Training data, model checkpoint, test areas, trawl mark density grids, and time-series slope data for automated trawl mark segmentation in the German Baltic Sea

This dataset accompanies the manuscript "Spatial extent of trawl marks in the German Baltic Sea Basins based on bathymetric grids" by Peter Feldens, Inken Schulze, Elisabeth Seidel, Svenja Papenmeier, Aicha Naumann, Daniel Oesterwind, Jacob Geersen, and Mischa Schönke (Leibniz Institute for Baltic Sea Research Warnemünde / Thünen Institut für Ostseefischerei) submitted to the Journal "Anthropocene". The data support a study that uses a U-Net convolutional neural network to segment trawl marks from multibeam echo sounder (MBES) bathymetric slope grids across approximately 1,069 km² of seafloor in the south-western German Baltic Sea. Study areas include Mecklenburg Bay, the Arkona Basin, Kiel Bay, and the Fehmarn Belt. Bathymetric data were collected between 2016 and 2025 at 1 m resolution (interpolated to 0.25 m) and provided primarily by the German Federal Maritime Agency (BSH).Download links for the BSH data are given in the paper manuscript and are available from https://marine-data.de/ . The dataset is organized into four folders, which are here available as zip files:Training_Data/Contains the image patches and masks used to train the U-Net segmentation model, together with the trained model checkpoint. Images are 256×256 pixel PNG patches of slope data covering training areas in Mecklenburg Bay, Kiel Bay, Fehmarn Belt, and the Arkona Basin (German Baltic Sea). The dataset comprises 2,318 patches, each augmented 6 times (rotation, flips, brightness/contrast changes, elastic transformations, random cropping). Masks are binary: white pixels indicate trawl marks, black pixels indicate background. The model checkpoint (checkpoint_epoch13_mbes.pth) is for a U-Net with a ResNet-34 encoder (Segmentation Models PyTorch library), pre-trained on ImageNet. Test_Areas/Contains five independent test areas (Test1–Test5) not included in the training data, used for model validation. Each test area includes a slope GeoTIFF (.tif with .tfw world file and .prj projection file), a manually annotated binary mask GeoTIFF, and a shapefile (.shp, .shx, .dbf, .prj) defining the test area boundary. Test areas are distributed across Mecklenburg Bay, the southern Arkona Basin, Kiel Bay, and the Fehmarn Belt. Grid_Densities/Contains trawl mark density grids as GeoPackage files (.gpkg) for each investigation area: Mecklenburg Bay (MB), Arkona Basin areas 1 and 2 (ARK, ARK2), Kiel Bay (KB), and the Fehmarn Belt. Fehmarn Belt split into a reference area (REF; towards west) and a Marine Protected Area (MPA; towards east). For the MPA area, densities for 2025 are included as well, establishing a 2-year time series. Density values are computed within a regular 25 m × 25 m grid and normalized to 0–1, where 0 indicates no trawl marks and 1 indicates complete coverage. Polygons smaller than 100 m² were filtered to reduce noise as described in the paper. Time_Series/Contains slope GeoTIFFs for the Fehmarn Belt Marine Protected Area (FB-MPA) from 2024 and 2025, used to assess temporal changes in trawling intensity. The data were recorded on Elisabeth Mann Borgese cruises EMB345 (cruise report: doi:10.48433/cr_emb345) and EMB369 (cruise report: doi:10.48433/cr_emb369 ) in 2024 and 2025. These time-lapse datasets document seafloor regeneration timescales and trawl mark density in the two survey years.