Detection of Boulders in Side Scan Sonar Mosaics by a Neural Network
Boulders provide ecologically important hard grounds in shelf seas, and form protected habitats under the European Habitats Directive. Boulders on the seaﬂoor can usually be recognized in backscatter mosaics due to a characteristic pattern of high backscatter intensity followed by an acoustic shadow. The manual identiﬁcation of boulders on mosaics is tedious and subjective, and thus could beneﬁt from automation. In this study, we train an object detection framework, RetinaNet, based on a neural network backbone, ResNet, to detect boulders in backscatter mosaics derived from a sidescan-sonar operating at 384 kHz. A training dataset comprising 4617 boulders and 2005 negative examples similar to boulders was used to train RetinaNet. The trained model was applied to a test area located in the Kriegers Flak area (Baltic Sea), and the results compared to mosaic interpretation by expert analysis. Some misclassiﬁcation of water column noise and boundaries of artiﬁcial plough marks occurs, but the results of the trained model are comparable to the human interpretation. While the trained model correctly identiﬁed a higher number of boulders, the human interpreter had an advantage at recognizing smaller objects comprising a bounding box of less than 7 × 7 pixels. Almost identical performance between the best model and expert analysis was found when classifying boulder density into three classes (0, 1–5, more than 5) over 10,000 m 2 areas, with the best performing model reaching an agreement with the human interpretation of 90%.