Large-Scale Mapping of Boulder Distribution in Acoustic Backscatter Data of the Baltic Sea by Neural Networks

The identification of marine cobbles and boulders (stones) based on acoustic remote sensing is important for the detection, delineation and for an ecological assessment of important seafloor habitats. Due to the large areas involved and the required high-resolution data, a manual interpretation is not feasible. In recent years, automated methods for stone detection were developed. However, these developments were only applied in comparatively small proof of concept areas, and a common barrier to practical implementation by authorities is the required upscaling. This case study aims to apply automated methods for boulder detection based on convolutional neural networks to larger areas, by identifying and validating boulder densities over several hundred km² in the western Baltic Sea in acoustic backscatter data and derived datasets. The use of distributed training sites of less than 0.5 km²  in size is proposed to improve the model capacity to adapt to variations of boulder appearance in remote sensing data related to local geological variation and survey conditions. Distributed validation sites of similar size are suggested to provide quality control during reprocessing with adapted models. Current limitations for the automated identification of individual boulders in backscatter data are demonstrated, which can be caused by survey geometry, data quality or obstacles and seafloor with similar acoustic characteristics.