Review of ELE-01226-2022, “Integrating underlying autotrophic biomass dynamics into
predictions of river ecosystem productivity”

Blaszczak and colleagues conducted a modeling study to incorporate algal biomass dynamics into predictions of riverine gross primary productivity via a latent biomass state, where biomass can be scoured during high-discharge events with subsequent growth and recovery. They found that predictions using the latent biomass model performed similarly to a simple time series model without the biomass component, and that the latent biomass model was able to successfully recreate hypothesized hysteresis dynamics following storm events as primary productivity is temporarily decoupled from light availability due to scouring away of algal biomass.

The motivation for including additional process representation in the simple time series model structure is well laid out and the methods are sound. The results are clear and the figures are well-prepared and nicely communicate the take-home messages. Model predictive performance was robustly assessed, and the manuscript is well-written and a pleasure to read; the Methods and model descriptions in particular are exceptionally clearly written, which I very much appreciated. Overall, I think with some re-working of the framing and interpretation, this carefully-conducted and well-communicated study is a valuable contribution. Specifically, the study falls within a current topic of debate within ecological predictive modeling as to the value of representing ecological processes in models for increased ecological understanding vs. simply pursuing the model with the best predictive skill, regardless of interpretability.

Major comments:

My one major comment is that in the Conclusions, the authors state that “we demonstrate the ecological insights into the resilience of productivity and disturbance thresholds that can be inferred from a model with latent biomass dynamics without losing predictive capacity” (lines 400-402). However, while the latent biomass model does successfully reproduce hysteresis dynamics for an individual storm event (which is beautifully communicated in Fig. 4!), fundamentally, incorporating the latent biomass component doesn’t appear to substantially improve model predictions compared to the simple time series model (so predictive capacity is not lost, but it is also not gained). To my mind, this undermines the conclusion that addition of the latent biomass component provides ecological insight into the resilience of productivity and disturbance (though one could argue the insight is that scouring and recovery of algal biomass may not be as important to riverine GPP as previously thought), since incorporation of this process doesn’t improve GPP predictions. As another way of stating it, how much inference should be drawn from the fitted parameters in the process equations of the latent biomass model (eqs 5 and 7) if those equations don’t substantially improve model predictive skill compared to a simple linear model? In the Conclusions, the authors suggest that incorporation of additional process representation in a model tailored to one specific river site might improve predictions (lines 390-394), and this may be true. However, I tend to agree more with the authors’ point that additional model complexity can lead to additional complications in terms of data needs and challenges in scaling to a large number of ecosystems (lines 394-398), whereas the potential benefit of added complexity in terms of predictive capacity and understanding is putative and hasn’t yet been achieved by the addition of latent biomass in the current study.

One possible way of addressing my comment would be to consider adjusting the framing and interpretation of results, as other researchers have suggested or found through modeling studies that simpler models often outperform (or perform just as well as) more complex ones (e.g., Ward et al 2014, Currie 2019). I have provided some additional references that might give food for thought on this topic under minor comments in the Discussion.

Currie, D. J. (2019). Where Newton might have taken ecology. Global Ecology & Biogeography, 28, 18– 27. https://doi.org/10.1111/geb.12842

Ward, E. J., E. E. Holmes, J. T. Thorson, and B. Collen. 2014. “Complexity Is Costly: A Meta-Analysis of Parametric and Non-parametric Methods for Short-Term Population Forecasting.” Oikos 123: 652–61.

Minor comments:

Line 4: add comma after “rivers”

Line 5: The phrase “disturbances impart signals of ecological memory on productivity time series” struck me as somewhat jargon-y and a little difficult to understand at first; is there a way to simplify the phrasing here? The mechanism which follows (“by modifying underlying standing stocks of biomass”) does help with understanding but it would still be nice to simplify the phrasing. The concept is nicely explained on lines 20-21 in the Introduction so maybe some of that language could be borrowed.

Lines 64-69: It might be helpful to some readers to formally define the concepts of latent state/state-space modeling rather than leaving those terms to parentheticals; also, this could be another place to reference Hobbs and Hooten 2015, in case there is a reader who would like to learn more.

Hobbs, N. Thompson, and Mevin B. Hooten. Bayesian Models: A Statistical Primer for Ecologists. STU-Student edition, Princeton University Press, 2015. JSTOR, http://www.jstor.org/stable/j.ctt1dr36kz. Accessed 14 Feb. 2023.

Lines 234-235: Here and throughout, I appreciate the authors’ candor regarding their assumptions

Line 363: river should be pluralized

Line 372: add comma after rates

Line 385: I’ve pasted some possible interesting papers along this line of thinking below; of these two, I think Rastetter would be the best option for supporting this statement.

Rastetter, E.B. Modeling for Understanding v. Modeling for Numbers. Ecosystems 20, 215–221 (2017). https://doi.org/10.1007/s10021-016-0067-y

Tredennick, A. T., G. Hooker, S. P. Ellner, and P. B. Adler. 2021. A practical guide to selecting models for exploration, inference, and prediction in ecology. Ecology 102:e03336.

Lines 391-398: Another option (since it seems the goal in this section is to discuss improving forecast skill) could be to fit a model to the residuals of the S-TS and/or LB-TS models

Fig. 2: there are currently no x-axis labels on insets; one possibility might be one-letter month abbreviations?

Appendix A.1: Would be helpful to have information here to help the reader understand how to interpret tradeoffs in the values of the evolving information state, where the previously best-performing model becomes the worst-performing model and vice-versa (this happens many times in the inset plots in Fig. 2). As written, the text addresses how models accumulate weight but not how they lose weight.

I thank the authors for their careful attention to my comments and those of the other reviewers. I think the decision to reframe the manuscript to focus on modeling for understanding greatly strengthens it, and the additional explanation of terminology associated with state-space modeling will be helpful to future readers. I also like the addition of the cross-river comparison panels to Fig. 2 and I think this addition facilitates a stronger Discussion. I have no further comments for the authors to address.