Dynamic Inference of Regional Traffic States in a Dense Waterway Using AIS Window Sequences

Abstract

Regional vessel monitoring requires both a description of current traffic and an estimate of its near-term evolution. This study evaluates a lightweight approach to dynamic traffic-state inference using sequences of Automatic Identification System (AIS) observations. A lower Yangtze waterway case area is divided into spatial cells, and vessel-level observations are aggregated into 10-minute windows. The inference target is the proportion of observed vessels with mean speed below 2 kn in a future window. A gradient-boosted tree model combines current traffic summaries with two preceding windows. Experiments use previously extracted observations representing 1,806,306 AIS messages and 6,014 vessel identities over approximately 4.7 days, with chronological training, validation and test periods. On 7,094 test samples, the 10-minute forecast achieves a mean absolute error of 0.0719, compared with 0.0810 for persistence and 0.0762 for a current-window model. These correspond to relative reductions of 11.2% and 5.7%. Accuracy for three operational traffic states reaches 81.15%. The advantage in continuous prediction remains at 20- and 30-minute horizons, although abrupt changes remain difficult. The results support adding short histories to vessel-window monitoring in this case area. The inferred states describe observed speed composition and should not be interpreted as independently validated congestion or collision-risk levels.