Overview

This manuscript describes the development of a new operational hourly, km scale precipitation analysis for the entire Australian continent, titled 'BRAIN'. The authors adapt an existing method - Optimal Interpolation - to blend gauged precipitation, a satellite precipiation estimate and precipitation estimates from radar. The use of a robust existing method in OI is in my view a sensible choice for an operational system, and the authors make well-judged simplications to enable its deployment at continental scale (e.g. the use of single radar grid cells to estimate error characteristics). The authors use stringent verification measures, including strict cross-validation and comparisons against a range of existing products that are in wide use in Australia. These are high hurdles to clear, yet the authors are able to demonstrate that BRAIN clearly outperforms existing alternatives. In addition, the manuscript is concise and clear, and in general a pleasure to read. No continentally consistent precipitation product at this temporal or spatial resolution currently exists for Australia, and when BRAIN is operationalised it promises to be a game-changing innovation for a wide range of scientific, industrial and environmental applications. I essentially recommend that the manuscript be published as is, subject to technical corrections.

I leave it to the authors to determine which of my recommendations they wish to act on, of which I will note only one here: I thought too much emphasis was put on the performance of BRAIN shown in Figure 7, which shows only two isolated events. I prefer the authors to refer to their robust assessments of performance in FIgures 5, 6 and 8 (and in the very useful Supplementary materials) than on Figure 7.

I congratulate the authors on what I believe will be a highly impactful product.

Specific comments

Introduction is lucid, economical and comprehensive. It clearly outlines 1) the promise of this product, 2) the challenges in establishing a sub-daily rainfall product in Australia 3) the range of methods available 4) a clear justification for the authors' use of optimal interpolation for their BRAIN product and 5) how the authors will adapt OI for their product.

242 "the "true value"" suggest "the "true value", T,"

L243 "the "true value"" - it's not clear from the equations below or the prose how the true values (T_k or T_i) are defined; accordingly the equations for errors are not explicit. Please describe how the 'true value' is estimated or defined (in prose is fine). I appreciate that OI methods are well known, but for those of us (like me) who are not very familiar with them, it would be nice to have this spelled out. 

L251 "in normalised form" should this be "in standardised form"? Normalisation can imply a normalising transformation. 

L255-267 "This system of linear equations is solved to determine the optimal weights 𝑤_𝑖, which are used to obtain the best precipitation estimate at each target grid cell by optimally combining satellite, radar, and gauge data." The equations above use a single term (O) to describe observations from gauges or radar, and weights vary with i (location), not data source. What happens when both radar and gauges are available? I assume these have quite different error characeristics, so I couldn't follow how they can be all thrown in together. Oh - never mind, this is explained below (Lines 271-317). It might be good to either avoid the statement "are used to obtain the best precipitation estimate at each target grid cell by optimally combining satellite, radar, and gauge data" at this point, or foreshadow the coming explanation of how you actually combine different observational data sources below, or both. The authors can add a version of this statement at the end of Section 3.2 if they wish.

L272 "while background-to-observation error correlations are assumed to be zero, as different sensor systems are considered independent" no change here, merely a comment: this is interesting, as it's possible to envisage cases where these are not independent (though I think the use of satellite as background vs gauges/radar means the authors are reasonably safe in this assumption). 

L285 "only the nearest radar grid cell to each target grid cell is retained as a supplementary observation" Might be good to (very briefly) note how 'nearest' is defined here: is it possible to have more than one 'nearest' cell? Never mind, this is also covered below!

L321 "except for gauges located outside the radar range where no radar information is available" I think this is saying that gauges outside the radar range are used to correct satellite data. Is this correct? (And if not, why not?)

L335 "For gauged locations, this task is relatively straightforward, as gauge observations can be used as ground truth to estimate the error characteristics of the background and observational inputs" Excepting that gauges are point data and satellite/radar are areal averages, which will be smoother in time and space. How did the authors account for areal averaging to compute errors against gauges?

L352 "Specifically, the Australian domain was divided into tropical, subtropical, and temperate subregions." It would be nice to see these regions on a map; could simkply be added to Figure 3a or similar.

Evaluation strategies: outpeforming both IDW and AGCD is a high bar to clear; I commend the authors for this kind of rigour.

L393 "The spatial match between the gridded data and gauge locations was established using nearest interpolation." I'm not sure what was being interpolated here - please clarify.

L407 "direct comparison with global precipitation products would be neither fully fair nor particularly informative." Fair enough - I assume the authors are suggesting BRAIN would substantially outperform these global products. As the verification in this study is already very rigorous, and you can't do everything, this is fine. I think it would nonetheless be interesting to compare BRAIN to these global products in a  future study if possible, as it would reinforce confidence in BRAIN for Australian users.

L547 "BR-SRG clearly outperforms AGCD and IDW in representing fine-scale spaital structure and intensity". I understand the need to show example predictions to illustrate how BRAIN functions in comparison to other products. I do not think it's appropriate to use these figures to describe relative performance, in particular because e.g. BRAIN does not outperform AGCD in out-of-radar regions on average according to Figure 6. Suggest this sentence be reworded to "BR-SRG shows fine-scale structure and intensities not present in AGCD and IDW."

L550 Figure 7. A few things to improve about this figure: 1) it would be nice to put the dates for the two rows in the y-axis label of the left-most panel. 2) I could not see the 'x' symbols described in the text; please choose a symbol that is legible. 3) I do not think it's good practice to present additional results on rainfall totals at gauges in the caption. This information could be presented in small text boxes within each panel; alternatively an additional panel could be added to each row with a bar chart showing the values for each product.

L561 "underestimates peaks"; as per the previous comment, this does not reflect e.g. bias statistics shown in Figures 5 & 6. It's my view that readers are drawn to examples like these, and they may use them to infer that these results are generally true, rather than refer to the much more robust statistics presented in Figs 5 & 6. I think that all that can be said is that BRAIN shows much greater spatial detail than the other products. This detail is supported by the verification metrics shown in Figures 5 & 6.

L565 "AGCD (Fig. 7e) fails to capture the peak at Cape Tribulation". 1) International readers may not know where Cape Tribuation is and 2) The totals by AGCD on and off the coast around Cape Tribulation are higher than either BR-SRG or IDW, so I do not think this comment accurate reflects what's shown in the figure. Again, I suggest avoiding making inferences about performance from figure 7. Figure 7 is useful in showing (1) the additional spatial detail shown by BRAIN over both AGCD and IDW and (2) that the spatial detail is consistent across the continent, while for IDW it is not. In my view conclusions on performance should be drawn only from the robust assessments shown in Figs 5 & 6.

L616 "Future development of BRAIN will focus on improving input data quality, refining error characterisation, and enhancing spatial and temporal resolution." (1) It would be nice to give an indication of how computationally demanding a re-estimation of the OI is when new predictor products are used or existing predictor products change. (2) Will the use of e.g. additional satellite products violate the assumption of error independence between products that is crucial to the simplifications pre

 

 

Typos etc.

L39 "Satellites-based" should be "Satellite-based"

L50 "collocation-based approach provides" should be "co-location-based approaches provide"

L51 Suggest a sentence break before 'However', and suggest "However, both performance-based and co-location-based approaches typically..."

L58 Suggest a sentence break before 'While', and suggest 'While offering high flexibility, ML-based approaches have...' 

L59 "In addition, all above" should be "In addition, all the above"

L75 "data source used for background field" should be "data source used for the background field"

L251 "By expanding and expressing the above equation" suggest "By expanding and expressing Equation 4"

L264 "is adjusted weight accounting" should be "is the adjusted weight accounting"

L337 suggest deleting "due to the absence of gauge observations" - it's redundant.

L394 "For the trialled implementation" suggest "For the trial implementation"

L414 "mean bias error" suggest deleting "error" as it's redundant

L604 "providing a consistent and flexible rainfall analysis capability" Suggest deleting "capability"

L607 "for gridded radar source" suggest deleting "source"

The manuscript presents an hourly, at 2 km resolution blended rainfall dataset (BRAIN) for Australia, derived from satellite, radar, and gauge observations. Given the current lack of a national hourly gridded rainfall product, this dataset addresses a clear need. The manuscript is generally well written, and the figures are clearly presented.

However, the methodology used in the study, including the Optimal Interpolation (OI) and Bias Correction, was developed and applied in other studies for the same purpose of generating blended rainfall datasets from satellite, radar, and gauge data (e.g., Yu et al. 2020 (blended data in the UK), Xia et al. 2025 (blended data in China)). There is no significant improvement in this study when the method was applied to Australia.

In addition, the use of inverse distance weighting (IDW) interpolation as a benchmark raises some concerns. Describing IDW as “the operational method in Australia for estimating hourly rainfall” may be somewhat misleading. IDW is a relatively simple interpolation technique, typically used for observing purposes rather than for producing authoritative rainfall datasets. Established Australian products such as AGCD, SILO, and ANUClimate employ more sophisticated interpolation methods, albeit primarily for daily timescales. As such, comparisons against IDW may not provide a sufficiently robust benchmark. The conclusion that BRAIN outperforms gauge data using the IDW interpolation is not convincing and can mislead users of the dataset.

Unfortunately, I suggest rejecting the manuscript for publication in HESS at this stage. I hope that the authors will revise the method and validation processes after the trial. While the dataset has potential value, I suggest submitting the manuscript to another dataset journal.

I attached detailed comments and suggestions to the PDF file.

General Comments:

This paper presented a multi-sensor quantitative precipitation estimation (QPE) product intended for operational implementation in Australia. It incorporates a geostationary satellite QPE from Himawari (10-min, 2-km), radar precipitation rates (5-min, 1-km), and gauge observations (15/30-min), aggregates the data to the hourly accumulations separately, and then merges them into a 1-hr,  2-km national QPE product via a statistical interpolation scheme. The satellite QPE, given its full coverage for the continent, served as a background field and the radar and gauge data are treated as observations in the interpolation scheme. To assure computational efficiency for operational applications, the authors limited 1 nearest radar data and 20 nearest gauges for a given grid point and developed analytical functions for the various error characteristics. The new product was evaluated on the 1- and 24h scales using 2-years (2022-2023) data, and the impact of each input data source was analyzed. The new QPE showed an overall improvement over the existing operational products in resolution, accuracy, and/or coverage.

 Overall the paper is well-written, and the data, methodology and results were presented clearly.  However, the interpretation of the results needs some clarifications. I have two major comments: 1) There is a lack of description of the “IDW” hourly gauge-based QPE, even though it's a key reference; 2) The interpretation of the hourly and daily validation statistics needs clarification and some artifacts in the product should be discussed (please see attached document).  I’d recommend a major revision of the paper before it can be accepted for publication. 

Specific Comments: Please see attached document