See attached file

Review of manuscript "Improved rain event detection in Commercial Microwave Link time series via combination with MSG SEVIRI data" by Andreas Wagner, Christian Chwala, Maximilian Graf, Julius Polz, Llorenç Lliso, José Alberto Lahuerta, and Harald Kunstmann.

OVERALL ASSESSMENT

I've read your well-written manuscript on improved rain event detection in commercial microwave link (CML) data with interest. The topic is highly relevant to improve wet-dry classification, and of the most important steps in CML rainfall retrieval. Especially for the Global South, where CMLs have the largest potential for improving rainfall information. Although some earlier work employs geostationary satellite data for rain event detection for CML data, this study is based on a much larger dataset with much wider coverage (3901 CMLs in Germany over a 4 month period). Moreover, it not only evaluates the performance of satellite data as an auxiliary-data-based (ADB) method, but also compares it with the performance of time-series-based (TSB) methods. Finally, also the combined use of the ADB and TSB method is evaluated employing a new method making use of liberal and conservative detection depending on the threshold. To conclude, this manuscript is a useful and innovative contribution to the field of CML rainfall estimation. I find it quite surprising that the satellite-based ADB methods have similar performance as TSB methods, especially because of representativeness errors (differences in sampling volume, parallax, et cetera) and inaccuracies in satellite precipitation probabilities. Below, I provide some suggestions, corrections and recommendations.

SCOPE, APPLICABILITY AND OUTLOOK

The combined use of a satellite-based ADB method and a TSB method gives (slightly) better results compared to TSB methods only. The analyses are based on data from Germany. Below a couple of thoughts that could be incorporated to better frame the manuscript in the introduction or that could lead to additional recommendations:

1. This study is of course relevant for CML rainfall estimation in countries where TSB methods based on high temporal sampling can be applied.

2. Now the manuscript seems mostly relevant for Germany, but Germany could also be seen as a testbed with relatively good gauge-adjusted radar reference data. Hence, this study is especially relevant for the Global South, where usually no other auxiliary (near real-time) data exist (ground-based weather radars and rain gauge data are sparse). In the Global South, part of the CML network is in rural areas, where densities may be too low to apply a "nearby-link" approach. In addition, often 15 min data are available, which limits the applicability of TSB methods. Because the performance of satellite-based rain event detection is evaluated and shown to be useful, this manuscript is especially relevant for the Global South.

3. One could recommend to test a combined satellite-based ASB method and compare it to the "nearby-link" approach, and also to combine both methods and evaluate its performance. This would especially be relevant for the Global South.

4. The authors state that "best results are usually achieved with radar data". These data are available for Germany. One could recommend to explore the use of radar data for rain event detection and also combine it with a TSB method. Naturally, when no radar data are employed, the results for CML rainfall estimation in Germany solely show the performance of CML, which is relevant (testbed), and this already provides good rainfall estimates. So this manuscript is not only relevant for other regions, but also for Germany as such. Results for Germany could improve, though, when radar data would be employed for wet-dry classification.

5. What do the authors expect in terms of performance of satellite-based ADB methods when data from the new Meteosat Third Generation (https://www-cdn.eumetsat.int/files/2020-04/pdf_mtg_info-pack_v12.pdf) would be employed? Its higher spatial and temporal resolution is beneficial and closer to the spatiotemporal resolution of CMLs, although parallax will still give rise to representativeness errors, especially in the mid-latitudes. Perhaps that more channels will allow for improvement in precipitation probability estimation. It could at least be worth mentioning MTG around L. 384.

METHOD AND RESULTS

1. Why did the authors choose these MSG products? The use of a geostationary product is clear because of its 15 min, or better, temporal resolution. But other products exist.

2. L. 145: "This baseline is the last dry time-step of the TRSL time-series" suggest that the baseline is based on only 1 data point. In not, please clarify which period preceding a rainfall event is considered for the computations, and how many dry time steps / data points are needed for its computation.

3. L. 154: Good agreements with what? Clarify, e.g., the CML rainfall estimates with ITU parameters compare well with reference data.

4. L. 159: Standard deviation of what variable? TSL, RSL, TRSL?

5. L. 233: "the largest differences": add "between the chosen probabilities".

6. L. 263: "behaved very similarly" seems especially the case for most graphs for p01, but larger differences are found for p10 and p20.

7. General remark: the readability of figures would be improved if the variable name and unit would be added to the vertical axes (Figures 3, 4, 5, 7).

8. You could consider using the present tense instead of the past tense when describing results.

9. L. 296-297: rephrase a bit, because the ADB method itself is completely independent from the CML time series, but it is about the effect of CML time series after applying a rain event detection method.

10. Caption Figure 6: make clear that the green line is a CML-based rainfall intensity.

11. L. 320-322: I find these lines difficult to follow. Can you explain more clearly how PC10all is obtained?

12. L. 325: it seems that one selection is made from the best of six combinations. Do the "three data sets" represent TSB only (CNN), ADB only (PC10), and the TSB & ASB combination (PC10all)?

13. L. 332: it seems at most ~0.15, and especially an improvement with respect to RS and a slight improvement with respect to CNN.

14. Caption Figure 7: "Same as Fig. 4" should be "Same as Fig. 5".

15. Figure 9: make square plots, so that the horizontal and vertical axes get the same scale and the dashed x-y line gets an angle of 45 degrees. Then it becomes more obvious that pc10all outperforms cnn. It would facilitate the comparison of performance between methods.

16. The authors could provide mean and/or median values of the metrics in Figure 9 for both datasets in each figure, to summarize the performance of each method. And if I understand correctly, this is already presented in Figure 7 for MCC. So, the reader could be pointed again to Figure 7 to emphasize that it contains the MCC metric for CNN, PC10 and PC10all (which are shown in Figure 8) and to help the reader to find the overall results.

17. I probably missed it, but what is the time interval or duration for which the rain event detection is evaluated? Is it 15 min or 1 h? When comparing to radar data, it is beneficial to use a somewhat longer duration, such as 60 minutes, to limit representativeness errors (e.g., differences in sampling volume, time it takes for precipitation to fall from the radar sampling volume to the Earth's surface, advection of precipitation). This is at least common for rain gauge versus radar comparisons. In the case of CMLs, one could argue that its scale is somewhere between the point and radar grid cell size, making it less vulnerable to representativeness errors.

REMAINING MINOR REMARKS:

- L. 28: A tipping bucket represent one important and frequently used type of rain gauge, but these are probably not considered the best rain gauges given limitations in sampling of low- and high-intensity rainfall related to the tipping bucket volume and the number of possible tips. So stating that these tipping buckets provide "usually provide the best point measurements of precipitation" will not generally be true. This statement could be weakened a bit.

- For instance, at L. 44 "precipitation intensity" is used, whereas CMLs are typically useful for estimating "rainfall intensity". In case "precipitation intensity" is kept, I suggest to mention once in the introduction that CMLs and the employed algorithm are typically suited for rainfall estimation, i.e., liquid precipitation estimation.

- L. 116: "No specific device for precipitation measurement is on-board": could be made more specific (space-borne radar and/or radiometer).

- L. 147: "radome" is typically used when referring to the protective cover of weather radars. Perhaps "cover" is a more appropriate word for CML antennas.

- Could the authors add some information on the availability of the radar, CML and satellite data sources (e.g., expressed as a percentage)?

- Data availability: are the geostationairy satellite data publicly available? If so, please add where the data can be retrieved.

- Titles in the reference list: the use of capital letters is not consistent, e.g., Steiner et al. (2004) versus Wang et al. (2012).