Evaluation of the AMSR2 Ice Extent at the Arctic Sea Ice Edge Using an SAR-Based Ice Extent Product

Passive microwave (PM) and synthetic aperture radar (SAR) observations are essential tools for providing long time series of sea-ice cover information, including sea-ice concentration (SIC) and sea-ice extent (SIE). Large uncertainties have been revealed in PM SIC/SIE products in the marginal ice zone (MIZ) and during the melting season, where fusion with SAR data could be effective for improving accuracy due to its high spatial resolution and ability to preserve detailed ice distributions. A comprehensive comparison of PM and SAR ice cover products is needed for better data fusion. This study evaluates one of the PM SIE products, the advanced microwave scanning radiometer 2 (AMSR2) SIE product retrieved with the arctic radiation and turbulence interaction study (ARTIST) sea ice (ASI) algorithm, using a neural-network-based SAR SIE product throughout the year 2019. First, we present key results of three assessment parameters, including the overall accuracy (OA), error-of-ice (EI), and ice edge location distance (LD), and then estimate the optimal SIC segmentation threshold for AMSR2 ASI SIE. Based on OA and EI, the annual average SIC threshold of 12.24%, winter average of 9.25%, and summer average of 16.43% are obtained and regarded as optimal by excluding cases with large uncertainties. Second, the AMSR2 ASI SIE product is found to perform better in identifying thin ice and melt ponds, while the SAR NN SIE product has better detection of brash ice and frazil ice. We introduce a parameter of sea-ice fragmentation fraction (IFF) to analyze the primary impact factors behind the different performances. It is found that the ratio of LD to IFF could distinguish the aforementioned different ice conditions, thus providing hints for combining the complementary advantages of the two SIE products during data fusion.

changes in terms of area, heat flux, wave energy, climate prediction, and ship navigation. SIC represents the grid-wise fraction covered by sea ice, whereas SIE indicates the presence or absence of ice within each grid/pixel in this article. Integrated SIE rather represents the total grid/pixel area covered by sea ice in a given image. At present, various algorithms and products based on satellite observations have been developed and released for monitoring ice cover [1], [2], [3]. Passive and active microwave measurements (mainly radiometer and scatterometer data) have been widely used for estimating SIC due to their long time series and large spatial coverage. Meanwhile, high-resolution synthetic aperture radar (SAR) images are increasingly used for deriving SIE with different ice classification algorithms, and as one of the multisource data for expert interpretation of sea-ice charting [2], [4].
Overall, the accuracy assessments of passive microwave (PM) SIC algorithms or products have been evaluated based on the average bias for SIC of 0%-100% and that for low SIC (e.g., near-0% or below-15%) and high SIC (e.g., near-100% or over-85%) [5], [6], [7], [8], [9], [10]. Various error sources have been investigated and discussed, including the emissivity and physical temperature of sea ice, atmospheric effects, melt ponds, and thin ice [5]. The accuracies of SIC algorithms under different net ice surface fractions (ISFs, i.e., fraction of the ice surface excluding melt pond and open water) [6], [9], the impacts on truncated 100% SIC and the performances of weather filters [7], [8] were also studied. Studies have shown that the uncertainties of PM SICs tend to be larger near ice edges and during summer melting periods [11], [12], [13], [14], [15], which is up to 20%-40% [6]. This is mainly due to the limitation of algorithm and the variable radiometric characteristics of sea ice under such conditions. For instance, the near-90-GHz brightness temperatures are very sensitive to atmospheric factors such as cloud liquid water and integrated water vapor, which is much more pronounced over ice regions with the distributed melting water features [5]. To address these challenges, algorithm upgrades and the fusion of additional satellite data have been proposed as potential solutions for improving accuracy [3], [6], [16].
With sufficiently high resolution and the capability of working all-weather all-day, SAR observations have been preferred for ice-water classification. Over the past few decades, various algorithms have been developed for SAR-based SIE retrieval, as reviewed in [2]. However, most of these algorithms have been proposed in regional experiments [4], [17], [18], [19]. This  There is not a long time record of dual-pol SAR that would be needed for long time series of Arctic ice monitoring using SAR sensors (in comparison to PM). The limited coverage/availability of SAR observations, as well as the large backscatter variability of ice surface characteristics under different ice types, making it challenging to apply these algorithms globally [2], [20].
To combine the advantages of the high spatial resolution of SAR and the pan-Arctic coverage of PM data, as well as their different sensitivity to the ice/water surface features, several studies have proposed approaches for SAR-PM SIC retrieval in their experiments [21], [22], [23]. Most of them use neural network algorithms to train and establish the SAR-PM-based SIC approaches [21], [22], [24]. Others aim to assimilate the SAR binary ice/water results with PM SIC [20]. Nevertheless, compared to the data fusion of PM and optical data [e.g., moderate resolution imaging spectroradiometer (MODIS)] that has been routinely used for producing SIC products (e.g., the 1-km advanced microwave scanning radiometer 2 (AMSR2) SIC product provided by the University of Bremen [25]), the fusion of PM and SAR data still needs further improvements.
Recently, a neural-network-based SAR ice/water classification product was released (hereafter called SAR NN SIE) with a 400-m spatial resolution, using 28 000 Sentinel-1 images of the Arctic in 2019 [26]. Such a large-scale and highresolution SAR SIE dataset has the potential to improve the accuracy of SIC estimates when combined with the PM data. However, a comprehensive comparison of the SAR SIE and PM SIC products is needed for effective data fusion. It remains unknown what are the pros and cons of the PM and SAR products, and how these products could be better combined for their complementary advantages, especially in the marginal ice zone (MIZ).
The daily AMSR2 ASI SIC products obtained from the University of Bremen are provided with spatial resolutions of 1 km (fusion with MODIS data, as mentioned), 3.125, and 6.25 km, with relatively high accuracies [27]. The daily 6.25-km AMSR2 ASI SIC product has been used for various purposes [26], [28], [29]. The evaluation of these products has been limited to short-term and regional studies due to the lack of independent validation data [13], [29], [30]. In this study, we aim to evaluate the daily 6.25-km AMSR2 ASI SIC product using the large-scale SAR NN SIE dataset in 400-m resolution provided by [26], with a particular focus on its performance in monitoring SIE in MIZ.
The thresholds for extracting SIE from PM SIC data differ with the retrieval algorithms and products due to their varying performances [15], [31], [32], [33], [34], [35], [36]. The thresholds also depend heavily on the ice edge characteristics that vary with different seasons and regions, e.g., whether it is compact or diffuse [1]. Some studies investigated the sensitivity of Arctic SIE to SIC threshold [37], [38]. For the annual Arctic SIE minimums, a change of the threshold from 15% to 35% can lead to an integrated SIE change of more than 10% in magnitude [38]. To determine the optimal threshold for the AMSR2 ASI SIE product and estimate the SIE uncertainty resulting from the commonly-used 15% threshold, we calculated its best-matched threshold with the SAR NN SIE data as a baseline.
The major work of this study is to evaluate the accuracy and determine the optimal ice concentration thresholds for the AMSR2 ASI SIE product by comparing it with the SAR NN SIE product along ice edges. Previous studies have shown that none of the PM SIC (SIE) algorithms can be uninfluenced by the weather effects and melted/wet ice conditions [9]. Likewise, the SAR SIE algorithms face challenges related to the identification of the ice and water with similar backscatter features in the MIZ, particularly in summer [2]. As such, this study not only presents the main results of evaluation parameters but also discusses the sensitivity of these two products to various error sources and summarizes their performances in different ice/water conditions across different periods and regions. Our ultimate goal is to explore the potential of data fusion for the AMSR2 and SAR SIE products or of SIC/SIE retrieval algorithms by blending AMSR2 and SAR (Sentinel-1) data to address the limitations of single PM or SAR-based algorithms.
II. DATA AMSR2 on the JAXA satellite GCOM-W1 has been one of the representative PM sensors for providing operational sea ice observations since 2012. It is a follow-on sensor of AMSR-E onboard the NASA Aqua satellite. The daily AMSR2 (and AMSR-E) SIC product accessible at the University of Bremen (https://seaice.uni-bremen.de/data/, last access: 11 February 2023) is based on the ASI algorithm [39]. The ASI algorithm mainly uses the brightness temperatures of 89-GHz channel at both horizontal (H) and vertical (V) polarization to differentiate sea ice and open water. The 89-GHz channel has a smaller footprint size (6 × 4 km 2 ) than lower-frequency channels and thus has a higher resolution (gridded 6.25 km). However, it is more susceptible to atmosphere factors, particularly, integrated water vapor and cloud liquid water. A bulk correction for atmospheric opacity, the so-called weather filter, is applied during the SIC production (https:// seaice.uni-bremen.de/fileadmin/user_upload/ASIuserguide.pdf, last access: 11 February 2023). The AMSR2 ASI SIC product (Arctic, version 5.4) is available from 2012 to the present [40], but only the data of 2019 was used in this study considering the availability of the SAR ice/water classification dataset.
The SAR ice/water classification dataset used in this study is produced based on Sentinel-1 Extra Wide (EW) mode dual-polarization (HH and HV) data using a deep learning architecture of U-net neural network [26]. It provides maximum time and spatial coverages for Sentinel-1 derived sea-ice observation over the Arctic region (72.5 • N-83 • N) in 2019, while keeping the original SAR imagery size of around 400-km width with a resampled pixel resolution of 400 m. An absolute mean difference of 5.55% was obtained from the comparison between AMSR2 and SAR SIC, while an overall accuracy (OA) of 93.98% was obtained from the comparison between the Interactive Multisensor Snow and Ice Mapping System (IMS) and the SAR NN SIE in the whole Arctic Ocean [26].

III. METHODS
The procedure described here begins with image selection and data processing of the SAR NN SIE product. It then proceeds with data processing and parameter calculation of AMSR2 ASI SIE and ends with detailed analyses on the evaluation results in Section IV. Melt ponds are very likely to be recognized as open water due to the low SAR backscattering value. The SAR SIE product is thus essentially close to the SAR ISF product, whereas the AMSR2 ASI SIE product could keep better recognition on the ice-covered areas with high melt pond fractions despite the possibility of underestimated SIC [6]. Such difference in melt pond recognition can cause additive errors regardless of the true ice edge locations. Therefore, for the SAR ice/water classification product, details such as melt ponds, narrow ice leads, and broken ice floes were largely removed to extract the major outer ice edges. The Pan-Arctic AMSR2 SIC data were then clipped and further processed based on the scenes that are consistent with the SAR NN SIE product. Overall, the basis of data processing includes: 1) focusing on SIE difference along the main sea-ice packs (see Sections III-A and III-B) and 2) considering the different ice distributions within each SAR image (around 400-km width), which motivates the introduction of new quantitative evaluation parameters to show the error fractions over sea-ice pixels instead of all pixels of individual SAR images (see Sections III-C-III-E). Fig. 1 shows the AMSR2 ASI SIC product and the mosaic SAR NN SIE images acquired on 24 July 2019 (during the peak melt season). Compared to the ice edges from AMSR2 ASI SIC (black lines), SAR NN SIE shows superior details in depicting ice edges due to its high resolution; however, it has limitations on spatial coverage. In general, visual inspection indicates that ice edge differences between these two products are largest in the Bering Sea, East Siberian Sea, and East Greenland Sea, with differences spanning dozens to hundreds of kilometers width. Fig. 2 illustrates the major data processing on an individual image. During the processing, we use "grids (6.25 × 6.25 km)" to describe AMSR2 ASI SIC and "pixels (400 × 400 m)" to describe SAR NN SIE. Details are provided below.

A. SAR NN SIE Data Processing
As described earlier, the processing of SAR NN SIE data involves two aspects: 1) image selection and 2) filtering and smoothing of sea-ice edges. For image selection, only images with an ISF of 5%-95% were retained, which indicates the presence of ice edges. Quality screening was then performed to eliminate the SAR NN SIE images with large errors caused by the thermal noise removal processing used in [26]. These errors usually appear as "straight lines at the ice edge" with a certain length (as shown in Fig. 11). In this study, images with straight edge lines longer than 300 pixels (120 km) were excluded. In total, 6972 SAR NN SIE images were used for comparison.
For filtering and smoothing of sea-ice edges, two steps were taken. First, based on the primary SAR NN SIE product [see Here, any filled region encircled by a closed contour of sea ice pixels is identified as a polygon. As shown by the black lines in Figs. 1(b) and 2(f), where three types of polygons: large ice packs, small water regions (e.g., ice leads) within large ice packs, and small ice floes beyond the ice packs can be extracted. Pixel-based ratio of area to perimeter (R a2p ) describes the size and shape of each polygon. It is dimensionless. A larger value of R a2p means a smaller flattening of one polygon or a larger total pixel number of one polygon. The authors started with a water grid of 6.25 × 6.25 km surrounded by pack ice in the AMSR2 ASI SIE image, which is equivalent to about 16 × 16 pixels of 400 × 400 m in the SAR NN SIE product. Assuming the 16 × 16 pixels to be a square, R a2p has the largest value of 4.0. Otherwise, R a2p keeps definitely below 4.0 provided that the polygon width is less than 16 pixels regardless of the length of its long side. To remove the small water polygons or ice floe polygons limited in the size of about one AMSR2 grid width, the R a2p threshold of 4.0 was determined.

B. AMSR2 ASI SIE Product Processing
The processing procedure for AMSR2 ASI SIE includes four major steps: 1) clipping the pan-Arctic AMSR2 ASI SIC product according to the SAR image; 2) removing the spurious SIC around the land; 3) resampling the clipped image to the grid spacing of 400 m using the nearest neighbor method; and 4) extracting the AMSR2 ASI ice extent where the ice concentration is higher than the threshold (e.g., 15%, 0%, and the newly calculated) and thereafter detecting the AMSR2 ASI ice edge.
Due to the mixed pixels of open water and land (with similar brightness temperatures to sea ice) within the grid of several kilometers, spurious SICs can be produced along coasts during the retrieval from radiometer data [41]. Although the AMSR2 Comparing (e) and (f) to (d), the variations of AMSR2 ASI SIE error-pixel numbers (thin lines) and their polynomial fits (red and blue thick lines) to SIC thresholds are plotted in (g). Thresholds of 22% and 4% are separately selected at the minima. SAR image ID: S1A_EW_GRDM_1SDH_20190104T062302_20190104T062406_025324_02CD5A_AB31.
ASI SIC product has excluded the grids of land fractions with some degree of expansion by using the GMT5 land mask and coastlines, the spurious SIC around land still remains (see the regions of around 10 • E to 15 • E along the coasts in Fig. 2(a) and (e). The existence of spurious SICs tends to cause opposite variations during the threshold selection as explained in Section III-E.
Therefore, a polygon-based approach was introduced to remove the spurious SICs. Based on the clipped AMSR2 ASI SIC product, the first step was to extract potential spurious ice polygons around lands (appear small, thin, and long) with ice concentration above 0%. The potential spurious ice polygons were selected based on three criteria: average SIC less than 40%, total grids (6.25 × 6.25 km) number less than 100, and R a2p value less than 1.2, where the thresholds were determined empirically. Next, the Euclidean distance between the potential spurious ice polygons and the land pixels (white and gray regions in Fig. 2(e) and (f) was calculated. If the distance was less than two grids, the ice polygon was removed as spurious sea ice around lands. Fig. 2(e) and (f), respectively, shows the clipped AMSR2 ASI SIC images before and after removing the spurious sea ice around lands, while the latter is also overlapped with black ice edges extracted from Fig. 2(d). We can see that the spurious SICs demonstrated by the basically zero sea-ice fractions in Fig. 2(a) along the west coast lines of the Svalbard Islands are accurately removed. However, due to the constraints on individual small ice polygons, part of the spurious SICs that remain connected with the large ice packs can still be reserved [see the marked yellow box in Fig. 10(f)]. Therefore, we largely but not overly removed the spurious SICs along the coasts of the AMSR2 ASI SIC products. As a reference, the area fraction of the removed spurious sea-ice polygons around lands to the nonland area in Fig. 2 is 0.067, and the overall average area fraction of removed spurious sea-ice polygons in the 6972 processed AMSR2 SIC images is 0.012.

C. Calculation of Sea-IFF
To analyze the relationship between the accuracy of AMSR2 ice extent and ice melt conditions, a new parameter called the sea-ice fragmentation fraction (IFF) was introduced. It is defined as the fraction of the removed ice edge details, i.e., the proportion of the pixels where the primary SAR NN SIE [S pri , Fig. 2

(b)] and the processed SAR NN SIE
differ, relative to the total number of integrated ice pixels IFF = n S pri xor S pro n ice,S pro (1) where XOR denotes the logical exclusive OR operation and n ice,S pro is the number of total ice pixels in S pro . The different regions in the numerator (1) consist of narrow inner open water polygons and outer small ice floes that are discarded during the SAR NN SIE processing procedure. IFF reveals the fraction of open water holes and ice fragments distributed on a complete piece of sea-ice "cloth." A larger IFF value indicates more small water details identified by SAR images mainly in the interior sea-ice regions. IFF is proportional to sea ice fragmentations but cannot represent the distribution density of larger-scale ice floes above several kilometers. The IFF value in Fig. 2 is 0.23. Similar to the concept of net ISF proposed in [6], the IFF parameter tends to indicate the net water fraction (including open water and melt ponds).

D. Calculation of Evaluation Metrics
After the data processing, the SAR and AMSR2 SIE products were matched in resampled grids/pixels of 400-m resolution. Three evaluation metrics were then introduced to quantitatively analyze the AMSR2 ASI SIE validated on SAR NN SIE: the OA, error-of-ice (EI), and ice edge location distance (LD). Although the pre-proposed integrated ice edge error (IIEE) [42] and integrated ice area error (IIAE) [43] have been used in global ice cover analysis in the Arctic and Antarctic, the IIEE or IIAE needs to be further divided by the integrated ice extent of each individual SAR image when the comparison is conducted on a given satellite image, which covers a limited geographic region. This is done to ensure that the evaluation metrics are not affected by the differences in ice fraction or ice edge length among the individual SAR images. Thus, based on the pixel numbers of overestimated (n O ) and underestimated (n u ) ice extent in one SAR image, the following parameters are defined: where n * represents the number of pixels of overestimation (n O ), underestimation (n U ), total image (n all ), ice extent (n ice ), and ice edges (n edge ) for the processed AMSR2 ASI SIE image. IIAE is the integrated extent/area of over-and under-estimated ice pixels, i.e., (n O +n U )·0.4 2 km 2 . In contrast, S ice represents the integrated extent/area of total ice pixels, i.e., n ice · 0.4 2 km 2 . Assuming that ice edges extracted from the processed AMSR2 and SAR ice extents are essentially parallel, the LD should be the ratio of IIAE to the length of ice edge, i.e., n edge ·0.4 km. OA shows the overall classification accuracy, whereas EI accounts for the proportion of IIAE to the integrated ice extent of individual SAR images. The components (n O /n ice ) and (n U /n ice ) of EI are called errorof-overestimation (EO) and error-of-underestimation (EU), respectively.

E. Optimal Threshold Selection
To estimate the best-matched threshold, the ice extent is first segmented from the AMSR2 ASI SIC product using thresholds ranging from 0% to 100%. The pixel numbers of overestimated and underestimated ice extent, n O and n U , are then calculated with respect to the AMSR2 ASI SIE and SAR NN SIE products. Eventually, the ice concentration threshold is selected when the number of error pixels (n O + n U ) reaches the minimum. Fig. 2(g) shows the number of error pixels varying with SIC thresholds for the AMSR2 SIC images before and after removing the spurious SICs around lands, i.e., Fig. 2(e) to (d) and (f) to (d), respectively. Because it mainly varies from 0% to 40% with a several kilometers width [see Fig. 2(e)], the spurious SIC around lands contributes exactly to the n O fraction when compared to SAR NN SIE with an ice-clean coast. Due to the significant length of the spurious SIC around lands, the AMSR2 ASI SIC will use larger thresholds to exclude it entirely from the AMSR2 ice extent. Thus, the location of the minimum of the (n O + n U ) curve turns out to be larger [see Fig. 2(g)]. In other words, if the spurious SICs around the land are retained, the best-matched thresholds will be 22% higher than the 4% obtained when it is removed. The processed AMSR2 ASI SIE and SAR NN SIE products [e.g., Fig. 2(f) and (d)] after the processing introduced in Sections III-A and III-B were finally used for the evaluation.

IV. RESULTS
In this section, we segmented the AMSR2 ASI SIE product using the 15% threshold to perform a quantitative evaluation and calculated the best-matched thresholds for all individual images to estimate their influence on the evaluation results. Temporal variations, spatial distribution of the evaluation metrics, and analyses for different large-error conditions were sequentially performed, from the perspectives of the Atlantic and Pacific Sectors, winter and summer of the Arctic Ocean. Note that the summer refers to the main ice melting period from May to September, while the winter refers to the other months that are dominant with the freezing process.
A. Evaluation of AMSR2 ASI SIE Product 1) Overall Comparison: Fig. 3(a) shows the OA and EI metrics calculated for individual AMSR2 ASI SIE (using 15% threshold) compared to the SAR NN SIE images within the 400-m pixel resolution. The results show that the OA values are predominantly above 0.8, among which more than 79% (in winter) and 52% (in summer) are above the annual average OA value of 0.90. The maximum and minimum daily average OA is found in January and July, respectively (0.97 and 0.82). On the contrary, EI means the over-and under-estimated ice extent error normalized by the integrated ice pixels of each clipped AMSR2 ASI SIE image. In Fig. 3(a), the distribution of EI shows higher dispersion than OA, with values mainly ranging from 0 to 0.4 and an annual average of 0.20. The difference in daily average EIs between winter and summer reaches 0.27 (max: 0.35 and min: 0.08). The seasonal variations of OA and EI both indicate that the SIE error increases in summer. Fig. 3(b) shows the time series of IFF, EI (including EO and EU), and LD values. IFF mainly describes water distributions on the ice surface, with an annual average of 0.15 and a seasonal increase in summer. EI and IFF change almost synchronously with a correlation coefficient of about 0.69. Note that the minimum of integrated AMSR2 SIE is found on 15 September 2019, when the IFF and EI both show a slight concave decrease. However, EO and EU show rather different seasonal variations. EO is higher than EU by over 0.1 before and after the peak-melt (August) periods, whereas EU is slightly higher than EO in the deep winter (January to March). This contrast indicates the distinct possible overestimation and underestimation patterns in summer and winter. Besides, the daily average LD also shows an obvious increase from a minimum of 9.5 km in January to a maximum of 26.9 km in August. The LD values of individual images vary from 0 to over 100 km, with a similar scattering distribution mode as EI. Due to the dominant (more than 75%) low values (less than 20 km), an annual average LD of 16.71 km is obtained. The corresponding average IIAE is 8644 km 2 in one SAR image size.
2) Categorized Comparison: Based on the complementarity of EI and OA and the seasonal variation of overestimated and underestimated amplitudes revealed in Fig. 3, all the images were categorized into four groups to illustrate the proportion and magnitude of different errors. The images were categorized as follows: 1) high accuracy group: OA ≥ 0.9 and EI ≤ 0.2; 2) underestimation dominant group: (OA < 0.9 or EI > 0.2) and EO-EU ≤ −0.1; 3) overestimation dominant group: (OA < 0.9 or EI > 0.2) and EO-EU ≥ 0.1; and 4) others. Fig. 4 shows the number of images for the four categories. With decreasing OA in summer, the number of images in Groups II-IV (i.e., large error groups) all increases and reaches the maximum in August (September for Group III). The monthly total number of selected images also increases by more than two-thirds with the persistent opening of ice edges  in the Arctic basin. Therefore, although the number of images in Group I remains more than 200 throughout the year, the proportion of Group I decreases largely from winter (63.0%) to summer (34.0%). Consistent with Fig. 3(b), Groups II and III show a large fraction difference at advance-melt (May to July) and end-of-melt/early-refreezing (September to November) seasons. During these periods, the number of images in Group III (percentage: 24%-28%) is more than twice that of Group II (5%-13%). The number and proportion of images in the four groups show a near-symmetrical distribution centered on the peak-melt season (August).
In addition to the number difference (see Fig. 4), complementary effects of IFF and LD variations are found between Group II and III (see Fig. 5). First, Group II (blue plots) generally presents higher LD values than that of Group III (light gray plots), especially before the peak-melt. Group III exhibits a clear monthly increase in IFF values, particularly with higher variation than Group II after the peak-melt (August to December). The contrast of LD and IFF values between Group II and III is maximized in August, during which the  3) New Thresholds: Based on the pixel number of overand under-estimated ice extent, the best-matched segmentation thresholds for AMSR2 ASI SIE were selected individually in each SAR image. Fig. 6(a) and (b) shows the distributions of thresholds in the four groups, with 0% occupying a large proportion: 954/3488 (number of images of bar 0%/total) for Group I, 253/1070 for Group IV, 467/913 for Group II, and 55/1501 for Group III (see the embedded diagrams). In Groups I and IV [see Fig. 6(a)], the thresholds mainly vary between 0% and 50%, because the underestimation and overestimation errors of AMSR2 ice extent are essentially equivalent. In contrast, the large underestimation of AMSR2 SIC/SIE in Group II causes the thresholds to fall into the low extreme values, while the large overestimation in Group III makes the thresholds tend to be high [see Fig. 6(b)]. Fig. 6(c) shows the seasonal variation of thresholds in the four groups, which all show an increase in summer and a consistent minimum in March. The Arctic SIE reaches its maximum in March and most of the sea ice is directly connected to the land, thus images along the ice edges in the east side of Greenland were mainly selected for comparison. The minimums suggest that the AMSR2 SIC has an overall underestimation in March (actually the whole winter). Conversely, the higher thresholds during summer indicate that the overestimation has a larger amplitude than the underestimation. However, a decline of the thresholds appears in August for Group I and II, consistent with the sharp increase of EU during the same period in Fig. 3(b). Fig. 6(d) shows the monthly average thresholds with positive and negative standard deviations of Group I. From the perspective of the AMSR2 ice concentration retrieved with the ASI algorithm excluding weather effects and other large uncertainties, the average threshold of Group I is more suitable as the overall optimal threshold for segmenting AMSR2 SIE. Further discussions are given in Section V-A. As a result, we adopted the annual average ice concentration threshold of 12.24%, the seasonal average thresholds of 16.43% in summer, and 9.25% in winter from Group I as the global optimum.
The daily integrated SIEs calculated, respectively, by the usual-used 15% and the newly-obtained annual/seasonal average thresholds are then further compared. In Fig. 6(e), the integrated SIE differences (blue line) remain relatively stable around 6.0 × 10 4 km 2 , representing an average proportion of 0.62% of the total Arctic SIE and 4.80% of the total MIZ SIE, when the annual average threshold of 12.24% is used. Subsequently, Fig. 6(f) shows that the integrated SIE difference can be relatively larger (with an average of 1.2 × 10 5 km 2 ) when using the 9.25% threshold in winter compared to that (with an average of 3.0 × 10 4 km 2 ) when using the 16.43% threshold in summer. The former, i.e., the integrated SIE difference in winter, represents a proportion of 1.27% of the total Arctic SIE and 11.36% of the total MIZ SIE [see Fig. 6(f)]. Based on the new thresholds, the annual average OA and EI values range from 0.90±0.02 and 0.20±0.03, respectively. The small changes in the average OA and EI suggest that the significant underestimation and overestimation patterns have not changed substantially when using the new thresholds. Therefore, they are unlikely to affect subsequent analyses for the evaluation of AMSR2 ASI SIE segmented by 15%. With the distinguishable spatial distributions of dominant overestimation and underestimation, Fig. 8 further presents the seasonal variations in terms of image numbers, IFF, and LD distributions, of the four groups in the Pacific and Atlantic sectors of the Arctic Ocean segmented by longitude lines 9 • E and 90 • W. First, during most of the winter months, no images are found in the MIZ of the Pacific sector [see Fig. 8(b)], which could be attributed to the extensive high SIC distributions and ice-land connected pattern. Second, the Atlantic sector is greatly overestimated during the entire melting season, as shown in Fig. 8(a), according to the larger number of images in Group III than in Group II from May to November.

B. Spatial Distributions in the Atlantic and Pacific Sectors
In Group III, the absolute EO-EU is between 0.1 and 1.0, the LD values are below 60 km, and the IFF values vary up to 1.12. Conversely, the Pacific sector shows dominant underestimation, with a more favorable number of images in Group II than in Group III from July to August. In Group II, the absolute EO − EU is 0.1-0.84, the LD range is 0-100 km, and the IFF values are below 0.5. Lastly, it can be observed that the extreme LD and IFF could be attributed to the segmented Pacific and the Atlantic sector, respectively, upon comparing Fig. 8 to Fig. 5.

C. Examples of Analysis in Winter and Summer
The variation of uncertainties suggests that the AMSR2 ASI SIE estimation is influenced by different factors in winter and summer, with frazil ice and newly grown thin ice being important in winter, and melt ponds, brash ice, small ice floes, and melted thin ice being significant in summer. This study also selected six examples, two for winter and four for summer, with relatively large EI values and different EO/EU amplitudes, large IFF or LD values, to further compare and analyze the possible error sources under different conditions (see Figs. 9 and 10). Table I  This study notes that the difference in observation time and spatial resolutions can impact the evaluation of AMSR2 ASI SIE. The latter could lead to large error in IIAE over extremely fragmented ice floes in summer, while the former does not have the same effect. The acquisition time of the SAR images in the examples falls in between the observation periods of the AMSR2 swath data. As the AMSR2 ASI SIC is a daily gridded product that resamples the ice concentration estimates from all the AMSR2 data collected within the calendar day using the nearest neighbor method (https://seaice.uni-bremen.de/filead min/userupload/ASIuserguide.pdf), it assumes that the ice conditions are the same during the comparisons between the SAR NN SIE and AMSR2 ASI SIE products.
1) Winter: Enhanced atmospheric activities, including winds and cyclones [44], can partly explain the major weather effects in winter, though new sea-ice growth generally shows increasing thickness with little influence from surface melting. Similar to the study in [45], larger errors in winter are found mainly in frazil ice and thin ice areas as shown in Fig. 9. First, Fig. 9(a) (which refer to the image pair of (a) and (b), hereafter the same) shows the normal condition where there is good agreement between the AMSR2 and SAR SIEs along most of the ice edges, except for the lower-right frazil ice regions [the same as the right ice edge regions in Fig. 9(c)]. Frazil ice is too thin to effectively alter the emissivity [45], resulting in its substantial underestimation in the AMSR2 ASI SIE product.
Besides this, the AMSR2 ASI SIE tends to overestimate the refreezing of thin ice in ice leads or melt ponds compared with the SAR NN SIE product [see Fig. 9(c)]. The fact is that the 89 GHz used in the ASI algorithm has been reported to have high accuracy for thin ice identification because of its small penetration depth and consequent insensitivity  to ice thickness [5]. Conversely, thin ice surfaces from the Sentinel-1 SAR images are typically underestimated because of the low backscatter, which is close to that of calm open water [4], [46]. In our experiment, the relative overestimation over thin ice and underestimation over frazil ice of AMSR2 ice extent are found with extensive persistence throughout the winter period. Generally, ice thickness continues to increase as winter proceeds, thus the area of thin ice decreases. Therefore, more overestimations with large amplitudes occurred at the refreezing-beginning (October to December), whereas the amplitudes and proportions of underestimation and overestimation turn toward the same in deep winter (January to April) [see Figs. 4 and 8(a)].
2) Summer: During the ice melting process, different surface conditions such as wet/melting snow, melt ponds, ice leads, and fragmented ice floes gradually appear, leading to changes in the surface temperature and emissivity variability [6]. We summarized four main ice conditions that were observed during the melt-advance (May to July), peak-melt (July to August), and end-of-melt (September) periods to explain the large uncertainty in AMSR2 ASI SIE estimation during summer (see Fig. 10).
Each image in Fig. 10 separately illustrates the impact of a specific error source. For instance, in Fig. 10(a), the black disks of low backscatters representing melt ponds in pack ice regions during the advance-melt period are recognized as open water in the SAR ice extent data. However, the AMSR2 ASI SIC product shows around 100% SIC (i.e., compact ice extent) in these regions. This could be due to the grid-wise counteracted PM brightness temperatures of melt ponds (reducing effect) and wet snow/ice (increasing effect) [6], [45]. In addition, the tie points (samples of 0% or 100% SIC) used for SIC retrieval algorithms may have included some melt pond information [6].
The SAR NN SIE product performs better at depicting melted brash ice with rough textures [see Fig. 10(c)], whereas AMSR2 ASI SIE more accurately identifies the thin ice (melted ice or new ice) with smooth surface [see Fig. 10(g)]. In Fig. 10(g), the prominent thin ice zone along ice edges appears during the end-of-melt or early-refreezing periods, though the causes for overestimation in Fig. 10(g) are exactly similar to those in Fig. 9(c). In contrast, the brash ice that largely intersects with open water in Fig. 10(c) occurs frequently during the advance-melt or peak-melt periods. Although 89 GHz is very effective for thin ice identification, it is also sensitive to atmospheric factors such as liquid water, ice clouds, and atmospheric integrated water vapors [5], [28]. Higher surface wind speeds and enhanced heat flux/mass transfer may have introduced more significant atmospheric effects in Fig. 10(c) than in the closed, smooth thin ice in  Fig. 10(g), leading to the underestimation of melted brash ice in the AMSR2 ice extent product. In addition, another error source could be the gradient ratio thresholds used during weather filter in ASI algorithm.
Subsequently, Fig. 10(e) shows ice floes with more homogeneous and much brighter surfaces in the SAR image, without any obvious identification errors appearing in either products. However, overestimated fuzzy ice concentrations (above 30%) are found in the northeast open ice area with floe sizes of several kilometers. This is attributed to the coarse resolution (6.25 km) of PM data with possible occurrences throughout the year, though the frequency of this overestimation is likely increased during summer melting with an increased number of such open ice floes.
3) General Analysis: As mentioned at the beginning of Section IV-D, due to the time difference between the AMSR2 and SAR observations, factors such as ice dynamics, ice growth/melting also affect the comparison and evaluation of the two SIE products. The SIE difference caused by ice drift (without considering the convergence and divergence) is thought to exhibit as near-parallel edges, which theoretically have equal distances even under various ice conditions in the four groups. For instance, the pack ice at the Fram Strait (in Group I) and the open ice in the Beaufort Sea (in Group II) may experience equal wind speeds, leading to near-equal ice edge displacements during the observation time intervals. Therefore, the LD introduced by ice drift is thought to be no larger than the annual largest LD (33.10 km) of Group I (the high accuracy group, performing as near-parallel edges). Moreover, it is difficult to separate the impact of various factors on the comparisons, especially factors more correlated with regional weather.
The examples in Figs. 9 and 10 (see Table I) correspond well with the temporal and spatial variation characteristics of over-estimation and underestimation revealed in Figs. 3 and 8. With respect to the AMSR2 ASI SIE, frazil ice is mainly underestimated throughout the entire winter [see Fig. 9(a) and (c)], while the refreezing of thin ice within ice leads contributes to larger overestimation in early winter [see Fig. 9(c)]. In contrast, the melt ponds and thin ice zone near edges, respectively, explain the significant majority of large overestimation rather than underestimation before and after peak-melt [see Fig. 10 We also emphasize that the superiority of AMSR2 for identifying thin ice (melted or new refreezing) has an impact on the high EO-EU values in both the Pacific and Atlantic sectors during the end-of-melt/early-freezing periods (September to November in Fig. 8). The Pacific sector accounts for a large proportion of open ice edges during summer melting and refreezing begin periods, where winds/storms and increased fetches could create high sea states and increase the air-sea fluxes of heat and momentum [47]. The ice characteristics in Fig. 10(d) and (h), corresponding to the extreme values in Fig. 8(b), suggest that thinner first-year ice predominates in the Pacific sector and has experienced much stronger basal and lateral melting or more rapid ice refreezing than the thicker multiyear ice predominating in the Atlantic sector.
In Table I, high IFF values (larger than 0.2) are found in the SAR images of more small ice or water polygons, i.e., long ice edges. Conversely, high LD values occur more in regions with brash ice, thin ice, or newly grown frazil ice. We further calculated the ratio of LD to IFF and found that this ratio could distinguish the conditions in Figs. 9(a) and 10(c) and (g) (high ratios) from those in Figs. 9(c) and 10(a) and (e) (low ratios). This provides us a great approach for data fusion of the AMSR2 ASI and SAR NN SIE products. When the images have higher ratios of LD to IFF (e.g., larger than 200 km), we could choose the relatively overestimated one between the two products as the accurate SIE. Otherwise, by identifying that the larger IFF values are caused by small water polygons [see Fig. 10(a)] or ice polygons [see Figs. 9(c) and 10(e)], we can then choose the AMSR2 SIE in case Fig. 10(a) and the SAR SIE in other cases. In addition, when the SAR images mainly consist of large error regions, such as brash ice, thin ice, or small ice floes, extremely high LD and IFF values were recorded, respectively, for Group II in Fig. 5(a) and Group III in Fig. 5(b).

A. New Thresholds
In previous studies, a threshold range of 10%-15% was considered to have excluded the influence of weather effects on the NASA team (NT) algorithms [48]. Unlike the NT algorithm, the ASI algorithm that uses 89-GHz PM data is much more sensitive to weather factors [5]. To reduce the threshold errors caused by large uncertainties, including weather and other effects, we first evaluated the AMSR2 ice extent segmented by the 15% threshold and then selected images with high accuracies (Group I: OA > 0.9 and EI < 0.2) to calculate the overall optimal threshold. In Group I, the images of AMSR2 ASI SIE and SAR NN SIE show parallel ice edges with LD no more than 33.10 km (the annual largest LD in Group I) [see the left of Fig. 9(b)]. The consistent ice distribution pattern of the two products in Group I suggests that the SIEs are correctly identified. The only difference between the two SIEs is exhibited in the near-parallel ice edges, which vary with the SIC thresholds and could fit better with adjusted thresholds. In this condition, the SAR NN SIE is considered to be "real" and the annual average thresholds acquired in Group I are assumed to be optimal. On the contrary, the images in Groups II-IV show large inconsistencies between the two SIE products due to various ice conditions, as exemplified in Figs. 9 and 10 [except Fig. 9(a)]. Under such conditions, ice edges in the two SIE products intersect with each other to a certain extent. Neither AMSR2 ASI SIE nor SAR NN SIE could be regarded as "real." These images are therefore not be used for the optimal threshold calculation.
Consequently, we obtained an annual average threshold of 12.24%, with seasonal average thresholds of 9.25% in winter and 16.43% in summer. The 9.25% threshold indicates a general underestimation of AMSR2 SIC/SIE in winter, especially from January to March [see Fig. 3(b)]. In contrast, the 16.43% threshold demonstrates the higher accuracy of the 15% threshold for AMSR2 ASI SIC in summer. Our results are consistent with the average ice concentration of edge pixels, 10.5% and 13%, which are obtained by comparing the AMSR-E/AMSR2 ASI SIC products with pseudo-ship observations in the Antarctic [13], [30].
During the processing procedure, we implemented two steps to improve the accuracy of our analysis. The first was the removal of spurious SIC along coasts in AMSR2 ice extent, and the second was the detail smoothing of SAR ice extent while retaining main ice distribution outlines. These steps are crucial in optimizing the AMSR2 ASI SIC thresholds based on the SIE validation data as accurately as possible and may be useful for other PM SIC products/algorithms.

B. Statistical Trends of AMSR2 ASI SIE Evaluation
For the evaluation of AMSR2 ASI SIE, in addition to the OA and EI, the sea-IFF and ice edge LD were also introduced to assess the SIE uncertainties under different ice conditions.
The results showed an annual average LD of 16.71 km (corresponding to an average IIAE value of 8644 km 2 ), while the individual LD values ranged from 0 to 100 km [see Fig. 3(b)]. Assuming that the Sentinel-1 images cover two-thirds of the Arctic ice edges, about 30 Sentinel-1 images (6972/365/0.66) are needed to cover all open ice edges that are unconnected with the coastlines [see Fig. 1(b)]. Thus, the integrated IIAE values can be around 259 320 (8644 × 30) km 2 , which is slightly smaller than the extent bias in the order of 5 × 10 5 -1 × 10 6 km 2 [32], but still reasonable. Furthermore, the LD parameter is an average distance description, which is different from the edge displacement score proposed by [49], indicating the maximum distance between two ice edge lines. The maximum distance of Fig. 10(h) calculated by the algorithm [49] is 97.60 km, while the LD has a value of 41.86 km. The significant advantage of LD is its calculability under extremely complex ice edges composed of fragmented ice floes or permeable melt ponds, whereas the ice edge lines from the two products are most likely to be confused, and difficult to find the corresponding maximum distance point pairs [see Fig. 10(a) and (e)].
In the time series (see Fig. 3) and spatial distributions (see Fig. 7 As analyzed/discussed in Section IV-D, the different spatial and temporal distributions of dominant EO and EU errors concern highly with the ice conditions that change considerably under various atmospheric and oceanic states. Further work is needed to investigate the precise physical mechanisms behind these differences. Due to the limited capacity, we put emphasis on summarizing the characteristic phenomena and better usage for the summarization in this article.

C. SIE Products Fusion Using the Ratio of LD to IFF
The MIZ is a crucial area for studying the complex seaice-air interaction under rapid thermodynamic and dynamic changes. Previous studies have shown that neither the PM SIC nor SAR SIE products/algorithms can accurately extract ice cover information in the MIZ, especially during summer. Therefore, many studies have proposed multisource data fusion as an effective means of improving the estimation of sea-ice phenology parameters [3], [5], [16]. The two main objectives of data fusion for SIC/SIE products are to improve accuracy and to improve spatial resolution. Several methods have been developed for the data fusion of PM and visible/infrared [e.g., MODIS, visible infrared imaging radiometer suite (VIIRS)] SIC products [25], [50], [51], which can improve the blended resolution to 1 km. However, few studies have focused on the fusion between PM and SARbased SIC/SIE products. Our results suggest that although PM SIC products typically underestimate ice concentration in thin ice and melt pond regions [5], [9], [45], the underestimation magnitudes do not significantly affect the PM SIE estimations based on the 15% threshold. In other words, the AMSR2 product (arctic radiation and turbulence interaction study (ARTIST) sea ice (ASI) algorithm) shows supplementary observation abilities for thin ice and melt ponds compared with the SAR SIE product, while the SAR NN SIE product can identify frazil ice and brash ice better than the AMSR2 ASI SIE product. We propose an approach for using the ratio of LD to IFF (see Table I) to enable data fusion of AMSR2 ASI and SAR NN SIE products in different ice conditions. Typical examples (see Figs. 9 and 10) indicate that it is feasible to identify cases and perform data fusion based on a fixed threshold of the ratio of LD to IFF. Further work can be done regarding data fusion, however, is not included in this article due to space limitations.

VI. CONCLUSION
This article utilized a batch of SAR NN SIE data (6972 images) to evaluate the accuracy and threshold of the AMSR2 ASI SIE product along ice edges throughout the year 2019. Compared to previous studies, we comprehensively analyzed the spatial and temporal characteristics of the AMSR2 ASI SIE uncertainty and its sensitivities to different primary impact factors.
The OA and EI are the basic evaluation metrics for the AMSR2 SIE product. The annual average OA and EI calculated based on individual SAR images are 0.9 (daily averages ranging from 0.82 to 0.97) and 0.2 (daily averages ranging from 0.08 to 0.35), respectively. Based on the OA and EI values, we divided all images into four groups: Group Ihigh accuracy, Group II-underestimation dominant group, Group III-overestimation dominant group, and Group IVthe others. The annual average ice concentration threshold of 12.24%, and seasonal average thresholds of 9.25% in winter and 16.43% in summer, calculated from Group I, are thought to be the overall optimal thresholds for the AMSR2 ASI SIC/SIE product, excluding the weather effects and other large uncertainties. We calculated the integrated SIE difference between using thresholds of 12.24% and 15% to be around 6.0 × 10 4 km 2 . This represents a proportion of 0.62% to the integrated Arctic SIE and 4.80% to the integrated MIZ SIE. When using thresholds of 9.25% and 15% in winter, the integrated SIE difference and its proportions are more than twice as larger as the above.
The overestimation (EO) and underestimation (EU) errors of AMSR2 ASI SIE show distinct temporal and spatial distributions. First, a larger uncertainty, i.e., an increased number of images of Groups II, III, and IV, is found in summer (May to September) than in winter (other months). However, Group III shows a much larger (around twice) increase/proportion than Group II within the advance-of-melt (May to July) and endof-melt/early-refreezing (September to November) periods. During the summer melting of July to August, the dominant overestimation (Group III) and underestimation (Group II) are, respectively, located in the Atlantic and Pacific Sectors of the Arctic Ocean.
The parameters of sea-IFF and ice edge LD are newly introduced to partly describe different ice/error conditions. We obtained the annual average IFF and LD values of 0.15 and 16.71 km, respectively. The IFF has a positive correlation of 0.69 to EI. Thus, dominant overestimation (Group III) and underestimation (Group II) could both have relatively high IFF values. In contrast, the dominant underestimation One example of SAR-based ice edge (blue lines) image that has "straight lines at the ice edge" (red lines) caused by the thermal noise removing processing. The Sentinel-1 EW HV-polarized backscatter values are shown as background in gray. The ice edges (blue lines) were extracted from the SAR-based ice extent provided by [26]. Image ID: S1A_EW_GRDM_ 1SDH_20190902T084303_2019 0902T084408_028840_03449C_6414.
usually has large LD values (up to 100 km), and the dominant overestimation does not necessarily.
The temporal and spatial distributions of AMSR2 SIE uncertainties can be related to various atmospheric and oceanic processes. Our results suggest that the AMSR2 ASI SIE product showed better observation abilities for thin ice and melt ponds than the SAR NN SIE product. In contrast, SAR NN SIE performed better identification on melted brash ice and newly grown frazil ice. We also found that the ratio of LD to IFF can distinguish these different ice conditions well. Therefore, the fusion of AMSR2 and SAR SIE products is realizable by taking the data with relatively higher accuracy in different conditions.
The complementary strengths of AMSR2 and SAR SIE products revealed in this article show good potential for the combined use of both in a better SIE product. Optimized use of the multichannel observations, correction of the atmospheric influence on high-frequency brightness temperatures [3], selection of the characteristic features of SAR images [2], and sectionalized/layered neural networks for special conditions are all crucial for accurate SIE estimates in MIZ during melting periods. In contrast, the distinct performances under various circumstances provide useful information for improving PM and SAR SIC/SIE products. APPENDIX See Fig. 11.