Summary:

            This article analyzes two different methods of disaggregating precipitation data to sub-hourly resolutions and compares it to high-resolution radar data over Germany and South Korea to evaluate its performance. Statistical metrics between WayDown and LetItRain overall show decent agreement to radar data, particularly in the mean and variance. Both methods overestimated extremes at low thresholds and underestimated extremes at higher thresholds, except for LetItRain, which showed comparable values at low thresholds in Seoul. The disaggregation models were then applied to climate projection data, where they were compared to radar data change factors and extremes. The modeled trends were generally opposite of the radar data in Germany for many statistics, while they appeared more similar in Korea. The authors concluded that applying disaggregation models to climate projection data should be done with caution.

Overall, this is an interesting paper, but some methodological questions need to be answered prior to publication.

General comments:

• I was confused by the methods applying the disaggregated precipitation models to climate projection data, which I hope can be clarified by some additional explanation.
  1. I do not understand how you obtain the RCP2.6 and RCP8.5 radar data. Is this what you refer to in lines 242–243 and lines 251–252 as the “1x1 km interpolated station-based RaKlida dataset” and “the regional climate model was downscaled to 1 km grid by the PRIDE model”? If so, stating that it is RCP2.6 and RCP8.5 radar data is misleading, because it is really just downscaled climate model data and not the observed radar data.
  2. I assume you are using the same downscaled data to ingest into WayDown and LetItRain as RCP2.6 and RCP8.5 radar data, correct? If so, please state that explicitly in the methods.
  3. Since you are using different models and different downscaling methods for climate projections and Germany and South Korea, I do not think it is fair to compare how well the disaggregation models compare against each other in those two locations because they have different model data they are ingesting. However, when you compare the “change factors” within each location, that is OK.
• Please improve the resolution of the figures, notably Fig.5’s legend, Fig. 6 and 7.

Specific comments:

• Line 144: When you say “binary 5-minute precipitation”, I am guessing you are referring to whether it rained or not? If so, please make that clear.
• Figure 2: Please make this figure caption clearer. I recommend referring to the appropriate panel after the corresponding text. I am guessing “daily value of 15 mm” refers to the left panel?
• Line 202: Do you calibrate the model using the raw future climate data? Please state what you use for calibration.
• Line 232–235: I am glad you use an algorithm to correct reflectivity, but I am wondering if you considered the impact of beam-blockage due to the mountains in South Korea? Did you account for this?
• Figure 6: I recommend labeling the five separate events with text or plotting the five different events in five different colors. Also, how do you determine which events count as separate, as some have multiple peaks in precipitation?
• Lines 295–300: Please refer to the subplots in the above figure to make the text easier to interpret.
• Line 389–390: I recommend putting more detail in the methods to explain how you obtained the 1000-year time series.
• Line 445–447/Fig. 13: Are you using the same time period to compare the radar data to RCP2.6 and RCP8.5?

Technical corrections:

• Line 28: Please remove “of” after “understanding”.
• Line 94–97: I recommend making these two sentences one sentence by stating …”driving variables, and which are not depending…”.
• Line 141: Please remove “i.e.”
• Line 304: You are missing a word between “of” and “means”.
• Line 505: Please remove “the statistics” after “For LetItRain” as this is redundant.

The manuscript addresses the topic of disaggregating daily precipitation at a sub-hourly time scale (useful for urban hydrology) by using two different models. The first one (WayDown) combines a first-order Markov chain with a 2D empirical beta copula, while the second one (LetItRain) is based on a double Poisson process and a gamma probability distribution. The models are validated by using 10-min radar data retrieved from 10 stations in Germany and South Korea, with different climates. Comparisons between the two models are made in terms of skill in reproducing various statistics over different time scales, also including the frequencies of different threshold values and extreme quantiles. Also, the change factors for various statistics are calculated by applying both models to climate projection data at daily and 10-min scales and compared to the corresponding ones computed based on radar data.

The manuscript is interesting and within the scope of the journal. It is well written and structured. From a methodological viewpoint, some clarifications are needed to better understand the reliability of the proposed approaches and the achieved results.

Specific comments follow.

Major comments

The authors distinguish between “disaggregation models”, for which the sum of disaggregated precipitation data is similar or equal to the original coarse precipitation values, and “stochastic models” aiming at reproducing statistics of the original precipitation time series. I do not agree with the terminology used to distinguish the two models (disaggregation versus stochastic), because both models are stochastic, as recognized by the authors at L 259. The main difference is in the statistics the two models preserve from the original dataset. Therefore, I suggest changing the terminology to prevent any misunderstanding.

Validation of both models is done by comparison of simulated and original radar datasets aggregated at the daily scale. In particular, several statistics at the monthly and annual scale for the whole and non-zero time series are used. The authors underline that a straightforward comparison of disaggregated time-series cannot be possible, as the WayDown model only keeps daily precipitation sums consistent with the input data (LL 259-264). This sounds like a major shortcoming of this model compared to the LetItRain model, given the first research question of the study. I suggest to better explain in the conclusions why this model has been chosen rather than other ones available in literature.

Concerning the validation in terms of consistency of extreme precipitation frequencies and magnitudes, it is not clear how the 10-min precipitation extremes are extracted from the reference and disaggregated datasets. As related to the frequency, the number of 10-min events exceeding given threshold values are calculated and divided by the corresponding number of events in 100 years. How overshooting events in 100 years are calculated? How is it ascertained that selected 10-min extremes are independent from each other (i.e., they don’t belong to the same events)?

Both models are applied to simulate 5-min precipitation data corresponding to the RCP scenarios 2.6 and 8.5. As far as I understand, WayDown is directly applied to the climate projection datasets, while LetItRain uses the change factor for the mean value and linear regressions between precipitation statistics to obtain parameters for future precipitation generation. In particular linear regressions between the mean and the variance, and between the mean and the proportion of wet periods are used, whereas for high-order moments, i.e., covariance and skewness, historical values are used from the original dataset, as the corresponding linear regressions are not suitable in these cases. Given that the authors state that there are indications thar high-order moments of precipitation will change in the future (LL 199-201), I wonder if the authors have tried to apply other techniques for developing nonlinear relationships (e.g., neural networks). If not, a motivation should be provided since, as the authors admit (LL 495-499), when statistics of the current period are preserved and assumed for the future , the application of the disaggregation models on the climate projection data should be done with caution.

Minor comments

LL 22-23: Clarify the meaning of “ensemble median” and “ensemble variability” in the abstract.

L 265: “for each station the time-series of 1000-year length equivalent were generated”. Do you mean 1000 time series of length equivalent to the observed series or do you generate 1000 years of disaggregated data for each station from which you sample several synthetic datasets of the same length of the observed data?

LL 276-278: “Corresponding disaggregated events were randomly picked from the models’ ensembles …”. This sentence is misleading. I suggest replacing “models’ ensembles” with generated or synthetic series.

Why change factors of the main statistics for the daily scale are represented in Fig. 12 for radar data only? Why the 10-min change factors for radar are not reported? Check the figure legend.