AI for Fisheries Science: Neural Network Tools for Forecasting, Spatial Standardization, and Policy Optimization

This paper explores the application of neural network tools, specifically LSTMs, CNNs, and Reinforcement Learning, to address key fisheries science challenges in forecasting, spatial standardization, and policy optimization, while acknowledging their tradeoffs and outlining future research directions.

The Gist

Imagine you are the captain of a massive fishing fleet. Your job is to catch enough fish to feed the world without wiping out the ocean's population. But the ocean is a chaotic, shifting place. Fish grow at different rates, they move to new areas, and the weather changes everything. Traditionally, captains have used old, rigid maps (statistical models) to navigate these waters. These maps are good, but they struggle when the terrain suddenly changes.

This paper is a "Food for Thought" proposal suggesting that we upgrade our navigation tools. Instead of just using old maps, the authors are testing Artificial Intelligence (AI)—specifically "Neural Networks," which are computer programs designed to think a bit like a human brain.

Here is a breakdown of the three main experiments (case studies) the authors ran, explained with simple analogies:

1. Predicting Fish Growth: The "Crystal Ball" vs. The "Smart Watch"

The Problem: To manage a fishery, you need to know how big the fish will be next year. Traditionally, scientists just take the average size of fish from the last five years or use a fixed formula (like a growth chart for humans). But fish don't always follow the chart; sometimes they grow fast, sometimes slow, depending on the environment.

The AI Test: The authors taught a computer two ways to guess the future size of fish:

• The Old Way: A simple average or a fixed formula.
• The New Way (LSTM): A special type of AI called a "Long Short-Term Memory" network. Think of this like a smart watch that doesn't just look at your current heart rate, but remembers your entire workout history, your sleep, and your diet to predict exactly how you'll feel tomorrow.

The Result: The "Smart Watch" (AI) was generally better at predicting fish sizes, especially when the environment was changing. However, when the fish were growing in a very predictable, steady way, the old formulas were just as good.

• The Catch: The AI is great at guessing what will happen, but it's a "black box." It can tell you the fish will be big, but it can't easily explain why (unlike the old formulas which give you a clear biological reason).

2. Mapping the Ocean: The "Photo Album" vs. The "Puzzle Solver"

The Problem: Scientists take surveys to count fish, but they can't check every square inch of the ocean. They only check specific spots, leaving huge gaps in the data. They need to fill in those gaps to create a complete map of where the fish are.

The AI Test: They tried using Convolutional Neural Networks (CNNs). These are the same AI tools used to recognize cats in photos. The idea was to treat the ocean like a photo: if you see fish in one spot, the AI should guess there are fish in the neighboring spots.

• The Old Way (tinyVAST): A sophisticated statistical method designed specifically to handle "puzzle pieces" (sparse data) and fill in the missing spots based on how fish usually move.

The Result: The "Photo Album" AI (CNN) struggled. It was designed to look at complete pictures, not scattered puzzle pieces. It couldn't fill in the gaps as accurately as the specialized "Puzzle Solver" (tinyVAST).

• The Lesson: Just because an AI is powerful doesn't mean it's right for every job. You wouldn't use a photo-recognition tool to solve a math equation; similarly, you need the right tool for sparse ocean data.

3. Setting the Catch Limit: The "Rulebook" vs. The "Video Game Player"

The Problem: Every year, managers have to decide: "How many fish can we catch?" Usually, they follow a strict rulebook (e.g., "If the population is X, catch Y"). This rulebook is static and doesn't learn from mistakes.

The AI Test: They used Reinforcement Learning (RL). Imagine a video game player who is trying to beat a high score.

• The Player (AI): Makes a move (sets a catch limit).
• The Game (Ocean): Responds (fish population goes up or down).
• The Score (Reward): If the player catches a lot of fish and the population stays healthy, they get a high score. If they overfish, they lose.
• The AI plays this "game" thousands of times, learning from its mistakes until it discovers a perfect strategy that no human ever thought of.

The Result: The AI player discovered some very weird, non-linear strategies. Sometimes it suggested catching nothing one year and double the usual amount the next, but it kept the fish population healthier and the total catch higher than the traditional rulebook.

• The Catch: The AI's strategy was so efficient but so strange that it might be hard to convince fishermen and politicians to trust it. It works, but it doesn't "feel" intuitive.

The Big Picture: What Does This Mean?

The authors are saying: "Don't throw away your old maps, but bring a GPS."

• The Good: AI can handle complex, messy data better than old methods. It can find patterns we miss and discover new, highly efficient ways to manage fisheries.
• The Bad: AI is a "black box." It gives answers but often can't explain the "why." In fisheries management, we need to understand why a decision is safe to trust it.
• The Future: We need to treat AI like a new, powerful tool in the toolbox. We need to test it rigorously against our old tools to see when it helps and when it fails.

In short: The ocean is changing fast. Our old ways of managing it are getting rusty. AI offers a shiny new way to navigate, but we need to make sure we understand how it works before we let it steer the ship.

Technical Summary

1. Problem Statement

Fisheries science faces unprecedented challenges due to shifting funding priorities and dynamic marine environments. Traditional stock assessment methods rely heavily on parametric, process-based models (e.g., Generalized Linear Models, von Bertalanffy growth curves) and statistical regression. While these methods facilitate inference, they are often constrained by rigid assumptions regarding data distribution and may struggle to capture complex, non-linear, and high-dimensional spatiotemporal patterns inherent in fisheries data.

The paper argues that Artificial Intelligence (AI), specifically Neural Networks (NNs), offers a less constrained alternative capable of learning latent patterns, handling non-local dependencies, and approximating complex relationships without explicit parametric assumptions. However, the application of advanced NN architectures (beyond basic machine learning) to foundational fisheries problems remains underexplored.

2. Methodology

The authors present a "Food for Thought" paper introducing three specific neural network subclasses and applying them to three foundational fisheries case studies. The methodology involves comparing NN approaches against established statistical benchmarks using simulated and real-world data scenarios.

A. Neural Network Architectures Used

• Long Short-Term Memory (LSTM): A type of Recurrent Neural Network (RNN) designed for sequential data. It utilizes "gates" and a cell state to manage long-term dependencies, making it suitable for time-series forecasting where past values influence future states.
• Convolutional Neural Networks (CNNs): Originally designed for image processing, these use convolutional filters to extract spatial features. The paper explores their ability to interpolate spatial latent variables based on geographic neighborhoods.
• Reinforcement Learning (RL): An agent-based approach where a neural network (the agent) learns a policy by interacting with an environment (the fishery system) to maximize a reward signal (e.g., sustainable catch), distinct from the static rules used in traditional Management Strategy Evaluation (MSE).

B. Case Study Designs

1. Case Study 1: Stock Demography Forecasting (Weight-at-Age)

   • Goal: Predict fish size-at-age.
   • Data: 300 simulated 30-year datasets based on Bering Sea pollock (Gadus chalcogrammus) life history. Two scenarios: time-invariant growth and time-varying growth (AR1 + trend).
   • Comparison: LSTM and simple NNs were compared against:
     • Simple 5-year mean ('mean-5').
     • Parametric von Bertalanffy Growth Function (VBGF).
     • VBGF with random effects (VBGF-RE).
     • Gaussian Markov Random Field (GMRF) with 3D AR1 effects.
   • Metric: Root Mean Squared Error (RMSE) via 10-fold cross-validation and retrospective forecast peels.

2. Case Study 2: Spatio-Temporal Standardization

   • Goal: Standardize survey data to produce annual abundance indices and continuous biomass surfaces.
   • Data: 20x20 spatial grid over 12 years with 50% random sampling (sparse data).
   • Comparison:
     • CNN: Processes spatial coordinates year-by-year.
     • CNN-LSTM: Incorporates temporal lag (previous 3 years) and binary masking for missing data.
     • tinyVAST: A state-of-the-art spatio-temporal statistical framework (Vector Autoregressive Spatio-Temporal).
     • Design-based estimator: Simple expansion of sampled means.
   • Metric: RMSE of log-biomass/log-density and visual inspection of interpolated surfaces.

3. Case Study 3: Policy Optimization via Reinforcement Learning

   • Goal: Discover optimal harvest policies to maximize catch while maintaining spawning biomass.
   • Data: Eastern Bering Sea pollock fishery simulation.
   • Comparison:
     • RL Agent: Learns a policy dynamically based on survey biomass and catch history.
     • Traditional MSE: Uses a static Harvest Control Rule (HCR) within a Management Strategy Evaluation framework.
   • Scenarios: Tested under "Constant recruitment" and "Ricker stock-recruitment" dynamics.
   • Metric: Median trajectory of spawning biomass and realized catches over projection periods.

3. Key Contributions

• Benchmarking NNs in Fisheries: The paper provides the first rigorous comparison of advanced deep learning tools (LSTM, CNN, RL) against modern statistical fisheries methods (GMRF, tinyVAST, MSE) across three distinct domains: forecasting, spatial interpolation, and policy optimization.
• Hybrid Architectures: It demonstrates the feasibility of combining CNNs (for spatial patterns) and LSTMs (for temporal dynamics) to handle complex spatio-temporal fisheries data.
• Policy Discovery: It introduces RL as a viable alternative to static HCRs, showing that agents can autonomously discover non-linear, adaptive policies that outperform traditional rules in complex, dynamic environments.
• Identification of Limitations: The authors explicitly highlight the "black box" nature of NNs, noting their lack of inferential capability (causal understanding) compared to process-based models, and their sensitivity to data sparsity.

4. Results

• Forecasting (Case Study 1):

  • Time-Invariant Growth: The LSTM model consistently achieved the lowest RMSE for hindcast and 1-2 year projections.
  • Time-Varying Growth: The GMRF (statistical model) performed best for hindcast and 1-year projections. However, for 2-year projections, the LSTM and simple mean performed comparably or better than GMRF.
  • Insight: NNs excel at pattern recognition in stable systems but may struggle to separate observation error from latent process variation without explicit structural constraints.

• Spatial Standardization (Case Study 2):

  • Performance: Neither the CNN nor CNN-LSTM outperformed tinyVAST in terms of RMSE for log-biomass or log-density.
  • Spatial Patterns: While CNNs identified general high/low biomass areas, tinyVAST was superior at recovering small-scale regional patterns.
  • Insight: CNNs, designed for complete images, struggle with the sparse, irregular data typical of fisheries surveys. tinyVAST remains the gold standard for index standardization in data-sparse contexts.

• Policy Optimization (Case Study 3):

  • Constant Recruitment: The RL agent learned a precautionary rule (zero catch below ~10,500 mt biomass) that maintained higher spawning biomass than the MSE scenario, despite large catch fluctuations.
  • Ricker Recruitment: The RL agent discovered a non-linear policy that stabilized biomass near 21% of unfished levels while sustaining catches (3.5 million mt) significantly higher than the MSE scenario (1–1.5 million mt).
  • Insight: RL can discover complex, adaptive policies that maximize yield without needing to step through an updated estimation model at every time step, offering potential efficiency gains.

5. Significance and Future Directions

• Complementary Role: The paper concludes that NNs should not necessarily replace process-based models but serve as powerful complements. They offer superior predictive performance in specific contexts (e.g., time-series forecasting) and novel capabilities (e.g., autonomous policy discovery).
• Research Roadmap: The authors identify critical research gaps:
  • Causal Inference: Developing methods to ensure NNs can distinguish correlation from causation, which is vital for policy decisions under novel conditions (e.g., climate change).
  • Data Sparsity: Investigating how to adapt CNNs for sparse, irregular fisheries data (e.g., using Vision Transformers or interpolated meshes).
  • Uncertainty Quantification: Expanding the vocabulary of fisheries science to include AI-specific uncertainty metrics (e.g., epistemic vs. aleatoric uncertainty).
  • Stakeholder Communication: Addressing the "black box" issue where RL-derived policies may be unintuitive, potentially hindering stakeholder buy-in.

Conclusion: This paper serves as a foundational "Food for Thought" for the fisheries community, demonstrating that while NNs are not yet a panacea for all fisheries modeling tasks, they hold significant promise for enhancing forecasting accuracy and discovering robust, adaptive management strategies in a rapidly changing ocean.