Efficient Deep Learning for Maritime Surveillance: Real-Time Anomaly Detection and Self-Supervised AIS Trajectory Forecasting

Automatic Identification System (AIS) telemetry from hundreds of thousands of vessels defeats hand-crafted alert thresholds. We study two-stage screening on a 90-day per-vessel AIS history (December 2025–March 2026) from a commercial aggregator. The first stage, TSMixer v2, a 36 K-parameter all-multilayer-perceptron (MLP) classifier trained to reproduce weather-aware rule labels, reaches over three seeds a precision–recall area under the curve of 0.995 and an F1 of 0.967 on a vessel-grouped split. In an external, recall-type check with kinematic-only (no-weather) labels, vessels on UK or US sanctions lists were 1.53 times as likely as matched controls to exceed the controls’ 90th-percentile flag share (95% CI 1.31–1.76), although mean shares did not differ. The second stage forecasts 24 position reports from 48: DefenseTSMixer, regressing absolute coordinates, errs by 229 km out of sample against 20 km for persistence; a motion-gated hybrid of persistence and a displacement-target Transformer reaches 17.1 km on a larger test set. A per-class Robust-Z scorer with a temporally disjoint reference flags fewer rows than a global one (17.7 versus 26.4%), but not with a motion-matched reference (11.6 versus 12.0%); its flags agree weakly with the rule labels and are screening signals, not validated anomaly evidence.