the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 01 Oct 2026
Brief communication: GNSS water-vapour fluctuations read the air mass of the 2024 Rio Grande do Sul and 2022 Recife floods hours ahead of the heaviest rain
Abstract. Satellite-positioning antennas measure the water vapour above them, and the minute-to-minute fluctuations of that measurement, usually smoothed away, describe how the moisture is arranged. We read them at eleven Brazilian antennas during two floods with little or no radar coverage, in Rio Grande do Sul in 2024 and at Recife in 2022, against six-hourly rain gauges. Over a season they separate wet from dry periods 6 to 12 hours ahead, with an area under the receiver operating characteristic curve of 0.80 to 0.87. In both floods the antennas rose into their top 5% hours before the worst rain.
- Preprint
(9856 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 07 Dec 2026)
-
RC1: 'Comment on egusphere-2026-5548', Anonymous Referee #1, 27 Oct 2026
reply
-
AC1: 'Reply on RC1', gael kermarrec, 28 Oct 2026
reply
Author comment on RC1, egusphere-2026-5548
I thank the referee for a careful reading. The comments converge on one point, which the revision now makes explicit from the first page: the quantity read in this communication is the fluctuation of the wet delay, described by two spectral parameters, not the amount of water vapour. The referee's comments and my answers follow in order.
1. Novelty, and the two missing references. Both are now cited and distinguished in the introduction. Sapucci et al. (2019) defined the jump of GPS-PWV before intense rain, and Gouveia et al. (2026) found such jumps about one hour before the heaviest rain of the 2024 episode at POAL and SMAR, two of the antennas used here. Both read the amount of vapour, smoothed in time, in the hour before the rain. This communication reads a different quantity from the same record, the texture of the fluctuations (σ and λ, defined below), together with the amount, and reads it against a whole season rather than against the preceding hours of the same day. The lead time is correspondingly different: 6 to 12 h over the season, measured as an AUC on 52 wet and 1040 dry six-hour periods in Rio Grande do Sul, against one hour for the jump. The introduction now states the three points on which this reading differs: the quantity (the spectrum of the 30 s fluctuations, which no smoothed product carries), the reference (a state judged against the dry season of the same antenna, so that a level fixed in advance has a measured false-alarm rate) and the range (6 to 12 h over a season). The conclusions present the jump and the seasonal state as two readings of the same record at two ranges.
2. Conversion of ZWD to IWV. There is none, and the revised Sect. 2.2 says why. The wet delay is used directly in millimetres of delay; the conversion to integrated water vapour is a multiplicative factor close to 6.5 mm per kg m⁻², which cancels in the standardisation of every index by the antenna's own median and interquartile range. The hydrostatic part is removed by model inside the positioning solution, where the wet part is estimated as a random-walk parameter.
3. Definition of σ and λ. Sect. 2.2 now gives the spectral model, Eq. (1): a Matérn spectrum of fixed smoothness ν = 5/6, flat below a corner frequency λ and falling as f^−8/3 above it, plus a white-noise floor N. The parameter σ is the standard deviation of the turbulent fluctuations, in millimetres of delay; λ is the corner frequency, in s⁻¹, the rate at which the moist structures crossing the antenna renew themselves. Both are fitted by maximum likelihood (Sykulski et al., 2019) in sliding windows of 62.5 min after removal of trends slower than 25 min, then standardised per antenna. Physically, λ falls and σ rises when the moist structures grow larger and more vigorous, as they do before convection (Kermarrec et al., 2025). The fitted parameters, the code from the parameters onward and the outputs are deposited (DOI 10.25835/v40ouls8).
4. Why 30 s, and how signal is separated from noise. The 30 s sampling is not chosen for the lead time; it is what makes the estimation of σ and λ possible. Separating the bend of the spectrum from the flat floor requires fluctuations resolved at periods from one minute to half an hour, which a 62.5 min window holds as 125 samples. An hourly product contains none of this band and a 5 min product only part of it. Receiver and estimation noise are flat in frequency and are absorbed by the floor N of Eq. (1), so that σ and λ describe the turbulent part of the record alone. The lead time of hours is the averaging time of these parameters before a rain period, a different quantity from the sampling. The revised Sect. 2.2 makes this distinction.
5. Warnings throughout the period; no distinction between dry, ordinary and extreme periods. The shaded band of Fig. 2 is not a warning. It marks the top 5 % of the dry-season values of the same antenna, a level reached by construction on one dry period in twenty. The alert is the persistence of the state in that band for two consecutive periods, and it fires once in each episode; the revised Fig. 2 marks these two alerts with triangles. The distinction between dry, ordinary and extreme periods is the subject of Table 1 and Sect. 3.1, measured over the whole season with the flood windows excluded: wet (≥ 10 mm in 6 h) against dry (< 0.1 mm) periods are separated with an AUC of 0.80 to 0.87, and the declared rule fires on 16 % and 11 % of the dry days. Extreme against ordinary wet periods are not separated (AUC 0.38 to 0.60), which Sect. 3.1 states. That vapour scores highest against a wet-or-dry truth is expected, since that truth is the quantity of water; texture measures whether the column is organising rather than how moist it is, carries two thirds of the skill of the amount on its own, and is the index that still moves over the saturated column of Recife. Sect. 3.1 and the conclusions now say this. That the state sits high through most of both flood windows is the observation: these are the days on which the dangerous air mass was present.
6. Decrease of IWV after the rainfall maximum. It is visible at PERC, which falls during the deluge of 27 to 28 May and returns to the 98th to 100th percentile once the rain has passed. The revised Sect. 3.2 names this fall as the drying that heavy precipitation imposes on the column, and the return as the sign that the air mass had not left. This is also why the two parameters matter beside the amount: over a saturated column the amount can no longer rise, whereas the arrangement of the moisture keeps varying.
7. Physical processes and reading of the figures. The revised introduction states the mechanism (moist structures growing and strengthening before convection, read in the spectrum as a fall of λ with a rise of σ), and Sects. 3.2 and 3.3 now tie each feature of Fig. 2 to the gauge record: the onset at the 92nd to 94th percentile, the crossing before the heaviest periods, the fall during the Recife deluge and the single alert per episode.
8. Why the departure from the seasonal median rather than the jump. Because the two answer different questions at different ranges. A jump is a change within the hour before the rain, read against the preceding hours of the same day; it says that rain is imminent. The seasonal percentile is a state read against the dry season of the same antenna; it says how unusual the air above the antenna is at that moment, so that a level can be fixed before an episode and its false-alarm rate measured on the dry periods of the season. The first gives one hour, the second gives 6 to 12 h with a measured cost. The revised Sect. 2.3 and the conclusions state this contrast.
Changes to the manuscript: introduction rewritten in its second paragraph (definitions, Sapucci et al. 2019 and Gouveia et al. 2026) with the three points of novelty stated in the third; Sect. 2.2 rewritten with Eq. (1), the treatment of noise and the role of the 30 s sampling; one paragraph added to Sect. 2.3 on the seasonal reference; one sentence added to Sect. 3.2 on the fall of PERC; one paragraph added to Sect. 3.3 on what the band and the alert are; two sentences added to Sect. 3.1 and to the conclusions on what texture reads that the amount does not; Fig. 2 revised with the alert markers; one sentence added to the conclusions on the two ranges. Two references were removed to stay within the limit of twenty.
Citation: https://doi.org/10.5194/egusphere-2026-5548-AC1
-
AC1: 'Reply on RC1', gael kermarrec, 28 Oct 2026
reply
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 245 | 123 | 42 | 410 | 48 | 44 |
- HTML: 245
- PDF: 123
- XML: 42
- Total: 410
- BibTeX: 48
- EndNote: 44
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Title: Brief communication: GNSS water-vapour fluctuations read the air mass of the 2024 Rio Grande do Sul and 2022 Recife floods hours ahead of the heaviest rain
Author: Gaël Kermarrec1
MS No.: https://doi.org/10.5194/egusphere-2026-5548
General Comment.
This manuscript describes the evolution of integrated water vapor (IWV) retrieved from Global Navigation Satellite System (GNSS) sensors during two major flood events in Brazil: one in Rio Grande do Sul in 2024 and another in Recife in 2022. In both cases, the IWV-GNSS measurements showed promising potential for providing warnings up to five hours in advance. The use of GNSS-derived IWV for nowcasting is well established and has proven effective. Its main advantages include the relatively low cost of the instruments, ease of operation, and the availability of measurements from existing geodetic networks, making this approach particularly valuable for operational weather monitoring and early warning systems. The use of this approach is not new, as several studies have investigated the potential of GNSS-derived integrated water vapor (IWV) for nowcasting intense rainfall events. For instance, a 2018 paper introduced the concept of an “IWV-GNSS jump”: Global Positioning System precipitable water vapour (GPS-PWV) jumps before intense rain events: A potential application to nowcasting”. This relevant study is not cited in the manuscript (https://doi.org/10.5194/egusphere-2026-5548). In addition, the 2024 extreme rainfall event in Rio Grande do Sul has already been investigated using GNSS measurements in the study “Analysis of atmospheric variations using GNSS signal as atmospheric sensor (PWV-GNSS) in the extreme rainfall events in Rio Grande do Sul (Brazil) in 2024”, which is also not cited. These studies should be discussed to better establish the novelty and contribution of the present manuscript.
I have several concerns regarding the methodology. The zenith wet delay (ZWD) is obtained by subtracting the zenith hydrostatic delay (ZHD), estimated using surface pressure and a representative tropospheric temperature model, from the total zenith tropospheric delay. However, the manuscript does not explain how the ZWD is converted into IWV. The text then states that “σ and λ are estimated in sliding windows” without defining these parameters or explaining how they are calculated and used in the proposed technique. This lack of methodological detail makes it difficult to assess the robustness and reproducibility of the results.
I am also concerned about the use of a 30-second temporal resolution. Although GNSS data can be processed at such a high temporal resolution, the authors should clarify how they distinguish meaningful atmospheric signals from noise and measurement variability. More importantly, since the main result is the potential for a five-hour lead time, it is unclear why such high temporal resolution is essential to the proposed approach. The authors should better justify this choice and demonstrate its added value for nowcasting.
The main novelty of the proposed technique appears to be the use of the seasonal median derived from nearby GNSS sensors as a reference value, together with the calculation of percentiles to quantify deviations from this baseline. However, the examples presented show warnings throughout the entire period, without distinguishing between periods with no rainfall, regular rainfall, and intense rainfall events. Although the jump in IWV is clearly visible in the figures (blue line), particularly during the most intense events, the proposed warning approach does not appear to differentiate adequately between these different conditions. The authors should also consider that IWV may decrease after rainfall reach it maximum intensity, as water vapor condenses into liquid water and is subsequently removed from the atmosphere through precipitation.
The manuscript would benefit from a clearer explanation of the underlying physical processes, supported by more detailed interpretations of the figures and their relationship to the observed rainfall events. In particular, the authors should clarify why deviations from the seasonal median provide a more effective warning indicator than the IWV jump itself, which captures the temporal evolution of individual events more directly.