Marine ecosystems are facing significant challenges from intensified fishing, pollution, climate change, and biodiversity loss. Ecosystem risk assessments are vital for informing effective policies and management decisions. Traditional approaches, such as Ecosystem Models (EMs) and Marine Spatial Planning (MSP), often rely on expert knowledge, which introduces subjective assumptions. This study evaluates six unsupervised methods — four clustering algorithms (Multi K-means, Fuzzy C-means, X-means, and DBSCAN) and two machine-learning models (an Artificial Neural Network, ANN, and a Variational Autoencoder, VAE) — to assess marine ecosystem risk in the Mediterranean Sea automatically, using open-access data from 2017 to 2021. Each method generated five annual high-risk maps based on ecosystem variables, including fishing effort, species richness, depth, coastal proximity, oxygen levels, net primary production, and thermohaline circulation intensity. Our quantitative analysis of 30 generated maps revealed pairwise similarities ranging from 72.2% to 95.9%, with Cohen’s Kappa scores between 0.46 (moderate) and 0.91 (almost perfect). All methods consistently identified high-risk hotspots in the Eastern and Western Mediterranean, the Tyrrhenian Sea, the Adriatic Sea, the Strait of Sicily, and the Aegean Sea. However, we also found discrepancies due to the different tendencies of the models to produce broader (precautionary) or more focused (conservative) risk assessments. Assessments by DBSCAN, ANN, and VAE were similar (∼90%) and broader, whereas X-means was more conservative. Multi K-means and Fuzzy C-means exhibited similar (∼92%) and more balanced results. These findings provide a data-driven foundation and practical guidance for developing Bayesian EMs and MSP with reduced reliance on subjective assessments.
Keywords: Risk assessment; Cluster analysis; Machine learning models; Marine ecosystems; Comparative analysis.
