LABPLAS Plastic Debris Classified Satellite Images for German Bight (2016-2024)

Here you find the satellite-derived maps (PNGs) of predicted locations of floating macroplastic aggregations between 2016 and 2024, for three areas of interest (AOI) in the German Bight, developed under the H2020 LABPLAS project. The maps are classified using the POS2IDON tool and two specific machine learning models (the UNET and XGBOOST models) and the true-color (RGB) images. The predicted locations of floating macroplastic aggregations are marked with large red dots in the classified maps.

The classification maps were created with POS2IDON pipeline, a tool to detect suspected locations of floating marine debris in Sentinel- 2 satellite imagery using machine learning. POS2IDON includes modules for data acquisition, pre-processing, and pixel-based classification using different machine learning models (e.g. Random Forest, XGBoost, Unet). The outputs of POS2IDON include classification maps for all the available Sentinel-2 imagery of a given region of interest and temporal period, specified by the user.

Using POS2IDON we performed a long-term analysis in three AOIs in the German Bight (North Sea), which is part of the LabPlas Case Study 1 area. To cope with potential false positives and retain the more relevant spatial and temporal patterns, POS2IDON was applied in the entire Sentinel-2 archive resulting in a massive processing of nearly 9 years of data (2016-2024). After running POS2IDON, classification images (for UNET and XGBOOST models) were obtained in the three AOIs. Subsequently, to decrease the number of potential false positives, automated post-processing procedures were applied, including the buffering of the clouds (from cloud mask and cloud class) and the removal of isolated MD pixels. Furthermore, images contaminated by clouds were automatically excluded using cloud cover percentage thresholds, and through visual analysis several images affected by sunglint and whitecaps were manually excluded. This resulted in a total of 95, 97 and 92 images for AO1, AO2 and AO3, respectively, that were used for long-term analysis and provided in this dataset.