Prediction of Alpine Foehn from time series of GNSS troposphere products using machine learning
- 1Institute of Geodesy and Photogrammetry, ETH Zürich, 8093 Zurich, Switzerland
- 2Swiss Federal Office of Topography (swisstopo), 3084 Wabern bei Bern, Switzerland
- 1Institute of Geodesy and Photogrammetry, ETH Zürich, 8093 Zurich, Switzerland
- 2Swiss Federal Office of Topography (swisstopo), 3084 Wabern bei Bern, Switzerland
Abstract. Remote sensing of water vapor using the Global Navigation Satellite System (GNSS) is a well-established technique and reliable data source for Numerical Weather Prediction (NWP). One of the phenomena rarely studied using GNSS are foehn winds. Since foehn winds are associated with significant humidity gradients between lee/luv side of a mountain range, tropospheric estimates from GNSS are also affected by their occurrence. Time series reveal characteristic features like distinctive minima/maxima and significant decrease in correlation between the stations. However, detecting such signals becomes increasingly difficult for large data sets. Therefore, we suggest the application of machine learning algorithms for detection and prediction of foehn events from GNSS troposphere products. The present study uses long-term time series of high-quality GNSS troposphere products from the Automated GNSS Network Switzerland (AGNES) as well as records of operational foehn index to investigate the performance of several different classification algorithms based on appropriate statistical metrics. The two best-performing algorithms are fine-tuned and employed on two years of test data. The results show very promising results, especially when reprocessed GNSS products are utilized. Detection- and alarm-based measures reach levels of 70–85 % for both tested algorithms and thus are comparable to those from studies using data from meteorological stations and NWP. For operational prediction, some limitations due to the availability and quality of GNSS products in near-real time (NRT) exist. However, they might be mitigated to a significant extend by provision of additional NRT products and improved data processing in the future.
Matthias Aichinger-Rosenberger et al.
Status: final response (author comments only)
-
RC1: 'Comment on amt-2022-33', Anonymous Referee #1, 17 Feb 2022
-
AC1: 'Reply on RC1', Matthias Aichinger-Rosenberger, 03 Mar 2022
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2022-33/amt-2022-33-AC1-supplement.pdf
-
AC1: 'Reply on RC1', Matthias Aichinger-Rosenberger, 03 Mar 2022
-
RC2: 'Comment on amt-2022-33', Anonymous Referee #2, 27 Apr 2022
The paper „Prediction of Alpine Foehn from time series of GNSS troposphere products using machine learning” shows first the selection of the ML methods and then usage of two of them on a GNSS tropospheric data set (tropospheric delays and gradients) to detect the foehn occurrences. It is a very new field of study as most of the GNSS meteorology research focuses on the precipitation/humidity parameters rather as foehn. Also the usage of the machine learning algorithms is interesting. I found the paper very well written. The only drawbacks of the paper are: 1. Sometimes a more extended discussion on the results is lacking, 2. The figures (especially Fig.5-10) could be made more interesting.
Specific comments
Title: since you always work on the past data (even with the NRT approach), it is rather a ‘detection’ than a ‘prediction’, so maybe the title could be changed accordingly.
Line 3: ‘lee/luv’ – a specific terminology, maybe worth explaining (at least in the Introduction, however ‘luv’ doesn’t appear anywhere else than the abstract
Line 68 ‘COSMO (Consortium for Small-scale Modeling).’ -> ‘Consortium for Small-scale Modeling COSMO)’; the full name should go before abbreviation
Line 90: This is not the exact formula from Rueger and I think there is a mistake there:
In Rueger there is a following formula:
N = 77.6890 Pd/T+ 71.2952 e/T+3.75463×10^5 · e/T^2
So if we substitute Pd=P-e then we get
N = 77.6890 P/T-6.3938 e/T+3.75463×10^5 · e/T^2
(should be a minus before the second term). However, I would recommend sticking to the original formulation as then you have a clear distinction between the dry and water vapor parts.
Figure 1: Would be nice to see the topography in this Figure to better visualize foehn
Section 4.1: I would recommend giving here at least very brief overview of the selected methods
Line 163: ‘(negative) maximum’ - > why not use ‘minimum’ here?
Fig.3 and Table 2 show exactly the same information, so I would recommend removing one of them, especially that Fig. 3 is not even addressed in the main text.
Figure 4: Make the foehn line more pronounced
Line 259: Would be good to comment here what the chosen parameters mean
Figure 5: Maybe you could add vertical lines so the reader can more easily compare the data for particular dates; also you do not comment this plot in the text
Figure 6 and 7: Maybe there is a way to plot them together for better comparisons of the two methods?
Line 284: A more detailed discussion about the features would be advantageous
Figure 9: Why not add here a line also of the match with GB (not only with the adjusted one); also it seems like the event of Oct 2020 was caught by the algorithm but in a different epoch – maybe it is something to look into
Line 312: Would be nice to see here more discussion on why you change the threshold and how it is done
-
AC2: 'Reply on RC2', Matthias Aichinger-Rosenberger, 03 May 2022
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2022-33/amt-2022-33-AC2-supplement.pdf
-
AC2: 'Reply on RC2', Matthias Aichinger-Rosenberger, 03 May 2022
Matthias Aichinger-Rosenberger et al.
Matthias Aichinger-Rosenberger et al.
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
202 | 89 | 13 | 304 | 4 | 4 |
- HTML: 202
- PDF: 89
- XML: 13
- Total: 304
- BibTeX: 4
- EndNote: 4
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1