Towards Explainable Deep Learning for Ship Trajectory Prediction in Inland Waterways

This study proposes an interpretable LSTM-based model for predicting ship trajectories in inland waterways that incorporates trained ship domain parameters to analyze attention mechanisms, revealing that while the model achieves competitive accuracy, its attention weights do not fully align with expected causal relationships between interacting vessels.

The Gist

The Big Picture: Predicting the Dance of River Boats

Imagine a busy river, like the Rhine in Germany. It's not just a flowing stream; it's a crowded highway where giant barges, small tugs, and ferries are all trying to get to their destinations without crashing.

The authors of this paper are trying to build a super-smart GPS that doesn't just tell a boat where it is, but predicts exactly where it will be in the next 5 minutes. Why? Because if you want to make river transport automatic (self-driving boats), the computer needs to know: "If that big barge turns left, will I need to turn right to avoid a collision?"

The Problem: The "Black Box" Mystery

In recent years, scientists have used Deep Learning (a type of AI) to get really good at this prediction. These AI models are like genius chefs who can cook a perfect meal, but they are also Black Boxes. You put ingredients in, and a perfect meal comes out, but you have no idea why they chose those specific spices.

The problem is: What if the AI is getting the right answer for the wrong reason?
Maybe the AI predicts a boat will turn left not because it saw another boat coming, but because it noticed the sun was setting. It works, but it's dangerous because it doesn't understand the real rules of the road.

The Solution: Giving the AI a "Social Bubble"

The researchers wanted to make the AI Explainable. They wanted to peek inside the Black Box to see how the AI is thinking.

To do this, they used a concept called a "Ship Domain."

• The Analogy: Imagine every boat has an invisible, personal bubble around it (like a force field). If another boat enters your bubble, you get nervous and react. If they are far away, you ignore them.
• The Innovation: Instead of giving every boat a fixed-size bubble (like a standard 100-meter circle), the researchers taught the AI to learn its own bubble size.
  • If two boats are heading toward each other (opposing), the AI learns to make the bubble huge because that's dangerous.
  • If two boats are moving in the same direction, the bubble might be smaller.

The AI has a "menu" of different situations (e.g., "Opposing boats," "Boats passing on the left," "Boats far away"). For each situation, it learns a specific number that represents how far away it needs to pay attention.

The Experiment: Three Different "Brains"

The team built three different versions of this AI to see which one was the most honest and accurate:

1. The "All-in-One" Chef (EA-DA): This model mixes everything together. It looks at the other boats and decides what to do all at once.
2. The "Simplified" Chef (E-DA): This model is a bit simpler, only looking at other boats at the very end of the thinking process.
3. The "Dual-Brain" Chef (E-DDA): This is the most interesting one. It splits the brain into two parts:
   • Brain A (Blind): Predicts where the boat would go if it were alone in the world.
   • Brain B (Attentive): Looks only at the other boats and decides how to adjust the path based on them.
   • Why do this? It's like having a driver who knows how to drive straight, and a co-pilot who only shouts, "Hey, there's a car coming!" This separation makes it easier to see if the co-pilot is actually doing anything useful.

The Results: The Surprise Twist

Here is where the story gets interesting.

• The Score: All three models were pretty good at predicting where the boats would end up (within about 40 meters of the actual spot after 5 minutes). They were all "smart."
• The Catch: When the researchers looked at the "Ship Domains" (the invisible bubbles) the AI learned:
  • The "All-in-One" and "Simplified" models learned something weird. They decided that for boats coming toward them, they should shrink their bubble to almost zero. They effectively ignored the oncoming traffic!
  • How did they still get the right answer? It turns out they were just guessing based on patterns in the data, not because they understood the danger of an oncoming boat. They were "cheating" by using other clues, not the interaction itself.
  • The "Dual-Brain" model was the only one that behaved logically. It actually grew its bubble for oncoming boats, just like a human captain would.

The Takeaway: Accuracy Isn't Everything

The main lesson of this paper is: Just because an AI gets the right answer, doesn't mean it understands the situation.

If you rely on the "All-in-One" model, it might work fine today. But if you put it in a new, weird situation it hasn't seen before, it might crash because it never actually learned the rule of "don't hit oncoming traffic."

By separating the "thinking" from the "reacting" (the Dual-Brain approach), the researchers proved that you need to check how the AI thinks, not just what it predicts.

What's Next?

The authors plan to use this "Dual-Brain" system to run "What-If" scenarios (Counterfactual Analysis).

• Example: "If that boat hadn't slowed down, would we have crashed?"
• This will help make river traffic safer and help humans trust the self-driving boats more, because they can see the logic behind the machine's decisions.

In short: They built a smart river navigator, realized some of them were just lucky guessers, and fixed the design so the AI actually understands the rules of the road.

Technical Summary

1. Problem Statement

Accurate ship trajectory prediction is critical for safety and traffic optimization in crowded inland waterways. While deep learning models have improved prediction accuracy, a significant gap exists in explainability.

• The Core Issue: Many "interaction-aware" models claim to account for ship-to-ship interactions (e.g., avoiding collisions), but it is often unclear how they do so. A model might achieve high accuracy by learning spurious correlations rather than genuine causal relationships regarding vessel interactions.
• Specific Challenge: In inland waterways, factors like river geometry, regulations, and fluid dynamics complicate trajectory prediction. Existing models often struggle to distinguish whether improved performance is due to actual interaction awareness or other factors.
• Goal: To develop a trajectory prediction model that not only achieves high accuracy but also provides interpretability regarding how and when the model considers interactions with other vessels.

2. Methodology

The authors propose an LSTM-based Encoder-Decoder architecture enhanced with learnable ship domain parameters and attention mechanisms. The approach is adapted from maritime shipping studies but tailored for the complexity of inland waterways.

A. Input Features

The model processes vessel states defined by:

• Longitudinal position: Waterway kilometer (wkm).
• Lateral position: Offset from the fairway center.
• Kinematics: Changes in position ($\Delta k, \Delta f$) between time steps.
• Context: Waterway curvature and direction.
• Output: Predicted positional distances for future time steps.

B. The "Learnable Ship Domain" Mechanism

Instead of using a fixed geometric ship domain (e.g., a fixed radius around a ship), the model learns a parameter tensor $S$ during training.

• Inputs to $S$: Discretized ship-to-ship relations:
  • $\Gamma$: Lateral distance.
  • $\Theta$: Relative direction (opposing, aligned, stationary).
  • $\Phi$: Rate of change in longitudinal distance.
• Function: $S(\Gamma, \Theta, \Phi)$ outputs a "domain value" (in wkm).
• Weighting Logic: The weight $w_{ij}$ assigned to a neighboring ship $j$ is calculated as:
  $$w_{ij} = \max(S(\dots) - \Delta_{ij}, 0)$$
  If the actual distance $\Delta_{ij}$ exceeds the learned domain value, the weight becomes 0 (the ship is ignored). If the distance is within the domain, the weight is proportional to the difference.

C. Model Variants

To investigate causality and explainability, three variants were developed:

1. EA-DA (Encoder-Attention-Decoder-Attention): Integrates attention in both the encoder and decoder. It fuses hidden states of all vessels using the learned weights.
2. E-DA (Encoder-Decoder-Attention): Removes attention from the encoder, applying it only in the decoder to isolate the impact of interaction awareness to the prediction phase.
3. E-DDA (Encoder-Dual-Decoder-Attention): The most distinct variant. It splits the decoder into two parallel paths:
   • BlindLSTM: Processes only the target vessel's own history (interaction-agnostic).
   • AttLSTM: Processes the weighted/attended hidden states of other vessels (interaction-aware).
   • Fusion: The final prediction combines outputs from both paths. This allows the model to predict a trajectory without interaction data if necessary, facilitating counterfactual analysis.

3. Key Contributions

• Explainable Architecture: The separation of the "selection of relevant ships" (via the learnable ship domain) from the "fusion of hidden states" (via attention) creates an intrinsically interpretable model.
• Causal Analysis: By using the E-DDA variant, the authors can explicitly determine if the model's accuracy relies on interaction data or if it learns to ignore it.
• Inland Adaptation: The study adapts maritime interaction models to the specific constraints of inland waterways (e.g., narrow channels, specific traffic rules).
• Critical Finding on "Interaction Awareness": The paper demonstrates that high prediction accuracy does not necessarily imply effective interaction modeling.

4. Experimental Results

• Dataset: AIS data from the Rhine River (km 595–611) spanning Jan 2021–Apr 2024 (3+ years).
• Metrics: Final Displacement Error (FDE) over a 5-minute horizon.
• Performance:
  • E-DDA achieved the best performance (Mean FDE: 38.4m, Median: 25.1m).
  • E-DA followed closely (Mean: 40.9m).
  • EA-DA performed slightly worse (Mean: 41.9m).
  • The interaction-agnostic benchmark (E-D) performed the worst (Mean: 45.1m).
• Interpretability Analysis (The Critical Insight):
  • E-DA & EA-DA: Despite good accuracy, the learned ship domain parameters showed counter-intuitive behavior. For opposing ships, the model decreased the domain value for negative distance changes (approaching ships), effectively ignoring them. The model achieved accuracy despite failing to properly model interactions, likely relying on other patterns.
  • E-DDA: This variant showed plausible behavior. The learned domain values increased for opposing, approaching ships, correctly indicating that these vessels are relevant to the prediction.
  • Conclusion: The E-DDA architecture successfully learned causal interaction rules, whereas the other variants achieved accuracy through non-causal correlations.

5. Significance and Future Work

• Beyond Accuracy Metrics: The study highlights that standard error metrics (like FDE) are insufficient for evaluating interaction-aware models. A model can be accurate but logically flawed (ignoring critical collision risks).
• Trust in Autonomous Systems: By ensuring the model's internal logic (attention weights) aligns with physical reality (ship domains), trust in autonomous inland shipping systems is enhanced.
• Future Directions:
  • Utilizing the architecture for counterfactual analysis (e.g., "What would the trajectory be if the other ship didn't exist?").
  • Refining the definition of encounter types.
  • Incorporating more sophisticated attention mechanisms.

In summary, this paper provides a framework for Explainable AI (XAI) in maritime trajectory prediction, proving that separating interaction modeling from prediction allows researchers to verify that a model is actually "thinking" about other ships, rather than just memorizing statistical noise.