Abstract
The Coupled Growth and Survival (CGS) model is a novel Bayesian neural network designed to simultaneously predict fish growth and survival within interconnected river systems. By integrating biologically linked processes and implementing "cohort fairness" mechanisms, the CGS framework significantly improves prediction accuracy and reduces computational runtimes from weeks to approximately 30 minutes compared to traditional ecological models.
Introduction
Understanding fish population dynamics is critical for effective conservation and fisheries management. However, modeling these dynamics presents significant challenges due to biological interdependence, environmental complexity, and data sparsity. Traditional models often treat survival and growth as independent processes, losing valuable predictive power. Furthermore, field data is often noisy, leading to models that perform well on some age groups (cohorts) while failing on others.
Existing Bayesian frameworks are biologically expressive but computationally intensive, often requiring weeks of simulation. Conversely, standard machine learning models are fast but frequently ignore fundamental biological constraints. The CGS model bridges this gap by integrating biological priors into a deep learning architecture.
The CGS Framework
The authors propose the Coupled Growth-Survival (CGS) network, which integrates biological priors into a deep learning architecture to mimic real-world ecological relationships.
Model Components
- Survival Model ($M_s$): This component treats survival as a Markov process, utilizing neural networks to estimate the probability of a fish surviving and migrating between river segments at each time step.
- Growth Model ($M_l$): An LSTM (Long Short-Term Memory) network predicts growth rates based on environmental features and the fish's previous body length.
- Coupling Mechanism: The predicted body length from the growth model is fed directly into the survival model. This mimics the real-world relationship where size influences survival, allowing the two tasks to act as regularizers for each other.
Cohort Fairness and Knowledge Constraints
To prevent the model from ignoring underrepresented age groups, the authors introduced an Adaptive Model Refinement strategy. A "global referee" monitors performance across cohorts; if the error for a specific cohort is high, the referee dynamically increases the learning rate for those samples. Additionally, the model is constrained by ecological knowledge, such as seasonal growth patterns, preventing biologically unrealistic predictions.
Data and Experimental Design
The study utilized a comprehensive real-world dataset from the West Brook watershed in Massachusetts, US, focusing on Brook trout (Salvelinus fontinalis). The dataset spans from 2002 to 2015 across four connected river reaches, encompassing 11,768 individual fish and over 23,000 survival and length observations across 13 cohorts.
Results and Discussion
The CGS model demonstrated significant improvements over baseline models:
- Accuracy: CGS consistently outperformed independent RNN models. For body length prediction, it significantly reduced Root Mean Squared Error (RMSE), particularly in scenarios with sparse training data.
- Fairness: The fairness-aware variant (CGS_fair-ps) reduced performance disparity across cohorts by over 60% compared to the baseline CGS model.
- Biological Validation: The model successfully identified ecological patterns, such as seasonal growth peaks and physical barriers to movement (e.g., a waterfall in "River 4"), without explicit prior instruction.
- Efficiency: By transitioning from MCMC-based Bayesian sampling to a backpropagation-based neural network, the researchers achieved a massive speedup—from weeks of computation to roughly 30 minutes.
Conclusion
The CGS framework represents a significant advancement in ecological modeling, demonstrating that biological coupling and fairness mechanisms can enhance both the accuracy and reliability of machine learning in conservation. Future work will explore multi-species interactions and finer spatial resolutions to provide better insights into micro-habitat preferences.
Summary updated