Predicting Barge Tow Size on Inland Waterways Using Vessel Trajectory Derived Features: Proof of Concept

Summary

This paper introduces a novel method to predict the number of barges being towed by a tug/towboat on inland waterways using Automatic Identification System (AIS) vessel tracking data and Machine Learning (ML). The study addresses the critical challenge that barges are non-self-propelled and unpowered, meaning they lack AIS transponders, creating a black box problem that hampers efficiency, compromises safety, and limits strategic planning across the waterway ecosystem.

The objectives of the paper are threefold: (i) to estimate the number of barges being towed by a tug/towboat with minimal error, leveraging both static vessel geometry and dynamic movement patterns derived from AIS data; (ii) to identify the most influential predictors of barge count; and (iii) to establish a foundation for future spatial generalization of barge count prediction models across diverse riverine environments.

The methodology integrates computer vision, spatiotemporal data fusion, trajectory analysis, and machine learning. Two complementary datasets are required: vessel trajectories from AIS data (including vessel IDs, locations, speeds, and directions) and satellite imagery from commercial sources (e.g., Planet Labs or Maxar). Although the aim is to use only AIS-derived features for prediction, an interim process eases manual sample labeling by applying pre-trained machine learning models (e.g., YOLO) to automatically detect vessel and barge objects from satellite imagery, which are then manually classified and matched to AIS records.

A spatiotemporal matching procedure associates vessels detected in satellite imagery with AIS records. For every satellite scene, AIS messages transmitted within two minutes before or after the image acquisition timestamp are selected. After vessels are identified in imagery using YOLO-based detection and bounding boxes are georeferenced, AIS records are grouped by vessel identifier (e.g., MMSI) and chronologically ordered to form paths. A detection is considered a valid match if its polygon intersects with an AIS trajectory and falls within the defined temporal window.

Once AIS metadata is linked to detected vessels, vessel trips are extracted by identifying stop locations using a density-based clustering algorithm with low-speed thresholds (e.g., <2 knots), minimum stop durations, and spatial clustering of AIS pings. These stops serve as origin and destination points, with each trip defined as the movement segment between two consecutive stop points. The trip that includes the satellite capture time is selected as the matching trip.

A comprehensive set of 30 AIS-derived features capturing vessel geometry, dynamic movement, and trajectory patterns was created. These features were grouped into categories such as speed and acceleration dynamics, directional stability, and operational patterns. The feature set was guided by the premise that vessels towing larger barge configurations tend to exhibit reduced maneuverability, smoother courses, and fewer directional changes. Features include vessel length, width, draft, mean/median/standard deviation of speed, interquartile range of speed, median absolute deviation of speed, maximum/minimum/range of speed, coefficient of variation of speed, percent time at low/high/optimal speed, entropy of speed distribution, mean positive/negative acceleration, standard deviation of acceleration, maximum deceleration, acceleration sign changes, standard deviation of course, entropy of course over ground, standard deviation of turn rate, total absolute course change, mean/standard deviation of difference between course over ground and heading, total trip time, distance traveled, direct straight-line distance, sinuosity index, vessel area, draft to length ratio, and various interaction/polynomial terms.

Six regression models from different algorithmic families were selected for comparison: three ensemble-based models (Random Forest Regressor, Catboost Regressor, AdaBoost Regressor), one kernel-based model (Support Vector Regressor), one regularized linear model (ElasticNet), and one classical count-based model (Poisson Regressor). Recursive Feature Elimination with Cross-Validation (RFECV) was employed for feature selection, iteratively training a model, ranking features by importance, and removing the least influential features while evaluating predictive performance via cross-validation. This procedure was applied separately across all six models, tailoring feature selection to each model's learning characteristics.

A stratified K-fold cross-validation approach was used with 2 folds due to the small dataset size (26 labeled instances). The data was stratified by the number of barges attached to each vessel so that each fold preserved the overall distribution of barge count classes. Model evaluation used Mean Absolute Error (MAE), defined as the absolute error in barge count estimation.

The case study covered a 240-mile section of the Mississippi River from Baton Rouge to the Gulf of Mexico, the nation's most active corridor for waterborne commercial traffic, with data gathered between January and April 2024. High-resolution satellite imagery (3 m) from Planet Labs was used, with a custom API processing AIS tracks and detecting spatiotemporal overlap with Planet imagery tiles. AIS data was recorded at 1-minute intervals across U.S. inland waterways, sourced from the Marine Cadastre Web Tool. Stop identification parameters included a maximum speed threshold of 1.0 knot, minimum stop duration of 60 minutes, and a spatial radius of 300m.

The final annotated dataset consisted of 26 labeled instances, each with an associated barge count, representing moving tows. Smaller barge configurations (e.g., two to four barges) were more frequently observed, while larger configurations (e.g., 12 or more barges) were less common, showing class imbalance which can influence prediction accuracy for less frequent barge counts.

Among the six regression models tested, the Poisson Regressor outperformed all others, achieving an average MAE of 1.920 barges across the two folds, using 12 of the 30 features. The Support Vector Regressor followed closely with an MAE of 1.922 barges using one feature (DUR SOGCV). AdaBoost ranked third with an MAE of 2.784 barges using two features (COG ENT, DIST KM). The Random Forest Regressor produced an MAE of 3.525 barges based on four features (ACC ZC, COG ENT, DIST KM, DUR SOGCV). The Catboost Regressor and ElasticNet exhibited the highest error levels, with MAEs of 4.148 and 4.445 barges respectively; Catboost employed 28 features, whereas ElasticNet used two (SOG PCT LOW, COG STD).

For the Poisson Regressor, in Fold 1 the model achieved an MAE of 2.482, overpredicting some zero-barge cases and underpredicting the single large tow (predicted 8.25 vs. actual 12). Fold 2 produced a lower MAE of 1.301, with most mid-range counts estimated within roughly ±1 barge of their true values. The results indicate that while the Poisson Regressor is well-calibrated for common, low-to-mid tow sizes, it still struggles with rare, high-barge outliers.

Feature importance analysis revealed that Entropy of Course Over Ground (COG ENT) and Trip Duration interacted with Speed Variation (DUR SOGCV) are jointly the most frequently selected features across all regression models. COG ENT quantifies the complexity of a vessel's heading profile; large barge tows, with their high inertia and limited turning radius, force tugboat operators to plan and execute only the most gradual, predictable maneuvers, smoothing out the heading signal, resulting in lower entropy. DUR SOGCV captures the interaction between trip duration and speed variability; longer tows with many barges require very consistent speed profiles to safely manage increased mass and drag, so trips that combine extended durations with tightly regulated speeds become strong indicators of high barge counts. Features not selected, such as maximum speed (SOG MAX) or mean positive acceleration (ACC POS MEAN), were likely excluded due to redundancy, lower predictive power, or misalignment with behavioral patterns.

The study concludes that the proposed approach provides a scalable, readily implementable method for enhancing Maritime Domain Awareness (MDA), with strong potential applications in lock scheduling, port management, and freight planning. The ability to predict the exact number of barges in an approaching tow in real-time would allow lock masters to optimize queuing, pre-position resources, and manage water levels, addressing the mismatch between modern tow sizes and legacy lock dimensions. Real-time data also provides value to public agencies like the U.S. Army Corps of Engineers, Departments of Transportation, and the Maritime Administration for planning, performance monitoring, and economic analysis.

Future work will expand the proof of concept to explore model transferability to other inland rivers with differing operational and environmental conditions, including the Ohio River (wider channel, higher traffic density), the Arkansas River (more constrained navigation), and the Columbia River (influenced by tidal patterns and hydropower infrastructure). Future work also aims to expand the dataset to include a broader range of vessel types, tow configurations, and geographies; incorporate environmental conditions such as river flow velocity and direction, wind speed and direction, water levels, turbulence, visibility conditions, channel width, and seasonal hydrologic variability; and explore more advanced deep learning architectures such as recurrent neural networks (RNNs) or graph neural networks (GNNs) to capture complex spatial and temporal dependencies.

Based on the paper, here are the specific improvements I can make to an AI system for barge count prediction, along with what the improved system can do:

• Improvement: Integrate the paper's finding that course entropy (COG ENT) and trip duration × speed variation (DUR SOGCV) are the most predictive features. Build a feature extraction module that computes these 30 kinematic features from raw AIS trajectories, with emphasis on directional stability and speed consistency metrics.

• What it can do: The system can automatically transform raw AIS pings into a rich feature vector that captures vessel maneuverability constraints imposed by tow size, enabling accurate barge count estimation without needing direct barge observation.

• Improvement: Implement the Poisson Regressor as the core predictive model, paired with RFECV for automatic feature selection. This combination achieved the best MAE (1.92 barges) using only 12 of 30 features, outperforming ensemble and kernel-based models.

• What it can do: The system can predict barge counts in real-time with minimal error, while automatically identifying and discarding redundant or noisy features, reducing computational overhead and improving generalization.

• Improvement: Build a two-stage pipeline: (a) satellite imagery + AIS matching using a 2-minute temporal window and polygon-intersection spatial matching, and (b) AIS trajectory reconstruction via density-based stop clustering (speed 60 min, radius 300 m). This pipeline generates high-quality labeled training data.

• What it can do: The system can automatically generate ground-truth labels for barge counts from satellite imagery, eliminating the need for manual annotation and enabling continuous model retraining as new data becomes available.

• Improvement: Incorporate stratified K-fold cross-validation (specifically 2-fold given the small dataset) that preserves the distribution of barge count classes. This prevents under-representation of rare large tows (e.g., 12+ barges) during training.

• What it can do: The system can produce reliable performance estimates even with imbalanced datasets, ensuring that predictions for uncommon but critical large tows are not systematically biased.

• Improvement: Add a feature importance analysis layer that ranks predictors by frequency of selection across models. This provides interpretability and guides future feature engineering efforts.

• What it can do: The system can explain why a particular barge count was predicted (e.g., low course entropy and high trip duration × speed variation indicate a large tow), making it useful for operators who need to trust and audit the model's decisions.

• Improvement: Design the model architecture to be spatially agnostic, with features that are normalized by vessel geometry (e.g., draft-to-length ratio, speed coefficient of variation) rather than absolute values. This allows retraining on new rivers (e.g., Ohio, Arkansas, Columbia) with minimal data collection.

• What it can do: The system can be deployed on any inland waterway with AIS coverage, adapting to local operational and environmental conditions (e.g., tidal influence, channel width, traffic density) with only a small set of labeled samples for fine-tuning.

1. Ingest raw AIS data (1-minute interval) and satellite imagery (e.g., Planet Labs 3m resolution) for any river segment.

2. Automatically match vessels between the two data sources using spatiotemporal fusion, generating labeled barge count samples.

3. Reconstruct vessel trips by identifying stops and segmenting trajectories, then compute 30 kinematic features per trip.

4. Predict barge count in real-time using the Poisson Regressor with RFE-selected features, achieving MAE < 2 barges.

5. Provide confidence intervals and feature importance explanations for each prediction.

6. Retrain continuously as new satellite imagery becomes available (e.g., daily updates), improving accuracy over time.

7. Deploy across multiple rivers with minimal adaptation, enabling lock scheduling, port management, and freight planning at a national scale.

This system directly addresses the paper's stated objectives: minimal-error barge count estimation, identification of influential predictors, and a foundation for spatial generalization.

Accurate, real-time estimation of barge quantity on inland waterways remains a critical challenge due to the non-self-propelled nature of barges and the limitations of existing monitoring systems. This study introduces a novel method to use Automatic Identification System (AIS) vessel tracking data to predict the number of barges in tow using Machine Learning (ML). To train and test the model, barge instances were manually annotated from satellite scenes across the Lower Mississippi River. Labeled images were matched to AIS vessel tracks using a spatiotemporal matching procedure. A comprehensive set of 30 AIS-derived features capturing vessel geometry, dynamic movement, and trajectory patterns were created and evaluated using Recursive Feature Elimination (RFE) to identify the most predictive variables. Six regression models, including ensemble, kernel-based, and generalized linear approaches, were trained and evaluated. The Poisson Regressor model yielded the best performance, achieving a Mean Absolute Error (MAE) of 1.92 barges using 12 of the 30 features. The feature importance analysis revealed that metrics capturing vessel maneuverability such as course entropy, speed variability and trip length were most predictive of barge count. The proposed approach provides a scalable, readily implementable method for enhancing Maritime Domain Awareness (MDA), with strong potential applications in lock scheduling, port management, and freight planning. Future work will expand the proof of concept presented here to explore model transferability to other inland rivers with differing operational and environmental conditions.