the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 11 Jun 2026
The Arctic Weather Satellite, introducing a new wavelength range for ice hydrometeor retrievals
Abstract. The first cloud property retrievals based on operational sub-millimetre measurements are presented, making use of the channels between 89 and 325 GHz of the Arctic Weather Satellite (AWS). The main quantities of the dataset are frozen water path (FWP) and the associated mass-weighted mean altitudes and particle sizes. In this first version, results are restricted to latitudes between 60° S and 60° N. The retrievals are based on detailed simulations of instrument observations. The actual inversion is made by a quantile regression neural network, and case-specific uncertainty estimates are provided.
Retrievals performed on simulations suggest that retrieved FWP values are essentially unbiased across a wide dynamic range, from 10 kg m-2 down to 40 g m-2. The associated mass-weighted mean altitude is also essentially unbiased for the entire relevant range of 2 km to 12 km. The particle size estimates, however, show a slight bias for sizes other than 400 μm. Comparisons with other datasets provide strong indications that these results also extend to retrievals from real observations; for example, local and zonal means match those of existing radar/lidar-based retrieval products.
The accuracy in FWP should be unprecedented among estimates based on passive satellite data, thanks to the new sensitivity afforded by sub-millimetre channels. The dataset complements cloud radar observations by providing a significantly broader spatial coverage. There is also potential to create a climate-relevant dataset, as the retrievals are directly applicable to the EPS-Sterna constellation, continuing the AWS observations up to 2045.
- Preprint
(2478 KB) - Metadata XML
-
Supplement
(650 KB) - BibTeX
- EndNote
Status: open (until 18 Aug 2026)
- RC1: 'Comment on egusphere-2026-2456', Anonymous Referee #1, 25 Jul 2026 reply
-
RC2: 'Comment on egusphere-2026-2456', Anonymous Referee #2, 04 Aug 2026
reply
The paper presents the first global frozen water path (product) from submillimetre measurements by the Arctic weather satellite (AWS). Retrieval products (FWP, median FWP height (Zm) and diameter (Dm) will be released after the acceptance of this publication. Each pixel-based retrieval comes with its specific uncertainty characteristics, making the data set highly valuable for a broad community. In general, paper includes thoughtful analysis of the retrieval performance (self-check) and comparisons with other available satellite products. The work is performed carefully; however, I think some more information on the underlying physical sensitivities and (different) information content of the data would make the paper even more valuable. Along the same lines, I include some suggestions on useful information for potential users.
General Comments:
1. Physical meaning of retrieved products.
The abstract states: “The accuracy in FWP should be unprecedented among estimates based on passive satellite data”. This is a strong statement that requires more argumentation. The manuscript needs to stress the complexity of ice particle size distributions (see Bartolomé García et al. 2024), transitioning from tiny particles (sub-visual cirrus) to large snow aggregates, and the information content of different retrieval products. For example, the selective sensitivity of the used data records (MODIS, CICM) that are affected by saturation in the VIS/IR signals.
As FWP is not a widely used variable, the difference to other product names (CIWP, see Stubenrauch et al., 2024) needs to be made clear. Anticipating certain user questions: Could you make a rough estimate of how much of the global average FWP is missed by conventional products and how much more your product (CHIP-AWS) can sense?
and give some recommendations (I don’t expect detailed answers) on
- How can one identify a “clear-sky” pixel and with which uncertainty?
- How many cirrus clouds will be missed?
- Do I also need to apply a threshold of 0.1 kg/m2?
- How much is a retrieval of 40 g/m2 worth?
- Is Dm basically only a measure of snowfall size?
2. Clearness and completeness of the provided information: At several instances (see my specific comments), more accurate formulations are needed. One example is the term bias, which is frequently used in the paper, but its actual meaning is not always clear, as truth is not really available.
Information on intercomparison data, e.g. MODIS, CLARA, is missing. The Discussion section is a bit confusing – here suddenly SPARe-ICE is appearing. More information on accuracy and limitation these data is needed. For the outlook the, authors should also consider intercomparisons with ground-based reference sites, e.g., ACTRIS/Cloudnet.
3. Length of the record: It is never really stated how long the data record for the presented analysis of CHIP-AWS is. I find it really important to have exactly a (full) one-year period (March to March) to avoid seasonal imbalances. MODIS or ERA5 could be used to assess the interannual variability.
Minor point, but in many instances: One sentence does not make a paragraph.
Specific Comments:
Abstract
(you should mentrion the product name (CHIP-AWS for branding)L4: Detailed simulations- here information on the blend of Cloudsat retrievals and ERA5 should be mentioned. For later: Any retrieval can only be as good as the underlying database. The authors are a bit short on information (and motivation) how they construct the database. Explain why not only ERA5 information is taken.
L11:” The accuracy in FWP should be unprecedented among estimates based on passive satellite data”. This is a strong message see major comment #1.
L12: “significantly” could be nearly anything – quantify
Introduction
L21 – first dataset, duration missing -> first full year data set
L34 225-1064 nanometer not micrometer😊; add (about 3mm) in parentheses after 94 GHz
L33- maybe refer to the figure in the Buehler paper for the different wavelength respnse
L35 – “spatial” is irritating – either vertical or radial
L39 – I would add climate and the changing water cycle as further motivatin
L43 – “applying machine learning” somehow sounds too fancy. ML is basically uses as regession
L47 – submillimetre vs far infrared?
L47 – for clarity mention the two conventional ones by name
L73 – “all hydrometeor classes” – hydrometeors haven’t been introduces yet. I think it is important to mention the arbitrary (model-dependent) distinction between ice and snow water content and the fact that kilometer-scale models introduce denser hydrometeor classes (graupel, sometimes hail) to enable the representation of strong convective events, i.e., thunderstorms.
L75 – “report” is not the right name – mostly they are even prognostic. Suggest to simply say “consider”
L82 – retrieval PRODUCT
L89 – I find the sentence confusing: “operational machine learning models for cloud retrieval have exclusively been trained on existing retrievals” . Maybe I misunderstood what is meant with “operational ML” (then it should be better explained) but tons of satellite retrievals are run on simulated measurements (mainly from atmospheric model data).
Similarly L90 “For reasons discussed below, this more complex approach could not be avoided” sounds weird.
- Sources and reference data
It would be helpful for the reader to get a rough idea of the retrieval and validation strategy before presenting the subsections to understand for which purpose they are needed for.
For example, I was first expecting that ERA5 output (including hydrometeors) would be used for the simulations.
L101: 145 over a +/ - ??degree range
L128: Give number of pressure levels. There is significantly more detail in the model layer output though I don’t believe it is important. More importantly, I am missing a discussing how physical consistency between thermodynamic state and ice hydrometeors is guaranteed? As an example could you have 20% relative humidity and high FWP?
L131: why is only cloud water content used? I would also recommend the inclusion of rain as light rain amounts are frequent and even present in cases of no surface precipitation.
L141: Why 2015? Maybe one could see from the longer-term MODIS, CCICs data set that this is an “average” year?
Subsections 2.3 and 2.4 don’t provide information on the uncertainty of the described products. Specifically what is the limitation of CCICs wrt high FWP?
CLARA and MODIS Data (and their uncertainties) are not described here and not mentioned in the data availability section.
- Data product
I would be curious how strongly the three main retrieval products are correlated?
L159 – see my main comment on duration of the data set for analysis – full year
L167 – provide also angle and swath width – then you don’t need it later
L188 – also Eriksson et al., 2020 don’t use the term spatial sensitivity but refer to vertical one..
L196 – “These restrictions are put in place to reduced the scope” -> These restrictions are put in place to reduce the complexity..
L197 - Variability can mean many different things, please explain your definition used here
L206 – at least mention that ERA5 is coarser
L214 – what does unreasonable mean – how can this be reproduced? I don’t expect a full description but maybe a link to a reference
L225 – How strongly are the variables correlated in the training, test and application? Do you see any changes? It is also very interesting to investigate the correlation with temperature, we also suggest including the zero degree line in the figures. One could also do a 2D histogram with zm and the height of the zero degree isotherm for consistency checks.
L230: There needs to be more discussion on the choice of the threshold and implications for analysis – when do we have a (precipitating) cloud?.
- Retrieval Method
L295 – for me it would have been helpful to mention early that a binning of 0.01 for the CDF would be used
What is the choice for the chosen output quantiles – why not 0.05 and 0.95?
L 313: I guess randomly chosen?
L320 This is a good opportunity to advise the user if a “trustworthy” FWP is retrieved..or what the likelihood is that this pixel has no FWP?
L330 – specifiy substantial
L333: THE mean estimate
- Results
L336 – agrees is a too strong word -> represents the observed atmospheric state
L340 – what is meant with accurate? Just the RT? Then I would say first a representative atmospheric state and then an accurate RT.
L343: best-performing – best understood surface interaction
L353: As a rule of thumb values below….
L358: There is some slight discrepancy at the coldest temperatures already seen here over ocean – so one could start here already start the discussion on intense convection
L360: You could also make a correlation matrix between different channels
L362 – specify where one can see this? Furthermore, intense convective cores – you didn’t describe beforehand why these are tricky, e.g. graupel, hail..
L374: I think the figures (going down to 1 gm-3) overemphasize the importance of the really low values. Looking at the helpful Fig 5.
L409: bias-free – how can you judge as truth is there; away -> higher?
L417: I am missing a discussion on the physical interpretation of Fig.8 Basically what can be seen what not?
Section 5.34. The MODIS data set was not introduced before. To my knowledge it was never constructed to include precipitating ice so there is no surprise about the difference
L439 “The CHIP-AWS estimates have no coastline artefacts that cannot be attributed to orographic features. “ That is confusing. Orography makes an artifact?
L442: cases : do you mean ocean and land? Maybe regimes is better: there is no clear indication where one can see the effect of the dry cases ?
L446: Better surface emissivities..there should be also some reference to the literature on that
L449 and following: I don’t understand the meaning of bias here because no truth is there.
L460: what is the correlation between DARDAR and 2C-ICE?
L469..and knowledge of the physics.
L470: In case of multi-layer clouds radar – lidar is not a good reference
L484 . Refer to where the retrieval accuracy is mentioned.
6 Discussion:
I got a bit confused with this chapter which seemed to be more like a summary than a discussion.
L487 and following: Which are the three orders of magnitude – be specific when one can “trust” the retrieval. As there are specific uncertainty values for each pixel, what would be the recipe for using them? What is meant with “entire relevant range”?
L489: “The comparison database values are radar retrievals” What is meant by that?..or is this the start of a new paragraph?
L491: Now the discussion goes to Spare-Ice which has not been introduced before - why is it not included with the MODIS and CLARA comparison?
L501: Why not ground-based estimates from Cloudnet? This could also provide information on zm.
L504: “Recent machine-learning approaches trained on radar retrievals have achieved similar results” Sounds very generic – where does the information come from – new architecture, new training??
- Conclusions:
L526: Simulations can mean anything – also atmospheric model simulations. In this sense the sentence in L534 should be reformulated” This is, to the best of our knowledge, the first simulation-based passive retrieval method”
L529: rather unspecific. Simulated Ta distributions agree with their climatological counterpart observed over one year globally.
L539: the effect of convective cores was never explained,
L534: “all three variable??
There should also be some link to passive MW snowfall and SWP retrieval (Camplani et al., 2024) which use partly similar principles.
Figures
not always all lines are explainedFigure 1: Explain CDF on first occurrence. Give date and time. How was the transect line chosen? It would be very helpful for interpretation to plot the zero-degree isotherm (from ERA5) into the Zm transect.
Figure 2: scaling of middle CDF/pdf is ns irritation – better go only to 1 or even lower FWP
Figure 6: Micrometer are missing, red and dashed lines need to be explained. The vertical scale of the MFE should be better visible for values of 100%
Tables
Table 1: main (top 3) variables. Why is the mirror angle not included?
Table 3: Why not polarisation and bandwidth?
Table 4: Are these really the best metrics? From these I would say it can only retrieve snow? What about relative error and only calculated above the threshold of 0.1 kgm-2? In the abstract you mention down to 40 gm-2 – but how trustworthy is such an estimate?
References
Bartolomé García, I., Sourdeval, O., Spang, R., and Krämer, M.: Technical note: Bimodal parameterizations of in situ ice cloud particle size distributions, Atmos. Chem. Phys., 24, 1699–1716, https://doi.org/10.5194/acp-24-1699-2024, 2024.
Camplani, A., Casella, D., Sanò, P., and Panegrossi, G.: The High lAtitude sNowfall Detection and Estimation aLgorithm for ATMS (HANDEL-ATMS): a new algorithm for snowfall retrieval at high latitudes, Atmos. Meas. Tech., 17, 2195–2217, 2024.
Stubenrauch, C.J., Kinne, S., Mandorli, G. et al. Lessons Learned from the Updated GEWEX Cloud Assessment Database. Surv Geophys 45, 1999–2048 (2024).
Citation: https://doi.org/10.5194/egusphere-2026-2456-RC2 -
RC3: 'Comment on egusphere-2026-2456', Anonymous Referee #3, 11 Aug 2026
reply
The manuscript "The Arctic Weather Satellite, introducing a new wavelength range for ice hydrometeor retrievals" by McEvoy et al. introduces a new frozen water path (FWP) product based on observations from the Arctic Weather Satellite (AWS), which provides the first operational Earth-viewing microwave observations at submillimeter wavelengths. The product is based on a pixel-wise retrieval algorithm using a database of reference hydrometeor properties derived from CloudSat Cloud Profiling Radar (CPR) observations. The dataset provides retrievals for the central portion of the AWS swath and is restricted to latitudes equatorward of ±60°. The retrieval also exhibits increased uncertainties over high-latitude land surfaces.
The manuscript is generally well written and provides an adequate characterization of the dataset and retrieval methodology. However, several statements regarding the accuracy, spatial resolution, and physical information content of the retrieval go beyond what can be established from the analyses presented in the manuscript. In particular, only limited comparisons against independent observations are provided, making it difficult to determine whether some of the additional small-scale variability seen in the AWS retrieval represents genuine FWP structure or retrieval noise. I therefore think that several claims regarding retrieval accuracy and resolved spatial structure should be removed, qualified, or reformulated.
Since these concerns primarily affect the interpretation and presentation of the results rather than the retrieval methodology itself, I recommend minor revisions.
Comments:
- Title: The current title may overstate the novelty of the wavelength range. Submillimeter observations have previously been used for atmospheric and hydrometeor observations from airborne sensors and limb-sounding instruments. I suggest clarifying that the novelty concerns operational Earth-viewing observations, if that is what the authors intend.
- l. 11: Given the limited independent validation presented in the manuscript, I do not think this claim regarding retrieval accuracy is sufficiently supported. I suggest removing or qualifying it. The claim also seems difficult to reconcile with the stated limitations of the retrieval over land surfaces.
- l. 45: Please elaborate on the statement that passive retrievals do not provide new physical insights. It is not clear to me what distinction is being made here between the physical information provided by passive retrievals and that provided by the proposed approach.
- l. 89: As I understand the methodology, the CHIPS-AWS retrieval is itself indirectly constrained by CloudSat-derived hydrometeor retrievals through the reference database used for its development. Please clarify this relationship when contrasting CHIPS-AWS with CloudSat-based products.
- l. 91: Please explain more concretely how the proposed approach enables improved physical understanding. What additional physical information can be inferred from the retrieval beyond the presence and magnitude of the retrieved ice hydrometeor signal?
- l. 194: Is the primary reason for these restrictions the difficulty of accurately simulating observations at high incidence angles over high-latitude surfaces? If so, I think this should be stated more explicitly.
- l. 205: I do not think an effective spatial resolution can be inferred directly from the nominal footprint size or sampling characteristics. Please quantify the effective resolution using appropriate test data, or reformulate this statement in terms of the nominal footprint or sampling resolution.
- l. 214: Please consider reformulating this sentence for clarity.
- l. 216 and following: Please specify whether (Z_m) is calculated along the vertical radar column or along the slanted AWS viewing geometry.
- Section 4.1: Please specify whether any spatial averaging or aggregation is applied to the reference FWP values before they are used in the retrieval database.
- l. 321: Please specify the number of neurons in each layer of the network.
- l. 330: Please provide a quantitative range for the differences described here as "substantial."
- Fig. 4: Please consider using frequency rather than channel names for the labels. This would make the figure easier to interpret without requiring familiarity with the AWS channel nomenclature.
- Fig. 6: The micron symbol appears to be missing.
- l. 404: The presence of greater spatial variability or finer-scale structure does not by itself demonstrate that the retrieval contains more information about actual cloud or FWP structure. Some of the additional variability in the CHIPS-AWS estimates also appears consistent with pixel-scale retrieval noise. I suggest avoiding an interpretation of increased spatial variability as increased physical information unless this can be demonstrated against an independent reference.
Citation: https://doi.org/10.5194/egusphere-2026-2456-RC3 -
RC4: 'Comment on egusphere-2026-2456', Anonymous Referee #4, 11 Aug 2026
reply
In this paper, the authors present their product for frozen water path (FWP) from the AWS satellite. This is an inherently valuable contribution, as it represents the first dedicated FWP product derived from a satellite with historic wavelengths. Importantly, the algorithm uses contemporary ML methods and cites recent algorithm literature, which is a requirement for any new retrieval method given today's rapidly advancing computing capabilities.
The authors present a strong evaluation of their product. They are transparent about the difficulties of performing an ideal evaluation, given the lack of proper overlap with an operational W-Band cloud radar. In the discussion, they make a well-supported argument that their simulated brightness temperatures are valid for this evaluation.
Most of my comments concern the results section. Some reorganization of the text is needed: several lines currently in figure captions belong in the body text, and some lines in the results section belong in the discussion. Specific examples are noted below, but in general, the results section would benefit from more objective and more descriptive discussion of the figures. Readers should be able to broadly visualize the figures and identify the key results from the text alone, and this is not currently possible.
Since my comments mostly concern the presentation of the science rather than the science itself, I recommend submission with minor revisions.
Major Comments
-
Occurrence fractions are not defined in the text and should be defined there rather than only in the figure caption. It is also unclear whether calculating bin occurrence is appropriate when FWP is measured on a logarithmic scale. Should these values be normalized by bin width?
-
What is the purpose of presenting the mean value here? No analysis of the mean appears in either subsection referring to Figure 5. If its relevance cannot be addressed in the results or discussion, consider removing it.
-
Section 5.3.2 would benefit from a rewrite and reorganization. The results are currently described in only two sentences, which seems insufficient to merit a standalone subsection. The following two paragraphs appear better suited to the discussion section.
-
More quantitative results are requested in Section 5.3.3. Specifically, what range of occurrence fractions qualifies as results that "must be considered good for passive observation"?
-
More quantitative results are also requested in Section 5.3.4. What is meant by "slight positive bias," and how should "areas of deviation" be interpreted? Some of these requests may seem minor if the level of description elsewhere were more quantitative, but as currently written, this section relies on subjective language without supporting numerical results.
Minor Comments
-
Section 3.1: Please clarify what is meant by "most probable" where it is defined.
-
Line 250: Please clarify that this refers to an ice-scattering model and a PSD model, to avoid confusion with a static PSD.
-
Line 362: Additional detail on the difficulties associated with convective cores would provide better context for why they are currently missed and why the authors plan to address them in future work.
-
Section 5.1.2: This section may not warrant a standalone subsection, as it does not present any visible results. The "ongoing work" sentence may fit better in the introduction.
-
Section 5.2: Consider splitting Figure 5 into two separate figures. The bottom row is not discussed until four subsections later and includes two variables that are not used or introduced in the top row.
-
Line 430: This information should be presented while the figure is still being described, prior to the discussion of results.
Citation: https://doi.org/10.5194/egusphere-2026-2456-RC4 -
Data sets
CHIP-AWS v1.0.0 2025 archive: Atmospheric ice mass properties Peter McEvoy, Eleanor May, and Patrick Eriksson https://researchdata.se/en/catalogue/dataset/2026-135/1?previewToken=859edaac-fe43-4f42-b1ad-f86a0d4dc4a2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 157 | 60 | 15 | 232 | 15 | 9 | 12 |
- HTML: 157
- PDF: 60
- XML: 15
- Total: 232
- Supplement: 15
- BibTeX: 9
- EndNote: 12
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
See the comments in the attachment.