Dear Authors,

I read your paper and is a first of its kind study towards the use of remote sensing products for landslide study in India. But, there are queries regarding the use of precipitation products and "forecasting".

Attached is the annotated file where I have provided detailed comments.

Thanks

This paper presents the first examination of different rainfall products for landslide forecasting in India. The products tested includes a variety of satellite products with a range of spatial resolutions and daily or hourly temporal resolutions, which are compared to a ground-based rainfall product at relatively coarse spatial resolution and daily temporal resolution.  The testing relies on an extensive, high-quality landslide inventory database across different regions of India, which provides a suitable foundation for the conclusion that the higher resolution, hourly satellite products provide the best performing rainfall ED thresholds. Not surprisingly, the ground-based database performs poorest when evaluated with standard ROC metrics since the spatial and temporal resolution is inferior.

The paper provides a useful approach and template for future comparisons across different regions, and additionally the conclusions are also useful for application and testing of more detailed, regional rainfall thresholds across India. Overall the paper will be of interest to readers of HESS. There are some important questions and details related to the ground-based dataset, which need to be provided and discussed to provide a complete analysis of the results and conclusions. I also recommend a few references and discussion points that would provide fuller context of this contribution, since not all datasets are equal or equally applied and considered across the world. Otherwise, following these revisions listed below by line number, the paper should ultimately be published in HESS. 

130. This is very intriguing, but also a bit unclear. Could you provide a little more detail about the accuracy classification?

We introduced a semi-quantitative “confidence” metric in the U.S. inventory (Mirus et al., Landslides, 2020, https://doi.org/10.1007/s10346-020-01424-4), which is used to reflect overall confidence in information source and detail (position, nature and extent, actual occurrence of landslide). However, our approach was rather ad-hoc based on subjective expert opinion. I like your idea of splitting the temporal and spatial accuracy and so I (and probably other readers) would also like to learn more.

Specifically, in looking at Figure 3 it seems your rating has more to do with precision/resolution, rather than accuracy, at least for the timing (3c), but if the exact location of the landslide is unknown it is unclear how did you determine the spatial accuracy (resolution?) component.  

Please add one or two sentences for each accuracy classification to describe how it was developed and assigned, and be sure to clarify whether you mean actually “accuracy” in both cases or if some other description is more accurate.

149-153. While I recognize that this is an impressive number of rain gages and there must be some variability between them, it would be useful to describe what general type of rain gages (i.e., tipping bucket ormanual?) and the typical location of the gages (i.e. in cities/towns/airports?). If there is not pattern, perhaps some general description of the objectives of the Indian government in deploying this network.

While the satellite measurements will have bias based on the measurement techniques, the pixel averaging of the ground-based measurements will be greatly influenced by the type and location of these point measurements, so some understanding of this would be helpful.    

159&162&183. Why is 0.1 degree listed as 12.5km, 10km, and then 12.5km again? Presumably there is some variability over the study area, but if so the appropriate range or average should be specified.  

114-199. This process is not entirely clear. Since this is not a standardized calibration approach, a little more detail is necessary to understand and repeat this process. Also, how was the 5mm median determined?

222. Can you provide slightly more detail about the 100 synthetic series? For example, are these synthetic time series of a potential landslide-inducing rainfall event?

276-277. Yes, but some further discussion of this acknowledgement is needed in the following section.

304. Typo: improvement

306. Yes, we also found this in the San Francisco Bay Area (Thomas et al., WRR, 2019, https://doi.org/10.1029/2019WR025577) in which a network of 96 tipping buckets (~2.5km spacing) was more accurate at capturing the orographic influence on rainfall intensity and cumulative amounts, when compared to the IMERG dataset. In our case the density of gages is much higher and includes measurements across a range of elevations relevant to landslide initiation. I suspect that the IMD used in your study does not capture rainfall measurements in steep, landslide-prone terrain, which may further contribute to the underestimate. Thus the degree to which ground-based measurements out- or under-perform relative to satellite products may depend very heavily on where the in-situ measurements are placed within the landscape.

315. For clarity, consider stating “finer” spatial resolution. I have found that some people mis-interpret “higher” resolution as “coarser” resolution.

317. I think it’s worth noting here that Leonarduzzi and Molnar (NHESS, 2020, https://doi.org/10.5194/nhess-20-2905-2020) found that the daily vs. hourly resolution is not always so clear, but depends also on the temporal resolution of the landslide inventory. In your case your inventory has fairly precise timing information (Figure 3c). This relates directly to your next point about the duration of landslide triggering events as well.