The HALO–(AC)3 aircraft campaign was conducted from 7 March to 12 April 2022 to investigate the evolution of air mass transformation processes during warm-air intrusions and cold-air outbreaks in the Arctic . In total, 59 flights with multiple research aircraft were realized, among them 17 flights with research aircraft HALO, which was based in Kiruna, Sweden. In Fig. , the locations of measurement are depicted, together with the campaign-averaged sea-ice concentration (SIC). In addition to all HALO tracks that were flown during the campaign (limited to the map extent; for a full overview see ).

Due to HALO's range of up to 9000 km, the measurements capture Arctic surface and atmospheric parameters on a regional scale while ensuring high spatial resolution when compared to satellite sensors like MODIS, AVHRR, or the Sentinel-3 Sea and Land Surface Temperature Radiometer SLSTR; The variability of different Arctic surface types (OW, IWM, TI, and SC) is highest in the region between the ice-free open ocean and the pack ice. Here, we focus on small-scale variability of the surface skin temperature resulting from the inhomogeneous distribution of Arctic surface types. Therefore, the following analysis will be restricted to the marginal sea-ice zone (MIZ), which is suitable for such investigations. The MIZ is defined as the region where the campaign averaged sea-ice concentration (SIC) average was between 10 % and 90 %. In addition, we include data where the SIC exceeded 90 % for less than 10 min. A set of remote sensing instruments was deployed on HALO , of which only those relevant to the analysis are briefly introduced here. To capture two-dimensional (2D) fields of TIR radiances, the VELOX (Video airbornE Longwave Observations within siX channels) TIR imager was operated in a nadir-viewing configuration. VELOX covers a spectral range of 7.7 to 12 µm, providing radiance measurements, which are converted to brightness temperatures . At a typical flight altitude of 10 km, the imager achieves a horizontal resolution of 10 m by 10 m per pixel, corresponding to a field of view (FOV) spanning an area of 5 km by 6 km. VELOX acquires images with a temporal resolution of 100 Hz.

The instrument is operated with six spectral filters resulting in six channels, of which two are redundant broadband channels (channel 1 and 4). The remaining channels are narrow-band, each centred on specific wavelengths. The uncertainty in the measurements is characterized by the net equivalent temperature difference (NETD) for each channel. The broadband channels have a NETD of 0.048 K, while the narrow-band channels show varying NETD values shown in Table . For HALO–(AC)3, the corrected brightness temperature data, resampled to 1 s temporal resolution, were provided by . To retrieve cloud cover, the HALO Microwave Package HAMP and the water vapour differential absorption lidar WALES were installed on HALO. In addition, 330 dropsondes were released during the campaign. We have restricted our analysis to cloud-free scenes in the MIZ. For this purpose, a cloud mask based on campaign-specific radar reflectivity and lidar backscatter coefficient thresholds was applied . To ensure the data quality, each scene was visually examined to confirm the absence of clouds.

Independent measurements of surface skin temperature were provided by MODIS sea-ice surface temperature IST; and sea surface temperature SST; . Both datasets were based on a split-window retrieval algorithm, which determined surface skin temperature from the measured brightness temperatures. For the respective surface types, MODIS channels 1 (0.645 µm), 2 (0.865 µm), 4 (0.555 µm), 6 (1.64 µm), 31 (11 µm), and 32 (12 µm) were used. The IST dataset were provided as swaths with a horizontal resolution of 1 km by 1 km, while the SST dataset is gridded with a horizontal resolution of 4 km by 4 km. Daily fields of SIC were provided by the assimilated MODIS/AMSR-2 SIC product, derived from a synthesis from MODIS and AMSR-2 . Depending on the combination of MODIS and AMSR-2, the fields of SIC have a 5 km horizontal resolution for all conditions and 1 km for cloud-free scenes. Satellite images in terms of spectral radiance with high horizontal resolution were obtained from the Sentinel-2 multispectral imager (MSI, hereafter Sentinel-2) data. To characterize the surface reflectivity, the red (0.664 µm), green (0.559 µm), and blue (0.492 µm) (RGB) channels were sufficient, which have a horizontal resolution of 10 m by 10 m. For the high latitudes reached by HALO-(AC)3 observations, the revisit time of Sentinel-2 is about 1 d, which enabled daily observations and allowed for collocation of the satellite observations with VELOX images . The Sentinel-2 data were accessed via the Google Earth Engine GEE; .

VELOX detects spectral radiances in the TIR wavelength range that are converted to TIR brightness temperatures, TB, characterizing the combined emission by atmospheric components and the surface. Actually, a significant contribution results from emission by atmospheric gases, although the spectral bands are located in the atmospheric window region. This atmospheric contribution has to be removed from the signal to derive the surface temperature, also called surface skin temperature. To correct for the atmospheric emission between the aeroplane and the surface, a split-window method (SW) is commonly applied . Adjusted to VELOX measurements, this approach can be formulated as follows: 1TS=asw+bsw⋅TB,5+csw⋅(TB,5-TB,6). TS represents the surface skin temperature, TB,5 is the brightness temperature measured with VELOX channel 5 installed on HALO in about 10 km altitude, which is least affected by water vapour absorption. The coefficients asw, bsw, and csw are empirically determined with a linear regression. Thus, TB,5-TB,6, represents the brightness temperature difference between channels 5 and 6, serving as a proxy for water vapour absorption. found that this brightness temperature difference observed in the Arctic region is also sensitive to other parameters, such as atmospheric inversion height or aerosol particles. Their proposed single-channel algorithm SCA; has been adapted to VELOX data as follows: 2TS=asca+bsca⋅TB,5. We have performed radiative transfer simulations (RTSs) to constrain the contribution of atmospheric absorption to the surface skin temperature for both retrieval methods. The RTSs were conducted with the radiative transfer library libRadtran; . The simulations were initialized with temperature and humidity profiles from dropsondes that were released from HALO during the campaign . The surface skin temperature was provided by MODIS , whereas ozone content was given by the ERA5 reanalysis data .

For the molecular absorption parameters, REPTRAN medium was chosen, along with the DIScrete ORdinate Radiative Transfer solvers DISORT; . To constrain the retrieval uncertainties as a function of the atmospheric total column water vapour concentration, the integrated water vapour (IWV) was varied from 0 to 50 kg m⁻². During the HALO–(AC)3 campaign, the integrated water vapour (IWV) was confined to values less than 10 kg m⁻² . To evaluate the two retrieval methods the total uncertainty δTS was calculated for both algorithms. Adapted from , who formulated the uncertainties for the MODIS IST retrieval, the total uncertainty of the VELOX retrievals was formulated as follows: 3δTSsw=(δTatmsw)2+(δTVEL,i)2,4δTSsca=(δTatmsca)2+(δTVEL,5)2,5δTvel,i=(δTvel,isys)2+(δTVEL,iran)2, where 6δTVEL,isys=A+B⋅TB,i,7δTVEL,iran=NETDB,i. The overall uncertainty of the surface skin temperature δTS was quantified as the square root of the sum of the squared uncertainties from the atmospheric correction and the uncertainty introduced by VELOX δTVEL,i, which was split into a random part δTVEL,iran equivalent to the NETD and a systematic part δTVEL,isys. The systematic uncertainty was parameterized based on the measured brightness temperature TB,i in channel i. The superscripts “sca” and “sw” correspond to the respective algorithms, while the index subscript i indicates VELOX channel i. The total uncertainty for both retrievals, depending on TB,5 and assuming a constant δTatm, is shown in Fig. . Due to the difference in considering only the NETD of channel 5 for the SCA and both NETDs of channel 5 and 6 for the SW, the SCA retrieval has a lower error across all temperature ranges.

To evaluate the sensitivity of the retrievals to IWV, the total uncertainty of both retrievals as a function of IWV is shown in Fig. . Below the IWV threshold of 10 kg m⁻², the SCA outperforms the SW, due to the reduced NETD only using one channel. Above this threshold, atmospheric absorption dominates the total uncertainty favouring the SW algorithm. In summary, the single-channel algorithm has a lower total uncertainty for IWV values below 10 kg m⁻², while the split-window algorithm is more suitable for more humid atmospheres. Therefore, the SCA is applicable in the Arctic region when low IWV concentrations are present.

As a consequence, we continue with the derivation of the regression coefficients asc,bsc of the SCA algorithm. For this purpose, the RTSs were performed with a temporal resolution of 1 s, resulting in simulated brightness temperature values for VELOX channel 5 TB,5,RTS at HALO flight altitude. These simulated brightness temperature values at flight altitude are linearly regressed against MODIS surface skin temperature. The resulting fit parameters for slope and offset serve as the single-channel coefficients: 8asc=9.051K9bsc=0.967K-1. In Fig. , MODIS surface skin temperatures TS,MOD were plotted against simulated brightness temperature TB,5,RTS. The regression shows a coefficient of determination of R2=0.99 and a root mean square error (RMSE) of 0.47 K. Substituting δatmsc=0.47 and NETD=0.473 K into Eq. (), the overall uncertainty of the SCA algorithm using RTSs with ERA5 IWV data was computed as 10δTS=1.1±0.3K, where the range of δTS reflects the error for different measurement conditions.

To distinguish different surface types, we have adapted established definitions . As no melt ponds were observed during HALO-(AC)3, this surface type was omitted. To illustrate the surface types, a Sentinel-2 true colour image is analysed in Fig. . All surface types applied in this study are present in this scene and characterized in Table .

The image analysis consists of three steps. First, the VELOX 2D-images are preprocessed. Next, a random forest (RFA) classification algorithm is applied for pixel-wise surface classification. Finally, a segmentation algorithm is used to identify and summarize areas of the same surface type.

In Fig. , the VELOX-retrieved surface skin temperature TS,VELOX is compared to the MODIS surface skin temperature TS,MOD, obtained from and , showing the coefficient of determination R2 to be equal to 0.96.

For this comparison, the instantaneous FOVs of the single VELOX pixels were combined by averaging to fit the MODIS pixel-size, allowing for a direct comparison between their two datasets. The RMSE was determined to be 2.0 K with a bias of 0.51 K and a mean average difference of 1.55 K. Furthermore, Fig. 8 indicates slightly higher values of surface skin temperature derived from VELOX, TS,VELOX, with respect to MODIS, TS,MOD, over sea ice and lower values over open water. As the dataset comprises multiple days, it is essential to provide information on the location of the data. To simplify this spatial information into a scalar, the data are grouped by their distance to the sea-ice edge (positive direction into the internal ice zone). As the individual pixels have been georeferenced, their relative distance to the nearest sea-ice edge (defined by campaign-averaged SIC values between 9 %–11 %) is computed. For this, the distance of each spatial segment centre to the temporally closest available AMSR-2/MODIS SIC pixel is calculated. The mean surface skin temperature coloured by surface types is plotted in Fig. a against the distance to the sea-ice edge.

In Fig. b, the mean surface skin temperature of all segments TS‾, weighted by their size, is shown. A clear separation between the TS‾ of different surface types is observed as expected. The values from Fig. b are displayed together with the corresponding error range in Table .

From the segmentation, we retrieve the corresponding segment size, mean temperature, and standard deviation of each segment. The results are illustrated in Fig. , showing the spatial distance of each segment centre to the nearest sea-ice edge plotted against the corresponding surface type. The fraction of the open-water surface type decreases from 40 % to below 5 % in the first 20 km, while the fraction of the snow-covered surface type becomes increasingly dominant when approaching the pack ice. The surface types thin ice and ice–water mix maintain relatively constant fractions of occurrence across all considered distances from the ice edge, with no clear trend observed.

When compared with the provided SIC from MODIS/AMSR-2, the computed RMSE and mean absolute difference (MAD) are 8 % and 5 %, respectively. For this comparison, the nearest available SIC data from the satellite product were matched with a similar FOV of the VELOX PLI. The errors result from the temporal mismatch between both datasets, as MODIS/AMSR-2 SIC is only available as a daily gridded product. When comparing the bias, i.e. the difference between VELOX SIC and MODIS/AMSR-2 SIC, an underestimation of 3 % or an overestimation of 5 % is observed, depending on whether only pixels classified as open water are considered open water or if pixels classified as both open water and ice–water mix are included. As shown in Fig. , the open-water fraction and the area fractions derived from VELOX agree within the given error range.

Since the segmentation enables the measurement of individual segment sizes, an analysis of the spatial structure of the data is performed. Here, we extend the concept of the floe size distribution FSD; to the segment size distribution NSSD, resulting in the following description: 11NSSD(xSEG)=C⋅(xSEG)β. Here, xSEG represents the segment size in units of m², C is an empirical constant, and β is the dimensionless power-law exponent describing the scaling of the distribution. The closer the exponent is to zero, the more NSSD favours large segments. This approach simplifies the complex spatial heterogeneity of the MIZ by expressing the scaling of NSSD using β, a single scalar value.

In Fig. , the segment size density NSSD (in units of km⁻²) for different surface types is displayed in a double-logarithmic graph as a function of the segment size, xSEG. In addition, the individual distributions are fitted with a linear model. The slope of each linear fit corresponds to the exponent of the power-law distribution, β. The different βi values computed for each surface type are shown in Table .

We conclude that, in addition to the different sea-ice types, the surface types ice–water mix and open water also follow a power-law distribution. The computed β values for, for example, the snow-covered sea-ice type are in the range of corresponding literature data, with values ranging from -0.91 to -2.9 . A key characteristic observed during our flights over the MIZ was the presence of large, contiguous areas of thin ice or snow-covered ice, which can appear as very large segments in our final classification (e.g. Fig. 7e). We interpret these as genuine physical features of newly forming ice in the MIZ at the time of observation rather than as segmentation artefacts. However, we acknowledge that the definition of a “segment” is sensitive to the processing methodology and that the scale of these large features can influence the resulting feature size distribution (FSD). To quantify this sensitivity, we performed an additional analysis by calculating the FSD statistics on the initial, finer-grained segmentation generated by SAM, before our final step of merging adjacent segments of the same class (an example of this initial stage is shown in Fig. 7c). The results, summarized in Table , demonstrate that the power-law coefficients change by 4 % to 8 %, depending on the surface type.

To gain more insight into the spatial heterogeneity within the MIZ, we fit the NSSD of the snow-covered segments to 10 km sized bins of distance to the sea-ice edge (in pack-ice direction).

In Fig. , the size power-law exponent β is shown as a function of the distance to the sea-ice edge for the different surface types. A linear trend is fitted only to the TI data, suggesting significance with a R2=0.76 and a p value less than 0.001. The increase in β from -1.6 to -1.3 reflects a physical characteristic of the MIZ. Closer to the sea-ice edge, a higher number of smaller segments is observed (more negative β) due to intensified floe breakup, whereas larger floes (less negative β) become more prevalent further into the MIZ, where ocean wave propagation is more attenuated .