the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
GNSS Radio Occultation Climatologies mapped by Machine Learning and Bayesian Interpolation
Abstract. Global Navigation Satellite Systems (GNSS) radio occultation (RO) is a space-based remote sensing technique that measures the bending angle of GNSS signals as they traverse the Earth's atmosphere. Profiles of the microwave index of refraction can be calculated from the bending angles. High accuracy, long-term stability, and all-weather capability make this technique attractive to meteorologists and climatologists. Meteorologists routinely assimilate RO observations into numerical weather models. RO-based climatologies, however, are complicated to construct as their sampling density is highly non-uniform and too sparse to resolve synoptic variability in the atmosphere.
In this work, we investigate the potential of machine learning (ML) to construct RO climatologies and compare the results of a ML construction with Bayesian interpolation (BI), a state-of-the-art method to generate maps of RO products. We develop a feed-forward neural network applied to COSMIC-2 RO observations and simulate data taken from the atmospheric analyses produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). Atmospheric temperature, pressure and water vapor are used to calculate microwave refractivity at 2, 3, 5, 8, 15, and 20 km geopotential height, with each level representing a different dynamical regime of the atmosphere. The simulated data are the values of microwave refractivity produced by ECMWF at the geolocations of the COSMIC-2 RO constellation, which fall equatorward of 46 latitude. The maps of refractivity produced using the neural networks better match the true maps produced by ECMWF than maps using BI. The best results are obtained when fusing BI and ML, specifically when applying ML to the post-fit residuals of BI. At the six iso-heights, we obtain post-fit residuals of 10.9, 9.1, 5.3, 1.6, 0.6 and 0.3 N-units for BI and 8.7, 6.6, 3.6, 1.1, 0.3 and 0.2 N-units for the fused BI&ML, respectively. These results are independent of season.
The BI&ML method improves the effective horizontal resolution of the posterior longitude-latitude refractivity maps. By projecting the original and the inferred maps at 2 km iso-height onto spherical harmonics, we find that the BI-only technique can resolve refractivity in the horizontal up to spherical harmonic degree 8 while BI&ML can resolve maps of refractivity using the same input data up to spherical harmonic degree 14.
- Preprint
(1919 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
EC1: 'Comment on amt-2023-205', Peter Alexander, 23 Oct 2023
If the authors submit a final version of the manuscript, then it will be necessary to include some evidence (a figure or a reference) showing that the current spatial and temporal density of RO leads to a need for better coverage when approaching specific goals.
Citation: https://doi.org/10.5194/amt-2023-205-EC1 -
RC1: 'Reply on EC1', Anonymous Referee #1, 02 Dec 2023
Comments on “GNSS Radio Occultation Climatologies mapped by Machine Learning and Bayesian Interpolation”
General comment:
This paper is concerned with constructing GNSS RO-based climatologies by machine learning (ML) method, and proposes three kinds of approaches: Bayesian Interpolation (BI), a feed-forward neural network (Multilayer Perceptrons, MLPs), and the combination of BI and ML (BI &ML) where the ML is applied to BI residuals. Applications of these methods to real and simulated COSMIC-2 RO data indicate that, the maps of refractivity produced by the MLPs better match the true maps than those by BI, and BI & ML yields the best GNSS RO refractivity maps. The methods are novel and the results exhibit the potential for producing GNSS RO climatologies.
specific comments:
“BI & ML” is about learning on residuals, which is a key strategy suggested by the authors. I think it is better to provide some reviews for learning on residuals in the introduction.
technical corrections:
For the convenience of readers, Figure 3, Figure 4, Table 1 and Table 2 had better be indicated for COSMIC -2 RO data.
Citation: https://doi.org/10.5194/amt-2023-205-RC1 -
AC2: 'Reply on RC1', Endrit Shehaj, 24 Apr 2024
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-205/amt-2023-205-AC2-supplement.pdf
-
AC2: 'Reply on RC1', Endrit Shehaj, 24 Apr 2024
- AC1: 'Reply on EC1', Endrit Shehaj, 24 Apr 2024
-
RC1: 'Reply on EC1', Anonymous Referee #1, 02 Dec 2023
-
RC2: 'Comment on amt-2023-205', Anonymous Referee #2, 25 Jan 2024
- AC3: 'Reply on RC2', Endrit Shehaj, 26 Apr 2024
-
RC3: 'Comment on amt-2023-205', Abhineet Shyam, 30 Jul 2024
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-205/amt-2023-205-RC3-supplement.pdf
- AC4: 'Reply on RC3', Endrit Shehaj, 31 Jul 2024
Status: closed
-
EC1: 'Comment on amt-2023-205', Peter Alexander, 23 Oct 2023
If the authors submit a final version of the manuscript, then it will be necessary to include some evidence (a figure or a reference) showing that the current spatial and temporal density of RO leads to a need for better coverage when approaching specific goals.
Citation: https://doi.org/10.5194/amt-2023-205-EC1 -
RC1: 'Reply on EC1', Anonymous Referee #1, 02 Dec 2023
Comments on “GNSS Radio Occultation Climatologies mapped by Machine Learning and Bayesian Interpolation”
General comment:
This paper is concerned with constructing GNSS RO-based climatologies by machine learning (ML) method, and proposes three kinds of approaches: Bayesian Interpolation (BI), a feed-forward neural network (Multilayer Perceptrons, MLPs), and the combination of BI and ML (BI &ML) where the ML is applied to BI residuals. Applications of these methods to real and simulated COSMIC-2 RO data indicate that, the maps of refractivity produced by the MLPs better match the true maps than those by BI, and BI & ML yields the best GNSS RO refractivity maps. The methods are novel and the results exhibit the potential for producing GNSS RO climatologies.
specific comments:
“BI & ML” is about learning on residuals, which is a key strategy suggested by the authors. I think it is better to provide some reviews for learning on residuals in the introduction.
technical corrections:
For the convenience of readers, Figure 3, Figure 4, Table 1 and Table 2 had better be indicated for COSMIC -2 RO data.
Citation: https://doi.org/10.5194/amt-2023-205-RC1 -
AC2: 'Reply on RC1', Endrit Shehaj, 24 Apr 2024
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-205/amt-2023-205-AC2-supplement.pdf
-
AC2: 'Reply on RC1', Endrit Shehaj, 24 Apr 2024
- AC1: 'Reply on EC1', Endrit Shehaj, 24 Apr 2024
-
RC1: 'Reply on EC1', Anonymous Referee #1, 02 Dec 2023
-
RC2: 'Comment on amt-2023-205', Anonymous Referee #2, 25 Jan 2024
- AC3: 'Reply on RC2', Endrit Shehaj, 26 Apr 2024
-
RC3: 'Comment on amt-2023-205', Abhineet Shyam, 30 Jul 2024
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-205/amt-2023-205-RC3-supplement.pdf
- AC4: 'Reply on RC3', Endrit Shehaj, 31 Jul 2024
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
657 | 191 | 49 | 897 | 39 | 35 |
- HTML: 657
- PDF: 191
- XML: 49
- Total: 897
- BibTeX: 39
- EndNote: 35
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1