General comment

This manuscript present workflow and data for harmonizing and joining several major water quality data basis on international and national level to a consistent dataset. The authors carefully describe their workflow and present selected examples on the spatio-temporal coverage of their data sets and suggest further improvements. This is a good example and a showcase what problems scientists have to face when joining datasets. The authors are very careful in making their workflow reproducible for scientists who want to work with the data, e.g. by flagging but not removing outliers. The resulting dataset is unique and very helpful for water quality researchers. I would consider the data as high quality with a few drawbacks mentioned below. While I very much appreciate the work the authors put into the data, I have some concerns regarding the manuscript itself. Overall I found the introduction rather weak. The authors build the justification for their work mainly on water quality modelling with a special focus on machine learning. For me this is to narrow. Any research rely on data, be it the mere statement of a global or regional statistics of a certain constituent, data-driven exploration of controlling factors behind observed patterns or statistical or even mechanistic models. I moreover miss a reference to the need for open data and the FAIR principles. Considering the workflow I greatly miss the handing of values below the limit of detection/ quantification. This is a major issue for all water quality studies and is left entirely open here. I hope the authors can make use of these general comments and the more detailed specific comments below.

Specific comments

Abstract

Line 1-4: I find the introduction rather weak. Some more specific words would be helpful to define the problem, to say what is already there and to what extent this manuscript is going beyond the status quo.

Line 2: It is not immediately clear that "current" study is referring to this manuscript. Consider rewording.

Introduction

Line 13: I disagree here. Why is water quality modelling an integral part of ecosystem health monitoring? Why starting directly with modelling? Models need a conceptual basis that arises from data and interpretation of mechanisms behind the observations. I think there is great value in the data analysis itself before making the step to modelling.

Line 21: What model inputs do you mean here? Please specify. Is it rather calibration data? Or input in terms of drivers (such as precipitation drives a rainfall-runoff model)? Or do you mean parameterization for processes such as a nutrient uptake?

Line 28-33: Here you seem to jump to hydrological models (quantity, not quality) without further mentioning that. To what extent can this be transferred to water quality models?

L34ff: I see the point for putting emphasis on machine learning approaches. However, this seems to be the major justification for large-scale or large-sample datasets and the major outlet of these data. Here I disagree - ML is one possibility but surely not the only justification for the need for water quality data with a wide spatial and temporal coverage.

Line 40: Do you really mean large-scale here? Or rather a wide spatial coverage?

Data

Line 71f: What dataset "that had been previously collected from the other sources" are you referring to? Is that showing up later in the manuscript? Maybe introduce that in the preceding section?

Table 2: Would it make sense to state the date of download here? The databases are still actively fed with new data, right?

Method

I miss a description how the detection limit was handled. This is very crucial. Often numbers such as <0.01 mg/L  are given. For some constituents this is rather the rule than the exception.

Line 120: Are outliers just detected (flagged?) or also removed? Why are time series characteristics derived before removing duplicates? Are duplicate station or duplicate samples ment?

Table 4: For me this table does not make sense. Why not stating all conversion information but only an example? The meaning of x1 to x2 should be mentioned in the header.

Line 155: Surely this decision was made wisely. However, it is not clear for the reader why one km is the threshold here. Can you explain that? In a headwater catchment of 2 km2 size a shift of the sampling stations by 1 km would probably not treated as one joint size, right?

Line 206: This is maybe not the best example/ justification. If an agricultural spill has a long-lasting effect on water quality, it would not be an outlier but create more than one elevated values. Consider adjusting the example here.

Line 210ff: What about physical impossible values? Negative concentrations, water temperature above 100°C or concentrations above the solubility of a constituent? Would it make sense to check and remove those as (if they occur) they will influence the percentiles/ quartiles?

Results

Table 6: Attribute parameters such as "upstream basin area" and "Drainage region" have not been mentioned before. However, this is a very crucial information for researchers working with the data. Choosing sites according to their catchment sizes is common with the consequence that this information needs to be reliable. Was that just taken over from the original datasets?

Table 7: Number of digits is relevant here! Is that fixed for all constituents to the same value? If yes, is should be e.g., 0.010 for median phosphorous, not 0.01. This decision should be part of the text.

Line 267: Table reference is missing.

Line 291ff: Usually concentrations are considered to be lognormally distributed. A skewness does thus not purely result from outliers. You should mention that non-normal distribution rule here.

Figure 4 is less informative than figure 3 but basically shows the same. For me this is not necessary. Potentially both figures can be combined.

Line 312: What is the question mark referring to?

Figure 7: Here I realize the point of the physical possible range of parameters. Oxygen can, to the best of my knowledge not be much higher than 15 mg/L - sure there can be oversaturation. But I would claim that 502.75 mg/L is impossible. Would those value not hamper your outlier statistics?

Line 330ff: This is a good idea to come up with suggestions for improving datasets. However, I strongly recommend to make this part of the problem definition in the introduction and the objectives of this study.

Line 352ff: The spatial lack of data is for me not directly connected to the adoption of machine learning. Yes, those techniques may can come up with an estimate. However, data-driven models are only as good as their training data coverage. For a prediction under boundary conditions/ landscape properties that have not been sampled you will need a mechanistic approach. The whole discussion here is a bit misleading. I don't think the solution for spatial gaps is not gap filling but rather measurements or incentives for countries to share data.

 General comments

Virro et al. describe aggregating and harmonizing five national, continental and global datasets that can be used for global water quality models. Among the five selected datasets the GEMSTAT itself is a global database containing harmonized data from the contributing countries. The authors follow the ETL approach (Extract-Transform-Load) to bring the different data sets together. The methods are well described and data, metadata and scripts are available at the given websites. The authors suggest in their conclusions to transform the set of CSV files to a relational database in future what would further improve the usability of this dataset. I encourage the authors to do it. Also I like the idea to develop an online dashboard for GRQA. From a global modelers perspective the derived GRQA can be easily used. The authors selected 42 specific parameter relevant for modelling nutrients (water temperature, oxygen, nitrogen, phosphorous, carbon compounds ...). Still, GRQA cannot solve the problem of data scarcity in Africa, Asia and South America and also the suggested machine learning cannot help here.

In general, the paper is well written and supports the publishing of GRQA. One issue for me is that it was not mentioned that WATERBASE is already integrated in GEMSTAT and how the authors dealt with this. There must be a lot of duplicate data from this fact. But the given procedures should have found all the doubled data.

I support the publication of this paper. Some minor issues are listed below.

Specific comments

line 55: please list the parameter used here “most important water quality parameters” or refer to your table 7

line 71: “... only eight parameter matched the data set” – you mean one of the 42 selected parameters? Pls clarify.

line 73 – 74: please give this numbers (how many parameter matched the set, site count range, mean time series length per site, average number of observations per site) for all data sets in a table and refer only to the parameters that were used

line 99: “WATERBASE has the shortest timeseries ... 2008 - 2018” - as far as I remember there are nutrient data available starting with 1992; please check again the disaggregated data; e.g. see the graphs given under this link

 https://www.eea.europa.eu/data-and-maps/daviz/rivers-nutrient-trend-4#tab-dashboard-01

The graph is based on data from WATERBASE.

line 125: “introduction: nutrients, carbon, sediments and oxygen”-Please refer to table 7 here.

line 217: I understood that WATERBASE was included into GEMSTAT; please see https://www.waterandchange.org/en/european-water-quality-monitoring-data-in-gemstat-database-undergoes-major-update/

Also here the time frame is given with 49 years for WATERBASE; please check and write a short paragraph how you dealt with this issue.

Figure 2: Why is the percentage of observations (1898 - 1970) not given her? If no data exist before 1970 it should be mentioned here. The colors in the legend don’t correspond to the colors in the plot. E.g., I can only guess what dark green is. Also I wonder what I can learn from this plot. Maybe the authors change this plot and show the time series available per continent instead. That is actually where I’m interested in as global modeler.

line 222-228: I wonder how can the same station be reported to different databases with different position information? In case it is the same station then both stations should contain the same time series? And how are you dealing with stations that are very dense (<1km) but not the same? For example at a tributary that is going parallel and there is one station in the tributary and one in the main stream before tributary is coming in. Did you find such cases and how did you deal with it?

line 296-297: It is not clear to me why the IQR test outliers are removed from the plot? For illustrative purpose – what does it mean? The plots would not look much different wouldn’t they? So the outliers are not shown or are removed from the data? And actually for TSS the outliers can be really data e.g. before the crest of a flood or at the beginning of discharge reach a region specific threshold TSS can get extraordinary values.

Figure 4: e.g. TSS – 18.9% outliers removed from the plot – this 18.9% refers to all data? So it could be that 18% outliers are in GEMSTAT and the rest in the other datasets? Do I understand it right? please clarify in the figure text.

line 307 – 313: Please delete this paragraph. The focus of this paper is on merging data together rather than assessing the content. Therefore I would remove lines starting with “DOC concentrations are lower ...” and ending with “point sources”. Even if Figure 7 is interesting from the content perspective I would remove it from this paper because it is not the main focus. In my opinion it is sufficient to mention that this kind of plots are available for all 42 parameters in the GRQA dataset.

line 320: What do the authors mean by “using the wrong form”? This is not clear to me. You should have checked the forms before using it in your codes and I’m sure you did. So how can it happen that a wrong form was used? Please explain.

line 322: Why are units in GLORICH in µg/l a problem during conversion process? The authors transformed it to mg/l. I cannot see a problem here. Please clarify the discussion.

line 352: I doubt that ML methods alone can help to fill the gaps in Africa and Asia without support of any measurements in the regions. So, the argument given in line 359 contradicts the statement given in line 352. Maybe a combination of remote sensing techniques and water quality modelling (including ML) could help. Please revise.

line 380: “The dataset is expected to have yearly updates after publishing” – are you sure that this is realistic? I would rather advise to remove this sentence. It is not necessary to state something like this in a paper, you can just do it if the capacity is available.

line 383 – 384: I like very much the idea to put GRQA in a database and to make it accessible via an online dashboard.

Technical corrections

line 268: table reference is missing

line 269: given numbers of sites per parameter (“15 (POP) up to 90792 (pH)”) don’t correspondent with the numbers given in Table 7. Please clarify.

line 312: reference is missing