Evaluating textural descriptors for automated image classification of stony reefs in turbid temperate waters

The rise of machine learning (ML) techniques has made automatic image classification increasingly relevant and essential for marine biologists. Despite advancements in computational power and growing interest in the field, underwater image analysis remains a significant challenge, especially in highly turbid environments. This study is the first to assess the potential of texture descriptors for classifying benthic species and habitats using turbid underwater imagery. Underwater images were collected in SE Baltic Sea reefs (4.4–42.2 m depth) using a drop-down camera. A total of sixteen textural descriptors were tested, of which three were selected for the CatBoost ML model image classification task. The model's performance was evaluated using annotated images provided by field experts. Among these, the MRELBP (Median Robust Extended Local Binary Pattern) algorithm achieved the highest overall performance. For individual classes, the best image classification results were achieved for large blue mussels by the LMP (Local Morphological Pattern) algorithm (F1 score: 0.72 ± 0.18) and small blue mussels (F1 score: 0.66 ± 0.13) by MRELBP. For lithological classes, sand was classified with the highest accuracy by MRELBP (F1 score: 0.69 ± 0.23). Model coverage estimates were acceptable in 49 % of the images, with blue mussels being the most suitable for evaluation. The results demonstrate textural descriptors capabilities in classifying real-world underwater images.