We accessed Level-1b brightness temperature (TB) data (version 2 calibration) of the Global Change Observation Mission first Water (GCOM-W1) Advanced Microwave Scanning Radiometer 2 (AMSR2) directly from the Japan Aerospace Exploration Agency (JAXA) Global Portal System (https://gportal.jaxa.jp/gpr/, last access: 26 February 2021). For this study, we used the brightness temperatures at both vertical and horizontal polarizations of the Ka (36.5 GHz) and W (89 GHz) imagery channels. Table 1 summarizes the spatial resolution of the microwave imagery channels of the AMSR2 and CIMR missions.

We use AMSR2 data from two periods: from 1 October 2019 to 31 December 2020 for the Northern Hemisphere (NH; 15 months) and from 1 June to 31 August 2016 for the Southern Hemisphere (3 months). These two periods are selected because they include winter freezing conditions in both regions and the summer melt season in the Northern Hemisphere, as well as to maximize the number of available on-ice drifters available for validation, as covered in the next section.

To validate satellite-based sea-ice drift vectors, we access GPS trajectories for on-ice drifters in the Arctic and Antarctic (Fig. 1). In the Arctic, buoys are in the Beaufort Gyre and in the Transpolar Drift. In the Antarctic, all buoys are in the central Weddell Sea. The colours in Fig. 1 represent the time of the observation records within the two periods.

A variety of buoy types are included in our validation data, but we are only concerned with three pieces of information per trajectory record: timestamp, latitude and longitude. Most buoys report positions on an hourly basis. The ice-tethered profiler data were collected and made available by the ice-tethered profiler programme (Toole et al., 2010; Krishfield et al., 2008) based at the Woods Hole Oceanographic Institution (http://www.whoi.edu/itp, last access: 1 June 2020). A variety of other buoys were accessed from the data portal http://seaiceportal.de (last access: 1 June 2021, Grosfeld et al., 2016), including all Antarctic buoys, and the buoys deployed at and around the Multidisciplinary drifting Observatory for the Study of Arctic Climate (MOSAiC) site (Krumpen et al., 2020; Nicolaus et al., 2021).

The sea-ice concentration (SIC) data from European Organization for the Exploitation of Meteorological Satellites (EUMETSAT) Ocean and Sea Ice Satellite Application Facility (OSI SAF) was accessed and transformed into an ice/no-ice mask using a threshold of 40 % SIC: grid cells with a SIC above 40 % are considered as ice-covered and can be used for the computation of sea-ice drift vectors. The Interim Climate Data Record “v2” based on Special Sensor Microwave Imager Sounder (SSMIS) data is used (OSI-430-b). Algorithms and processing chains are described in Lavergne et al. (2019).

The sea-ice motion tracking methodology implemented here, including the quality control steps, is very similar to that described by Lavergne et al. (2010) and implemented in the operational chains of the EUMETSAT OSI SAF (Lavergne et al., 2016). We recall below three unique features of this processing chain.

First, it implements the continuous maximum cross-correlation (CMCC) motion tracking algorithm. The CMCC stems from the well-known MCC (Ninnis et al., 1986; Emery et al., 1986) but implements a continuous optimization of the cross-correlation function (rather than a discrete optimization in MCC). The continuity is enabled by on-the-fly linear interpolation of image pixels. This continuous optimization strongly reduces the “quantization noise” present in many MCC-based sea-ice drift products. Lavergne et al. (2010) documented how the CMCC-based ice motion vectors were more accurate than those based on MCC (see also Hwang, 2013; Sumata et al., 2014).

Second, for a given microwave frequency, the information content of both the vertically and horizontally polarized images is combined within the optimization of the cross-correlation function. In practice, and following Lavergne et al. (2010), the solution sea-ice drift vectors are at the maximum of the sum of two cross-correlation functions: one from the vertically polarized imagery and one from the horizontally polarized imagery. The reader is referred to the discussion in Lavergne et al. (2010, Sect. 3.2) for a discussion of this approach. For the remainder of our paper, despite mentioning only the microwave frequency, we do use both polarizations in the motion tracking.

Third, it implements an iterative quality control step to detect and correct a few questionable (a.k.a. “rogue”) vectors. Those vectors are at the maximum of the cross-correlation function but point in a direction completely different from the neighbouring vectors. All block-based motion tracking algorithms need such a quality control step (e.g. Girard-Ardhuin and Ezraty, 2012; Haarpaintner, 2006; Tschudi et al., 2020), but most authors remove the rogue vectors, and the vector field has missing data cells. The quality control step of Lavergne et al. (2010) both detects the questionable vectors and – most of the time – corrects them, reducing the occurrence of data gaps.

As in Lavergne et al. (2010) and in the EUMETSAT OSI SAF sea-ice drift product (Lavergne et al., 2016), we process sea-ice drift vectors with a grid spacing of 62.5 km on two polar stereographic grids (Arctic and Antarctic).

The sea-ice drift product run at the EUMETSAT OSI SAF based on the algorithms of Lavergne et al. (2010) is processed from daily gridded maps of satellite signal: daily averaged maps of TB from the JAXA AMSR2 and United States Department of Defence (DoD) SSMIS, and daily averaged backscatter coefficients (corrected to 40∘ incidence angle) of the EUMETSAT Advanced SCATterometer (ASCAT). These daily averaged maps aggregate satellite signals from 00:00 to 23:59 UTC for a given day, and their valid time is 12:00 UTC. We name this type of sea-ice drift product a daily map product (noted DM).

It is noteworthy that, in addition to averaged maps of satellite signal from which DM sea-ice drift vectors are computed, we prepare the corresponding maps of daily averaged satellite sensing time: the average of the observation time from all the swaths during the 24 h period of the daily averaging (Fig. 2 in Lavergne et al., 2016). This mean sensing time is an important piece of information attached to each DM sea-ice drift vector: even though they are from daily maps with valid time 12:00 UTC, the DM vectors will be used and validated taking into account these space-varying mean start and stop times.

For this study, we prepared both DM products and swath-to-swath products (noted S2S). S2S sea-ice drift products are processed from swaths of satellite signals that are gridded individually, thus without daily averaging. For a given day, there are typically 12 to 14 such individual gridded orbits per satellite sensor, each having a different valid time separated by approximately 100 min. The sea-ice area at the overlap of two such individually gridded maps is where S2S drift vectors are computed (see Fig. 2). When preparing the individually gridded swaths of brightness temperatures, we also prepare the corresponding grids of sensing time for later use in the validation.

We prepare two periods of S2S and DM ice drift products from GCOM-W1 AMSR2 36.5 GHz TB imagery: from 1 October 2019 to 31 December 2020 for the Northern Hemisphere and from 1 June to 31 August 2016 for the Southern Hemisphere (see Sect. 2.1). In both periods, we process sea-ice drift vectors with durations (a.k.a. time spans) ranging from ∼ 100 min (the separation time between two consecutive orbits) to 48 h. The duration of an ice drift vector is the difference between the timestamps of the start and stop image from which motion is estimated (these images are either individual swaths for S2S products or daily maps of DM products).

A comparison between any two data sets first requires deciding on a collocation strategy in space and time. To collocate sea-ice motion vectors is somewhat different from collocating other geophysical variables in that the reference data source (a time series of GPS records from the ice drifting buoy or ship) is not directly comparable to what the satellite product measures (a net Lagrangian displacement vector over a time period). For this validation exercise, we follow the approach of Lavergne et al. (2010), Hwang (2013), and others to compute equivalent net Lagrangian displacement vectors from the in situ trajectories as part of the collocation step. In short, for each satellite vector to be validated, the in situ record closest to the start position of the satellite vector is selected, and – if within a suitable geographical radius – the in situ net Lagrangian displacement vector is computed from its GPS records using the start and end time of the satellite vector. We rely on a nearest neighbour approach in both space and time domains (nearest position, nearest time) without interpolation.

For a collocation pair to enter in the collocated matchup database, the start time of the in situ displacement has to be within plus or minus 3 h of the start time of the nearest satellite drift vector, the duration of the in situ vector has to be within plus or minus 1 h of that of the satellite vector, and the distance between the start positions of the in situ and satellite vectors has to be less than 30 km. To avoid over-representation in the case of buoy clustering (e.g. buoy arrays or the MOSAiC site), only the closest buoy to a satellite vector is kept. In addition, the collocated matchup database is filtered so that no directly neighbouring satellite vectors co-exist in it in order to reduce the effects of the correlation lengths stemming from the satellite retrieval.

For the S2S product, the start and end times of the drift vectors are the time stamp of the individual satellite swaths. For the DM products, they are the space-varying mean observation (overflight) times at the start and end images, meaning that the satellite product is not used as if starting and stopping at 12:00 UTC everywhere in the product grid.

The swath of CIMR (> 1900 km) will be larger than that of AMSR2 (1450 km). To study the impact of a wider swath on the characteristics of a future S2S sea-ice drift product from CIMR, we simulate some of its orbits and swaths. CIMR is to fly along a sun-synchronous dawn–dusk orbit with a 98.7∘ inclination. Additional orbit and instrument parameters relevant to the simulations are in Donlon et al. (2020).

We simulate 2 consecutive days of CIMR orbit and swath coverage using an ad hoc tool relying on the Earth Observation Mission Customer Furnished Item Software (EO CFI SW) libraries (http://eop-cfi.esa.int/index.php/mission-cfi-software/eocfi-software, last access: 23 October 2020). In practice, the orbit propagation is performed using the mean Keplerian orbit propagation mode available in the orbit library. The spacecraft attitude is modelled by applying local normal pointing and yaw steering law, and the instrument swath edges are defined in terms of directional look angles. Finally the zone visibility functions were used to compute the coverage mask of the instrument swath over the Northern Hemisphere and Southern Hemisphere grids separately. The visibility library includes functions to calculate the intersection of the swath points with the Earth ellipsoid, which are invoked internally by the zone visibility routine.

The coverage mask (binary map) of an individual swath is then used to mask a daily averaged map of AMSR2 TB (brightness temperature) to create a simulated CIMR TB swath that enters the sea-ice motion software. Our simulated CIMR swaths are thus only for studying the impact of the wider swaths and not the better spatial resolution.

Given the full daily coverage of GCOM-W1 AMSR2 (except the observation gap at the pole in the Arctic), the number of potential DM sea-ice drift vectors is a function of the sea-ice extent only. Conversely, the number of potential S2S sea-ice drift vectors varies with sea-ice extent and with the area of overlap between the individual swaths. This area of overlap depends on the relative orientation of two swaths and thus on the time difference between the orbits. Consecutive orbits will allow for a larger overlap than, for example, orbits that are 4–8 h apart. Due to the orbit configuration of the GCOM-W1 mission, two orbits that are roughly 24 h apart will have large overlaps.

Figure 3 shows the spatial distribution of the number of S2S drift vectors that start and/or stop in the period of 1–2 December 2019 in the Arctic (left) and in the period of 15–16 August 2019 in the Antarctic (right). The latitude dependency is clearly visible in both the Northern Hemisphere (NH) and Southern Hemisphere (SH) grids (the parallels at ±75∘, ±60∘ are drawn). For the AMSR2 instrument, the regions poleward of ±75∘ correspond to the areas with most inter-swath overlap. In total 110 441 S2S ice drift vectors are computed in the period of 1–2 December 2019 in the NH, while only 1729 DM vectors are available from the same 2 d. Similarly, 74 799 S2S ice drift vectors are computed in the period of 15–16 August 2019 in the SH, but only 3617 DM vectors are available from the same 2 d. The sea-ice extent in the NH on 1 December 2019 was 10.8 million km2, and it was 17.9 million km2 in the SH on 15 August (according to the OSI SAF “v2” SIC data; see Sect. 2.3). This difference is reflected well in the larger number of DM vectors computed in the SH than in the NH. Conversely, the higher number of S2S vectors in the NH than in the SH is controlled by the different latitudinal distribution of the sea-ice cover as illustrated in Fig. 3. In the NH, the majority of S2S vectors are computed in a band between 75 and 80∘ N (up to 90 S2S drift vectors per grid cell), and their number slowly decays towards 85∘ N with spatial patterns that are typical of satellite swath geometry. The sharp decay poleward of 85∘ N is due to the polar observation hole of the AMSR2 instrument with a width of ∼ 0.5∘ (see Fig. 2, left panel). Two factors amplify this decay. First, there is less overlap between swaths at very high latitude (edges of the swaths). Second, the motion tracking algorithm works with sub-windows (e.g. 11×11 image pixels), and this limits the retrieval of drift vectors in the immediate vicinity of areas with missing image data (e.g. the polar observation hole, sea-ice edge, coastal region, etc.). In fact not a single S2S vector (nor DM vector) is retrieved north of 88.5∘ N from the GCOM-W1 AMSR2 data. The transition from ∼ 90 S2S drift vectors per grid cell to 40 and fewer equatorward is visible on both the NH and SH maps. In the SH, most of the sea-ice cover is south of 75∘ S, which is the case all year round. Still, even at the outskirts of the SH sea-ice cover, we have 5–15 S2S vectors – compared to a single DM vector – per grid cell. The numbers above pertain to the configuration in which drift vectors are processed at a grid spacing of 62.5 km: reducing the grid spacing would directly increase the number of DM and S2S vectors.

The first advantage of adopting an S2S ice drift processing for microwave radiometer satellite data is thus to take full advantage of the individual swaths and get access to many more sea-ice drift vectors than in the DM configuration.

These S2S sea-ice drift vectors however have very different characteristics compared to the DM products, especially in the time and duration domain. The histograms in Fig. 4 show the distribution of the S2S and DM vectors from the period of 1–2 December 2019 (NH, top row) and 15–16 August 2019 (SH, bottom row) in terms of start time of the drift (left), end time of the drift (middle) and duration (right). We first note that the distributions of DM vectors are not a single value at 12 h (start), 36 h (stop) and 24 h (drift duration) but that we report some variation around those values. As introduced earlier, the start and stop times of DM vectors are the space-varying mean observation time of all the swaths during the 24 h period of the daily averaging. Despite the variations, the distributions of start time, stop time and duration are concentrated around expected peak values at 12, 36 and 24 h, respectively. Noticeably, the duration of most DM vectors falls within plus or minus 2 h of 24 h. Nevertheless, these mean times associated with the DM ice drift vectors are values averaged over several overlapping swaths, and they sometimes do not provide a faithful representation of the time characteristics of the ice drift product, e.g. when a region of the sea-ice cover is observed twice in the early morning and once in late afternoon.

Conversely, the temporal information attached to S2S drift vectors is much more accurate and has significantly different distributions in Fig. 4 when compared to those of DMs. The distribution of start time (left) and stop time (middle) covers the whole range in the 48 h period, with peaks around 0, 18 and 40 h past 1 December at 00:00 UTC (start time) and 25 and 48 h past 1 December at 00:00 UTC (stop time). These peaks correspond to an increased number of S2S drift vectors being available at the end of the 48 h period and is thus a function of many parameters including the orbit and swath characteristics of the satellite mission (in our case GCOM-W1 AMSR2) and the extent and geographic repartition of the sea-ice cover, both of which combine into overlap characteristics. The S2S histograms in Fig. 4 should thus be seen as an illustration of a general pattern (repartition over the whole period, with peaks). The S2S duration histograms (Fig. 4, right panels) are also controlled by the characteristics of the orbit, swath width and the sea-ice cover. They document that a wide spectrum of drift durations are recorded by S2S drift products, with peaks near 0, 24 and 48 h in the NH and more spread in the SH. These peaks generally correspond to when the swaths overlap most.

All in all, the short analysis conducted here documents that S2S and DM ice-drift products have distinct characteristics. S2S ice-drift products offer many more vectors, and this number varies with latitude. The two types of products cover the temporal domain differently. The S2S products have a broad but not homogeneous coverage in start times, stop times and duration. Our next step is to investigate if the accuracy of the S2S ice-drift product is better, similar or worse than that of DM ice-drift products. We assess this accuracy against collocated in situ drifter GPS trajectories. Results are reported in the next section.

We collocate all S2S and DM sea-ice drift vectors with on-ice drifters trajectories (Sect. 3.3) for both the Northern Hemisphere and Southern Hemisphere (Sect. 3.2). Here we analyse the statistics from this validation.

Table 2 summarizes the validation results of all DM and S2S sea-ice drift vectors with a duration of 24 ± 42 , thus the majority of DM vectors and a subset of all S2S vectors (right panels in Fig. 4). The selected DM and S2S drift vectors thus correspond roughly to the same drift duration. In both the Northern Hemisphere and Southern Hemisphere, S2S vectors result in better validation statistics than DM vectors. The reduction in the RMSE values (σ) is significant, e.g. a reduction by ∼ 30 % in the NH and SH when adopting an S2S algorithm. The bias remains very small in the NH and is reduced in the SH. As expected, the number of matchup samples N is larger for S2S validation, and the increase with respect to DMs is larger in the NH than in the SH because most on-ice buoys are from the central Arctic Ocean with many swath-to-swath overlaps (Fig. 2).

The validation statistics are somewhat worse in the SH than in the NH. This is probably due to the lower number of validation data points and the fact that some are on the outskirts of the sea-ice cover in dynamical regimes that are challenging for motion tracking from coarse-resolution radiometry imagery (Fig. 1).

These numbers mean that a typical 24 h drift displacement vector can be measured with an uncertainty of typically 1 km in each component by an S2S product even though the original imagery is at a rather coarse resolution (the -3 dB footprint of the AMSR2 36.5 GHz channels is 7×12 km; see Table 1). As expected, the uncertainty of DM sea-ice drift vectors is larger (∼ 30 % increase). In the next section, we repeat the validation experiment but this time we consider the full range of durations, from ∼ 100 min to 52 h.

Figure 5 documents the evolution of the validation statistics (bias and RMSE) with the drift duration for the same S2S and DM drift vectors based on GCOM-W1 AMSR2 Ka-band imagery. The bias (μ) and RMSE (σ) in dX and dY components of the S2S vectors are plotted for drift durations ranging from ∼ 100 min (drift computed from two consecutive swaths) to 52 h. Bias and RMSE of DM drift vectors are plotted for 24 and 48 h drift durations separately. The values reported in Table 2 correspond to the conditions around 24 h.

Figure 5 confirms that the RMSE obtained for S2S drift vectors is smaller than that of 24 and 48 h DM vectors, as well as at intermediate drift durations below 52 h. The RMSE of S2S vectors increases regularly with the drift duration, from ∼ 0.5 km for drift duration of ∼ 100 min to ∼ 2 km for drift duration of 52 h. In the NH, the bias stays small for the whole range of drift durations and is smallest around the 12, 24, 36, and 48 h marks, a repeat cycle that we will study in more details in the next section. In the SH, the same cycle is observed but with a stronger amplitude. As already reported in Table 2, bias and RMSE of both S2S and DM vectors are larger in the SH. In any case, we confirm that the S2S drift processing can be advantageous for passive microwave radiometry missions (with respect to the current state of the art which is all based on DM processing) as S2S brings many more drift vectors (Sect. 4.1), and those drift vectors are more accurate (this section).

The two last sections focused on two 3-month winter periods in the Arctic and Antarctic. Here, we present monthly validation results covering October 2019 to December 2020 (15 months) in the Arctic. Our main objective is to investigate if the S2S approach helps the retrieval of sea-ice drift vectors during the Arctic summer melt season. Due to surface melt and increased wetness in the atmosphere, the tracking of sea-ice drift from passive microwave instruments has traditionally been a challenge during summer. While Kwok (2008) has shown that imagery from the AMSR2 mission can be used to track summer sea-ice drift (using a DM approach), the accuracy when compared to buoy trajectories was shown to be much reduced.

Figure 6 shows monthly validation statistics for several DM and S2S products obtained from the AMSR2 36.5 GHz imagery. The 48, 24 and 18 h drift products were prepared and validated following Sect. 3.3. Figure 6 confirms that the validation statistics of drift vectors with shorter durations (e.g. 18 and 24 h) are better than those of vectors with longer durations (48 h), both in terms of RMSE and bias and for the whole 15 month period. This was already noted in Sect. 4.3 for the period of October–December 2019. Figure 6 also confirms that, for most of the year, S2S drift vectors reach better validation statistics than DM vectors. This is true for all the winter months (October–April). However, the better accuracy of S2S drift vectors is not apparent during the summer months (May–September) when DMs achieve (slightly) better results. Validation results during summer are indeed worse than during winter, but the main driver for the worsened accuracy in summer seems to be the duration of the drift vectors (24 h vs. 48 h) and not the adoption of an S2S vs. a DM approach.

During summer in the Arctic, the atmosphere is wetter and contributes significantly to the brightness temperature recorded at 36.5 GHz, effectively hiding more of the surface emissivity. The surface emissivity is also more variable in time because of the sub-daily cycles of melting and freezing (early and late in the summer season) and the direct impact of weather systems travelling over the sea ice. It is thus not a surprise to see better validation statistics with shorter than longer drift durations since a shorter duration will increase the chance of tracking the same surface emissivity patterns with less chances for a change happening in the time between the two images.

When conducting the same investigations with the 18.7 GHz imagery of AMSR2 (not shown) we found roughly the same results, but the validation statistics were slightly worse than those obtained with 36.5 GHz throughout the year. The 18.7 GHz microwave frequency is emitted from deeper in the sea ice and snow medium and is less affected by the atmosphere so that one would expect more stable surface emissivity patterns available for sea-ice motion tracking (Kwok, 2008). However, the coarser resolution of the 18.7 GHz frequency channels (Table 1) works against this property by blurring the emissivity patterns.

We note that, even if DM vectors seem to validate better than S2S vectors during the summer melt season, adopting the S2S approach still gives many more vectors per day than the DM approach.

Geolocation accuracy of the satellite images from which drift vectors are computed is a key component of the uncertainty budget. Even high-resolution satellite images (such as those from SAR) will give poor sea-ice drift vectors if the images are not correctly geolocated.

To simulate the impact of geolocation accuracy on the S2S drift vectors, we purposely misregister the GCOM-W1 AMSR2 36.5 GHz imagery to the locations of the 18.7 GHz (Ku-band) imagery. Because of how the Ku- and Ka-band radiometer feeds are arranged in the focal plane of the AMSR2 instrument, they do not exactly point at the same locations on Earth during the rotation of the reflector antenna (Maeda et al., 2016). The Ku-band fields of view (FoVs) are systematically further ahead in the flight direction with respect to the Ka-band FoVs. The offset between the positions ranges from 750 m to 1 km in the flight direction, with a mean value of 917 m. For comparison, the spacing between successive scans of both microwave frequency channels is 10 km, and the spatial resolution of the Ku-band FoVs is 18 km (Table 1). The geolocation error introduced is thus of an order of magnitude less than both the spacing and resolution of the imagery channel.

We repeat the whole S2S and DM processing but this time by remapping the Ka-band imagery misregistered to the Ku-band locations. A visual analysis of the new sea-ice drift maps does not identify obvious problems with the new vectors (not shown).

Figure 7 is a repeat of Fig. 5a but with the Ka-band imagery using the geolocation of the Ku-band FoVs. Comparing Fig. 7 to Fig. 5a demonstrates the impact of geolocation errors on the accuracy of S2S and DM drift vectors. The accuracy of DM vectors is only marginally impacted by the geolocation error. Their bias is still very close to zero, and their RMSE is similar to those reported in Fig. 5 (around 1.5 km for 24 h drift duration and 1.9 km for 48 h). Conversely, the accuracy of S2S drift vectors is strongly impacted. Especially the bias of the dX (respectively dY) component of the vectors varies from -1.5 to +1.5 km (respectively -0.5 to +0.5 km), with sharp transitions between positive and negative values around the 12 and 36 h drift durations. Away from these transitions, i.e. for drift durations of 0, 24 and 48 h, the bias is much smaller and close to zero, as was the case in Fig. 5a. The RMSE of S2S vectors (the width of the shaded areas) is similar to that obtained in Fig. 5a across the full range of drift durations.

As documented in Fig. 7, the impact of geolocation error on the bias of the S2S drift vectors has a clear dependency on the duration of the drift vectors, with a repeat period of 24 h. This pattern is explained by the angle formed between the flight directions of the two orbits from which the S2S drift vectors are computed.

Geolocation accuracy is of importance for all satellite-based products but will have an exacerbated impact on sea-ice drift products. Indeed, if both satellite images from which the drift vectors are computed have a geolocation error, and the two geolocation errors are in opposite directions, the drift vector will be strongly affected. Figure 8 gives a schematic view of the impact of a constant geolocation error in the flight direction (which is very close to what we simulated in Fig. 7).

In Fig. 8, we illustrate how the relative angle between two swaths sustaining the computation of an S2S vector has an impact on the magnitude of the retrieval error in the presence of a constant geolocation offset along the flight direction. The four panels correspond to four different relative angles. When the two swaths are close to parallel (θr≈0), the geolocation error does not result in significant drift errors. This corresponds to time separation between the AMSR2 swaths of ∼ 100 min, ∼ 24 h and ∼ 48 h. When the two swaths are in opposite flight directions (θr≈180), the errors on the drift components are maximum, which is the case for time separations of ∼ 12 and ∼ 36 h. Other relative angles (two lower panels) make intermediate contributions to the errors. There is a direct link between the situations illustrated in Fig. 8 and the variation in the biases with drift durations in Fig. 7 in the case of a strong geolocation offset in the flight direction. We already noted a similar cycle of the biases in the NH and SH validation exercises in Sect. 4.2 (Fig. 5) which might result from a small residual geolocation error of the AMSR2 36.5 GHz imagery in the flight direction. In Sect. 5, we will discuss what the implications are for the retrieval of S2S drift vectors from the CIMR mission.

Before closing this section on geolocation errors, we discuss the fact that the DM drift vectors seem much less affected by the artificial geolocation error introduced for Fig. 7 than the S2S vectors are (compare the error-bar symbols of Figs. 5 and 7 and especially that the bias is zero in both). When building daily composite maps of averaged TB, the impact of the geolocation is smeared but not cancelled. Indeed, and especially at lower latitudes, locations on the sea-ice cover are often observed by sequences of subsequent orbits, which means a constant geolocation error from each orbit will still result in a non-zero mean geolocation error in the daily averaged map. Nevertheless, the bias in drift components of DM vectors is close to zero because of the repeat cycle of the observations of (in our case AMSR2) orbits: a location on the sea-ice cover will be observed on average roughly at the same time of the day but 1 or 2 d apart. This repeat cycle of the satellite orbit ensures that the non-zero mean geolocation error is similar at all locations in the daily maps and thus that the impact on the DM drift vectors is small (θr≈0).

All in all, we conclude from this section that S2S drift vectors are more sensitive to geolocation errors than DM drift vectors are. The retrieval of accurate sea-ice drift vectors from individual swaths puts stringent requirements on the geolocation accuracy of passive microwave missions if all swath overlap pairs should be processed. Even seemingly limited geolocation errors such as those simulated in this section will result in inaccurate (biased) estimates when orbits overlap in opposite flight directions.

The key principle of motion tracking algorithms (including those for sea-ice drift) is to identify local intensity patterns on one image and track it on another image. The accuracy of the motion vectors will thus be better if the satellite images offer intensity patterns that are clearly defined and stable with time. The sharpness and stability of satellite microwave images of the sea-ice surface depend both on the spatial resolution achieved by the FoV and the microwave frequency and polarization of the imagery channel being processed. Indeed, channels (or instruments) with coarser spatial resolution will provide blurred images of the microwave emissions, with fewer or weaker intensity patterns to track. Channels with higher sensitivity to stable characteristics of the surface (e.g. sea-ice type, snow depth, etc.) and/or lower sensitivity to the varying atmosphere above the sea ice (e.g. cloud liquid water path, air temperature) will offer sharper and more stable intensity patterns to track.

We investigate the impact of resolution and microwave frequency of the imagery by running an experiment where GCOM-W1 AMSR2 Ka (36.5 GHz) and W (89.0 GHz) imagery channels are first resampled on a pan-Arctic grid with spacing of 3.125 km. Such a fine spacing is an oversampling of the true resolution of both microwave channels (see Table 1). The remapping of each channel is done independently and accounts for the width of the FoV. Swaths are resampled separately, and S2S drift vectors are computed and collocated against buoy data, and validation statistics (bias and RMSE) are extracted for the same 3-month period as in Sect. 4.1. Only the ∼ 24 h S2S vectors are computed. This experiment is then repeated for purposely coarsened versions of the images. Bi-linear coarsening kernels at factors of 2 (6.25 km), 3 (9.375 km), 4 (12.5 km), 5 (15.625 km) and 6 (18.75 km) are applied to the original 3.125 km grid spacing images. S2S drift vectors with durations of ∼ 24 h are computed and collocated to buoys, and their validation statistics are extracted.

Figure 9 plots the evolution of the validation statistics (bias and RMSE) of the AMSR2 NH ∼ 24 h S2S drift vectors with the width of the coarsening kernel. The best image resolution is to the right of the x axis.

As expected, the RMSE from both 36.5 and 89 GHz imagery drops with finer spatial resolution (from left to right in the plot) and the bias is reduced towards zero. This behaviour stalls at the 6.25 km kernel, and no further improvement of RMSE is observed when processing the full-resolution 3.125 km images. In fact a slight worsening can be observed for the 36.5 GHz channels.

Thanks to their higher microwave frequency, the GCOM-W1 AMSR2 89 GHz imagery channels have a distinctly better resolution than the 36.5 GHz channels (Table 1). However, this finer resolution does not seem to translate into a much better accuracy of the S2S drift vectors as the RMSEs and biases reported in Fig. 9 are very similar for both microwave frequencies. Here we can hypothesize that the finer spatial resolution of the 89 GHz channels is balanced by the relative lack of strength (and/or temporal stability) of intensity patterns to track over a ∼ 24 h period. Notably, the 36.5 GHz imagery is emitted from deeper in the ice and snow media, is more sensitive to sea-ice type and snow depth (rather stable in time), and is less sensitive to the atmosphere above the ice. The 89 GHz is emitted closer to the top of the ice and snow media and is thus more sensitive to temperature changes. It is also much more sensitive to liquid water path in the atmosphere that can easily blur or hide intensity patterns at the surface. We note that the lack of improvement of the RMSE metric from 6.25 to 3.125 km could also be due to our rather simple approach to resampling the 89 GHz channels to the 3.125 km grid. However, more advanced gridding techniques (e.g. Backus–Gilbert) could also be challenged by the lack of sufficient overlap between neighbouring 89 GHz fields of view. The slight increase in RMSE seen for the 36.5 GHz channels from 6.125 to 3.125 km also indicates that overly oversampling the native spatial resolution of the imagery channels is not beneficial and that it seems more efficient for sea-ice drift retrievals to aim at an optimum resampling resolution for each channel rather than the finest possible spacing.

Still, the spatial resolution of the images does not only drive the accuracy of the drift vectors but also the overall resolution of the vector field. Indeed, a finer image will allow motion tracking algorithms to be run with smaller image sub-windows and thus allow for the creation of a denser vector field. For example, the finer resolution of the AMSR2 89 GHz channels allows for tracking smaller areas down to 70× 70 km and processing denser vector fields (e.g. 31.25 km; Ezraty et al., 2007). Increasing the grid spacing to 31.25 km was attempted in our study but did not result in significantly different results for either the 36.5 GHz or 89 GHz imagery (not shown). In fact, the result was a slightly larger RMSE for both frequency channels and at all coarsening kernels. This might be because tracking smaller image sub-windows leads to tracking fewer intensity patterns per vector. As is true for the retrieval of many other geophysical variables, increasing the spatial resolution often leads to increases in the noise level. The increased RMSEs might also reflect that most of our validation data are from the central Arctic Ocean and therefore mostly sample drift events that are coherent over large spatial scales. The higher-resolution drift vectors might be more accurate in cases when sharp spatial gradients of the drift field are observed, either in the event of sea-ice deformation or in places with stable velocity gradients such as in the East Greenland Sea. However, these conditions are not well enough represented in our validation data set to have a positive impact on the overall statistics.

The wider swath of CIMR with respect to AMSR2 (and to all other conically scanning radiometer missions) should result in larger areas of overlap between individual swaths and thus in more S2S sea-ice drift vectors. A larger number of S2S vectors should be observed towards lower latitudes (needed to better sample the core of the sea-ice cover in the Southern Hemisphere) and higher latitudes (needed to monitor the sea-ice cover close to the North Pole).

We simulate 48 h of CIMR orbit and swath coverage (Sect. 3.4) and enter these swaths in our S2S sea-ice drift processing chain. We repeat the analysis of spatial distribution presented in Sect. 4.1 with AMSR2 data.

Figure 10 plots the spatial distribution of the number of S2S vectors obtained from the simulated CIMR orbit data. To ease comparison with AMSR2, we use the same sea-ice cover as in Fig. 3 (1–2 December 2019 in the NH, 15–16 August 2019 in the SH). The total number of S2S vectors for the Arctic coverage is 143 101 for CIMR (compared to 110 400 for AMSR2 and 1729 for the DM product). Over Antarctic sea ice, the number is 108 601 for CIMR (compared to 74 737 for AMSR2 and 3617 for the DM product). The wider swath of CIMR thus results in 30 %–40 % more S2S vectors compared to AMSR2.

When comparing Fig. 10 (CIMR) to Fig. 3 (AMSR2) in the Northern Hemisphere, a striking improvement is seen towards the North Pole. In Sect. 4.1 we explained how the polar observation hole of the AMSR2 imagery, despite being of only 0.5∘ latitude, has a wider impact on the coverage of S2S vectors near the pole. The CIMR orbit and swath width are specifically optimized to allow for full sub-daily coverage of the polar regions including the poles and thus “no hole at the pole” (Donlon et al., 2020). Our analysis shows that the number of S2S vectors from CIMR reduces towards the pole but that there are vectors all the way to the pole on a daily basis.

We underline that the actual number of S2S vectors from CIMR will be larger than discussed here simply because the increased spatial resolution of the Ku- and Ka-band imagery (Table 1) will allow for a denser vector field (target 25 km grid spacing; Donlon et al., 2020) than we processed (62.5 km grid spacing).