Abstract High‐resolution mapping of seafloor habitats has wide applications for fisheries, conservation efforts, offshore infrastructure planning, mineral extraction, and scientific modeling. This study leverages existing widespread single‐beam acoustic data and machine learning to create habitat maps for the Gulf of Alaska at a spatial resolution of less than 30 m. A convolutional neural network was trained using concurrent underwater images as the ground truth. To extract habitat type from the images, a semantic segmentation based on a random forest model was used to classify the interspersed substrate types. For the acoustic data, two models were trained: a full model using five transducer frequencies, 18, 38, 70, 120, and 200 kHz, and a reduced two‐frequency model using data at 38 and 120 kHz. The acoustic inputs were normalized and scaled to reduce the effects of varying bottom depths and vessel speeds. Model outputs are seabed rugosity and proportions of substrates: sand/mud, gravel, cobble, boulder, and bedrock. Both models have similar performance, with errors of about 12% in compositions. Comparing results across different survey years shows an agreement of 95% at similar locations. The resulting substrate maps appear to correctly identify known geographic features in the Gulf of Alaska. Descriptions of co‐occurrence of substrate types, their spatial distribution, and correlations in the study area are also provided.
Agarwal et al. (Thu,) studied this question.