| 24 Aug 2026
Thirteen Winters (2012–2025) of Penetration-Aware Arctic Sea-Ice Emissivity and Emission Temperature from AMSR2
Abstract. Passive microwave observations enable long-term, all-weather monitoring of Arctic sea ice, but their physical interpretation depends on how surface emission is represented. In conventional satellite emissivity products, emissivity is typically retrieved under a skin-emission assumption, even though low-frequency microwave radiation over sea ice can originate from a deeper, non-isothermal subsurface layer. This study presents a new satellite-derived dataset of low-frequency sea-ice emissivity representative of the microwave-emitting layer and the corresponding emission temperature for wintertime Arctic sea ice. The dataset is generated using a simulation-trained retrieval framework in which penetration-aware sea-ice surface emission parameters are combined with reanalysis atmospheric profiles in radiative transfer calculations to construct forward-consistent training data, and neural-network inversion models are trained separately for emissivity and emission temperature to retrieve each variable independently from top-of-atmosphere brightness temperatures.
With the developed algorithm applied to Advanced Microwave Scanning Radiometer 2 (AMSR2) observations, a 13-winter record (2012/2013–2024/2025) of penetration-aware emissivity and emission temperature is produced over the Arctic Ocean north of 70°N under high sea-ice concentration (Kang, 2026; https://tinyurl.com/5n6843r8). Because neither emissivity nor emission temperature has a direct observational reference, the retrieval is evaluated in observation space via radiative-transfer closure. Brightness temperatures forward-simulated from the retrieved parameters agree closely with AMSR2 observations across all winters (r ≈ 0.99, spread of about 1 K), demonstrating radiative consistency despite training on simulation-based data. Comparable agreement is obtained at the 23.8 GHz horizontally polarized channel, which is not used in the retrieval, supporting cross-channel physical consistency. The dataset is temporally stable across winters and is intended to support radiative-transfer applications, satellite retrieval development, sea-ice characterization, and polar data assimilation.
Overall I found this manuscript was informative and well written. The results are fairly good with observation space biases that are small for this application area. I find the term `penetration-aware’ not quite appropriate because the manuscript is about passive emission, whereas when the reader hears `penetration’ they may be thinking of an active signal. I think `depth aware’ would be more appropriate.
I think the authors should Improve explanation of the contribution including how this is different from previous work by the author, in particular the 2025 paper where the ANN model was introduced. The data (instrument and frequencies are different), in addition to other items.
The current approach is developed more so for frequencies that are very relevant for sea ice monitoring, and uses AMSR2 data. The relevance of the work is a little vague (e.g. “radiative-transfer applications, satellite retrieval development, sea-ice characterization, and polar data assimilation”). If more context could be given that would help. For example, the frequencies used here (around 18, 23 and 36 GHz) for dry conditions (as is investigated here) are not normally of interest for NWP because they are not that sensitive to the atmosphere, but could be used for sea ice forecasting. In this regard, the following reference could be added,
Scott, K. A., M. Buehner, A. Caya, and T. Carrieres, 2012: Direct Assimilation of AMSR-E Brightness Temperatures for Estimating Sea Ice Concentration. Mon. Wea. Rev., 140, 997–1013, https://doi.org/10.1175/MWR-D-11-00014.1.
since it is an earlier study that developed a method for direct assimilation of brightness temperatures for sea ice forecasting, when radiative transfer models were a bit less developed than they are now
On line 50 It is stated that other methods assume the emission originates from the surface skin. It would be helpful to the reader to state here what the underlying assumption is more clearly. Is it that a skin temperature is used explicitly in the other methods? Or an air temperature that is correlated with skin temperature, or are there other details.
It would also be good to clarify the rationale for machine learning here. As I see it, the use of ML allows emissivity and emitting layer temperature to be retrieved using only AMSR2 observations. The 1D Lagrangian model can calculate these variables as well, but requires physics based inputs rather than brightness temperatures (which may be spatially sparse), and is probably computationally more expensive than ML. In this sense the lower part of Fig 1 is a forward problem, whereas ML is used for the inverse problem (upper part).
Data used: Granule data are used for the brightness temperatures (Level-1R), which I think is swath data (several overpasses each day), while a daily SIC product is used. Is this the case? If so how are the data consolidated?
Line 171-172 It is stated microwave and infrared temperatures are nudged into the model - these two data sources would retrieve temperature at different depths (infrared is a skin temperature and microwave being from the emitting layer), how can they be used consistently in the nudging framework and how can the microwave temperature be used without knowledge of the emitting layer depth? What are the explicit assumptions and has the sensitivity to these assumptions been tested?
I think the argument that the emissivity is a property of the sea ice/snow (although it is both an electromagnetic and structural property) does make it a useful quantity to have knowledge of, but I am not sure how useful that is here because of the treatment of snow (see below)
Line 292 it is stated that dry snow is nearly transparent at these frequencies. But the 18.7 - 36.5 vertically polarized brightness temperature difference is used for snow depth in the literature, due to scattering at 36.5 GHz (which is not a feature at 18.7 GHz due to the larger wavelengths), so I don’t follow the statement that dry snow is nearly transparent. It also raises the question of what surface type emissivity is being retrieved for 36.5 GHz, which may be related to the RTM used.
Treatment of snow - Snow is considered in the 1D Lagrangian model, but the RTM used assumes no snow, so the simulated TBs are not considering the impact of snow. Isn’t this inconsistent?
Line 187 What radiative transfer model was used to diagnose emissivity and emitting layer temperature? In addition to being an important piece to understand the scientific context and underlying assumptions, it would also help interpret the first-year/multi-year ice parts of the manuscript (it does not seem that the ice type is given, so it must emerge from the physical modelling for the simulated emissivity, is that correct?). Wouldn’t the emitting layer depth be different for the different ice types and also the temperature at depth? Is that what the emitting layer temperature range is showing (Fig 3).
Table 3 - sigma is the referred to as the retrieved-reference error spread - it is not an error, it is a standard deviation of the difference. The values in the table also look much smaller than the widths of the histograms in Fig 4 suggest.
Line 295 Where it says the cross-channel coherence plays a role in constraining Te - where does this happen in the method? In data assimilation, there could be interchannel error correlations, for example.