Here we present the authors response to CC1: 'Comment on egusphere-2023-1755', John Ding, 23 Aug 2023, about Nash-Sutcliffe efficiency and its logNSE and sqrtNSE variants.

We decided to use the logNSE for both calibration and evaluation of the model. This choice was made because logNSE is more sensitive to low flows (Oudin et al., 2008) and since we are experiencing draughts recently in Paraná State, Brazil, this sounded like a good choice. However, we also used the following other criteria to evaluate the model: Pearson Correlation Coefficient (R), Nash-Sutcliffe Coefficient (NSE) and sqrtNSE.

Table 2 shows the median calibrated NSE value of 0.621 for 26 watersheds (from validation set) out of a total of 126 in State of Paraná, Brazil. This information wasn’t totally clear in the manuscript, and we have rewritten it in the new version that will be uploaded soon.

Thanks for correcting Equation 1, the leading expression "1-" was truly missing and the new version of the manuscript is corrected.

The main idea of the article was to compare regionalization techniques, not to compare model performance metrics. We agree that according to Cinkus et al. (2023) and Mizukami et al., (2019) the choice of performance metric matters for model evaluation, that’s why we have chosen 4 metrics: R, NSE, logNSE, and sqrtNSE. However, we wanted to see which regionalization procedure for GR4J parameters was more efficient, and we showed that for Paraná catchments the Similarity approach was slightly better.

To use an autoregressive (AR) model, as suggested, we would need to know the flow in the watershed where we suppose we don’t have flow data. So, we don’t think it is a good idea to be implemented as extra work for this research/manuscript, maybe some future work.

Thanks for the suggestions, we will consider them for a future study.

References:

Oudin, L., Andréassian, V., Perrin, C., Michel, C., & Le Moine, N. (2008). Spatial proximity, physical similarity, regression and ungaged catchments: A comparison of regionalization approaches based on 913 French catchments. Water resources research, 44(3).

Cinkus, G., Mazzilli, N., Jourde, H., Wunsch, A., Liesch, T., Ravbar, N., Chen, Z., and Goldscheider, N.: When best is the enemy of good – critical evaluation of performance criteria in hydrological models, Hydrol. Earth Syst. Sci., 27, 2397–2411, https://doi.org/10.5194/hess-27-2397-2023, 2023.

Mizukami, N., Rakovec, O., Newman, A. J., Clark, M. P., Wood, A. W., Gupta, H. V., and Kumar, R.: On the choice of calibration metrics for “high-flow” estimation using hydrologic models, Hydrol. Earth Syst. Sci., 23, 2601–2614, https://doi.org/10.5194/hess-23-2601-2019, 2019.

Here we present the authors response to RC1: 'Comment on egusphere-2023-1755', Anonymous Referee #1, 06 Oct 2023.

We are grateful for the comments and contributions made by Anonymous Referee #1. However, we respectfully disagree with the argument that the manuscript is more of a technical report than a scientific paper. We have proposed a scientific study to compare hydrological regionalization methods for ungauged basins and to understand watershed physiographic-climatic similarity. Here we present the response to each specific comment:

1. We have shown by the comparison between three different regionalisation approaches that with a fewer number of donor basins the technique based on physiographic-climatic similarity is more accurate for prediction of daily and Q95 reference flows in basins from Paraná state, Brazil. When increasing the number of donor basins, the method based on spatial proximity has a comparable performance to the method based on physiographic-climatic similarity. In that sense and because there are usually not many donor basins available, we can state that the physiographic-climatic similarity method is superior to the others for the study site.
2. Thanks for the comment, we agree that some scientific issues and innovations were not clearly presented in the introduction and conclusion. Nevertheless, the additional knowledge our current work brings to the literature is that we have characterized 6 different types of catchments in Paraná State, based on catchment descriptors. Gathering data (39 descriptive indices divided in 4 categories: physiographic, climatological, land use / land cover, and soil type) and grouping the watersheds based on physiographic-climatic indices is an innovative contribution to scientifically understand the hydrology of one of the water richest regions of the world. The transfer of GR4J constants and model parameters between basins of the same cluster was made with a rigorous criterion since we have divided the clusters into 100 training basins and 26 validation basins. One major innovation is that we have used techniques from machine learning (K-mean and Random Forest methods) to perform the regionalisation procedure. The results obtained in our work are probably valid for other hydrological models since they are based on the watersheds physiographic-climatic similarity. The verification of this hypothesis is a future work that is mentioned in the last paragraph of the conclusion. These innovations and scientific contributions were added to the introduction and conclusion of the new version of the manuscript.
3. Random Forests (RF) are an ensemble learning technique that combines the predictions of multiple decision trees to make more robust and accurate predictions. They incorporate two key mechanisms to reduce overfitting: i) Bagging (Bootstrap Aggregating): Random forests train each decision tree on a random subset of the data, with replacement. This means that each tree sees a slightly different portion of the data, reducing the likelihood of any single tree overfitting to the training data; ii) Feature Randomization: In addition to using bootstrapped samples of the data, random forests also consider only a random subset of the features (variables) for each split in the decision tree. This further reduces the tendency of the model to overfit to specific features (Hastie et al., 2009). The original article by Breiman (2001) states: “use of the Strong Law of Large Numbers shows that RF always converge so that overfitting is not a problem”. All RF models applied in our study have used 80% of the data to the training set 20% of the data to the validation set. We used 1,000 decision trees and we limited the number of relevant variables to those who received weights (importance) greater than 0.01. The maximum tree depth was also limited: nodes were expanded until all leaves were pure or until all leaves contain less than two samples. This information was reinforced in the new version of the manuscript.
4. The choice of training and validation sets shown in Figure 4 is a common practice in machine learning. We have decided the percentage allocated to each set to ensure that our regionalisation generalizes well to unseen basins. This procedure is a precaution that few regionalisation works adopt. In summary: we have used a typical choice of 80% of the data to the training set and we have allocated 20% of the data to the validation set.

Some of the research questions that our study can answer are:

1. Which is the best streamflow regionalisation method for Brazilian watersheds with humid subtropical (Cfa) and oceanic (Cfb) climates?
2. If there is physiographic-climatic similarity between two basins (one with flow data and another without flow measurements): can we transfer hydrological model constants from the monitored catchment to calculate flow in the unmonitored basin?
3. Are the catchments from Paraná State similar? Is it possible to group them into homogeneous clusters with the same hydrological characteristics?

The extensive work on data processing performed by the authors provided a State-of-the-Art dataset for our study site, adding novelty to the field of hydrological characterization and classification. Our research shows that basins with similar physiographic-climatic characteristics are more prone to be chosen as candidates for flow regionalization.  We provided a proof-of-concept that basins without flow monitoring can have a good approximation of streamflow if other physiographic-climatic indices are provided. Furthermore, we have shown that machine learning algorithms perform better with physiographic-climatic indices as inputs. All these arguments will be incorporated in the new version of the manuscript.

With respect, we find the reviewer’s feedback superficial. None of the points addressed by the reviewer hint to misinterpretations in our work. The scientific value is discussed above. Given the journals aims: “HESS encourages and supports fundamental and applied research that advances the understanding of hydrological systems…”, which we cover, we grade the final statement of the reviewer questionable.

References:

• Breiman, L. (2001). Random forests. Machine learning, 45, 5-32.
• Hastie, T., Tibshirani, R., Friedman, J., Hastie, T., Tibshirani, R., and Friedman, J. (2009). Random forests. The elements of statistical learning: Data mining, inference, and prediction, 587-604.

Here we present the authors response to RC2: https://doi.org/10.5194/egusphere-2023-1755-RC1, Juraj Parajka, 20 Oct 2023.

We are grateful for the comments and contributions made by Juraj Parajka. We agree with the points highlighted by RC2 and we will incorporate the suggestions provided, according to the following:

The introduction will be improved to highlight the research gaps and how our study goes beyond the existing literature, which consists of the following points:

1.    The construction of a State-of-the-Art dataset for 126 watersheds in Paraná State containing for each catchment: streamflow, precipitation, evapotranspiration daily data and 39 descriptive indices divided in 4 categories: physiographic, climatological, land use / land cover, and soil type.

2.    The need to better understand regionalization techniques in subtropical climate, which has a very distinct and specific runoff generation mechanisms, compared previous studies.

3.    The classification of Parana catchments into 6 groups based on physiographic-climatic similarity. Our results show that despite being basins located in the same region of Brazil, there are heterogeneities related to relief and land use, which are important to hydrological processes.

4.    The use of machine learning methods to improve the transfer of hydrological model constants between similar basins.

The study of low flows and droughts is critical in the context of water availability in Brazil, river dams and reservoirs are used for hydroelectric generation (70 % of Brazilian energy sector), to provide drinking water for population, irrigate crops, and distribute water for industrial use. The Paraná basin, a major hydroelectric producing region with 32 % (60 million people) of Brazil's population, recently experienced the most severe drought since the 1960s, compromising the water supply for 11 million people in São Paulo (Melo, et al. 2016). The focus of high flows is also important in the region because of the recurrent occurrence of floods (Stevaux et al. 2009). The influence of climate change in streamflow regime is unclear for Paraná state and reinforce the need for further research development. Blöschl et al. (2019) have shown that changing climate both increases and decreases European river floods, however we haven’t found any study analysing data from the south of Brazil. We agree that the description of the seasonality of low flows can help to understand the hydrological processes in the study region, but our focus was motivated by the recent droughts that occurred in the south of Brazil (Cunha et al. 2019) and the impact it was on the population and on the economy of the country.

We agree with the suggestion to change the structure of the manuscript to Introduction ->Data (Study region)-> Methods->Results-> Discussion and-> Conclusions. This includes reorganizing the catchment descriptors into the Data section, put regionalisation methods and GR4J model description into the Methods section and enlarge the Discussion with our novel contributions. We can also focus more the review on the research objectives: prediction of daily runoff hydrographs and low flows.

We thank the suggestion to interpret in more detail the results from the context of dominant runoff generation processes and how these are described by different regionalisation methods, however we think this might be considered a future study for our research group.

References:

Melo, D. D. C., Scanlon, B. R., Zhang, Z., Wendland, E., & Yin, L. (2016). Reservoir storage and hydrologic responses to droughts in the Paraná River basin, south-eastern Brazil. Hydrology and Earth System Sciences, 20(11), 4673-4688.

Blöschl, G., Hall, J., Viglione, A., Perdigão, R. A., Parajka, J., Merz, B., ... & Živković, N. (2019). Changing climate both increases and decreases European river floods. Nature, 573(7772), 108-111.

Stevaux, J. C., Latrubesse, E. M., Hermann, M. L. D. P., & Aquino, S. (2009). Floods in urban areas of Brazil. Developments in Earth Surface Processes, 13, 245-266.

Cunha, A. P. M., Zeri, M., Deusdará Leal, K., Costa, L., Cuartas, L. A., Marengo, J. A., ... & Ribeiro-Neto, G. (2019). Extreme drought events over Brazil from 2011 to 2019. Atmosphere, 10(11), 642.