the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 20 Jul 2026
High resolution rainfall estimation from Commercial Microwave Links in the city of Niamey in Niger
Abstract. Commercial microwave links (CML) from cellular phone networks have become a source of opportunistic information on rainfall. In this study we use a dense network of 144 CMLs over the city of Niamey, in Niger, to estimate rainfall with a 15-minute resolution. The validation of these estimates with rain gauges shows that for links located within 2 km of a gauge the correlation of 15-minute time series is around 0.85, and the overall bias is low (0 to 10 % depending which link length and frequencies are considered). The CMLs are well suited to detect and quantify the intense rainfall rates associated with the mesoscale convective systems which bring most of the rain – and often floods – in the region. These good results are obtained with a simple and easily transferable processing chain. The study confirms the interest of CMLs in otherwise data scarce areas.
- Preprint
(2429 KB) - Metadata XML
-
Supplement
(1066 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-2962', Anonymous Referee #1, 29 Aug 2026
-
RC2: 'Comment on egusphere-2026-2962', Anonymous Referee #2, 18 Sep 2026
Review of the manuscript "High resolution rainfall estimation form Commercial Microwave Links in the city of Niamey in Niger" by Modeste Kacou, Marielle Gosset, and Inigo Reizabal.
Many studies have shown that commercial microwave links (CMLs) in cellular telecommunication networks can be employed for rainfall estimation. It has potential to complement rainfall estimates from rain gauges and ground-based weather radars. The potential of CMLs to improve rainfall information and its spatiotemporal resolution is highest in (virtually) ungauged regions. These regions are often in the (sub)tropics, where rainfall estimation is even more challenging due to high spatial rainfall variability. This manuscript is relevant, since relatively few studies have addressed CML rainfall estimation in those regions. I do have some remarks, though, for instance, concerning the (underpinning of) methods, the layout of some of the figures and the wording.
A general comment: CML rainfall estimation is often challenging given the sparse reference data and high spatial rainfall variability. This study would benefit from analyzing more rainfall events, but perhaps these were the only strong(er) rainfall events for the selected May - August period? Or reference data were not available for other events? Is there any particular reason why these events were selected? In addition, the selection of rainy events, raises the question whether any dry time intervals were analyzed? An important aspect of CML rainfall estimation is their ability to prevent non-zero rainfall estimates during dry periods. These seem not present or underrepresented in the current study.
Another important thing that stands out (L. 229 - 230): it seems that wet-dry classification is performed by using the closest rain gauge. If yes, this creates a dependency on the rain gauge data and limits the general applicability of the methodology, especially in terms of coverage, but perhaps also for operational applications, since rain gauge data are often not available in real-time. Please also discuss why other wet-dry classification methods were not considered.
To conclude, the manuscript needs moderate revision, and helps to advance CML rainfall estimation in data sparse regions.
Section 2
- The term "sub links" is used once and not defined. Please define CML and sub link, and clearly mention in results in case multiple sub links from the same path are employed (then you have less unique link paths).
- L. 137: "For some links, the transmitted signal levels were also provided": for the other links transmitted signal levels were not provided. Did the mobile network operator confirm that transmitted signal levels should be nearly constant? Did they confirm that automatic transmit power control was not used?
- L. 141: please underpin why you use mean RSL, e.g., by employing Van der Valk et al. (2023).
- L. 166: "Note that Rain Gauge ...": remove this line, since this is already clear from Table 1.
Section 3
- L. 185: Please provide more details on the employed DSDs, especially which country/region and if climate is similar to the study area.
- L. 215-217: I agree that the baseline matters, but what probably matters more is a correct wet-dry classification. This prevents non-zero rainfall estimates during dry periods. This is also beneficial for obtaining an accurate baseline, since it should be computed from dry intervals only. Hence, I suggest to give more attention to the importance of wet-dry classification and be a bit more specific.
- Section 3.2: "raw attenuation" is not attenuation, but the signal power. It only becomes attenuation when compared to transmitted signal level (Eq. 3) or when the received signal power is subtracted from the reference level. Hence, it would be better to called it, e.g., Praw & Pbase. Also be aware of the minus sign, since adding a minus sign to Abase is not defined in the text. Eq. 5. should become Arain = - P + Pbase.
- L. 248: CMLs (paths) or sub links (possibly multiple links on the same path)?
- In my opinion, Figures 3 & 4 provide a nice visualization.
- L. 297: what was the maximum rainfall rate encountered in the rain gauge data and based on what temporal resolution? I assume these were tipping bucket gauges, which probably have difficulties in measuring more than 100 mm/h anyway. Emphasize that both the longer length of time intervals, 15 min, as the path length (typically a few hundred meters to kilometers), will lead to lower mean CML rainfall intensities on average compared to, e.g., 5-min point values from rain gauges.
- L. 298: correlation and median of what?
- L. 298, L. 322: And how did you determine the distance? Using the middle of the links?
- L. 319: this also assumes that other sources of error do not play a role, e.g., attenuation due to refraction, blockage by obstacles (similarly stated at L. 363-365, but this "other sources" should also be mentioned here).
- L. 322: "29 CML-gauge pairs": can you clarify whether "only" 29 pairs instead of the earlier mentioned 144 CMLs are used in this study? Stated differently, clarify where results are based on 144 CMLs? (perhaps I just missed this).
- L. 325: "the intersect parameter". You probably mean "intercept parameter". Note that the fitting line is not present in Figure 5a-b, whereas it is described on L. 325.
- L. 329: probably the "* 100%" is missing, also in Figure 5c. Otherwise the relative difference would have a median value of 5 for 0-1 km links. Moreover, the caption of Figure 5c mentions "percentage of residual attenuation".
- Figure 6: a difficulty with WAA models is that one usually does not know whether zero, one or two antennas are wet, especially given the expected high spatial rainfall variability in a tropical climate. This study assumes 2 antennas (L. 346). All this should be discussed a bit more around the already good discussion starting at L. 355.
- L. 385: the improvement can also be due to the correction for WAA, since this also reduces attenuation during dry intervals.
- Figure 9 caption: mention that these are 15-min rain rates.
Section 4:
- The metric KGE is used, but has not been (briefly) explained.
- L. 423-425: the bias is better for Figure 9e,f, but worse for Figure 9d.
- L. 427: which "set of algorithm parameters" should be considered? I.e., related to which processing steps.
- L. 462: the detection of low rain rates may also be a challenge for tipping bucket gauges (e.g., 0.5-mm resolution).
- L. 504-505: "robust information on the mean rainfall over Niamey can be derived from the ensemble of all links": can you underpin this? Because correlation between CML rainfall estimates is high?
- L. 507: in the manuscript you only refer to supplement figures S1 and S11, whereas figures S1 - S11 exist. Add references to the other figures, where appropriate.
- L. 519: "south-east of the city": it seems the western part of the city.
- Figure 14: use the same axes for all figures.
- Figure 14: what kind of interpolation has been done and what is the difference between b) and c) & d)? The area with rainfall values is much smaller for (b). Moreover, c) & d) provide 1-km resolution, and b) shows values for a kind of districts. Why?
- Figure 14: quite some extrapolation in c) & d).
- Figure 14: use the same color scale for all figures.
- Figure 14b: why the rain gauge value of 76.04 mm (Table 1) has not been plotted in b)?
Conclusion:
- "which is unusual under such latitude": to estimate rainfall at 15-min resolution is common when using CMLs. You probably mean to say that 15-min rainfall information is uncommon "under such latitude", because few or no rain gauges or ground-based weather radars are operational to provide such information.
- Out of curiosity: do the rain gauges provide real-time information? I find it interesting to read that the mobile network operator was able to provide CML data in real time.
- L. 580: replace "Global Microwave Data Initiative" by "Global Microwave link Data collection Initiative".
L. 50, L. 208: Note that the reference to Leijnse et al. (2007) is not correct. The authors refer to:
Leijnse, H., Uijlenhoet, R., and Stricker, J. N. M.: Hydrometeorological application of a microwave link: 1. Evaporation, Water Resources Research, 43, https://doi.org/10.1029/2006WR004988, 2007.
However, that paper is about evaporation, not precipitation estimation. The appropriate for reference for L. 50 is:
H Leijnse, R Uijlenhoet, JNM Stricker, Rainfall measurement using radio links from cellular communication networks. Water Resour Res 43, W03201 (2007).
This may also be the appropriate reference for L. 208, but perhaps it is:
Leijnse, H., R. Uijlenhoet, and J. N. M. Stricker (2007b), Hydrometeorological application of a microwave link: 2. Precipitation, Water. Resour. Res., 43, W04417, doi:10.1029/2006WR004989.
Typos / grammar / equations:
- L. 41: replace "problemic that" by "problematic given that".
- Caption Figure 1: replace "blue et green" by "blue and green".
- L. 146: replace "the the" by "the".
- L. 160-161: replace "when the rain gauge are compared with the CML estimation" by "when the rain gauge estimates are compared with the CML rainfall estimates".
- L. 186: replace "very closed" by "very close".
- L. 263: replace "CML estimation" by "CML rainfall estimation" (i.e., include the name of the variable that is estimated).
- L. 334: replace "frequency range" by "frequency ranges".
- L. 460: replace "very small links" by "very short links".
- L. 461: "25-30%": it seems "21-34%". "about 35%": it seems "36%"
- L. 463 - 464: The opposite is true. WAA correction reduces the computed value for attenuation as used to estimate rain intensity.
- L. 497: replace "for each 10 events" by "for each of the 10 events".
- L. 520: replace "rainfall are seen" by "rainfall is seen".
- L. 551: replace "lack of operational" by "lack of an operational".
- L. 560: replace "CMLS technique" by "CML rainfall estimation technique".
- L. 564: replace "ftp" by "FTP".
- L. 578: "proposed solution was implemented in real time": you mean the CML data acquisition? The rainfall estimation itself was not done in real-time, I suppose.
- Use italic fonts for variables and normal font for text (subscripts, e.g. Arain in Eq. 2, where "rain" should be in normal font). Check all equations, L. 180, 193, 205, 314, 329, 350. Basically, check the entire manuscript. In addition, please use subscript for A0 and Amax in Eq. 10. Be consistent with the multiplication sign: a dot in some equations, a * in Eq. 10.
- One reference is duplicate: "Overeem, A., Leijnse, H., and Uijlenhoet, R.: Measuring urban ...".
References:
Van der Valk, L. D., Coenders-Gerrits, M., Hut, R. W., Overeem, A., Walraven, B., and Uijlenhoet, R.: Measuring rainfall using microwave links: the influence of temporal sampling, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2023-1971, 2023.Citation: https://doi.org/10.5194/egusphere-2026-2962-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 198 | 87 | 31 | 316 | 60 | 20 | 23 |
- HTML: 198
- PDF: 87
- XML: 31
- Total: 316
- Supplement: 60
- BibTeX: 20
- EndNote: 23
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Modeste Kacou
Iñigo Reizabal
This paper illustrates that in regions that suffer from lack of meteorological information, the cellular phone network can provide rainfall measurement and mapping. This is illustrated in Niamey, the capital city of Niger using data provided by a telecom operator during 2017, a season marked by many intense rainfall. We analyse the microwave attenuation and transform it into rainfall maps, every 15 minutes, showing very high rainfall rates related to intense tropical convection.
This paper illustrates that in regions that suffer from lack of meteorological information, the...
Review of the manuscript „High resolution rainfall estimation from Commercial Microwave Links in the city of Niamey in Niger“ by Kacou et al.
# Summary
This paper presents the processing and analysis of CML data from the city of Niamey for ten strong rain events. A new method for determining the baseline is introduced. Filtering of outliers and WAA estimation is also done. An interesting comparison of rainfall estimates before and after filtering and WAA correction is shown, which provides valuable insights. Further statistics are provided and rainfall maps are shown, but only very briefly. In general, the work behind this manuscript is very relevant because studies on CML rainfall estimation in Africa are still limited. The paper is also suitable for HESS. However, the manuscript has several weaknesses, which will require a major revision.
# Major general comments
1. No justification of new methods: The authors have introduced a new method for determining the baseline level. They write that they have also tried other methods. It remains unclear what the advantages and disadvantages of the different methods are, though. A method intercomparison is probably beyond the scope of this manuscript, which provides its main value by doing an analysis for a new CML dataset in a new region in Africa. But it should be discussed in detail why specific existing methods did not perform adequately and why the newly introduced method was a better choice. Also, the limitations of the new method should be discussed. Of course, if the authors wish, they could also add a quick exemplary comparison to other methods. Similarly, it should be discussed in more detail how filtering and WAA estimation is similar or different to existing methods.
2. Limited interpretability of the results: The authors pick ten rainy days with strong rain events. Because of this, it remains unclear how their processing would perform over longer continuous periods which included longer dry periods with higher chance for false-positive rain rates and rain events with lower rain rates. I think the analysis can be kept like it is, but the interpretation of the results has to be done more carefully. Claims about real-time and operational applicability cannot be derived from this type of analysis. Furthermore, the section on rainfall maps is very short and misses required details. This section should be extended. In addition, in my opinion a short comparison to satellite rainfall products should be added, e.g. using IMERG-early as an „almost real-time“ product. This comparison would put the performance of the CML rainfall estimates into perspective.
3. Mediocre presentation quality: The writing is not on the level required for HESS. Many sentences contain grammatical errors or would benefit from improved phrasing. The writing in the whole manuscript needs to be revised with the help of an experienced author. In addition, the figures have mixed quality. Figure 3 and all the similar ones are nicely made, but e.g. Figure 5 looks like copy-paste of the subplots. The Figure style is very different in Figure 10, 11 and 12. Figure 11 is particularly hard to read. Also Figure 14 should be improved. See my specific comments below.
# Minor general comment:
There are many long lists of references for which it is unclear what the benefit for the reader is. E.g. L207 has 9 references which probably all contribute something different to this sentence. Another example is L178 which has four reference of which two are the historical standard reference, but two newer ones, in particular one from 2025, are added, with unclear motivation. What do these reference provide in addition to the two older ones? I would suggest to improve the usage of references in the manuscript.
# Specific comments
L58 and following: This paragraph on state-of-the-art is missing relevant recent work. The work from Kumah et al. 2021 (https://doi.org/10.3390/rs13163274) and 2022 (https://doi.org/10.1016/j.atmosres.2022.106357) from Kenya should be mentioned. Also the work from Blettner et al. 2025 (https://doi.org/10.1175/JHM-D-24-0157.1) needs to be briefly described here.
L105: Was polarisation data provided or not? If not, did you assume that all links are vertically polarized? What is the error in rainfall rate if polarisation is wrong?
Fig1a: What is the meaning of the different line widths?
L133: It is not clear how the CMLs along the same paths are different. Do they have different frequencies or different polarisation? There cannot be a CML with same frequency and same polarization along the same path. It is also not clear from the text above if you count sublinks as CMLs. Are there some paths with only one sublink? That would be very uncommon. See Fig 3 of Andersson et al 2022 (https://doi.org/10.5194/essd-14-5411-2022) for the suggested definition of sublinks and CMLs.
L138: „…TSL are constant, very stable…“ What does it mean that TSL is very stable? Is it constant or not? Also „…as in most studies…“. Not sure if there are really more studies where TSL is assumed to be constant. I would tend to say no. Maybe just formulate this a bit more carefully.
L141: „For the present work we used only the mean RSL…“. This is interesting and should be elaborated on. Other large studies with 15-minute data, notably the country-wide analyses from the Netherlands, use min and max. Why did you not use min RSL? It contains valuable information about the most extreme rain rate during the 15-minute interval. Also, is the mean RSL biased because of the distribution of RSL values within the 15-minute aggregation window? Please make these things clear in the text.
Table 1: Is the maximum rain rate based on the data aggregated to 15-minutes, which is used for comparing to CML rain rates, or based on the original 5-minute resolution?
L170: The technique for CML rainfall estimation has been introduced already in the intro section with appropriate references. I do not understand why two new references are needed here. In particular the first one, which seems to be in French, is fitting for a HESS manuscript.
L184: How was polarisation accounted for? Isn’t Mie theory only working for spherical scatterers. If you assumed spherical rain drops, what are estimated errors when comparing the T-Matrix calculations for oblate spheroid rain drops and considering vertical and horizontal polarisation?
Table 2: Would be helpful to provide the values for a and b from ITU-2003 or ITU-2005 here as comparison.
L224: The method from Schleiss and Berne 2010 was not developed for 15-minute data. The nearby-link approach, used e.g. in Overeem et al 2011, is a much better example.
L226 and following: „This method was chosen for its skill…“ How did other wet-dry methods perform, e.g. the nearby-link approach which seems very well suited for such a dense urban CML network? Did you also check performance regarding false-positives during longer dry periods? What to do if there is no nearby rain gauge and one cannot focus the analysis on specific rain events? Is this method robust enough for real-time application where it continuously has to deal with potential false-positives during dry periods but also needs to be sensitive to the onset of smaller rain events? Please discuss this here in much more detail. In particular why you did not choose one of the existing methods?
Figure 2 and related text on baseline determination: How would the baseline look like for an event with longer duration? Event 2 has 11h duration. Isn’t there the risk to pick the wrong value during such long rain events. If there is a bit of fluctuation before and after the event, the baseline might not be from the dry period before the event. Is the 24-hour mode a rolling window value or is it calculated from the 24 hours before the event? If this is the case, how is the onset of the event set?
Section 3.3: Please mention other methods from the literature to filter out problematic links or time series. Why did you not choose these methods?
L345: How is the equation applied during CML data processing? It depends on rain rate. But WAA will influence the CML-derived rain rate? Would it be better to have the function depend on A_rain?
L352: What is the reasoning behind having an initial WAA value? The other models you cited (Schleiss, Lejinse, Kharadly and Ross) do not have that? Could the 1dB quantisation be an effect that requires this initial value? Or is it because of the 15-minute temporal resolution? Would be good to explain this with one or two sentences.
L370: How fast is this? Does that mean that, for every data point where CML attenuation is converted into rain rate, a numerical solver has to be run? If it is a bit slow, couldn’t this be improved with a precomputed look-up-table for each CML frequency (maybe at 1 GHz steps)?
L407: Not clear what „base“ means in „CML base estimation assessment“ here.
L408: Probably should be „section 3“ and not „section 2“
L445: Given that the whole analysis is carried out for selected days with strong rain events, the results in this section do not provide a real assessment of the method’s performance. The tradeoff between POD and FAR is much easier to balance when only selected periods with a lot of rainy data points are used. Maybe POD could still be interpreted, but FAR is not reliable if long dry periods are not included in the analysis. Please either adjust this section or maybe leave it out completely because the numbers are not helpful if the analysis is not robust enough.
L460: What are the reasons for the high FAR for „very small links“, which probably should be „very short links“?
Figure 11: This figure should be improved. It is really hard to read. The alignment of the text and info, the small font sizes, etc. In addition, I find it a bit hard to quickly interpret the data with the different bin sizes on the x-axis. Wouldn’t it be easier to show PDFs with an equal bin size, and maybe in addition a CDF for the accumulations to highlight the contribution to the overall rainfall sum?
L487: It would be good to comment here again on the detection limit given the 1 dB quantization. Is this also the effect of WAA that maybe pushes some of the CML rain rates, e.g. from 2-5 mm/h to 0 mm/h for the short CMLs? Please comment.
L500: Probably should be „Figure S10“ and not „Figure S1“.
Figure 12: Instead of mentioning the correlation between the two gauges in the text, wouldn’t it be good to add this gauge-gauge correlation to this plot, e.g. as a coloured marker per event?
L503: Is it shown somewhere that „correlations are lower when considering links which are further away from the gauges“?
L508: „…average of the gauges, over Niamey area“ sounds as if there were more gauges in the area. But here, you probably mean that it is the average of the two or three gauges available, correct? Should be clearer in the text.
L510: „…suggesting that the median could be used as a robust estimator of the average rainfall over the city and the interquartile as an indicator of the variability.“ Since we only have data from two gauges here and we do not see the variability of the gauges, I would not draw conclusions about the „average rainfall over the city“ and the suitability of quartiles to indicate rainfall variability. Looking at the map in Figure 1 and knowing that the gauge at the airport is only available for one event, there could be higher or lower rainfall in the eastern half of the city, which the gauges would not see, but the CMLs. On the contrary, some erroneous CML rainfall estimates might lead to high variability of the CML rainfall estimates that do not necessarily say something about the rainfall variability within the city. Please adjust this section.
Figure 13: I actually find that Figure S11 would fit directly here in the manuscript. It is the only figure where you show all analyzed events. If you use two columns (with shared y axis) and five rows it would be easily readable and fit on one page. I would also suggest to draw the gauge data as individual lines to show their level of agreement or disagreement. To reduce clutter in the figure, just use a shaded area for the q25-q75 region and just plot the q50 as line.
L516 and following: How was the map in Figure 14 c and d produced? Is this an interpolated map from CML rainfall rates? If not, what is it based on?
Figure 14: This figures needs to be improved. The subplots are not aligned. Also, if subplots shared the same axis, you should leave out the duplicated axis tick labels and axis labels. The plot titles should make it clear that the top row shows rainfall accumulations for one full day and the plots below show rainfall rates. Maybe you could also show the full sequences of rainfall maps for this event. This would be interesting. Because these maps will be read only qualitatively and not quantitatively, they could also be a bit smaller so that 24hours fit on one page, e.g. as 4x6 subplots.
L516 and following, again: Overall, I find this section quite short, given that this is the first and only time you show rainfall maps. To me, this could be more interesting and important than the POD and FAR analysis on which I commented above.
L536: „…which is unusual under such latitude“. I do not understand why this would be unusual. There have been CML rainfall analysis with 15-minute data in Africa.
L545: „…the total amount of rain brought by each intensity class is well represented“. This is not true. 2-5km for short links shows strong underestimation of CML rainfall.
L558 and following: There are too many technical details and longish explanations of the meteorological situation. For a conclusion section this should, in my opinion, be shortened. Also, the text till the end of the conclusion section should be divided into several logically separated paragraphs.
Bibliography: There are some duplicated entries. Please check carefully.
# editorial comments
L34 „in most of the earth surface...“ here the „in“ does not seem correct
L36 „Combining both sensors…“ probably needs a comma after „sensors“ and after „time steps“, also not sure if „infra-hourly“ is a correct English term
L37 Should be „Radar-based“
L39 „singularly“ does not seem to be the correct word here
L41 „This observational gap is all the more problematic that…“ this does not seem to be a correct English formulation.
L51 I have never heard the term „liaison“ in this context. Maybe wrong.
L55 „…CMLs use for rainfall…“ does not seem to be correct English.
L71 has to be „carried out“
I stop with the detailed editorial comments here. Even above, I already skipped some minor errors. Please do a very thorough proof-reading and improvement of writing when preparing the revision.