| 25 Sep 2026
Using 1 second LEO clock corrections for the Fengyun-3E radio occultation retrievals
Abstract. Clock offset constitutes a major error source for Global Navigation Satellite System (GNSS) radio occultation (RO) products, particularly for stratospheric observations. The adoption of high-rate clock correction strategy can effectively mitigate clock offset induced errors. In this study, high-rate (1-s) undifferenced (UD) Low Earth Orbit (LEO) clock corrections are used to retrieve RO observations from the Fengyun-3E (FY-3E) satellite. The retrieval results are compared with those derived from the conventional 30-s clock correction scheme and the single-difference (SD) scheme currently adopted in the official data processing. The results demonstrate that the 1-s LEO clock correction scheme improves the overall quality of FY-3E RO retrievals. For bending angle, the most prominent improvements are found in the altitude range of 40-60 km. Globally, its systematic difference is roughly 3% smaller than that of the 30-s scheme and 1% smaller than the SD scheme. Such superiority further propagates downward to refractivity and dry temperature. For refractivity, the largest improvements are also found in the altitude range of 40-60 km with systematic differences for the 1-s UD scheme reduced by approximately 0.5% relative to the other two schemes. Standard deviations of the 1-s scheme remain consistently smaller across the complete height range. Standard deviations of dry temperature are also improved with standard deviations about 0.5% smaller than SD and 1% to 2% smaller than the 30-s scheme in the altitude range from 0-40 km. The results in this study demonstrate the capability of using 1-s clock correction scheme for further promoting Fengyun GNSS Occultation Sounder-II (GNOS-II) RO observations. The results here can provide support for future operational processing and archive reprocessing of all Fengyun GNSS occultation datasets.
For the FY-3E radio occultation products, this paper investigates the improvements in the statics of BAs, refractivity and temperature resulting from the use of a 1-s LEO clock correction over the use of a 30-s clock correction or a single-difference approach. The paper lacks a discussion of the methods used to obtain the 1-s LEO clock solutions, and the conclusions are based on the analysis of a very limited number of occultations.
Major comments:
In a UD approach, a high-rate GNSS clock product is needed in order to compute the LEO clock solution. The authors mention using the CODE GNSS precise orbit and clock products for the GPS constellation, without specifying which product type. Independent of the actual product stream used, there is still an important point to clarify: one can always obtain a 1s LEO clock solution, but one would expect this solution to introduce noise, if the GNSS clock product has a lower rate (see, e.g., the discussion of V_SAMPL in the LEO PCF example of the Bernese manual). To obtain a meaningful 1s LEO clock solution, the authors should use 1s GNSS clock products. (this kind of products are provided, e.g., by JPL: https://sideshow.jpl.nasa.gov/pub/JPL_GNSS_Products/RealTime/). Starting from a 1s GNSS clock product the authors could make the case that the 1s LEO clock solution is indeed superior to the 30s clock solution even when using the 5s GNSS clocks (e.g., by comparing the LEO solutions obtained with the 1s GNSS and the 5s down-sampled GNSS clock product). Please note, however, that any such conclusion would not apply to all constellations, since the short-time behavior of the GNSS clock hardware depends both on the constellation and on the H/W version (see, e.g., the two papers by Yao et al. 2023, and Padovan et al. 2025 quoted in the manuscript).
The number of profiles entering the analysis is just over 1k. The results are presented both globally, and in 5 latitudinal bands, where each band’s analysis would be based on roughly 200 profiles. These numbers are too low to make any conclusive statement on the superiority of one method over another. Provided a good case can be made that the 1s LEO clock solution can be reliably obtained (see previous comment), then the analysis should be based on at least some few thousands of profiles.
Additional comments:
Ln 52: “UD has emerged… in recent years.” UD has been established for quite some time. It has been used for the Metop satellites since 2006, and it is the baseline approach for most RO missions, including commercial ones.
Ln 90-94: You should provide some info on the CODE product used, and add its reference (in addition to the reference in the data availability).
Ln 97-101: Please specify whether in the comparison the background is interpolated at the BA location or the closest grid point is taken.
Ln 120: Please specify which relativistic correction you are including.
Ln 136: Provide at least a reference for the statistical optimization (see also comment to Fig. 8).
Ln 179: What does “the aft receiver clock products” mean? There is only one clock solution for the receiver, with its values corrected for the h/w delay.
Fig. 2: Given the horizontal scale, the impression is that the 1s clock is much noisier. Having in addition (or as a replacement) similar plots with, say, a 20 min horizontal timescale, would be informative to assess the clock behavior at RO-relevant timescales.
Ln 182/183: Do not up-sample the clock solution: You are adding no information, since the resampled series is completely determined by the input one. The only comparison you can make for the red curve is at averaging times >= 30s, where both clock series have data (the green curve can have data at averaging times >=1s).
Ln 225: The characterization of 3% and 1% improvement only applies to a subset of the full 40-60km range. If a number must be quoted, then use an average over the height range (possibly weighted). Also, one cannot assess the standard deviation above 50km, since it’s off scale.
Fig. 7: In the text you claim that “statistical errors are too noisy for characterization”, but in the figure you present the mean of the bias and noise. They may well be too noisy because you don’t have enough profiles or because the signal up there is too small for the profiles to be meaningful.
Ln 260ff and Fig. 8: The figure is the statistically optimized version of Fig. 7, and unless a description is provided on the way the optimization is performed, it is difficult to assess the results of Fig. 8. (See also comment for Ln 136.)
Ln 270: as already pointed out before, the numbers quoted (2%) only refers to a limited height range, not the entire 40-60km interval.