Segmenting submarine boulders with Random Forest and nnU-Net: from bathymetric rasters to survey-scale size distributions

In the soft-bottom-dominated southern Baltic Sea, geogenic hard substrates are biodiversity hotspots. The lack of open-source, pixel-level segmentation approaches for boulders in MBES data has made it difficult to advance the ecological understanding of boulder fields in the Baltic Sea. This study developed and evaluated two segmentation approaches. A deep learning model (nnU-Net) and a classical machine learning model (APOC-25, based on Random Forest classification) were trained to segment boulders using multibeam echosounder (MBES) bathymetry data at scale. Both models were trained on manually annotated data. Additionally, a pipeline for extracting per-boulder dimension metrics from the resulting segments was developed. The trade-off between annotation effort and segmentation performance was assessed. The nnU-Net model achieved the best overall performance, with a recall of 0.77, a precision of 0.84, and an F1 score of 0.8. APOC-25, trained on a sparse subset of labelled pixels, reached a recall of 0.73, a precision of 0.51, and an F1 score of 0.60. Upscaling the input data improved detection performance for both models. Additional analyses of areal coverage bias and detection accuracy were performed, and predictions were compared against side-scan sonar (SSS) annotations at two sites. This study provides open-source tools for systematic boulder mapping, enabling the generation of ecologically relevant baseline data and facilitating habitat mapping, monitoring and restoration.