Assessing the Interpretability–Performance Trade-Off of Artificial Neural Networks Using Sentinel Fish Health Data

Round 1

Reviewer 1 Report

The article examines the trade-off between interpretability and predictive accuracy of neural network (NN) and regularized regression (EN) models in environmental effects monitoring, using the health of trout-perch near the Oil Sands Region in Alberta, Canada, as a case study. While neural networks offer higher predictive accuracy, their complexity reduces interpretability, a critical factor in ecological studies. A hybrid method combining EN for variable selection and NN for modeling is proposed, providing a balance between accuracy and interpretability. This hybrid approach proves to be the best choice for the specific application of the study, recommending more congruent and consistent data collection to improve environmental monitoring analyses.

Comments:

Abstract and Introduction (sections 1 and Abstract): It would be beneficial to expand the discussion on the importance of interpretability in ecological models by offering concrete examples of how a lack of interpretability can affect environmental management decisions. This would strengthen the argument on the importance of finding a balance between accuracy and interpretability.

Materials and Methods: In the description of the data and analysis techniques, providing additional details on the variable selection process could help readers better understand how variables were chosen for the analysis and their ecological relevance.

Results: The presentation of the results could be improved by including more visualizations, such as comparison charts between models, that clearly illustrate the differences in predictive performance and interpretability among the analyzed methods.

Discussion: While the discussion does a good job linking the results to the study objectives, it could be enriched with a more detailed comparison to previous works, especially studies that have used similar approaches in different environmental monitoring contexts. For example, you work as Real-time on-road monitoring network of air quality DOI 10.3303/CET1974041.

Conclusion: The section could benefit from a deeper reflection on the long-term implications of using hybrid NN-EN models in environmental monitoring. Consider adding suggestions for future research that might explore further combinations of modeling methods or applications in different environmental contexts.

Author Response

Please see attached document.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper reviews the modeling of fish biological parameters with water quality data. The author explores three different approaches: Elastic Net (EN), standard Neural Network (NN), and hybrid Neural Network (NN hybrid). Generally, the paper is well-written and structured. However, it does not seem scientifically sound to me, as the authors merely compare various models and assess their performance, and the conclusions appear obvious. Nonetheless, I agree that this type of analysis could be of interest to certain audiences.

To enhance its relevance to the intended audience, I would suggest the following changes, primarily in the methodology section and discussion section:

• - Title: The current title is too ambitious for a paper that fits fish biological variables with water quality data using NN and EN models (already published).
• - Main Objective: Define more clearly in the introduction section. Also, justify the need for this study based on previous references, as similar approaches already exist in the literature.
• - Methodology: Please define the set of variables used for fitting purposes (water quality) for the model (the ones defined in lines 115-116?), include the code (which would significantly enhance the paper), the libraries, and the version of R used. In its current form, reproducibility is not possible.
• - Results and Discussion: No comments are made on hyperparameter optimization and the structure of the final neural network.
• Expanding the discussion to consider the trade-off between interpretability and performance is crucial. This discussion could consider the practical implications for environmental monitoring and management, enriching the paper's contribution to the field. For example, how does the increased accuracy of NN models compare against the loss of interpretability in real-world monitoring scenarios?
• The discussion on integrating these modeling approaches with decision-making processes in environmental management is missing. Assessing not only the models' predictive performance but also their utility in informing effective conservation strategies would provide a significant contribution.

Author Response

Please see the attached document.

Author Response File: Author Response.pdf