Predicting barge tow size on inland waterways using vessel trajectory-derived features: proof of concept

Inland waterways move substantial freight by barge, yet tow configuration, which is important for berth and lock scheduling, remains difficult to observe in near real-time because Automatic Identification System (AIS) transponders are carried by towing vessels rather than barges. This proof-of-concept study evaluates whether barge count can be inferred from AIS-derived towing-vessel trajectories. High-resolution satellite imagery along the Lower Mississippi River was used to label tow instances and link observed barge counts to corresponding AIS trajectories. For each matched trip, 39 features describing vessel geometry, speed and acceleration behavior, course and heading variability, and path efficiency were computed. Recursive feature elimination with cross-validation identified model-specific feature subsets, and six regression models were evaluated. Under a custom two-fold internal validation of 21 observations from non-singleton tow-size classes, the Poisson and support vector regressors produced nearly identical mean absolute errors of 1.920 and 1.922 barges, respectively. Five singleton observations were retained during training but excluded from held-out error estimation because their classes could not be represented in both folds. The Poisson regressor retained 12 features, whereas the support vector regressor retained one interaction feature. Course entropy and the interaction between trip duration and speed variability were the most frequently selected features across models. The findings provide preliminary evidence that tow size may be inferred from towing-vessel behavior even when barges are not represented in AIS messages. Broader validation using larger datasets, additional river corridors, nested model evaluation, and partial-trip trajectories is required before operational deployment.