the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Assimilation of Spaceborne Radar Observations in the Goddard Earth Observing System (GEOS) Model
Abstract. Spaceborne radar observations provide valuable information on the vertical structure of precipitation and hydrometeors, making them an important resource for numerical weather prediction and model evaluation. The Joint Effort for Data assimilation Integration (JEDI) is the unified and configurable data assimilation system developed and stewarded by the Joint Center for Satellite Data Assimilation (JCSDA) on behalf of its U.S. federal partners being adopted in NASA’s Goddard Earth Observing System (GEOS) suite of products. This study describes a new spaceborne radar data assimilation capability developed for use in JEDI, including both the radar forward operator and the associated data assimilation components. This new capability is then demonstrated using observations from the Global Precipitation Measurement (GPM) Dual-frequency Precipitation Radar (DPR). The radar forward operator simulates reflectivities that agree well with observations, with mean differences typically below 2 dB during the August 2023–January 2024 test period. Consequently, a large number of DPR observations can be successfully assimilated into GEOS. Case studies of tropical cyclones show that DPR assimilation can substantially improve the analyzed distribution of precipitation and hydrometeors and, in some cases, improve atmospheric state variables such as wind, pressure, temperature, and water vapor. These results demonstrate the potential of spaceborne radar observations to improve the representation of precipitating systems in GEOS and, ultimately, the quality of future reanalyses and weather forecasts.
- Preprint
(52707 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-4086', Anonymous Referee #1, 24 Oct 2026
-
RC2: 'Comment on egusphere-2026-4086', Anonymous Referee #2, 28 Oct 2026
Review of “Assimilation of Spaceborne Radar Observations in the Goddard Earth Observing System (GEOS) Model” by Moradi et al.
28 October 2026
Recommendation: Major revision
Synopsis:
This manuscript describes the implementation of a spaceborne radar reflectivity assimilation capability within the Joint Effort for Data assimilation Integration (JEDI) framework, built around the Community Radiative Transfer Model (CRTM) active-sensor module and its tangent-linear and adjoint counterparts, and demonstrates it in the NASA Goddard Earth Observing System (GEOS) 3D-FGAT system using Global Precipitation Measurement (GPM) Dual-frequency Precipitation Radar (DPR) Ku-band (Channel 1) and Ka-band (Channel 2) observations. The manuscript outlines the spatial averaging of DPR observations onto 25-km bins, a reflectivity-dependent observation error model, a variational bias correction (VarBC) scheme using height and scan-angle predictors, and quality control (QC) criteria. The capability is then evaluated using observation-space departure diagnostics and two tropical cyclone case studies (Typhoon Khanun and Cyclone Midhili) compared against ERA5 and IMERG, followed by a single forecast example for Typhoon Khanun.
The underlying technical development is an important and timely contribution: a community-supported, multi-frequency radar forward operator with consistent tangent-linear and adjoint code inside a unified data assimilation framework is something many groups will benefit from, and I am generally supportive of the eventual publication of this work. However, in its current form, the manuscript reads more like a high-level system description with an attached demonstration than a rigorous data assimilation study. The methodological sections devote too much space to generic JEDI and GEOS infrastructure while leaving underexplained the crucial mechanics that determine how radar observations actually update the model state (specifically the control variables, the treatment of diagnostic hydrometeors, and the cycling configuration). More critically, I found several internal inconsistencies in the observation error and bias sections, the quality control excludes the very regimes where radar reflectivity carries vital information for initiating or suppressing precipitation, and the impact evaluation relies on two qualitative single-time analysis snapshots and a single 5.5-day forecast without any aggregate verification statistics over the multi-month cycling experiment. Therefore, my recommendation is major revisions. My detailed comments follow.
Major issues:
(1) The historical context provided in the introduction for radar data assimilation (L39–54) is incomplete and in places inaccurate. Direct assimilation of radar reflectivity and radial velocity in NWP models goes back considerably further than the “last decade” indicated on L39. Ensemble-based assimilation of ground-based radar observations was demonstrated in the early 2000s (e.g., Snyder and Zhang, 2003; Dowell et al., 2004; Tong and Xue, 2005; Aksoy et al., 2009), and radar observations have been assimilated operationally at multiple NWP centers for well over a decade, as documented in the Gustafsson et al. (2018) review that the authors themselves cite. Consequently, the statement on L46–48 that radar assimilation “has been limited worldwide” and largely confined to “research and experimental settings” contradicts the authors’ own reference. Similarly, L53–54 states that recent CRTM developments “enable the assimilation of radar observations,” even though radar reflectivity assimilation was clearly possible previously using other forward operators—including, for spaceborne radars, the operator of Fielding and Janisková (2020) that is also cited in the manuscript. What CRTM specifically contributes is a community-maintained, multi-frequency forward, tangent-linear, and adjoint operator integrated within the JCSDA/JEDI ecosystem. The introduction should be revised to state this contribution accurately while properly crediting the earlier body of work on both ground-based and spaceborne radar data assimilation.
(2) In Sections 3.2–3.4 and Section 4, the manuscript includes too much unnecessary technical detail on generic JEDI and GEOS workflow components—including IODA, UFO, SWELL (which, by the authors’ own admission on L186–188, plays a minimal role here), and the standard GEOS Incremental Analysis Update (IAU) procedure—while omitting or burying critical aspects of the radar data assimilation methodology. Sections 3.2–3.4 should be combined into a single, much shorter section that focuses strictly on the software and scientific modifications required to support radar reflectivity assimilation, and Section 4 (L190–232) should be substantially condensed by citing existing GEOS and JEDI publications. The space saved should be used to clearly explain the actual assimilation methodology and experimental setup:
(2a) What are the control variables of the 3D-FGAT variational analysis, and are hydrometeors included among them? It is not until the beginning of Section 8 (L388–393) that the reader is informed that cloud and precipitation fields in GEOS are diagnostic rather than prognostic variables. This is a fundamental property of the system that must be introduced and explained in Section 4. If hydrometeors are not control variables, then radar reflectivity innovations can only influence the analysis through the sensitivity of the diagnosed hydrometeors to the prognostic control variables. Conversely, if hydrometeors are control variables, then the static GSI background error covariance (GSIBEC) must contain hydrometeor error variances as well as cross-covariances between hydrometeors and the prognostic variables (temperature, water vapor, wind, and surface pressure) to produce the multivariate adjustments shown in Figures 9–10 (such as the several-hPa reduction in minimum sea level pressure). To my knowledge, standard static GSI background error covariances do not include such hydrometeor cross-covariances—which are normally supplied by the ensemble component in a hybrid EnVar framework, as the authors’ own remark on L221–225 implies. Please explain explicitly how radar reflectivity innovations update both the hydrometeor fields and the prognostic variables, ideally supported by showing the analysis increments in the control variables for one of the case studies.
(2b) How are the background hydrometeor water content profiles (L203) generated and interfaced with JEDI? Are they diagnosed by the GEOS moist physics outside JEDI and passed in as background fields, or diagnosed within UFO/GeoVaLs? How are the GEOS diagnostic species mapped onto the five hydrometeor species required by CRTM (snow, cloud ice, graupel, rain, and cloud liquid water; L158–159)—in particular, does GEOS diagnose graupel, and if not, how is that handled? Furthermore, if the variational analysis produces increments to hydrometeor fields, what happens to those increments when the diagnostic hydrometeors are recomputed from the prognostic state during the subsequent 6-hour IAU integration?
(2c) The cycling and solver configuration needs to be stated precisely. The text refers to a 6-hour cycle on L195, an “hourly” 3D-FGAT procedure on L220, and “3D-Var” on L329–330. Please use consistent terminology, specify the assimilation window, FGAT sub-window binning, and number of outer and inner loops, and confirm that the CTRL and DPR experiments are identical in every respect except for the assimilation of DPR reflectivity. In addition, on L194, please clarify the distinction between “model initial conditions (so-called restarts)” and “background (first-guess) fields” at the start of each cycle for readers unfamiliar with the GEOS IAU cycling structure.
(2d) The stated experiment period is inconsistent across the manuscript. Both the abstract and L234–235 state that the cycling experiments were carried out continuously from August 2023 through January 2024, whereas Figures 3–5 cover only the one-month period from 15 August to 15 September 2023, Figures 6–7 do not state a time period in their captions, and Typhoon Khanun is evaluated on 2 August 2023 (prior to 15 August). Please clarify the exact period run, the period used for each diagnostic, and the spin-up duration allowed—particularly for spinning up the VarBC coefficients of a newly introduced instrument.
(2e) Figure 7 indicates 223,120 observations over a one-month period for Channel 1, corresponding to roughly 1,860 observations per 6-hour cycle compared with the millions of conventional and radiance observations assimilated per cycle in GEOS. Please report the mean, minimum, and maximum number of assimilated DPR observations per cycle for each channel and compare them with the routine observation counts in the control experiment so the reader can properly contextualize the impact results in Sections 7–8.
(3) I have several major concerns regarding the observation processing, representativeness error, and observation error specification in Section:
(3a) The horizontal spatial averaging of DPR observations over 25 × 25 km regions (L102–103, L259–262) is superobbing, and I recommend using that standard terminology consistently and describing the procedure completely: Are the 25 × 25 km bins defined on a fixed global grid or along the DPR swath? Is the averaging—both horizontal and the vertical averaging onto model layers in UFO (L155–157)—performed in logarithmic units (dBZ) or in linear reflectivity factor (Z, mm⁶ m⁻³), which yield different means and error characteristics? Also, why is horizontal averaging performed outside JEDI prior to ingestion while vertical averaging is performed inside UFO? Most importantly, in Figure 5b, the observation counts per 25-km bin range only between 1 and 6. It is unclear from the text and caption whether Figure 5b displays the number of native 5-km footprints averaged into each 25-km superob or the total number of assimilated superobs per grid cell accumulated over the one-month period (15 August to 15 September 2023). In either case, this raises a critical issue: if the sample size per bin is only 1–6 points, those sample sizes are far too small for the within-bin standard deviation (σ_bin) to be statistically meaningful, and a minimum sample count must be enforced; if instead Figure 5b shows the accumulated superob count over an entire month, then most grid cells are sampled only once or twice in 30 days. Please resolve this ambiguity and report the distribution of native footprints per superob.
(3b) While DPR observations are averaged both horizontally (from 5 km to 25 km) and vertically (from 125-m range gates onto model layers), the representativeness error proxy σ_bin in Equations 4–5 accounts only for horizontal within-bin variability. Vertical reflectivity gradients within a model layer—especially across the melting layer—can be substantial; please explain why vertical subgrid variability is neglected. In addition, please state explicitly whether the observation error covariance matrix R is assumed diagonal, and discuss the implications of neglecting vertical error correlations between adjacent model layers in the same profile as well as inter-channel error correlations between collocated Ku- and Ka-band observations.
(3c) The coefficients in Equation 5 (σ_0 = 1 dB, α = 0.25, β = 0.02, and γ = 0.1) are introduced without justification. Please explain how these constants were determined (whether from prior studies, theoretical arguments, empirical fits, or sensitivity tests) and discuss the sensitivity of σ_obs and the resulting analyses to these parameter choices. Of all the terms in Equation 5, which contributes the strongest influence on the final σ_obs? Is it the reflectivity (y) term? Rather than the one-dimensional marginal histograms currently shown in Figure 3, it would be much more informative to provide a two-dimensional histogram of σ_obs against observed reflectivity (y) to show this functional dependence directly. Also, because the innovation term γ|y − H(x_b)| makes R depend on the background state, please clarify whether σ_obs is held fixed across inner-loop iterations during the variational minimization (otherwise the cost function is no longer quadratic).
(3d) More seriously, a back-of-the-envelope calculation shows that the observation error distributions plotted in Figure 3 cannot be reconciled with Equation 5 under the quality control bounds given in Equation 7. Specifically, substituting the maximum possible values indicated in Equation 7 into Equation 5’s error model (σ_obs = 1 + 0.25σ_bin + 0.02y + 0.1|y − H(x_b)|) yields upper-bound contributions of 0.25 × 6 = 1.5 dB from σ_bin, 0.02 × 70 = 1.4 dB from y, and 0.1 × 6 = 0.6 dB from the departure term. Summing these with σ_0 = 1 dB gives an absolute maximum possible value of σ_obs = 4.5 dB (and typical values of only 2–3 dB for y ≈ 25 dBZ, σ_bin ≈ 2 dB, and |y − H(x_b)| ≈ 2 dB). Yet L269–272 and Figure 3 report that σ_obs ranges from 4 to 11 dB with a peak near 6–7 dB for Channel 1, and from 4 to 9 dB peaking near 5–6 dB for Channel 2. Either the coefficients actually implemented in the code differ from Equation 5, or Figure 3 plots a different quantity. This discrepancy must be resolved.
(3e) Related to L273–275, the authors claim that “relatively conservative error values were adopted to limit observation impact” and note that these errors can be further tuned. However, if the errors in Figure 3 (6–7 dB) are the ones actually used, they exceed the standard deviation of the observation-minus-background departures reported in Figure 7 (4.8 dB for Channel 1 and 3.9 dB for Channel 2). Because the innovation variance in observation space is the sum of the observation and background error variances (σ_(O-B)² = σ_obs² + σ_b²), an observation error standard deviation larger than the total innovation standard deviation is statistically inconsistent with the innovations and implies that the observations receive very little weight. At the same time, without presenting sensitivity experiments or quantitative calibration diagnostics (such as innovation consistency ratios or Desroziers diagnostics), it is not possible to establish that the error specification is “conservative” in the first place or to gauge what further tuning would have contributed.
(4) The organization, physical interpretation, and diagnostic presentation of quality control and bias correction in Sections 5 and 6 (Equations 6–7 and Figures 4, 6, and 7) require substantial revision:
(4a) On L289–297, the quality control rejection criteria (Equation 7) are presented at the very end of Section 5, where it is noted that observations are rejected “before observation biases were calculated and applied.” Because quality control is independent of and performed prior to bias correction, it would be much clearer for the reader if the QC criteria were introduced in their own subsection before the discussion of bias correction. I then recommend starting a new section at L278 dedicated specifically to observation bias and bias correction, preceded by a clear explanation of the JEDI VarBC implementation (including how coefficients are initialized and spun up for DPR, and whether any anchor or constraint is applied).
(4b) The physical origin of the biases in Figure 4 and their relationship to radar attenuation need a much deeper explanation (L87–103, L130–143, and L280–288). First, Section 2.1 (L87–103) should state the wavelengths corresponding to Ku-band and Ka-band, and explain how those wavelengths affect attenuation. While Section 3.1 mentions that CRTM accounts for attenuation by rain, cloud water, and gases, the manuscript never explains the magnitude of attenuation expected at Ku- versus Ka-band wavelengths, nor whether attenuation by frozen/melting hydrometeors and multiple scattering (which can be significant at Ka-band in deep convection) are accounted for. Second, a critical piece of information is missing from Section 2.1: Which GPM DPR Level-2 reflectivity product is assimilated—the measured (attenuated) reflectivity or the attenuation-corrected reflectivity? Because CRTM simulates attenuated reflectivity (Z_a in Equation 1), assimilating the attenuation-corrected DPR product would double-count attenuation. Third, in Figure 4 and Figure 6, Channel 2 exhibits a bias of −1 to −3 dB and negative departures concentrated in the lowest 2 km, which looks very much like residual two-way attenuation at Ka-band. Please discuss on L280–281 whether the diagnosed bias is the result of attenuation, why Channel 1 (0 to 2 dB) and Channel 2 (−3 to −1 dB) have biases of opposite sign, and whether applying an unconstrained height-dependent VarBC to a well-calibrated active radar risks aliasing systematic GEOS model hydrometeor biases (such as underrepresented frozen hydrometeors and excessive cloud liquid water) into the observation bias correction.
(4c) In Section 6 (Figures 6 and 7), please clarify whether the simulated reflectivity H(x_b) is computed from the prior model background fields, and whether the statistics shown in Figures 6 and 7 are calculated before or after QC and before or after VarBC bias correction (noting that Figure 7 contains departures well beyond the 6-dB QC threshold of Equation 7, and curiously lists the exact same observation count, 223,120, for both channels despite their different sensitivity profiles). Please also check the sign convention across figures: Figure 4 shows a Channel 2 bias of −3 to −1 dB and Figure 6 shows Channel 2 mean differences mostly between −2 and 0 dB, yet Figure 7 reports μ_d = +1.7 dB for Channel 2 with the high-density core lying above the 1:1 line (indicating that Figure 7 plots simulated-minus-observed rather than observed-minus-simulated as stated in the caption and text). For Figure 6, please also report the number of observations in each FOV–elevation bin so the reader can assess how representative the averages and standard deviations are, particularly above 8–10 km. Finally, in Figure 7, why does the simulated Channel 2 reflectivity deviate from the observations much more strongly than Channel 1, appearing to saturate near 30–35 dBZ at higher observed reflectivities? Is this behavior caused by deficiencies in the model background hydrometeor profiles or by limitations in the CRTM forward operator and attenuation calculation at Ka-band?
(4d) The last two QC criteria in Equation 7 reject any observation where either y ∉ [15, 70] dBZ or H(x_b) ∉ [15, 70] dBZ. Imposing the 15-dBZ lower bound on both the observation and the background means that DPR observations are only assimilated where both the radar and the model background simultaneously contain precipitation. Consequently, the assimilation system cannot initiate precipitation where the model background is erroneously dry, nor can it suppress spurious precipitation where the model precipitates but DPR observes clear air or non-precipitating cloud—which are arguably two of the most impactful roles of radar reflectivity assimilation (see, e.g., Aksoy et al., 2009 regarding the assimilation of “no-precipitation” reflectivity observations). This major methodological limitation and its consequences for the case studies in Sections 7–8 must be explicitly discussed, and the claim in the abstract and conclusions that observed and simulated reflectivities agree within 2 dB should be qualified as applying only to this doubly conditional, post-QC sample. In addition, applying the gross check |y − H(x_b)| > 6 dB before bias correction (L289–290) when Channel 2 has a bias of up to −3 dB will reject observations asymmetrically, and a 6-dB threshold represents less than 1σ_obs if σ_obs is 6–7 dB as in Figure 3.
(5) In Section 7 (Figures 8–12), the evaluation of the impact of DPR assimilation on the analysis relies entirely on qualitative visual comparisons of two tropical cyclones at single analysis times, and several key features are either unexplained or inadequately benchmarked:
(5a) Because the authors conducted a continuous cycling experiment over multiple months, a system-level quantitative evaluation of the analysis is essential. In addition to DPR observation-minus-background (O − B) and observation-minus-analysis (O − A) departure statistics, the authors should present standard observation-space verification showing the background and analysis fits to all other routinely assimilated observation types in CTRL versus DPR (especially all-sky and clear-sky microwave radiances such as GMI, AMSR2, ATMS, and MHS, which are sensitive to the same hydrometeor and moisture fields).
(5b) In Figures 9–10 for Typhoon Khanun, by far the largest discrepancy between the GEOS analyses (both CTRL and DPR) and ERA5 occurs in the second row of Figure 9 (cloud liquid water content), where GEOS values are dramatically higher than ERA5 across the domain and assimilating DPR observations performs worst in reducing this discrepancy (in fact, contrary to L340–343, DPR appears similar to or even higher than CTRL over the eastern half of the domain). However, the authors make no attempt to explain why specifically cloud liquid water performed the worst in the assimilation, or through what physical or covariance mechanism Ku- and Ka-band reflectivities (which are largely insensitive to small cloud droplets) modify cloud liquid water. Equally important, there is no discussion of why such a large discrepancy in cloud liquid water is not apparent for Cyclone Midhili in Figure 11, where the Control experiment itself is already in much better agreement with ERA5 than in the Typhoon Khanun case. Note also that Figures 9 and 11 use very different colorbar ranges for the same variables (e.g., cloud liquid water spans 0–100 g m⁻² in Figure 9 versus 0–450 g m⁻² in Figure 11, and rainwater spans 0–450 g m⁻² versus 0–200 g m⁻²), which visually masks the true magnitude of differences between the two storms; consistent scales or explicit comments on the scale change are needed.
(5c) Using ERA5 as the sole quantitative reference for tropical cyclone structure and intensity (L326–333, L355–361) has important caveats that should be addressed. ERA5 does not assimilate spaceborne radar reflectivity (its hydrometeors are model-generated), its horizontal resolution (~31 km) is comparable to the C360 (~25 km) GEOS grid used here, and ERA5 is well known to substantially underestimate tropical cyclone intensity (for example, at 0900 UTC 2 August 2023, Typhoon Khanun’s best-track central pressure was several tens of hPa deeper than the ~983–992 hPa values shown in Figure 10 for ERA5 and GEOS). Likewise, IMERG incorporates GPM DPR/GMI retrievals and is therefore not fully independent of the assimilated DPR data. For tropical cyclone intensity and position—which should be evaluated across all tropical cyclones mentioned on L363–364 rather than just two snapshots—best-track datasets (e.g., IBTrACS/JTWC/JMA) should be used as the primary reference. In this context, please also revise L355, which refers to the two-dimensional MSLP and 850-hPa wind maps in Figure 10 as “cyclone intensity metrics”: intensity metrics (such as minimum central pressure or maximum 10-m wind speed) are scalar quantities by definition, not spatial fields. Several qualitative statements in the text also need quantitative support (for example, L357–358 refers to 850-hPa winds “exceeding 35 kt” even though the colorbar in Figure 10 saturates at 34.2 kt, and L371–372 states that the CTRL precipitation core in Figure 11 is “much smaller” than ERA5, which is not obvious visually).
(5d) To interpret the spatial analysis changes in Figures 9–12, the reader needs to know whether and where the 245-km DPR swath actually intersected Typhoon Khanun and Cyclone Midhili during the respective 6-hour assimilation windows (0900 UTC 2 August 2023 and 0000 UTC 15 November 2023), or whether the differences largely reflect accumulated increments from prior cycles. Please overlay the assimilated DPR swath footprints on Figure 8 (or Figures 9–12) and report the number of DPR observations assimilated in those cycles.
(6) Section 8 presents an extremely limited investigation of forecast impact that cannot support the broad claims made in the abstract (L10–14) and conclusions (L424–426). Only a single forecast verification time from a single initialization is shown for a single storm (Figure 13, Typhoon Khanun), and it is not clear how many analyses or forecasts were performed across the two storms. Even more problematically, the forecast fields shown in Figure 13 are valid at 0000 UTC 8 August 2023 from an initialization at 2100 UTC 2 August 2023—a forecast lead time of 123 hours (~5.5 days). It is well established that by 5–6 days lead time, tropical cyclone track and intensity errors grow large and nonlinearly and are dominated by synoptic-scale steering uncertainties and chaotic error growth; attributing a difference in storm position in a single 123-hour forecast to modest initial-condition changes is not scientifically meaningful (moreover, the “best track” position invoked on L403 is not even plotted on Figure 13, nor are MSLP or wind fields shown). In its current state, Section 8 is not useful at all and needs to become much more comprehensive. Given the six-month cycling period (August 2023–January 2024), the authors should present systematic, aggregated forecast verification statistics across all cycles and tropical cyclones in the period—including TC track and intensity errors as a function of lead time against best track, quantitative precipitation verification (e.g., bias, RMSE, or ETS against IMERG), observation-space forecast fits to DPR reflectivity, and standard global/tropical NWP forecast verification metrics. If such a systematic forecast evaluation is beyond the scope of the present study, Section 8 and the corresponding claims of forecast improvement in the abstract and conclusions should be removed.
Citation: https://doi.org/10.5194/egusphere-2026-4086-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 329 | 313 | 49 | 691 | 48 | 49 |
- HTML: 329
- PDF: 313
- XML: 49
- Total: 691
- BibTeX: 48
- EndNote: 49
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This study implements a spaceborne radar data assimilation (DA) capability within the Joint Effort for Data assimilation Integration (JEDI). Using observations from the Dual-frequency Precipitation Radar (DPR) aboard Global Precipitation Measurement (GPM), the authors conduct numerical experiments to investigate the impact of assimilating spaceborne radar observations on analyses and forecasts in the Goddard Earth Observing System (GEOS) model.
The topic is important and the results are interesting. However, I have several concerns regarding the manuscript structure, experimental design, description of the DA methodology, and presentation of the results. In particular, the manuscript should better emphasize the modifications introduced to enable radar DA, provide a more detailed description of the assimilation methodology, and present a more systematic and quantitative assessment of the impacts on analyses and forecasts. Therefore, I recommend major revision.
Major comments:
1. Description of the JEDI system and GEOS-JEDI framework
The general descriptions of JEDI and GEOS-JEDI systems are unnecessarily lengthy. These sections should be streamlined to emphasize the modifications introduced to enable spaceborne radar DA, rather than providing extensive descriptions of existing system components that are not specific to this study. In particular,
2. Experimental configuration and assimilation methodology
(1) What is the actual period covered by the continuously cycling experiments? The manuscript states that “The experiments are carried out continuously for the period spanning August 2023 through January 2024” (Lines 234–235), whereas the captions of Figures 3–5 refer to the period of “15 August to 15 September 2023”.
(2) What is the cycling frequency of this DA experiments? 6 hourly?
(3) Assimilation methodology: Given that this is a DA study, the assimilation methodology should be described more clearly and given greater prominence.
Similarly, it would be helpful to mark the DPR observations on top of both cyclones in Figure 8.
3. Observation error specification
4. Impacts of radar observations on analyses and forecasts