Using metocean repositories to support the creation of large datasets for AI applications

Non-collaborative vessels identification is a strategical interest both for the civil and military world. Indeed, boats involved in illegal activities can become unrecognizable from other receiving antennas by switching off their AIS, thus resulting as “dark vessels”.

Satellite imagery can in principle support the detection of dark vessels through the automatic identification of their wakes, but these patterns and their visibility are strongly influenced by meteo-marine conditions. In this direction, the UEIKAP (Unveil and Explore the In-depth Knowledge of earth observation data for maritime Applications) Project, funded by the Italian Ministry of University and Research, is developing an Artificial Intelligence (AI) system for the automatic identification of dark vessels from optical and SAR (Synthetic Aperture Radar) images. In particular, this contribution focuses on the creation of the dataset used for the training of the AI based on data from Marine Copernicus Ocean and other publicly available repositories. The AI requires an extensive satellite image dataset of vessels and related wakes to be trained, with an analogue dataset of similar wake patterns caused by external phenomenon and not by the vessel itself. Both are built with a number of 1500 optical images of wakes and 1000 non-wakes images in 4 different bands integrated with ancillary data. The same procedure is applied also for SAR images. A crucial role is played by Marine Copernicus data in assessing the environmental conditions that can control pattern formation on the sea surface and its visibility, supporting the interpretation of satellite images and the disambiguation of wake and wake-like patterns

In practice, for each image depending on its acquisition time and location, our algorithm retrieves the gridded fields of key atmospheric and oceanographic variables, computes derived quantities and stores the whole information in a self-explanatory and interoperable netCDF file, additionally generating 2D plots of the extracted variables. More specifically, hourly atmospheric fields (10 m wind components, total cloud cover, and total precipitation) at 1/4° spatial grid resolution are retrieved from the ERA5 reanalysis via the Copernicus Climate Service (C3S). Oceanographic quantities such as wave spectral parameters, surface current velocity components, potential temperature, surface and near-surface salinity and temperature (upper 10m) are collected from Copernicus Marine Service (CMEMS) with hourly to daily frequency and spatial resolution ranging from 1/24° to 1/5°. Wave steepness, surface and near-surface potential density anomaly and its horizontal gradients, as well as surface current convergence and near-surface buoyancy frequency are additionally computed in the process. Although the final aim of this operation has been conceived for a specific scope, the code can easily be used for a broader set of applications with different meteo-Oceanografic information and different regions and conditions.