This is a well written and logically presented study in which the authors reconstruct daily runoff over the entire land area of Switzerland for a 60-year period. I like the work and see no major issues, though I have some suggestions that I think would make the study’s contributions clearer. I also try to be complementary to the existing review.

Comments:

[1] My first comment is about the paper title. I expected a somewhat different study based on the title. The mentioning of a runoff monitoring product in the title suggests some type of derived data product, though the focus of the manuscript is the extensive development of neural network approaches to perform the reconstruction. I assume that this is when the authors say data-driven, though I do think a title that better represents the actual study content would be preferrable.

[2] The authors’ view of “traditional hydrological models” is overly narrow (lines 24ff.). While physically-based (pb) models, like the one previously developed for Switzerland (PREVAH), have a high computational demand and are rather data hungry, this is not the case for all hydrological models. In fact, much of hydrology uses rather parsimonious models (GR4J, HyMod, PDM…) which do not put a high demand on computational resources. It would be good if the authors either refine their statement to pb models or widen it to include a wider range of model complexities. Given that the simulation of daily runoff is done with such simpler models in many countries, I would suggest the latter.

[3] A similar point can be made about the data need of hydrological models which is discussed in lines 36ff. Several widely used hydrological models can be driven by precipitation and temperature only – if they are of the parsimonious type.

[4] The authors use squared error metrics for model calibration. They then use disaggregated components of such metrics for further analysis- which I like. What I missed in the analysis is any assessment of whether these components show any structure across Switzerland. For example, Gudmundsson et al. (2012, WRR) showed for example a strong correlation between bias errors and elevation differences for some comparable catchments to those used here. Did you look for any systematic biases (in the context of Fig. 4)?

[5] The relative NSE range shown in the legend of Figure 3 seems very small. Is the variability shown in the various small plots actually relevant?

[6] The other reviewers already made some comments and suggestions regarding the trend analysis performed, and the need to test a non-parametric strategy. I will not repeat his points in this review, but I believe that they are justified.

[7] Rather than the qualitative evaluation in section 4.2, is there not enough information in the 98 catchment differences to show where and when PREVAH is better/worse?

[8] Section 5.3 “The reconstruction of runoff back to the early 1960s for Switzerland is a novelty enabled by the reduced data needs of our deep learning-based approach.” But would the NN benefit from additional data?

[9] (lines 446ff.) The authors state that “A limitation in our approach was the reliance solely on air temperature and precipitation data for long-term reconstruction, excluding other meteorological factors like cloud-related effects, which could only be indirectly approximated by the model.” Can you name examples of hydrological models that consider cloud-related effects? Do you mean the consideration of sunshine hours? You could have used such information, couldn’t you?

[10] Also, the authors state that the “The assumption of static variables, such as land use and glacier coverage, being constant over time is a necessary simplification but introduces potential inaccuracies.” I am not completely clear why this is a necessary simplification. Why can changing forest cover and a limited contribution of melting glaciers not be included?

[11] Conclusions: “One of the major strengths of our approach lies in its computational efficiency, which opens up possibilities for contiguous near real-time monitoring and potentially forecasting of runoff.” And “…allowing for the rapid evaluation of thousands of scenarios that were not feasible with traditional physically-based models.” Here the authors state their assumption of “traditional physically-based models” which is not the same as traditional hydrological models. It would be good to clarify this difference in the Introduction section.

This review document is the same as that uploaded on 1 June 2024, when I mistakenly entered a community comment rather than a review comment.