No abstract

<p>WASP (Windrows AS Proxies) is a data processor, developed in the frame of the European Space Agency (ESA) OSIP Campaign, exploiting Copernicus Sentinel-2 L1C images to detect and catalogue the presence of filaments of floating marine debris with high probability of containing man-made litter. WASP takes advantage of the prototype EO data processor developed in the frame of ESA project  “Earth Observation (EO) Track for Marine Litter (ML) in the Mediterranean Sea” that successfully proved for first time that Copernicus Sentinel-2 data can detect the presence of marine litter accumulations as proxies of plastic litter content.</p><p>WASP puts significant effort in masking unneeded data that has been source of false-positive detections, including sun glint and clouds. Also, a new spectral analysis technique has been employed to identify the most promising Copernicus Sentinel-2 bands to be used in the detection of such filaments, which has also led to the construction of a novel spectral index WASP Spectral Index (WSI). This index enables the detection of filaments of floating debris.</p><p>The images processed using WSI are transformed into binary masks to be analysed by a deterministic object classifier, which looks at the geometry and shapes of the detections to identify ML windrows within them and separate them from background noise and/or false positives. This enables automatic processing and classification of the images, which makes possible to generate regional and/or local databases of remote-sensed floating debris, which can be exploited by means of geostatistics to support research and monitoring of marine litter in the environment.</p><p>These implementations are also supported with the introduction of advanced super-resolution techniques that are downscaling the spatial resolution of the bands to 10m, well beyond the simple interpolation, yielding better quality on the results.</p><p>In a preliminary assessment, the implemented proposed algorithm has proven to be successful in identifying windrows even when those are too thin to be visible in True Colour images by the naked eye. Nevertheless, some drawbacks/limitations have been found, principally associated to residual limitations when removing bad data, and with the special case of the problematic wave glint, well known in the Sentinel-2 data but of difficult solution.</p><p>Once the entire Sentinel-2 archive over the Mediterranean Sea is processed and following an in-depth analysis, a database of the identified proxies, including spatial and temporal patterns will be created over this initial region. The final EO product will be a map of on sub-mesoscale marine debris concentrations in the Mediterranean Sea based on Copernicus Sentinel-2. The product will consist on a census of these structures for each processed tile for the Mediterranean Sea, with potential for global scalability. Scientific research, cleaning activities and policy making on marine litter are only a few of the activities that could benefit from such a product.</p><p>This activity collaborates on the “Remote Sensing of Marine Litter and Debris” IOCCG taskforce.</p>

<p>Plastic pollution is widely recognised to be an emerging ecological disaster (Eriksen et al., 2014). While a steady increase in the amount of marine litter is being observed, plastics constitute some 60 to 80% of the total waste (Miladinova et al., 2020), which drift and settle through sinking and beaching. The Black Sea, a semi-enclosed basin with numerous litter inflows by huge watershed rivers, and with only one spillway at the Bosporus, is an ideal test area for the development of litter detection and tracking technologies. Although the occurrence of marine litter in the Black Sea is poorly known, with lack of data in the abundance of floating debris (Miladinova et al., 2020), remote sensing from space (RSS) is considered a promising tool for the observation of floating marine plastics because of its wide observation cover. However, success was only obtained i.in areas with huge accumulations of litter (canals, harbours and estuaries, e.g. rows of litter in the sea after flooding), and ii.with applying “Ocean Colour” RSS methods designed for the assessment of concentration of phytoplankton or other particulates, which are far-off fitting the needs of detecting and tracking scattered macro-litter patches or rows, though they could apply to micro-plastics.</p><p>Within the conventional framework of DCRIT (detection-classification-recognition-identification-tracking and targeting) and based on the classic methodologies derived from Multidimensional Signal Detection Theory (MSDT), we are currently developing a scheme to address the issue of recognising faint signatures of marine litter in Earth Observation (EO) data sets. Most of the RSS studies are focused on the detection of plastic using (a) its spectral signature over water through applying indices such as Normalized Difference Vegetation Index (NDVI) or Floating Debris Index (FDI) owing to the issue of EO pixel size greater than litter accumulation width, with (b) universal thresholds. In our case, we adjust the detection thresholds to the ‘a-priori’ information on litter presence, provided by a model, to the environmental andthe RSS observation conditions, balancing the probability of detection and false alarms using a Bayesian approach.The ‘detector’ is the heir of the binary classification algorithm developed by ARGANS Ltd on a grant by European Space Agency (ESA), which is abinary detector followed by a multi-label classification using a deterministic decision tree to distinguish natural from anthropogenic debris. The ‘a priori’ information is provided by a marine litter model deployed in the Black Sea, locating the main litter accumulation areas. Then, the posterior probability of the uncertain classification of pixels as plastic is the conditional probability that it is assigned considering the observation conditions and the plastics’ presence information coming from the model. To assess the confidence of detection, the Bayes theorem is combined with Receiver Operating Characteristic (ROC) curves. The latter ones can be used to assign higher probabilities to observations with a positive classification and lower probabilities to observations that do not. A further analysis combining both tools allows to improve the thresholds selection to classify pixels as plastic as a function of the background information.</p>

No abstract