Advancements in Multi-Sensor Fusion for Space-Based Coastal Hydrospatial Surveying Through AI and Cloud Computation
Abstract
Littoral ecosystem mapping is inadequate due to gaps in bathymetry data between the land-sea interface, caused by the hazards, cost, and speed of traditional bathymetric survey methods in shallow waters. In this study, we have developed a new methodology for space-based shallow-water mapping, using a fusion of earth-observation sensors to produce satellite-derived bathymetry (SDB) at scale. The system integrates ICESat-2 LiDAR data and Sentinel-2 multispectral imagery with cloud computing and deep learning to go beyond standard regression models to mitigate common SDB challenges including imagery artifacts, low-reflectance benthic types, and identification of extinction depth. With billions of ICESat-2 ATL03 data points now available, manual extraction of ICESat-2 space-based bathymetric LiDAR (SBL) data is unscalable. As such, a deep learning model was trained to automatically extract bathymetric photons. The model achieved 88.6% bathymetry classification accuracy and increased production rate by 1,500%. Multispectral Sentinel-2 imagery is used in a cloud-computing environment to create artifact-free multi-temporal image composites on a per image tile basis. Individual images are selected using metadata constraints, processed with a marine-specific atmospheric correction, glint corrected, and masked for clouds artifacts before compositing to a singular image. SDB data is produced using a multi-modal approach where four discrete depth-retrieval models are available for use including band ratio, random-forest, weighted multi-regression, and radiative transfer. This flexible strategy applies the most suitable method for a given locations water column conditions, benthic types, and SBL availability, to ensure optimal SDB accuracy. SDB data is post-processed using a hybrid AI routine which employs an unsupervised machine learning model and a semantic segmentation model to delineate and remove outliers and depth retrievals over optically deep waters. Imbricate and adjacent SDB surfaces are amalgamated to create a seamless, continuous surface. Final SDB is assessed for statistical accuracy and uncertainty. Applying this suite of technologies, SDB with <1m RMSE has been derived at multiple locations around the world, reaching depths of 29.95m, with over 200,000km² produced to date.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.G55D0272S