⚡ electrical engineering

A Gradient Boosted Mixed-Model Machine Learning Framework for Vessel Speed in the U.S. Arctic

This study employs a two-stage gradient boosted mixed-model framework on U.S. Arctic AIS data to demonstrate that distance to coast and bathymetric depth are the primary drivers of vessel speed, while effectively distinguishing between stationary and moving vessels to characterize navigational regimes.

Original authors: Mauli Pant, Linda Fernandez, Indranil Sahoo

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mauli Pant, Linda Fernandez, Indranil Sahoo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Arctic Ocean as a giant, icy highway where ships are the cars. For a long time, scientists trying to understand how fast these "cars" go have looked at the data as one big, smooth flow. But the authors of this paper realized that's like trying to understand traffic by averaging the speed of a car stuck in a parking spot with a car speeding down the interstate. It doesn't tell the whole story.

Here is a simple breakdown of what they did and what they found, using everyday analogies.

The Big Problem: The "Zero Speed" Trap

The researchers looked at a decade of data (2010–2019) from ships in the U.S. Arctic. They noticed something weird: more than half the time, the ships weren't moving at all. Their speed was exactly zero.

If you try to build a single math model to predict speed, treating "stopped" and "moving" as the same thing is like trying to predict the temperature of a room by averaging a block of ice with a hot stove. You get a number that doesn't make sense for either.

The Solution: They built a two-stage machine learning system, like a two-step security check:

Step 1 (The Gatekeeper): First, the computer asks, "Is the ship moving or stopped?" It looks at the conditions to guess if the ship is likely to be cruising or sitting still.
Step 2 (The Speedometer): If the answer is "Yes, it's moving," then the computer asks, "How fast is it going?"

The Ingredients (What they fed the computer)

To make these guesses, they fed the computer a mix of data, like ingredients in a recipe:

The Ship's Identity: What kind of boat is it? (A tugboat, a cargo ship, a cruise liner?)
The Ship's Mood: Is it "underway using engine," "at anchor," or "fishing"?
The Map: How far is it from the shore? How deep is the water?
The Weather: How strong is the wind? How much sea ice is there?
The Steering: Is the ship going straight, or is it zig-zagging?

The Results: What Actually Drives the Speed?

The computer learned some surprising things about what makes a ship stop or speed up in the Arctic.

1. The "Shoreline Rule" is King
The single biggest factor wasn't the ice or the wind; it was how close the ship was to the land.

Near the shore: Ships are almost always stopped or moving very slowly. It's like a busy city intersection where cars are constantly stopping, turning, and waiting.
Out at sea: Once the ship gets far enough from the coast, it speeds up and stays steady. It's like hitting the open highway where you can cruise without stopping.

2. Steering is a Speed Killer
The computer noticed that if a ship is changing its direction a lot (zig-zagging), it's likely moving slowly or stopped. If it's going in a straight line, it's likely moving fast. Think of it like a runner: if they are constantly turning corners, they can't run as fast as if they were running down a straight track.

3. The Ice and Wind Surprise
You might think sea ice and wind are the main reasons ships slow down. The study found that while they matter, they are actually secondary players.

The "Ice Risk" is real, but in the summer months (July–October) when they studied, the weather and ice weren't the primary reason ships were stopped. The location (near shore) and the ship's activity (turning vs. going straight) mattered much more.

4. The Type of Ship Matters

Fast movers: Passenger ships and tankers tended to go faster when they were moving.
Slow movers: Cruise ships, fishing boats, and dredgers tended to be slower, even when they were "underway."
The Anchors: If a ship's status said "at anchor," the computer knew with near certainty it was stopped (0 speed).

Why This Matters (In Simple Terms)

The authors say this two-step method is better than old methods because it separates the "parking lot" behavior from the "highway" behavior.

Old way: "The average speed in this area is 4 knots." (This hides the fact that some ships are stopped and some are going 8 knots).
New way: "In this area, ships are likely stopped because they are near the shore and turning. But if they do move, they will go at a moderate speed."

The Bottom Line

The study concludes that to understand ship speeds in the Arctic, you have to look at where the ship is and what it is doing (steering vs. cruising) first. The ice and wind are important, but they are not the main bosses of the ship's speed in the way we might have thought. By separating the "stopped" ships from the "moving" ships, the researchers got a much clearer, more accurate picture of how these vessels operate in the wild, icy north.

Technical Summary: A Gradient Boosted Mixed-Model Machine Learning Framework for Vessel Speed in the U.S. Arctic

Problem Statement
Understanding the drivers of vessel speed over ground (SOG) in the Arctic is essential for managing navigational safety and environmental impact, particularly as climate change increases seasonal traffic. Existing literature often relies on analytical models or treats vessel speed as a single continuous process. However, Arctic AIS data exhibits "strong zero inflation," with over 53% of observations recording zero speed. Treating zero and positive speeds as a single continuous variable obscures distinct operational regimes (stationary vs. active transit). Furthermore, previous studies frequently analyze environmental drivers (sea ice, wind, bathymetry) in isolation or fail to account for the non-independence of repeated observations from the same vessel, location, or time period. This limits the ability to characterize the nonlinear, multivariate drivers of Arctic navigation.

Methodology
The authors propose a two-stage machine learning framework utilizing GPBoost (Gradient Boosted Decision Trees with Grouped Gaussian Random Effects). This approach explicitly separates the modeling of zero-speed events from positive-speed events.

Data: The study analyzes 14.8 million AIS observations from the U.S. Arctic (July–October, 2010–2019). Covariates include sea ice concentration, wind components (along-track and cross-track), bathymetric depth, distance to coast, course over ground (COG), vessel group, and navigational status. Environmental data was collocated with AIS positions on a 0.5° grid using geostatistical interpolation (kriging) to ensure complete coverage.
Stage 1: Binary Classification: A GPBoost model with a Bernoulli likelihood and probit link predicts the probability of $SOG > 0$.
Stage 2: Conditional Regression: For observations where $SOG > 0$, a GPBoost model predicts the magnitude of speed. The response variable is the square root of SOG ( $\sqrt{SOG}$ ) to mitigate skewness.
Random Effects: Both stages incorporate grouped Gaussian random effects to account for correlation within specific vessels (MMSI), spatial grid cells, and daily time indices. This structure separates systematic covariate effects from vessel-specific, spatial, and temporal variability.
Interpretability: The authors employ SHAP (SHapley Additive exPlanations) values to decompose model predictions, quantifying the marginal contribution of each covariate to the probability of movement and the magnitude of speed.

Key Results

Model Performance:
- Classification: The model achieved strong discrimination between zero and positive SOG, with an out-of-fold Area Under the Curve (AUC) of 0.845.
- Regression: The conditional speed model explained approximately 77% of the out-of-fold variance ( $R^2 \approx 0.77$ ) with a Mean Absolute Error (MAE) of roughly 1 knot.
Dominant Drivers:
- Distance to Coast: Identified as the most influential predictor in both stages. Proximity to shore strongly reduces the probability of positive SOG and limits the magnitude of speed. The effect is highly nonlinear, stabilizing only beyond 40–50 km offshore.
- Change in Course ( $\Delta COG$ ): The second most influential factor. Small heading changes correlate with sustained transit and higher speeds, while large or frequent changes indicate maneuvering and reduced speed.
- Bathymetric Depth: A significant determinant, with deeper waters supporting higher speeds and a higher likelihood of movement compared to shallow coastal regions.
Secondary Drivers:
- Vessel Group and Status: These factors show systematic effects (e.g., Status 0 "underway using engine" strongly predicts positive SOG; Status 1 "at anchor" strongly predicts zero SOG) but are less influential than spatial and steering variables.
- Environmental Factors: Sea ice concentration and wind showed modest average influence on global SHAP rankings. The authors note that while these factors are critical in specific hazardous conditions, their effects are attenuated when averaged across the diverse operational contexts of the open-water season.

Significance and Contributions
The paper claims to advance the literature by introducing a unified framework that treats zero and positive speeds as distinct operating regimes rather than a single continuous process. By separating stationary behavior from active navigation, the study provides an interpretable characterization of Arctic vessel operations that distinguishes stopping/maneuvering patterns from sustained transit.

The integration of structured random effects within a gradient boosting framework allows for the resolution of nonlinear environmental responses while accounting for the correlated nature of AIS data. The results empirically characterize Arctic vessel operating regimes, suggesting that spatial context (distance to coast) and steering behavior ( $\Delta COG$ ) are the primary determinants of vessel speed, outweighing vessel type and environmental factors like ice and wind in a general sense.

This data-driven characterization offers a basis for speed management and corridor-level assessment, highlighting that navigational and ecological risks often co-occur in nearshore regions where vessels are more likely to be stationary or maneuvering slowly, rather than in offshore transit corridors. The authors conclude that while threshold-based ice risk metrics are informative, they cannot capture the complex, multivariate drivers of vessel speed observed in practice, necessitating the type of statistical characterization provided by this study.