Sea ice drift vector extraction based on feature matching using CS-1 Images

The study emphasizes the crucial role of sea ice drift in Arctic climate research and the protection of human activities. Employing C-SAR/01 (CS-1) imagery as the primary data source, the study assesses the effectiveness of the ORB feature matching algorithm for extracting Arctic Sea ice drift vectors. Additionally, it investigates the spatial distribution variations in the drift vectors extracted from the HH and HV polarization channels. The accuracy of the extracted drift vectors is validated using manually extracted sea ice drift data. Experimental results reveal a significant disparity in the number of drift vectors extracted from CS-1 HV polarization images compared to HH polarization images. The sea ice drift vectors extracted using the Oriented fast and Roasted Bried (ORB) operator demonstrate an average velocity error of less than 0.27 cm/s (0.31 km/d) and an average direction error of less than 5.66° in HV polarization images.


Introduction
Ice drift speed is an important parameter of sea ice, playing a significant role in navigation, offshore operations, sea ice model validation, and climate model improvement in ice-covered regions [1] .Satellite remote sensing is widely used for observing sea ice concentration due to its advantages of wide coverage, short revisit period, and relatively low cost.The sensors used for sea ice drift monitoring can be divided into two categories: optical and microwave sensors.Optical sensors, such as Advanced Very High-Resolution Radiometer and Moderate-Resolution Imaging Spectroradiometer, have been extensively used in sea ice monitoring [2][3] .However, in polar regions, harsh weather conditions and cloud cover affect the imaging quality of optical sensors.To extract sea ice information from optical data, cloud-covered areas need to be masked, significantly reducing the retrievable area, which is not ideal for practical applications.Microwave sensors, on the other hand, can penetrate clouds and are not affected by weather conditions, allowing for all-day, all-weather operations.Thus, microwave sensors have become the primary sensors for sea ice monitoring [4][5] .Based on their working principle, microwave sensors can be passive or active.Early sea ice information extraction was based on passive microwave sensors using microwave radiometer data, exploiting differences in ice emissivity to deduce the spatial distribution and calculate the vector of sea ice drift.Currently, internationally representative passive microwave remote sensing sea ice drift products include those from OSISAF and CERSAT-IFREMER, while the Polar Pathfinder based on passive microwave data is widely used for model validation and data assimilation studies [6][7] .Recent studies have shown that algorithm improvements and fusion techniques within the Arctic Ocean region can enhance the accuracy of sea ice drift products, and optimizing algorithms can greatly improve the rationality of ice velocity probability distribution [8] .
Currently, there are two main research methods for sea ice drift vector monitoring: regional tracking based on pattern matching methods and point feature tracking based on feature tracking methods.Template matching algorithms are sea ice simple and have high calculation accuracy [9] , but they are not resistant to rotations of sea ice, sensitive to image noise, and have low computational efficiency [10] .The sea ice drift products extracted from radiometers and scatterometers mentioned earlier are based on template matching algorithms.Feature matching algorithms were first proposed by Lowe [11] .These algorithms detect feature points in the main and auxiliary images, describe them with high-dimensional descriptors (such as corner points and extremal points), and match the feature points in the main and auxiliary images based on the descriptors.Feature tracking methods are not sensitive to image matching problems with translations, rotations, and affine transformations, and can overcome the issue of drift detection for rotating sea ice, which is the most critical advantage compared to pattern matching methods.The ORB operator, for example, uses the subsampling method to generate low-resolution images for scale invariance and employs the FAST algorithm to mark feature points in the image, leading to improved computational efficiency [12] .ORB operator was used to extract sea ice drift vectors from Sentinel-1 SAR images, enhancing operational efficiency [13][14][15] .However, the matching accuracy of this method is often low.Demchev proposed a method for filtering out the erroneous sea ice drift vectors formed by the remaining mismatched point pairs.This method utilizes a comparison among the vectors in the neighborhood and achieves good filtering results [16] .
Gaofen-3-02 satellite, CS-1 was launched on November 23, 2021.It has 12 operating modes and uses various imaging modes such as strip mode, spotlight mode, and TOPSAR.It also has dedicated modes, such as wave mode and global observation mode, designed for typical marine applications.The SAR payload imaging of CS-1 has a spatial resolution ranging from 1 to 500 meters and a swath width of 10 to 650 kilometers.It offers single-polarization, dual-polarization, and quadraturepolarization capabilities.Once GF-3-02 satellite is in orbit, it will join the existing Gaofen-3-01 (GF-3) satellite to form a constellation.This constellation will effectively decrease revisit intervals and meet the immediate requirements of users in industries such as marine, disaster management, meteorology, and water resources, by facilitating swift observations of land and sea.Additionally, the constellation will enable the utilization of interferometry techniques in different orbits and provide crucial support for the operational use of polarimetric SAR satellites in China.Although Gaofen-3 constellation SAR imagery has high resolution, only a few research works have applied its data [17][18] .
Hence, this paper employs the novel TOPW mode on CS-1 to perform sea ice drift extraction, catering to the quick and efficient requirement of observing sea ice movement.This advancement plays a crucial role in supporting the operational application of SAR satellites in China, thus providing vital backing.This work compares the performance of the ORB feature matching algorithm in extracting sea ice drift vectors from the HH and HV polarization channels of CS-1 imagery.The algorithm proposed by Li et al. is applied to improve the filtering efficiency of erroneous ice vectors while retaining correct vectors [19] .The accuracy of sea ice drift vectors extracted from CS-1 imagery is validated based on manually extracted data.

SAR image and preprocessing
The CS-1 satellite is a high-resolution, fully SAR satellite that features 12 distinct imaging modes in its design.In this study, we utilized the TOP Wide swath scan mode (TOPW) images from the CS-1 L2 dataset acquired on April, 2023, in the Arctic region.These images have a swath width of 500 km and a spatial resolution of 100 m, offering both HH and HV polarization channels.The two images provide extensive coverage with a time interval of approximately 2 days, allowing for the calculation of sea ice concentration.Table 1 presents detailed information of the experimental images, while Figure 1 illustrates the spatial locations of the images and the distribution of the manually extracted points within the images.In this study, we preprocessed the SAR image into a grayscale image ranging from 0 to 255 through the following procedure.
(1) Radiation correction: Equation (1) was used to perform radiation correction on CS-1 images.A backscatter coefficient was obtained for each pixel Where  0 is the corrected backscatter coefficient, and QualifyValue and KdB are the quality value and calibration constant, respectively; And   =DN 2 , DN is magnitude of CS-1 image.
(2) Multilook processing and speckle filtering: multilook processing can be used to acquire images at a specified resolution while suppressing speckle noise in images.In this study, the spatial domain average method was used for multilook processing.To further reduce the effects of speckle noise, Lee filtering was performed on the images after multilook processing.
(3) Regularization: The filtered σ0 image in this study was regularized through percentage truncation and stretching based on the following Equation ( 2

Extraction of sea ice drift vectors based on feature matching algorithm
The main steps of the ice drift vector extraction algorithm based on feature matching are as follows: (1) The earlier image in the image pair is selected as the reference image (Image 1), while the later image is considered as the target image (Image 2).ORB feature operators are employed to extract feature points from both the HH polarized channel and HV polarized channel images of each pair after pre-processing.
(2) The feature points extracted from the HH polarized channel image of the reference image are matched with the corresponding feature points from the HH polarized channel image of the target image.Sea ice similarly, the feature points from the HV polarized channel image of the reference image are matched with the respective feature points from the HV polarized channel image of the target image.The FLANN (Fast Library for Approximate Nearest Neighbors) matcher can be utilized for matching the feature points between the reference and target images.FLANN is an efficient algorithm for fast nearest neighbor search [20] .
(3) After obtaining the matched point pairs, the geographical coordinates are obtained based on the image coordinates.Combined with the imaging time interval, the sea ice drift vectors can be derived.Finally, the erroneous vectors are filtered out to ensure the accuracy of the results.

Error drift vector filtering algorithm
The Nearest Neighbor Distance Ratio test (NNDR) is commonly used in the feature tracking algorithm to retain correct matching point pairs and filter out erroneous ones.The NNDR test examines the ratio between the distance of the closest match and the distance of the second closest match.If the second closest distance is close to the closest distance, it indicates that the closest distance match is unreliable.Empirically, a distance ratio lower than 0.7 can filter out at least 90% of erroneous matches at the cost of 5% correct matches.Please note that the NNDR test is used as a reference criterion to select correct matching point pairs by considering the distance ratio between the closest and second closest matches.The NNDR test relies on the correlation between the descriptors of feature points and is independent of the characteristics of ice motion.Therefore, even after the NNDR test, a certain number of erroneous vectors may still remain.In response to the above issues, this study uses the filtering method.If the difference between a vector component and the mean value of all vector components exceeds 2 times the standard deviation (empirical value), it is considered incorrect match.Additionally, the surrounding vectors within a 10 km radius are examined for each vector.If there are 8 or more surrounding vectors, and the difference between the vector component and the mean value of these vectors does not exceed one standard deviation, it is considered a correct match.Otherwise, it is determined as an erroneous match.

Using manually extracted sea ice drift vectors as reference data：
The accuracy of the sea ice drift vectors extracted from the CS-1 images was investigated using manually extracted sea ice drift vectors (extracted by eye-catching feature tracking using the ArcGIS software) as reference data.The Root Mean Square Error (RMSE) in sea ice drift speed and sea ice drift direction were used as indicators of validation accuracy.Since the starting positions of the reference vectors differ from the vectors extracted from the CS-1 images, matching is required.The nearest neighbor method was employed for pairing.

Comparative analysis of different polarization channels
Figure 2. illustrates the distribution of sea ice drift vectors obtained using the ORB feature matching algorithm with different almost channels of the two images.The results from the HH-polarized and HV-polarized were combined and displayed, and all sea ice drift vector results underwent NNDR testing.
In each image matching, the maximum number of retained features for the ORB operator was set to 1000,000.From the figure, it can be observed that the HV-polarized (15,877 points) yields the highest number of vectors, while the HH-polarized (3045 points) extracts fewer feature points.Additionally, the sea ice drift vectors acquired from varying numbers of ORB measurements demonstrate a consistent spatial distribution pattern on the HV polarization.It can be observed that there is a certain difference in the number of sea ice drift vectors obtained from different polarization channels.Generally, the HV polarization channel image yields a higher number of sea ice drift vectors compared to the HH polarization.The abundant grayscale information in the images contributes to the enhanced detectability of sea ice features, thereby influencing the reliability of sea ice drift vectors.In this study, we opted to evaluate the degree of grayscale information richness in the images based on information contrast, which is insensitive to noise.
Figure 3. shows the information contrast calculated for the HH-polarized and HV-polarized images with 250 m resolution acquired in TOPW mode.Obviously, The HV-polarized images acquired have a higher mean information contrast and therefore richer grayscale information than the corresponding HH-polarized images.The characteristics of the information contrast of the HH-polarized and HVpolarized images are consistent with the reliability of the corresponding extracted sea ice drift, suggesting that the rich grayscale information in the images is responsible for the higher reliability of the sea ice drift data obtained from the HV-polarized images in TOPW mode.

Validation of extracted sea ice drift accuracy and spatial distribution analysis
Based on the findings in section 3, it can be deduced that the number of feature points extracted using HH polarization is limited, and the contrast in grayscale information is relatively low.Consequently, for the subsequent analysis, our focus will solely be on the results obtained from HV polarization.
The ORB method tends to generate a considerable number of incorrect vectors.However, by applying a filtering method, a greater number of accurate vectors can be retained.The filtering process removes variables with excessively large values or significant disparities in direction.Figure 4 illustrates the filtered outcomes of the algorithm, primarily for visual representation purposes.It can be observed from Figure 4 that the proposed algorithm utilizes the distribution characteristics of vector components, which closely resemble a normal distribution.Initially, the algorithm selects the correct vectors by calculating the confidence interval for high-speed vectors.Afterwards, for further verification, a computationally efficient neighbourhood comparison method is employed.The direction of sea ice drift demonstrates a consistent pattern with evident regional variations.In this research, the accuracy of drift vectors' velocity and direction obtained through the feature matching algorithm was compared to the reference data obtained manually by extracting target information.The outcomes are presented graphically in Figure 5.As depicted in Figure 5, the proposed method exhibits a high level of consistency with the reference buoys, indicating a strong agreement between the sea ice drift vectors extracted using the algorithm and the feature matching algorithm.The velocity error is approximately 0.27 cm/s (0.31 km/d), while the angular error between the two is 5.66°.This signifies that the extraction of sea ice drift vectors using the HV polarization in CS1 TOPW mode exhibits a high level of reliability with this approach.

Summary
The ORB feature matching algorithm efficiently extracts high-resolution and high-precision sea ice drift vectors from CS-1 TOPW mode images.In this study, a comprehensive investigation of the sea ice drift vector extraction technique based on the feature matching algorithm was conducted, leading to the following conclusions: (1) The CS-1 HV polarization images yield a greater number of sea ice drift vectors compared to the HH polarization images, which can be attributed to the richer grayscale information present in the HV polarization.
(2) This study applies the confidence interval filtering method and neighbourhood comparison filtering method.The filtering algorithm could rapidly filters out incorrect vectors.
(3) The validation of the sea ice drift vector extraction algorithm using manually extracted data indicates that the sea ice drift vector field extracted from CS-1 images exhibits high accuracy in terms of velocity and direction.
These points demonstrate that by utilizing CS-1 images, combining them with the ORB feature matching method, and applying appropriate error matching elimination techniques, it is possible to effectively and rapidly generate high-resolution Arctic Sea ice drift fields.This approach can provide accurate and detailed data support for environmental and climate studies in the Arctic region and serve as a data source for Arctic shipping safety.

Figure 1 .
Figure 1.Study area (The red box represents Image 1, the blue box represents Image 2, and the blue dots represent reference points).

Figure 3 .
Figure 3. Information contrast of HH-polarized and HV-polarized images acquired in TOPW mode.

Figure 4 .
Figure 4. distributions of the sea ice drift vector.

Figure 5 .
Figure 5. Assessment of the accuracy of extracted results using manually extracted sea ice drift vectors as reference data.(a) HV-polarized speed; and (b) HV-polarized direction; And black solid line represents the one-to-one line.

Table 1 .
Detail information of SAR images 3