Applying Deep Learning to Quantify Drivers of Long-Term Ecological Change in a Swedish Marine Protected Area.

In this study, we trained an object-detection model to classify 17 benthic invertebrate taxa in archived footage of a study site on the northern west coast of Sweden (a wall section of the Koster Fjord) within the Swedish marine protected area Kosterhavet National Park. The model displayed a mean average precision score of 0.738 and was applied to footage from 1997 to 2023, generating a dataset of 72,369 occurrence records. The dataset was used to quantify depth distributions and abundance trends of both individual taxa and functional groups over time. Depth distributions for 15 of 17 taxa occurred at depths ≥ 45 m. Distributions of 11 taxa aligned with empirical observations, and for the remaining six taxa, we propose expanded depth distributions in the area. Abundances over time significantly increased for eight taxa and decreased for five taxa, while the overall community structure throughout the study period shifted toward smaller, more heat-tolerant suspension feeders. We found that temperature preference and size were significant drivers of the observed abundance trends in individual taxa. Community structure was altered by the loss of large, heat-sensitive taxa to greater depths due to increased temperatures. We also observed a strong trend of increasing abundances in the remaining community, including six trawling-sensitive taxa, highlighting the effectiveness of the park's protective measures. To protect key cold-water species, we suggest that current fishery regulations of the national park should be expanded to deeper (colder) waters and that new marine protected areas should also be established in deep waters. Our study demonstrates the application potential of video surveillance combined with deep-learning technology, and we recommend the implementation of standardized video monitoring in marine ecosystem management.