DynIceData: a gridded ice–water classification dataset at short-time intervals based on observations from multiple satellites over the marginal ice zone

ABSTRACT High-resolution observations of short-term changes in sea ice are critical to understanding ice dynamics and also provide important information used in advice to shipping, especially in the Arctic. Although individual satellite sensors provide periodic sea ice observations with spatial resolutions of tens of meters, information regarding changes that occur over short time intervals of minutes or hours is limited. In this study, a gridded ice–water classification dataset with a high temporal resolution was developed based on observations acquired by multiple satellite sensors in the Marginal Ice Zone (MIZ). This dataset – DynIceData – which combines Sentinel-1 Synthetic Aperture Radar (SAR) data with Gaofen-3 (GF-3) SAR and SDGSAT-1 thermal infrared imagery was used to obtain observations of the MIZ with a range of temporal resolutions ranging from minutes to tens of hours. The areas of the Arctic covered include the Kara Sea, Beaufort Sea, and Greenland Sea during the period from August 2021 to August 2022. Object-oriented segmentation and thresholding were used to obtain the ice–water classification map from Sentinel-1 and GF-3 SAR image pairs and Sentinel-1 SAR and SDGSAT-1 thermal image pairs. The time interval between the images in each pair ranged from 1 minute to 68 hours. Ten-kilometer grid sample granules with a spatial resolution of 25 m for the GF-3 SAR data and 30 m for the SDGSAT-1 thermal data were used. The classification was verified as having an overall accuracy of at least 95.58%. The DynIceData dataset consists of 7338 samples, which could be used as reference data for further research on rapid changes in sea ice patterns at different short time scales and provide support for research on thermodynamic and dynamic models of sea ice in combination with other environmental data, thus potentially improving the accuracy of sea ice forecasting using Artificial Intelligence. The dataset can be accessed at https://doi.org/10.57760/sciencedb.j00001.00784.


Introduction
Over the past 40 years, as a result of global warming and the effects of Arctic amplification, the Arctic has warmed at a rate of 0.75°C every ten years, which is nearly four times the global average rate of 0.19°C per decade (Rantanen et al., 2022).Rising temperatures have caused the summer sea ice extent in the Arctic to decline by 12.6% per decade (Kwok, 2018).This rapid decline in sea ice cover has increased the commercial viability of trans-Arctic shipping routes: the area of 90-day safe navigation for open-water (OW) vessels increased by 35% between 1979 and 2018, and the navigable period of the Northeast Passage is about 92 days longer than it was 40 years ago (Cao et al., 2022).The Kara, Laptev, and East Siberian sectors have become more accessible in recent years because of the thinning of sea ice prior to the melt season and the enhanced positive polarity of the summer Arctic Dipole Anomaly (Lei et al., 2015).The rapid changes in sea ice that are occurring in the Arctic Marginal Ice Zone (MIZ) increase the risks to OW vessels with weak ice-breaking capabilities (Leppälä et al., 2019).Near-real-time high-precision monitoring of sea ice is, therefore, needed to mitigate risks to navigation, especially in the MIZ (Wang et al., 2020).
The area of multi-year ice in the Arctic has decreased by 2 × 10 6 km 2 in the last 40 years, and currently more than 70% of the ice is seasonal.Rapid changes can occur in this ice; the most rapid occur in the MIZ, mainly in the seasonal ice, and also at the transition from sea ice to open water (Yang et al., 2023).The ice is affected by the interactions between the ice, air, and ocean, which involve complex and variable thermodynamic and kinetic processes (Deng & Dai, 2022).For example, the sea ice in the MIZ grows and melts faster and moves more easily with the stage of sea ice thickness decreasing (Liu et al., 2021); it also breaks up more easily, meaning that waves from open water can penetrate.This causes ice floes to break up and other processes that can alter ice cover to occur.The ice in the MIZ is also more susceptible to the influence of wind and ocean currents (Kousal et al., 2022).It can, therefore, be inferred that the main forces driving sea ice change in the MIZ are dynamical rather than thermodynamical.The distribution of heat, mass, and momentum fluxes in the Arctic plays an important role in controlling the ice edge and its location (Yu et al., 2009).This means that fine-scale monitoring of ice floes is needed (Bateson et al., 2020): the higher the observation frequency and the shorter the time interval between observations, the better the detailed characteristics of the sea ice dynamics will be reflected in the data.Fine-scale and high-frequency monitoring are also very helpful to understanding the dynamics of sea ice in the MIZ and the physical changes that are occurring.
At the moment, there is still a lack of high-spatiotemporal resolution datasets that can be used to obtain the characteristics of sea ice change -including sea ice growth and decline, ice movement, and the opening and closing of ice leads -at a fine scale in a short time or that provide near-real-time (within one day) highfrequency observations (Cheng et al., 2022).The existing operational sea ice products for the Arctic are mainly based on passive microwave radiometer data; these meet the demands of near-real-time applications and can provide daily data for pan-Arctic sea ice parameters such as sea ice concentration (SIC) and sea ice movement.However, the spatial resolution of SIC products is as large as 3.125 km (Spreen et al., 2008), which is too coarse for the analysis and identification of small-scale sea ice characteristics and dynamics.The resolution of sea ice movement products is between 10 km and 25 km with a time interval between observations of one day (Meier, Markus, et al., 2017;Meier, Stewart, et al., 2017).However, sea ice can move quickly, and the dynamic range of the sea ice edge within one day can be large.This means that the spatial and temporal resolution of existing datasets that are based on passive microwave remote sensing do not have sufficient spatial and temporal resolution to accurately reflect the rapidly changing characteristics of sea ice in the MIZ (Huang et al., 2022).
Along with the development of satellite data techniques, marine remote sensing has entered the big data era.The use of multi-source remote sensing data can greatly increase the revisit frequency of Arctic observations (Li et al., 2020).The availability of highresolution remote sensing data provides a means of sea ice monitoring based on massive samples of short-term changes in sea ice.The sources of these data include optical, thermal infrared, and synthetic aperture radar (SAR) data acquired by multiple satellite sensors.SAR data, in particular, enable highly precise, high spatial resolution monitoring to be carried out in all weather conditions.The Sentinel-1 SAR satellite captures sea ice scenes by periodically visiting Arctic regions within 6 days (Ren et al., 2020).Sentinel-1A and Sentinel-1B share the same orbital plane with a 180° phase difference, and the use of their combined data provides an improved revisit time of 1-3 days in the Arctic (Karvonen et al., 2012).However, a temporal resolution of one day is still insufficient for the near-realtime monitoring of ocean and sea ice conditions required by shipping; in addition, due to an instrument abnormality, the Sentinel-1B satellite stopped delivering data on 23 December 2021, which has resulted in a reduction in observation frequency.The Gaofen-3 (GF-3) satellite, launched in August 2016, has 12 imaging modes and can also achieve high-frequency coverage in the Arctic region, with a revisit time of as often as one day (Zhao et al., 2021).The combination of the Sentinel-1 and GF-3 SAR data can greatly increase the revisit frequency and allow high-frequency monitoring of sea ice in the Arctic region.Thermal infrared remote sensing can be used to distinguish between ice and water due to the large temperature contrast between open water and sea ice (Paul & Huntemann, 2021;Paul et al., 2015), and this type of imaging is not affected by the time of day.The most widely used thermal infrared dataset is the Moderate Resolution Imaging Spectroradiometer (MODIS) ice surface temperature dataset, which has a 1-km resolution and can be used to identify sea ice leads with a width of up to 1 km on a daily basis (Hoffman et al., 2022;Willmes & Heinemann, 2015).Data acquired by the Thermal Infrared Spectrometer carried by the Sustainable Development Science Satellite 1 (SDGSAT-1 TIS), which was launched on November 5, 2021 and has a spatial resolution of 30 m (Guo et al., 2023), allows higher-resolution ice monitoring in the Arctic.Using a combination of images acquired by different sensors, high temporal resolution monitoring that allows discrimination between ice and water in the Arctic can be realized.
For navigation applications, accurate seasonal sea ice forecasts are needed to mitigate the risks posed by rapid changes in sea ice conditions.Sea ice modeling is concerned with the basin-wide evolution of ice conditions and describes sea ice across a range of spatial scales from medium to large.This modeling is based on the fundamental laws of thermodynamics and dynamics together with motion and ice conservation laws (Leppäranta et al., 2020).Data from measurement buoys, such as the MOSAiC (Multidisciplinary drifting Observatory for the Study of Arctic Climate) sea ice data timeseries, are also used for short-term ice forecasting and climate research (Lei et al., 2022).The thermodynamics of sea ice is well understood, but models are not yet able to completely describe the dynamics of sea ice in the MIZ (Liang et al., 2020).Sea ice parameters and change patterns are used as the input dataset for sea ice forecasts and in climate and ocean models; the accuracy of these largely depends on the detail given for the respective sea ice conditions (Andersson et al., 2021;Shu et al., 2021).However, most numerical models cannot directly generate details such as ice leads and areas of small crushed ice (Reiser et al., 2020).Coupled with the rapid changes that have been occurring in Arctic sea ice, its physical properties have also changed, and the uncertainty in the results of sea ice simulations has greatly increased (Li et al., 2020;Zhu et al., 2022).
Artificial intelligence (AI) deep learning algorithms that can be used to make predictions of sea ice are also being developed; these have advantages over traditional image information extraction algorithms based on physics or statistics.Improved observation data with high resolution and short time intervals have been used as input and training data to further increase the amount of input data and training samples used in models (Boutin et al., 2020).However, data-driven AI algorithms lack understanding of the physical mechanisms and process related to the dynamics of sea ice in the MIZ (Bennetts et al., 2022).It is difficult to simulate the physical processes involved in parameterized sea ice dynamics when modeling the ocean and climate (Zhu et al., 2022).Highresolution sea ice data can be helpful in producing accurate descriptions of sea ice movements and in determining the characteristics of small-scale sea ice changes.Data that describe sea ice changes that occur over short time intervals can add to our knowledge of sea ice, as the physical law information of the thermodynamics and dynamics of sea ice, further improves simulated forecasts of ocean and sea ice.
In this study we developed a dataset, DynIceData, to provide a high spatiotemporal resolution dataset of sea ice in the MIZ -data which until now have been missing.This dataset was formed from selected image pairs of Sentinel-1, GF-3, and SDGSAR-1 images of the MIZ and used to monitor sea ice over short time intervals during the period from August 2021 to August 2022.Object-oriented segmentation and thresholding methods combined with visual interpretation and manual correction were used to classify sea ice and water.Together, the selected image pairs comprise a gridded ice-water classification dataset covering short time intervals based on observations of the MIZ made by multiple satellites.The ice-water classification image pairs were segmented into 10 km × 10 km grid cells to analysis the characteristics of sea ice change in the short-time intervals, and 7388 grid samples are provided in this version of DynIceData.The temporal resolution of the dataset ranges from 1 minute to 68 hours.
Using the DynIceData data, a preliminary analysis of the characteristics of sea ice change and movement over different short time intervals was performed.The results could be used as sea ice change information and combined with artificial intelligence (AI) sea ice inversion models or forecast models with a high spatiotemporal resolution to produce results that are close to the real changes occurring in sea ice.

Definition of the arctic marginal ice zone
The Marginal Ice Zone (MIZ) in the Arctic, the zone where the sea ice and ocean meet, is defined as the area of ocean with a sea ice concentration ranging from 15% to 80% (Rolph et al., 2020); this zone extends about 100 km from the ice edge (Yu et al., 2009).To define the MIZ research area for this study, existing coarse-resolution SIC data, published by the University of Bremen in Germany (https://data.seaice.uni-bremen.de/)and which have a spatial resolution of 3.125 km, was used to map regions of sea ice.This dataset has the highest resolution among the existing operational SIC products that are based on radiometer data.It consists of pan-Arctic SIC data acquired daily since June 2002.
It was found that the areas with an SIC of more than 0% up to 70% covered the ice edge in the Arctic and formed a continuous distribution zone.And then, a 400-km wide buffer around this zone was established in order to include all of the sea ice in the MIZ, as shown in Figure 1.This buffer size was suitable for the period from May to December.The area of the MIZ was calculated for images acquired on different dates, and the images were screened for possible use in sample pairs.

Sentinel-1 and GF-3 SAR data and SDGSAT-1 thermal infrared imagery
The DynIceData dataset was produced using GF-3, Sentinel-1 SAR, and SDGSAT-1 TIS data.Details of the data used are shown in Table 1.Because of the smaller swath width of the GF-3 SAR and SDGSAT-1 TIS data, GF-3 SAR and SDGSAT-1 TIS images of the Arctic MIZ that overlapped with Sentinel-1 images were selected.The data in the areas of overlap were then regarded as image pairs, and the time interval between the acquisition of the images forming each pair was calculated.

Sentinel-1 and GF-3 SAR data
Both the Sentinel-1 and Gaofen-3 satellites carry C-band synthetic aperture radar (SAR) instruments.Sentinel-1 is composed of two satellites, Sentinel-1A and Sentinel-1B, which share the same orbital plane and have a revisit period in the Arctic of 1-3 days (Guillaume et al., 2022).In this study, the EW-mode Level-1 Ground Range Detected (GRD) product was used because of its large-scale observation capability and reduced revisit interval.These data are HH/HV-dual polarized with a 40-m resolution and 410-km swath width and use the WGS 1984 geographic coordinate system.Gaofen-3 (GF-3) is the only radar imaging satellite in China's Gaofen Special Project and was launched in August 2016 (Zhang et al., 2017).It has 12 imaging modes, which is the most of any satellite currently operating.GF-3 can produce high-resolution imagery with a wide mapping band at the same time (Zhang et al., 2022).We used the standard stripmap (SS)-mode Level-2 product, which has a 25-m resolution and 130-km swath width and can obtain dual-polarized measurements.The selected Gaofen-3 and Sentinel-1 image pairs covered the period from August 2021 to December 2021; this constituted a transition period during which rapid changes in the sea ice occurred and the distribution of ice in the MIZ was typical.The selected imagery covered the Arctic sea ice edge, mainly within the Kara Sea.A total of 23 such pairs was collected; details are given in Appendix S1.

SDGSAT-1 thermal infrared data
SDGSAT-1 is the first scientific satellite dedicated to serving the 2030 Agenda and the first Earth science satellite developed by the Chinese Academy of Sciences.SDGSAT-1 was launched in November 2021 (Guo et al., 2023) and is equipped with a thermal infrared spectrometer (TIS), which is mainly used to detect the spatial distribution of surface heat radiation.The TIS operates in the three bands B1 (8-10.5 μm), B2 (10.3-11.3μm), and B3 (11.5-12.5 μm), of which B2 is the least affected by stripe noise.In this study, therefore, L4 B2-band data were used.These data have an imaging width of 300 km, a pixel resolution of 30 m, and are based on the UTM projection.SDGSAT-1 thermal infrared data provide daily coverage of the Arctic region; the area of overlap between SDGSAT-1 thermal infrared data and superwide Sentinel-1 SAR data is mainly located in western Greenland, the Kara Sea, the Barents Sea, and the Beaufort Sea.The shortest time interval between the acquisition of the two types of data is 3 hours.
To produce DynIceData, after overlapping samples of SDGSAT 1 thermal infrared data and Sentinel-1 SAR data were selected, cloudy images were removed.This resulted in 33 pairs of overlapping images covering the period from July to August 2022, mainly of Greenland and the Beaufort Sea.The time interval between the acquisition of the images in each pair ranged from 13 minutes to 68 hours.The selected image pairs are listed in Appendix S2.
Figure 2 shows a frequency histogram of all the image pairs with different acquisition time intervals.It can be seen that the number of image pairs with short time intervals (less than 10 hours) is the largest, and that there is a range of time intervals over which the two types of imagery complement each other.

Methods
After the Sentinel-1, GF-3 SAR, and SDGSAT-1 TIS images had been preprocessed and the areas of overlap between them had been clipped, ice-water classification of these areas was carried out and gridded 10 km × 10 km sample pairs were produced.From this, the final DynIceData dataset was obtained.Details of this process are shown in Figure 3.

Processing of GF-3 and sentinel-1 SAR data
The GF-3 Level-2 SAR data were radiometrically calibrated, and in-orbit mosaicking was carried to obtain the backscatter map.The formula used to calibrate the GF-3 data is shown as Equation ( 1), where σ 0 is the radar backscattering coefficient (units: dB), P l is DN 2  (DN being the image gray value), QV is the maximum gray value in the image before quantization, and K is the calibration constant: The processing of the Sentinel-1 EW mode data included radiometric calibration, incidence angle correction, and reprojection to produce data that geographically matched the Gaofen-3 data.The formula used to radiometrically calibrate the Sentinel-1 data is shown as Equation ( 2), where K is the calibration constant: Since the amount of backscattering from sea ice and seawater is linearly correlated with the incidence angle, the wide incidence angle of Sentinel-1 was corrected to the central incidence angle of Gaofen-3.The formula used for this is shown as Equation (3).α denotes the center of the GF-3 imagery's angle of incidence, k is the normalization coefficient, which depends on the backscattering coefficient of sea ice type, generally set as the experience value of 0.2 with HH-polarization SAR data or 0.025 with HV-polarization SAR data (Park et al., 2018(Park et al., , 2019)).x 1 is the incidence angle at a given pixel, y 1 is the backscattering coefficient corresponding to the pixel before correction, and y 2 is the backscattering coefficient corresponding to the pixel after correction: The Sentinel-1 imagery that overlapped with the GF-3 data was then reprojected to the UTM projection of the GF-3 data and resampled to a resolution of 25 m.The Sentinel-1 images that had been overlaid with SDGSAT-1 thermal infrared images were reprojected to the same UTM projection as the thermal infrared data and resampled to a resolution of 30 m.The overlapping areas were then clipped and a Lee filter was used for spot denoising of the SAR images to obtain the final image pairs.

Processing of SDGSAT-1 thermal infrared data
The processing of the SDGSAT-1 TIS data included radiometric calibration and temperature conversion (Hu et al., 2022).The formula used for the radiometric calibration was L = DN × Gain + Bias, where L is the radiance incident on the sensor aperture (units: W/m 2 /sr/ μm) and DN is the calculated DN value after relative radiometric calibration.For the B2 band, the gain has a value of 0.003946 and the bias is 0.124622.The DN values were converted to radiance values by using the radiometric scaling coefficient; the radiance was then converted into the in-orbit effective temperature or radiance temperature using the Planck equation, which is given as Equation (4).In this equation, T is the brightness temperature (K), h is Planck's constant (6.626 × 10 −34 J•s), c is the speed of light (2.9979 × 10 8 m/s), k is Boltzmann's constant (1.3806 × 10 −23 J/K), λ is the central wavelength (μm), and L is the spectral radiance at wavelength λ (W/m 2 /sr/μm):

Ice and water classification
The SAR imagery was classified using object-oriented segmentation and thresholding.The classification of the GF-3 and Sentinel-1 SAR image pairs as ice or water was mainly performed using HV-polarized data because HV-polarized SAR data are less sensitive to waves on open water than HH-polarized data and can be more easily used to detect the difference between ice and water.However, the stripe noise in areas of low backscatter (such as open water) in HV-polarized data is obvious.The impact of this noise can be reduced by using object-oriented segmentation.
The next step was the use of object-oriented segmentation to segment and merge the SAR images.Segmented images were generated based on the spatial, spectral, and texture properties of the data using the following steps (Jin, 2012): (1) computing a gradient map from the image, (2) computing a cumulative distribution function from the map, (3) modifying the map using the selected scale-level value, and (4) segmenting the modified map using a watershed transform.A watershed algorithm (Roerdink & Meijster, 2001) sorts pixels by increasing the greyscale values; the pixels with the minimum value and the regions of the image with similar pixel intensities are assigned to different segments.The result is a segmented image where each region is assigned the mean spectral value of all the pixels that belong to that region.A gradient image can be computed using a Sobel edge detection method, where the highest pixel values represent areas with the highest pixel contrast.When applying the watershed algorithm to the gradient image, the image is progressively flooded starting with the lowest gradient values (the uniform part of the objects) until the areas with the highest values (the edges) are reached.
The scale level was computed from the normalized cumulative distribution function (CDF) of the pixel values in the image.A scale level of 40 was selected, meaning that the lowest 40% of gradient values in the gradient image were discarded.This resulted in the edge features being well segmented.
A merging algorithm was used to evaluate the spectral similarity of the segments and to merge the segments with similar spectral attributes.The Euclidean distance between the segments was calculated using the Full Lambda Schedule method with a scale level of 90.The algorithm then iteratively merged adjacent segments based on a combination of spectral and spatial information.Merging occurred when the algorithm found a pair of adjacent regions, i and j, such that the merging cost t i,j was less than a defined threshold lambda value (Robinson et al., 2002): Here, O i is region i of the image, |O i | is the area of region i, u i is the average pixel value in region i, u j is the average value in region j, and u i À u j � � � � is the Euclidean distance between the spectral values of regions i and j. lengthð@ðO i ; O j ÞÞ is the length of the common boundary between O i and O j .
The merging algorithm merged large areas of open water and sea ice that had a strong texture and output the image segmentation map and vector data.A segmentation diagram was used to classify ice and water using a thresholding method.
The threshold segmentation method that was used took advantage of the obvious difference between the values of the backscattering coefficient of areas of ice and water in the SAR imagery.These areas were determined as being either ice or water based on a threshold value as shown in Equation ( 6).In this equation, f(x,y) is the backscattering coefficient of the SAR image pixel, T is the threshold used for the icewater segmentation, and g(x,y) is the pixel value after the ice-water segmentation.The threshold was selected based on visual interpretation and the features of the image histogram combined with the distribution of ice and water in the HH/HV or VV/VH dual-polarized imagery (Zhang et al., 2022).Thresholding was also carried out based on the different brightness temperatures of ice and water in the SDGSAT-1 TIS imagery.
Finally, manual correction of misclassified and missing points was performed based on visual interpretation and segmentation vector diagrams; a land mask was also applied.

Sample gridding process
The area of overlap for each image pair was extracted and a 10 km × 10 km "fishing net" grid laid over this area.Figure 4 shows the process of producing sample granules using the grid.The 10-km cell size was designed to be consistent with the 10-km spatial resolution of the EUMETSAT (European Organization for the Exploitation of Meteorological Satellites) OSI SAF (Ocean and Sea Ice) SIC product, which is the product with the highest accuracy in the MIZ and which can provide good-quality monitoring of sea ice in areas of low concentration (Huang et al., 2022).
The ice-water classification results of image pairs were used to be segmented into grid cells, the different image pairs corresponded to different time intervals.Finally, the gridded ice-water classification cells obtained from pairs of images of the MIZ with different time intervals were obtained.

Data record formats
The DynIceData dataset that was produced in this study is publicly available at the website https://doi.org/10.57760/sciencedb.j00001.00784.The dataset file consists of two folders named "25 m" and "30 m".The 25 m folder includes GF-3 and Sentinel-1 SAR image pairs with different time intervals.These images were acquired between August and December 2021 and have a spatial resolution of 25 m.The area covered by these image pairs mainly includes the Kara Sea; the folder consists of 5443 data with a total volume of 98.3 MB.The 30 m folder includes SDGSAT-1 thermal infrared and Sentinel-1 SAR image pairs with different time intervals.These images were acquired in July and August 2022 and have a spatial resolution of 30 m.The area covered consists mainly of the Greenland Sea and the Beaufort Sea.Table 2 shows the details of the DynIceData dataset covering different parts of the Arctic.
The folders contain the classified gridded ice-water data of the Arctic MIZ.The files are named according to the format "Area_TimeInterval_Date_Grid Number_SatelliteName_SpatialResolution.tif"; a further explanation of how the naming was carried out is given in Table 3.For example, the file "karaSea_1h20min_20210904_21_GF3_25m.tif" in the 25 m folder belongs to a sample pair along with the file named "karaSea_1h20min_20210904_21_S1_25m.tif".These file names indicate that the location is the Kara Sea, the interval between the image acquisition times is 1 h 20 min, and the date is September 4, 2021.The former file is the GF-3 sample, and the latter is the Sentinel-1 sample.The spatial resolution is 25 m.

Examples from the dataset
The DynIceData dataset has a resolution of 10 km × 10 km.It was created using SDGSAT-1 thermal infrared and Sentinel-1 SAR image pairs with a resolution of 30 m and GF-3 SAR and Sentinel-1 SAR image pairs with a resolution of 25 m.The dataset is in TIF raster data format and uses a UTM projection coordinate system.Figure 5 shows gridded sample pairs of GF-3 and Sentinel-1 images that were used for ice-water classification in the Kara Sea.These image pairs have different time intervals of 53 s, 1 h 20 min, 3 h, 20 h, and 48 h, from which the distribution and movement of the sea ice and the freeze-thaw conditions can be seen in each case.

Validation of ice and water classification
The minimum time interval between the acquisition of the GF-3 and Sentinel-1 images was 53 s, meaning that there was almost no change in the areas of ice and water between the two, so the image pair of Sentinel-1 and GF-3 with 53 s time interval was used to evaluate the classification results.The classification results of the image pair were obtained by using the threshold method are shown in Figure 6 (b) and (d).Reference classification results were obtained by applying the k-means method combining the GF-3 and Sentinel-1 imagery.Figure 6 (f) shows the characteristics of ice and water clustering.
To quantify a comparison of the results of the GF-3 and Sentinel-1 classification with that of the k-means classification, the kappa coefficient and confusion matrix were used.As shown in Tables 4 and 5, the Sentinel-1 results have an overall accuracy of 98.20% and a kappa coefficient of 0.96; the GF-3 results have an overall accuracy of 95.58% and a kappa coefficient of 0.89.These results show that applying the threshold method to the GF-3 and Sentinel-1 images produced results with a very high accuracy.
Figure 7 shows the results of the classification of the SDGSAT-1 TIS imagery.The Sentinel-1 backscattering coefficient image for the same area is shown in Figure 7 (c); the time interval between the two images is 3 h.It can be seen from Figure 7 (b) that, by applying thresholding method to the SDGSAT-1 data, good classification of the ice and water was obtained; the ice edge is also well distinguished.This indicates that 30-m SDGSAT-1 data, which has the highest resolution of currently available thermal infrared imagery, can produce reliable ice-water classification results.6. Analysis of sea ice changes and dynamics

Fine spatial differences in multi-temporal sea ice classification
The DynIceData dataset has a higher spatial resolution (25 m or 30 m) and a shorter time interval between images (minutes to hours) than the kilometer-level daily SIC products.DynIceData is thus able to show fine sea ice features such as small melt pools and polynyas (with sizes of the order of meters) as well as rapid changes in the ice, including freezing, thawing, and movement of the ice.The spatial distribution of sea ice changes over different time intervals detected using GF-3 and Sentinel-1 image pairs is shown in Figure 8.These results were obtained by subtracting the sea ice and water classification results obtained from the earlier image from that of the later image.Changes such as movement of the ice and freezing and melting can be seen.
In the results shown in Figure 8, these changes occurred mainly at the sea ice edge, the edges of ice floes, in ice leads, and in melt pools.In the case of the 2-min time interval, no movement of the sea ice can be seen and the morphology of the edge area is unchanged; the only difference is at the edges of the ice floes.In the case of the 1-h time interval, within the dense concentration of large ice floes in the northern part of the image, there has again been hardly any movement; however, some northward movement of the ice can be seen in the southern part of the image where the ice flows are relatively small and scattered.For the image pair with a time interval of 3 h, cracking of the ice leads has occurred and northward movement of the sea ice can be seen; the land-fast ice has also begun to freeze.In the case of the 20-h time interval, the sea ice has moved a large distance.These results show that the sea ice that was some distance from the land had an obvious tendency to move away from the land, the ice floes were moving northward, and the land-fast ice frozen.The difference images shows that the displacement of sea ice increases with the time interval, which indicates that, over short time intervals, the motion of sea ice is mainly affected by dynamic factors.

Characteristics of changes in sea ice area over different time intervals
Figure 9 shows the changes in the total area of sea ice over different time intervals that were derived from gridded GF-3 and Sentinel-1 image pairs of the Kara Sea.The total change in area was calculated based on the area of sea ice gain and sea ice loss in each grid cell.The red line in the graph shows that the total change in sea ice area increases with the time interval.In the second column of Figure 10, the sea ice in the red area is drifting quickly under the influence of wind and ocean currents.The sea ice in this area is also susceptible to the interaction between waves and sea ice (Kousal et al., 2022); this can cause break-up of the sea ice and lead to large changes in ice movement within a short time (Manucharyan & Andrew, 2022).

Relationship between sea ice change and SIC at different time intervals
In second column of Figure 11, the sea ice marked in red is in a region of high SIC.In this case, the change in the SIC is a result of the sea ice colliding and being squeezed together.Thermodynamic factors also play a part as there is a strong exchange of heat and water (Yang et al., 2023); for example, the opening and closing of ice leads is affected by the temperature.Small-scale changes such as the development of melting pools and freezing will affect the total change in the ice area.
By determining the SIC that corresponds to the greatest change in the sea ice area, the location of the sea ice edge can be estimated.This can then be used to predict further subsequent changes in the SIC and sea ice area.

Discussion and applications
The DynIceData dataset provides standardized classified images of ice and water in the Arctic MIZ based on a 10 km × 10 km grid.In this study, the use of SAR data acquired by China's Gaofen-3 satellite and SDGSAT-1 thermal infrared data in combination with Sentinel-1 SAR data allowed ice monitoring with a high spatial (25 m and 30 m) and temporal resolution to be carried out.For the first time, the temporal resolution of this monitoring was as little as minutes or hours.Thus, the dataset that was produced provides a new means of studying rapid changes in polar sea ice.The DynIceData dataset can be used to study these rapid changes over long or short time intervals and to analyze changes in sea ice movement over different time intervals.Changes in the distribution of sea ice, freezing and thawing, and the movement of ice in key areas of the Arctic can also be studied.This provides support for the analysis of the relationship between sea ice and environmental factors and the study of sea ice patterns.As DynIceData consists of basic data on sea ice dynamics, it can also be used to support the development of dynamic and thermodynamic models of sea ice.
Because of its high spatiotemporal resolution, DynIceData provides a basis for the study of sea ice dynamics and thermodynamics as well as the inversion from the DynIceData and predictions of sea ice that have a high spatiotemporal resolution.Since current high spatiotemporal resolution data of the Arctic are spatially and temporally discontinuous, DynIceData could be used for the inversion of sub-kilometer SIC across the whole Arctic based on AI and physical models, thus providing high-resolution sea ice observations that make inversion of sub-kilometer SIC over the whole Arctic possible.The data could also be used to limit and constrain AI models and improve the accuracy and reliability of data inversion.A further use of DynIceData could be as an initial high-precision data field for sea ice forecasts as well as sample data for use in short-term sea ice forecasts with a high spatiotemporal resolution based on artificial intelligence.DynIceData could also be combined with SDGSAT-1 thermal infrared data, optical data, buoy data, and environmental data to form a rich, high spatial and temporal resolution sea ice data set for improved monitoring of changes in sea ice in the MIZ, sea ice inversion, and sea ice forecasts.Finally, the dataset could be used to provide information on the distribution of sea ice and to make predictions for use by commercial shipping sailing in the sea ice margin.This would help to mitigate the risks that sea ice poses to shipping and provide data for the real-time monitoring of sea ice in the Arctic as well as for other applications related to navigation risks and route planning.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Figure 1 .
Figure 1.Definition of the Arctic marginal ice zone based on an SIC of between 0% and 70%.

Figure 2 .
Figure 2. Frequency histogram of image pairs with different time intervals.

Figure 3 .
Figure 3. Flow chart illustrating the construction of DynIceData.

Figure 4 .
Figure 4. Details of the sample gridding process.

Figure 5 .
Figure 5. Examples of 10-km gridded sample pairs with different time intervals in DynIce Data showing ice-water classification in the MIZ.

Figure 6 .
Figure6.Classification results for GF-3 and Sentinel-1 images together with the reference results obtained using the k-means method.

Figure 8 .
Figure 8. Spatial distribution of changes in sea ice detected using image pairs with different time intervals.

Figure 10 and
Figure 10 and Figure 11 show the relationship between the SIC and change in sea ice area derived from the gridded GF-3 and Sentinel-1 image pairs of the Kara Sea with different time intervals.The SIC was obtained from classified ice-water images based on the earlier image in each image pair using the ratio of the sea ice area within each grid cell to the total area of the cell.It can be seen from the scatter plots in the first column of Figures 10 and 11 that the change in the area of sea ice first increases and then decreases as the SIC increases.The shape of the curves is close to that of a parabola, and the peak values, representing the maximum change in sea ice area, occur at an SIC of between 40% and 60%.The greater the time interval, the greater change in the sea ice area.

Figure 9 .Figure 10 .
Figure 9. Relationship between time interval and total change in ice area in the Kara Sea.The number of grid cell pairs is 315 at 53 s, 316 at 18 min, 241 at 52 min, 569 at 3 h, 238 at 20 h, and 239 at 45 h.

Figure 11 .
Figure 11.The relationship between the SIC and total change in sea ice area over different time intervals in a region of high SIC.

Table 2 .
Details of the DynIceData dataset covering different parts of the Arctic.

Table 3 .
Dataset naming format and description.