An Explainable AI Approach to Dynamic Vessel Speed Regulation for Marine Mammal Conservation

Ship strikes represent a primary threat to marine mammals, with nearly 92% of whale habitats overlapping with shipping routes, yet static speed zones fail to account for dynamic whale distributions. A Random Forest classifier was trained on 47,156 sperm whale observations comprising 23,578 presence records balanced with equal pseudo-absences from random sampling in areas with no sightings, incorporating sea surface temperature, mixed layer depth, primary productivity, sea level anomaly, bathymetry, and bathymetric standard deviation. The model was trained on an 80% random split with five-fold cross-validation applied to the training set, achieving 85.97% accuracy and 0.9387 AUC-ROC (cross-validation: 85.31% ± 0.24%). SHAP identified bathymetric standard deviation and bathymetric depth as dominant predictors, followed by mixed layer depth and primary productivity, aligning with established ecological understanding of sperm whale habitat preferences in deep, variable, productive environments. Predicted presence probabilities (0–1) were categorized into five risk tiers with corresponding speed recommendations: Very Low (<0.2, 20 knots), Low (0.2–0.4, 17 knots), Medium (0.4–0.6, 14 knots), High (0.6–0.8, 11 knots), and Critical (≥0.8, 8 knots), with thresholds justified by IMO-recommended 8–10 knot speed limits for whale protection zones. This simulation demonstrates a conceptual framework for translating environmental risk predictions into vessel speed advisories; however, real-time implementation requires integration with AIS infrastructure and is presented as a proof-of-concept. This research advances environmental technology by demonstrating proactive, explainable whale presence prediction while addressing regulatory transparency requirements for dynamic, evidence-based environmental regulation.