Operationalizing the Precautionary Approach in Marine Pollution Control: A Machine Learning Framework to Examine Score Sensitivity in Potentially Polluting Shipwreck Risk Assessment

Potentially Polluting Wrecks (PPWs) represent a critical global threat to marine ecosystems, requiring proactive risk management aligned with the Precautionary Approach of international environmental law. However, standard multi-criteria risk assessment frameworks frequently suffer from sensitivity to criterion weights, where maximum point allocations determine the relative contributions of criteria to final risk scores. The consequences of these allocations for management classifications require explicit examination. To examine this sensitivity, this study presents a data-driven diagnostic framework integrating machine learning diagnostics with multi-way sensitivity simulations, applied to a comprehensive national fleet dataset (n = 1196) with targeted diagnostics on the dominant fishing vessel cohort (n = 926). Unsupervised clustering and supervised architectures (Decision Tree, Random Forest, and Logistic Regression) were deployed to evaluate parameter discriminative power and fleet stratification. The diagnostics revealed that Marine Environmental Sensitivity (C5) has a strong association with the composite score and sensitivity-based classifications. Through bivariate simulations, target-reduction scenarios were identified within a predefined search grid. The illustrative scenario (S3)—applying a 20.0% weight reduction in environmental sensitivity with proportional redistribution—reduced the managed count from 247 to 167; fishing vessels comprised 71.60% of the 81 excluded vessels. This framework quantifies how criterion weighting and redistribution shape management classifications, providing a transparent basis for reviewing priorities within the existing shipwreck assessment system.