAR sensors with green lasers have been used to solve some of these problems, they come at a high cost in terms of their footprint and are prone to inaccuracies in turbid water. As shorelines undergo changes due to erosion, wetland loss, hurricane effects, sea-level rise, urban development, and population growth, consistent and accurate bathymetric data become crucial. These data play a significant role in comprehending and managing sensitive interfaces between land and water. Satellite-derived Bathymetry (SDB), which has been described by maritime and remote sensing researchers for over 50 years, emerges as a gap-filler, encompassing bathymetry extraction approaches using active (altimetry) and passive (optics) satellite sensors. In the past decade, advancements in sensor capabilities, computational power, and recognition by the International Hydrographic Organization (IHO) have propelled SDB to unprecedented popularity. This study explores the contribution of super-resolution in SDB for the first time in the shallow water zone of Horseshoe Island, Antarctica. Random forest and extreme gradient boosting machine learning-based regressors were used on Landsat-8 OLI images, which were atmospherically corrected by the ACOLITE algorithm and spatially enhanced twofold via the generative adversarial network for single image super-resolution (SRGAN). The bathymetry predictions with these two machine learning algorithms on SR images were benchmarked against previous studies in the same region and showed admissible results concerning the IHO standards. Furthermore, the results indicate that the bathymetric inversion performance of the spatially enhanced image via SRGAN is superior to the original multispectral image and pan-sharpened image in terms of the metrics observed, namely, root mean square error (RMSE), mean average error (MAE), and coefficient of determination (R²). Comparison between the original and SR image bathymetry inversion for the 0–15 m depth range indicate improvements of up to 0.13 m for RMSE, up to 0.30 m for MAE, and up to 11% for R². These results promise possible effective usage of super-resolution in SDB with satellite images such as Sentinel −2, which do not include a panchromatic band.