Review of the paper “A high-resolution map of diffuse groundwater recharge rates for Australia” by Lee et al.

            This paper presents an interested method to estimate groundwater recharge rates across Australia using chloride measurements. The text is well written and logically organized so that is easy to follow. The figures are excellent and ready for publication. I am not specialist of geochemistry or chloride but I found this study very relevant for hydrologist like me who is interested in groundwater processes. While I consider my following comments as minors, I think they must be addressed with attention before publication.

I am especially disappointed by the comparison between the presented product (from chloride) to the previous estimates from other studies (for example Moeck et al. 2020). Such previous studies report mean recharge estimates 5 times larger than the present study. You only attribute this drastic difference to “the spatial distribution of recharge point estimates, and the estimation of recharge values using different recharge estimation techniques” (line 470). I am sorry but I am not very convinced by this argument.

• Why you do not compare point estimates from Moeck et al. (2020) with your product at the same locations, exactly? For example, using a squatter plot, or the density function of each product. In other words, I find your comparison with existing previous products is not enough detailed. Please, improve this comparison.
• It seems that, naturally, Australian soils are largely affected by salt (Wicke et al. 2011; https://www.encyclopedie-environnement.org/en/zoom/land-salinization/). Perhaps my understanding is not very good, but it seems that only chloride range of 35-125 mg/L for groundwater is considered as normal. Your Figure 1 shows that at least the half of your data are superior to this range. Could your underestimation (of recharge estimates) thus be due to a natural large groundwater concentration in chloride (Clgw) making your equation 2 obsolete?
• Chloride concentration in groundwater could be significantly impacted by human activities like agriculture or industry. If anthropogenic chloride “flows” into groundwater, its concentration will be larger than in natural systems and then, because equation 2, your estimates will be biased and too low. Are you sure that your measurements are not drastically impacted by these processes?
• Line 550, you clam “Our CMB-based recharge rates are considerably lower than other studies including global water balance models (e.g., Döll and Fiedler, 2008; de Graaf et al., 2015; Müller Schmied et al., 2021). This is likely due to the fact that global water balance models estimate modern recharge.” … Ah ok, but why, I do not understand your explanation? Do you mean that your estimates do not account for modern recharge, so they account for what? In other words, what is the period of validity of your estimates?

So, even if I found this study very relevant and promising, I think that (1) the comparison with other product should be done more in depth, (2) a discussion about the salt-affected soils over Australia is perhaps relevant, and (3) a discussion about the impact of anthropogenic processes on groundwater chloride concentration (and then on your results) must be emphasized.

Biblio:

Wicke B, Smeets E, Dornburg V, Vashev B, Gaiser T, Turkenburg W & Faaij A (2011) The global technical and economic potential of bioenergy from salt-affected soils. Energy Environ Sci 4:2669-2681. https://doi.org/10.1039/C1EE01029H

I found the manuscript to be very interesting and, as a groundwater modelling practitioner, I expect it to be a valuable resource if published. I expect to use it as a source for initial model parameterisation of diffuse, rainfall derived-recharge fluxes and for providing a point of comparison and reference for groundwater models in Australia.

The document is well written and provides an excellent description of the methods used, the main findings and discusses interesting outcomes including the limitations in the approach.

I understand that point estimates of groundwater recharge have been obtained from chloride measured in groundwater bores by the Chloride Mass Balance method using gridded chloride deposition, runoff, and precipitation datasets. The point estimates have been integrated through a Random Forest analysis to produce a recharge model for the entire continent.

Although I have no experience or understanding of the Random Forest method, I assume that the R₅, R₅₀ and R₉₅ distributions illustrated in Figure 6 illustrate the uncertainty associated of the Random Forest analysis and do not include the additional uncertainty of the Chloride Mass Balance estimates used to obtain the point estimates. In my opinion the text would be improved by a clarification of this point.

I found the comparison to similar published studies in Section 4.2 to be of particular interest. I was surprised at the apparent discrepancy between the average point recharge estimates from the current study and those collated from other recharge studies in Australia (specifically Crosbie et al. (2010a) and Moeck et al. (2020)). The current study provides average point recharge estimates that are about 5 times lower than those obtained from the other studies. The text suggests that different distributions of data used to derive the recharge estimates and the different methods used to calculate recharge (including watertable fluctuation, catchment scale water budgets and other environmental tracers) may be the factors that explain these discrepancies. Without further discussion and examples, I find it difficult to accept that these issues can explain the magnitude of the discrepancy. For example, I find it unlikely that the spatial distribution of data used for the current and previous studies will be significantly different. I assume they all rely on measurements made in groundwater bores, the total population of which being the same for all studies. The discussion also calls into question the reliability of the Chloride Mass Balance method when compared to other recharge estimation techniques.

While not suggesting that significant revisions to the manuscript are necessary, I believe the paper would benefit from a more focussed, qualitative assessment of uncertainties included in the recharge distributions presented in Figure 6. This should not only address the uncertainty in the Random Forest model but also in the uncertainty associated with Chloride Mass Balance estimates themselves including the reliability of the datasets used to obtain the point estimates.

In this study, the authors used chloride mass balance (CMB) to derive long-term groundwater recharge rate estimates for Australia. A random forest model was built and tested using 17 relevant climatological, geologic, hydrologic, and static soil/vegetation variables as the predictors. The random forest model was validated using the point-scale CMB recharge rate estimates and the best-performing model was based on 8 of the 17 variables. The 8-variable model was used to generate the median, 5^th, and 95^th percentiles of recharge rate for the entire Australia. Overall, the manuscript is very well written. The experiments are set up in an organized and thoughtful way. I enjoy reading the discussion section where the authors provide guidance for practitioners to use the dataset. I only have some major and minor comments as outlined below.

Major comment:

Table 1: I have questions on the temporal evolution of these factors and the importance of the temporal component of the model. Depth to water table is a time-varying variable. Specify what value of depth to water table is used in this study.

CMB is a method that measures long term (hundreds to thousands of years) groundwater recharge rate. I notice the authors use different time periods for different input features. My two questions are 1) Are those periods the longest time periods with data availability? 2) If yes to question 1), the time periods of data availability are still time periods that cannot match up the residence time of chloride which is on the order of hundreds to thousands of years). How did the authors go about that? The authors can do a sensitivity analysis on using different time periods of input variable to test the sensitivity of their model results to the choice of input time periods.

Minor comments:

Line 105: “…we identified 17 different gridded datasets (Table 1).” Is the distance to coast a gridded dataset? If it is a gridded dataset, specify the spatial resolution in Table 1. Otherwise, change the wording on Line 105.

Table 1: The categories do not make total sense to me. Geology seems to belong to “Surface processes and hydrogeological” category. “geomorphological” can be changed to “soil properties”. For sand, silt, and clay fractions, the description states they are 100 to 200 cm interval. Does this mean the input features are for the 100 -200 cm vertical layer? If yes, justify why choosing a deeper soil layer instead of values for the entire soil column.

Line 205: Step (6) removes cases where estimated recharge equals or exceeds mean annual rainfall. Explain why and how did that happen. Could this be related to the errors underlying the estimation of recharge rates?

Line 220: What is the “typical practice”? Specify the name of the method.

Line 221: Each tree in the random forest model (the model) was trained on n randomly selected observations, with replacement (i.e., bootstrapping) from the training subset, where n is equal to the total number of observations in the training subset.

Line 306: “The recharge area of these deep systems is likely to be hundreds of kilometres away from the bore location, whereas our analyses assume recharge occurs within the 0.05° × 0.05° pixel from the chloride deposition map that contains the bore.” How does this influence the results or how do the authors address this question.

Line 314: “The mean recharge rate…”, do the authors mean “spatial mean recharge rate”?

Figure 3 and Line 320: Add the map of precipitation to Figure 3 to assist the comparison.

Table 2 and Line 344: The best-performing 7-variable grouping has a performance as good as the 8-variable grouping. The less the number of variables, the lower the computation cost and potential of over-fitting. Why not choose the 7-variable grouping?

Line 369: typo, “200 trees” should be “250 trees”

Line 445-459: Could the covariance/correlation between variables influence the feature importance of a specific variable? For example, precipitation, distance to coast, and elevation are correlated. Will using all three variables in the model introduce redundant information and potentially increase their explanatory power?

Line 450: The sentence in the parenthesis reads weird. Rephrase.

Line 515: Geology seems to be an important factor to explain the overestimation in the model. However, on line 450, the authors state that geology was not included in the highest-performing model because it cannot split the data due to low cardinality. It reads conflict to me. The authors should explain more.