Dear Authors,

I have read and carefully evaluated your manuscript "Temporal clustering of precipitation for detection of potential landslides", which falls within my area of expertise.

The manuscript presents original research upon a topic (detection of landslide triggering conditions from cluster analysis of precipitation time series) that is well within the aims of NHESS journal. Results and conclusions are original and significant. Therefore, in my opinion the manuscript deserves publication, provided you are ready to implement some minor revisions. 

GENERAL COMMENTS

1- English is generally good, but the manuscript would benefit from a careful and thorough revision. It seems that some parts (e.g. the discussion) were written in a rush, with typos and a flow of the text that can be improved.

2- The Methodology section misses a paragraph about the test site description. An appropriate test site figure is needed. Fig 1 could be used as well, but I suggest adding the meteorological stations used in the study and a clear subdivision showing which are the "Lisbon area" and the "North of Lisbon" area. Also, I suggest using two colors for deep and shallow landslides.

3- While reading the manuscript, I got the feeling of reading a "regional" paper, which is a shame because the analysis and the conclusions are in my opinion are interesting and relevant for a global readership. I therefore encourage to improve the introduction and the discussion, adding elements outside Portugal.  Some good points could be e.g.: - to acknowledge that the subdivision between single landslide events  and multiple (areal) landslide events is gaining consensus in a growing number of works (e.g. https://doi.org/10.1186/s40677-018-0105-5; https://doi.org/10.1016/j.geomorph.2017.07.032). - to discuss/introduce that the combined role of antecedent and peak precipitation is considered in recent works (https://doi.org/10.1007/s10346-023-02176-7; https://doi.org/10.1007/s10346-020-01505-4)

4- The validation part needs to be supported by data. I suggest adding some tables to compare the metrics of the three approaches (e.g. at lines 225-229)

5- I suggest placing figures just after they are mentioned in the text. Some of them are quite far away.

SPECIFIC COMMENTS

L20 - this example about slope gradient could be removed. I think everyone reading NHESS is well aware of that.

L89-91 - the two datasets are very different. How this difference reflects into the results? Does it add uncertainty? That could be discussed in the discussion.

L96- what if an event is composed by both deep and shallow slides?

L100- Please be clear on: what is the cell size (in meters) at this latitude.

L100- How do you deal with multiple effects in which landslides are scattered across different cells?

Section 2.2 - Please state clearly if thresholds are already published (Zezere 2015) or if you updated them. Also, showing some equation or graph would be nice. 

L124-126 - this assumption requires a reference or an explanation

L171 - this is why the threshold should be visualized with a graph or an equation (see my previous comment)

L181 forecast = forecasted

L186 - you do not consider TN in your metrics. I guess it is to avoid to have a distortion of the resulting values because of the numerical disproportion between TN and TP. Please, state it clearly in the text.   

L210 find = found

L233 -How do you combine? I guess by considering landslides only when the requirements of both methods are met... but this is not trivial, please state it clearly.  

 L246-248 - This sentence is ot clear. Is this an example? Or 5 occurrences are considered "almost always"?

L305-306 - Please, rephrase

  L327 - dataset of was = dataset was

Section 4.3 - You discuss these differences, but you should also mention if and how they affect the results. Moreover, if you consider it a limitation, you should mention it. (I think it is nice when papers address also the main limitations of the works)

In the manuscript titled “Temporal clustering of precipitation for detection of potential landslides”, a novel approach is attempted by temporal clustering of the precipitation data to detect landslide events and compared with the use of classical empirical rainfall thresholds. The new approach based on temporal clustering has shown an overall good power of detecting landslide events for both sites but there is no clear outperformance over the classic rainfall threshold method. The novel clustering-based method has a higher sensitivity despite a lower precision than the threshold-based method and it had a fairly different performances depending on the site. Further the performances of the two approaches combined is evaluated for the region of Lisbon in Portugal. The hybrid method results showed a equal to or lower TP and FP than the one of rainfall thresholds alone. I recommend acceptance of the article following the minor revisions and clarifications suggested below.

• In section 2.2., time series window 1 to 90 days is considered based on what condition?
• In line 109, “(iv) the precipitation total preceding the landslide events, for windows of 1 to 90 days ending the day of the event, is computed”.  Could this be explained?
• Section 2.4. “we computed the presence of rainfall clustering preceding the second event with the modified series”. How is the presence computed?
• Fig 9 legend symbols doesn’t match the given plot
• Section 4.2 Details about the connection of evaporation and precipitation clusters?
• Section 4.3. “In general, only newsworthy content is reported by newspapers, which means that landslides that caused human damage or occurred in an urban environment are usually highlighted. For this reason, only landslides with a rainfall threshold with a return period of more than 3 years were used. The main aim was to reduce the possibility of including landslides with a triggering factor other than rainfall (e.g. human activity). Landslides with critical rainfall combinations with a return period of less than 3 years were assumed not to have been triggered by rainfall”