Crespi, Enigl et al. study different rainfall datasets and test their potential for predicting geohazards. The test is conducted over a comparatively large are in parts of Italy and Austria. Results include how well the tested datasets and statistical descriptions of rainfall extremes identify storms and recommendations on how the which dataset should be used.

The strength I see in this study is more on the comparison of the datasets than on the testing of statistical thresholds to identify storms. Some of the datasets compared in this study are often used, also in other data-sparser regions of the world, making such a comparison useful. Which rainfall statistic is most powerful in predicting geohazards is a widely studied topic and I don’t think the authors do this in much depth in this study. Furthermore, the results and conclusions are not presented in a very accessible way. I therefore mainly recommend streamlining and restructuring to frame the research in the right context and make it more accessible (i.e. higher impact). Nevertheless, I congratulate the authors on the work they’ve done so far, which I find useful and with practical impacts. I list my main comments below and line-by-line comments further down.

• The paper cannot be read very fluently, and one often has to guess the intention of the authors with certain paragraphs/figures. For example, the methods around L180 on the rainfall datasets are a mix of methods, results and discussion. Likewise for the hazard catalogues, where trends are calculated and discussed in the methods (~L250). Also the discussion and conclusions could be better structured to better convey the key messages by adding subsections to the 3-page discussion that explicitly address the goals of the paper (testing the methods for extreme rainfall definition, testing different datasets, implications for practitioners).
• While I think the detection thresholds calculated from “areal mean”, local p99” and “anomaly” I think generally are meaningful statistics to use. But I don’t see much reasoning on why exactly these were chosen and there are not many references either in this part. Given that rainfall thresholds for geohazards has been a research topic for a long time, I miss the novelty compared to other studies or even just the justification for using exactly these statistics, while so many other statistics could be computed too (antecedent rainfall, multi-day cum. rainfall, …).
• The dataset comparison is conducted by comparing the hit rate, eg in Table 2, at an artificially set threshold of top 5% rainfall events. However, from my experience it is more common and interesting to compare the predictive power of these datasets to separate hazardous from non-hazardous dates at a range of thresholds. For landslide early-warning, it is almost standard to report receiver operating characteristics. You will easily find references on this and the statistics can be calculated from the data you have with the eg the scikit learn library in python (eg ROC-curve).

Specific comments:

L17: can you say more about the three definitions? Abstract readers will want to know the temporal scales of your analysis.

L21-24: Please specify in the abstract which data products you are testing. Now it only becomes clear that ERA5-Land is bad. But what is good? What do you mean by «high-resolution observation»?

L79-80: also, the cited papers all seem hydro/flood related but not landslides

L113: a short intro to this section and the reasoning on how you chose the datasets would be helpful here. Also, a table with key facts about th edifferent datasets would be very helpful

L181-196: These paragraphs are a mix of methods, results and discussion. Furthermore, I miss the link to your study. Could you say why showing these monthly means is important for your study on extremes?

L201: Please specify how you define “gravitational mass movement” as this is a very broad term. Does it include rock glaciers? deep-seated landslides or only shallow? Debris flows? Rockfall?

L250: these lines again seem like results to me. Unless they were taken from other studies, but then a citation should be enough.

L283-287: Do you have evidence from other studies to support these assumptions?

L363: should be “quantile” instead of “percentile”

Table 3: I’m having troubles understanding this table. I think it shows into which quantiles the 330 events fall. So each row should add up to 100%, which it doesn’t, probably due to rounding. What are “intersected hazards”? I couldn’t find a definition in the text and I don’t get why this number differs among datasets.

L695: can you provide an example for an application requiring “accurate description of precip fields”?

Dear authors please see my review comments in the attached file.