With this investigation we pursue the overall goal to obtain precise and reliable data on floating macro-litter regarding their quantity, position, accumulation zones, material properties, floating depth, and sources, which may serve as a basis for litter recovery, source elimination, and prevention of litter dispersal. To achieve this goal we will develop a remote sensing based detection and tracking system for large floating marine litter. The system will consist of components for remote sensing data analysis for litter detection (high resolution optical and SAR) and identification (hyperspectral), oceanographic forecasting, GIS analysis (overlay and merging of results, uncertainty analysis, and litter tracking), and a web interface for visualizing the results. As some components (oceanographic forecasting, SAR-based eddy detection) are already operationally in use, the main focus will be put on the application of machine learning and deep learning algorithms for plastic litter detection based on daily high-resolution optical (PlanetScope or SkySat) and SAR data as well as on the development of algorithms for litter tracking over time in daily remote sensing imagery. Our first approach, purely based on mono-temporal optical PlanetScope imagery analyzed with CNNs, has been tested in the Mediterranean Sea and was able to distinguish drifting objects from actively moving objects like ships and from stable objects like rocks or buoys (examples in Appendix, Fig. 3). This methodology will be further developed, expanded and tested in the Adriatic Sea in close cooperation with CNR-ISMAR and isardSAT. Hyperspectral satellite data will be acquired from accumulation zones predicted by the oceanographic forecast system. Spectral classification methods will be tested to identify plastic material based on spectral absorption features. In case of success this will be the first identification of offshore floating plastic marine litter from space.
