A sea ice extraction method based on multi-feature superpixel segmentation

Abstract

The application provides a sea ice extraction method based on multi-feature superpixel segmentation, comprising the following steps: step 1, obtaining a dual-polarization SAR image of a target sea area and performing image preprocessing to obtain a preprocessed image; step 2, extracting a backscattering coefficient feature, a normalized feature, a morphological feature set, an OTSU feature and a texture feature from the preprocessed image and fusing them to obtain a multi-dimensional data feature set; step 3, using an SNIC superpixel segmentation method to segment the multi-dimensional data feature set to obtain a superpixel block set; and step 4, inputting the superpixel block set into a supervised training machine learning classifier to classify sea water and sea ice to extract sea ice. The application can effectively improve the sea ice classification precision and robustness, can generate a sea ice distribution feature set with a spatial resolution of 40 meters, and can effectively support the monitoring of the interannual and intermonthly changes of the Arctic sea route sea ice.

Description

Technical Field

[0001] This invention relates to a sea ice extraction method based on multi-feature superpixel segmentation, belonging to the field of remote sensing image processing technology. Background Technology

[0002] Currently, sea ice information extraction in the Arctic shipping route mainly relies on optical remote sensing imagery or single-feature SAR data classification methods. These methods have significant technical limitations when dealing with the complex and ever-changing ice conditions in the region. Traditional methods based on optical remote sensing imagery heavily depend on the reflectivity differences between sea ice and seawater in specific bands (such as near-infrared). Although such methods can identify sea ice under certain conditions, their feature sets are relatively limited, and they are constrained by frequent cloud cover and polar night phenomena in the Arctic region, making continuous and effective all-weather monitoring difficult and failing to meet the high reliability requirements of sea ice information for navigation safety in the Arctic shipping route.

[0003] In comparison, single-feature SAR data classification methods have the advantage of all-day, all-weather imaging, but still have several shortcomings. First, the inherent speckle noise of SAR images severely interferes with the extraction of sea ice texture information, resulting in low classification accuracy based on a single backscattering feature. Second, the wide-scan characteristics of SAR systems cause backscattering intensity to vary significantly with the incident angle, leading to large differences in the appearance of the same ground feature under different imaging geometries, thus increasing the difficulty of accurately distinguishing sea ice from open water based on a single feature. Finally, while traditional threshold-based image segmentation methods are highly efficient in simple scenarios, their segmentation results are easily affected by image brightness and noise, and they are highly dependent on manually set thresholds, exhibiting poor generalization ability across different images and making it difficult to adapt to the complex and variable sea ice environment in the vast area of ​​the Arctic shipping route.

[0004] SAR images provide scattering features that effectively reflect the backscattering of radar beams, thus demonstrating advantages in describing the relationship between an object's surface and its surrounding environment. Gray-level co-occurrence matrix (GLCM), as an effective texture analysis tool, accurately characterizes the texture structure of an image by describing the spatial joint distribution of pixel gray values, including properties such as directionality, spacing, and amplitude, and is therefore widely used in remote sensing image analysis. However, although GLCM can effectively characterize the texture features of sea ice images, the multidimensional texture feature sets it generates are often high-dimensional and contain information redundancy. Directly applying these features may lead to low computational efficiency in classification models and susceptibility to overfitting.

[0005] In recent years, machine learning algorithms have been widely applied to sea ice classification tasks. However, the feature combinations used by these algorithms are often unoptimized and fail to systematically integrate multimodal features that comprehensively reflect the physical properties and spatial information of sea ice. Therefore, the accuracy and robustness of existing methods in complex scenarios still fall short of the high accuracy requirements for sea ice classification in Arctic shipping safety. Especially under complex weather and imaging conditions, how to achieve high-precision, high-resolution, and highly robust automated classification of sea ice and open water remains a core technical problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a sea ice extraction method based on multi-feature superpixel segmentation to solve the problems existing in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention provides a sea ice extraction method based on multi-feature superpixel segmentation, including: Step 1, acquiring dual-polarization SAR images of the target sea area and performing image preprocessing to obtain preprocessed images;

[0008] Step 2: Extract backscattering coefficient features, normalization features, morphological feature set, OTSU features, and texture features from the preprocessed image and fuse them to obtain a multidimensional data feature set;

[0009] Step 3: Use the SNIC superpixel segmentation method to segment the multidimensional data feature set to obtain a set of superpixel blocks;

[0010] Step 4: Input the superpixel block set into a supervised machine learning classifier to classify seawater and sea ice in order to extract sea ice.

[0011] In one specific implementation, the dual-polarization SAR image is a dual-polarization SAR image based on the HH and HV dual-polarization bands obtained by Sentinel-1 in EW imaging mode, where both HH and HV are backscattering characteristic bands.

[0012] In one specific implementation, the texture feature extraction method is as follows: Principal component analysis (PCA) is used to reduce the dimensionality of the GLCM texture features of the preprocessed image, and the three texture features with the highest contribution rates are selected: contrast, correlation, and variance.

[0013] In one specific implementation, the standardized feature is a corrected backscattering coefficient feature, which is obtained by correcting the backscattering coefficient feature of the preprocessed image using incident angle normalization, specifically as follows: Multiple pixels are randomly selected in the preprocessed image, and the following steps are performed on each randomly selected pixel to obtain the corresponding corrected backscattering coefficient value. The corrected backscattering coefficient values ​​of all randomly selected pixels constitute the corrected backscattering coefficient feature of the preprocessed image; Step a: Extract the backscattering coefficient value and incident angle value of the pixel, and construct a linear fitting function, which is: Step b: Construct the inverse function of the linear fitting function using the average incident angle. The inverse function of the linear fitting function is: Step c: Construct the backscattering coefficient equation after normalizing the incident angle to obtain the corrected backscattering coefficient value at the incident angle θ. The equation for the backscattering coefficient after normalizing the incident angle is: Where a and b are the slope and intercept of the linear fitting function, respectively. This represents the backscattering coefficient value when the incident angle is θ. yes The inverse function of θ ref It is the average incident angle, θ ref The value is 33 ° , This is the corrected backscattering coefficient value.

[0014] In one specific implementation, the morphological feature set is obtained by the following method: calculating the erosion features, dilation features, opening operation features, closing operation features, top cap transformation features, and bottom cap transformation features of the preprocessed image based on the GEE platform; calculating the contribution rate of each morphological feature of the preprocessed image using principal component analysis (PCA); and selecting the five morphological features with the highest contribution rates to form the morphological feature set. The five morphological features with the highest contribution rates are: erosion features, dilation features, opening operation features, closing operation features, and bottom cap transformation features.

[0015] In one specific implementation, the OTSU feature is a binary mask of the preprocessed image based on the OTSU algorithm.

[0016] In one specific implementation, the machine learning classifier is a random forest classifier.

[0017] In one specific implementation, the machine learning classifier is pre-trained with supervised training using labeled sample data before being used for actual classification; the labeled sample data is randomly collected from dual-polarization SAR images that have completed visual interpretation of sea ice, and the visual interpretation results are used as ground truth labels.

[0018] In one specific implementation, the parameter settings of the machine learning classifier during supervised training are dynamically adjusted through a preset parameter combination. The adjusted parameters include a classifier type selection strategy, a model training iteration control method, and an accuracy optimization mechanism, selecting the optimal parameter combination that achieves the best classification performance.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0020] This invention extracts and fuses backscattering coefficient features, normalized features, morphological feature sets, OTSU features, and texture features to obtain a multidimensional data feature set. The SNIC superpixel segmentation method is then applied to this multidimensional data feature set to obtain a set of superpixel blocks. This set of superpixel blocks is input into a supervised-trained machine learning classifier to classify seawater and sea ice. This invention effectively improves the accuracy and robustness of sea ice classification and can generate a sea ice distribution feature set with consistent spatial resolution (40 meters), effectively supporting the monitoring of interannual and monthly variations of sea ice in the Arctic shipping route. The core design and innovative functions of this invention include the following aspects.

[0021] 1. Multimodal feature fusion strategy: Breaking through the limitations of traditional sea ice classification that relies on single or a small number of features, it creatively fuses the backscattering coefficient features, normalized features, morphological feature set, OTSU features and texture features of the preprocessed image to construct a high-dimensional, complementary multi-dimensional data feature set, which fundamentally improves the separability of sea ice from open water.

[0022] 2. Correction of backscattering coefficient characteristics of dual-polarization SAR images by incident angle normalization: An incident angle normalization algorithm based on a cosine correction model is proposed to correct the backscattering coefficient characteristics of satellite wide-scan mode images. The algorithm uses 33° as the set average incident angle, which effectively eliminates grayscale distortion caused by incident angle changes in wide-scan dual-polarization SAR images and improves the consistency and robustness of features.

[0023] 3. PCA-based texture feature dimensionality reduction and optimization method: To address the issues of high dimensionality and redundancy in GLCM texture features, principal component analysis (PCA) is used to reduce the dimensionality of GLCM texture features in the preprocessed image. The three texture features with the highest contribution rates are selected: contrast, correlation, and variance, which improves classification efficiency while preserving key information.

[0024] 4. Cooperative processing of superpixel segmentation and feature fusion: The SNIC superpixel segmentation method is used to perform superpixel segmentation on the multidimensional data feature set, which improves the classification unit from pixels to regions, effectively suppresses salt-and-pepper noise in SAR images, and ensures regional consistency and boundary smoothness of classification results. Attached Figure Description

[0025] Figure 1 This is a flowchart of a sea ice extraction method based on multi-feature superpixel segmentation.

[0026] Figure 2 This is a comparison of sea ice classification results between the experimental group and the control group in Example 1 of the present invention based on summer images. (a) is the original image, (b) is the classification result image of control group 1, (c) is the classification result image of control group 2, (d) is the classification result image of control group 3, and (e) is the classification result image of the experimental group.

[0027] Figure 3 This is a comparison of sea ice classification results between the experimental group and the control group in Example 1 of the present invention based on winter images. (a) is the original image, (b) is the classification result image of control group 1, (c) is the classification result image of control group 2, (d) is the classification result image of control group 3, and (e) is the classification result image of the experimental group.

[0028] Figure 4 This is a comparison chart of the sea ice classification results of the experimental group in Example 2 of the present invention and the sea ice concentration data. Detailed Implementation

[0029] The present invention will now be described in detail with reference to the embodiments and accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0030] refer to Figure 1 A method for extracting sea ice using multi-feature superpixel segmentation includes: Step 1, acquiring dual-polarization SAR images of the target sea area and performing image preprocessing to obtain preprocessed images.

[0031] The image preprocessing includes one or more combinations of filtering, cropping, stitching, geometric correction, radiometric correction, atmospheric calibration, cloud masking, and land masking. Land masking is used to remove land areas from the image, reduce the influence of irrelevant features, and narrow the research scope; while cloud masking is used to remove cloud interference. The specific image preprocessing method needs to be selected based on the actual remote sensing image quality.

[0032] Step 2: Extract backscattering coefficient features, normalization features, morphological feature set, OTSU features, and texture features from the preprocessed image and fuse them to obtain a multidimensional data feature set. The multidimensional data feature set can fundamentally improve the separability of sea ice and open water.

[0033] Specifically, the dual-polarization SAR image is a dual-polarization SAR image based on the HH and HV dual-polarization bands obtained by Sentinel-1 in EW imaging mode. Both HH and HV are used as backscattering coefficient characteristic bands, which can reduce the image of polar night phenomena and clouds and fog, thereby optimizing the accuracy of sea ice extraction.

[0034] Preferably, the texture features include contrast, correlation, and variance. Specifically, the Gray-Level Co-occurrence Matrix (GLCM) effectively characterizes the texture structure of an image by describing the joint spatial distribution of pixel gray levels, including attributes such as directionality, spacing, and amplitude. However, the multidimensional texture feature set it generates suffers from high dimensionality and information redundancy, leading to low computational efficiency and susceptibility to overfitting in the classification model if used directly. Therefore, this invention employs Principal Component Analysis (PCA) to reduce the dimensionality of the GLCM texture features of the preprocessed image, selecting the three texture features with the highest contribution rates for classification. These three features are contrast (CON), correlation (COR), and variance (VAR). This approach preserves the most discriminative texture information while improving the robustness of the feature set and the computational efficiency of the classification model.

[0035] The formula for calculating contrast ratio CON is: .

[0036] The formula for calculating the correlation coefficient (COR) is: .

[0037] The formula for calculating variance VAR is: .

[0038] Where G represents the number of gray levels, p(i,j) represents the probability of a pixel pair with gray levels i and j occurring at a specific distance and direction in the image, x and y represent two classes of adjacent pixels, µ is the mean of p(i,j), and n is the absolute value of the gray level difference, representing the magnitude of the gray level difference between pixel pair i and j. x This represents the expected value of the grayscale value in the horizontal direction, reflecting the average level of the grayscale level in the horizontal direction, µ. y This represents the expected value of the grayscale values ​​along the column direction, reflecting the average level of grayscale levels in the vertical direction. x Indicates the dispersion of gray values ​​along the row direction, measuring the degree of deviation from the mean distribution, ɛ y It indicates the degree of dispersion of gray values ​​in the column direction, and measures the degree of deviation from the mean distribution.

[0039] Specifically, the standardized feature is the corrected backscattering coefficient feature, which is obtained by correcting the backscattering coefficient feature of the preprocessed image using incident angle normalization, as follows: Multiple pixels are randomly selected in the preprocessed image, and the following steps are performed on each randomly selected pixel to obtain the corresponding corrected backscattering coefficient value. The corrected backscattering coefficient values ​​of all randomly selected pixels constitute the corrected backscattering coefficient feature of the preprocessed image; Step a: Extract the backscattering coefficient value and incident angle value of the pixel, and construct a linear fitting function. The linear fitting function is: Step b: Construct the inverse function of the linear fitting function using the average incident angle. The inverse function of the linear fitting function is: Step c: Construct the backscattering coefficient equation after normalizing the incident angle to obtain the corrected backscattering coefficient value at the incident angle θ. The equation for the backscattering coefficient after normalizing the incident angle is: Where a and b are the slope and intercept of the linear fitting function, respectively. This represents the backscattering coefficient value when the incident angle is θ. yes The inverse function of θ ref It is the average incident angle, θ ref The value is 33 ° , This is the corrected backscattering coefficient value.

[0040] The incident angle refers to the angle formed between the radar beam and the vertical line to the ground. For Sentinel-1 EW imaging mode images, the backscattering intensity of sea ice and open water surfaces is closely related to the change in incident angle. As the incident angle increases from near to far, the image brightness gradually decreases. This phenomenon leads to sea ice and open water having the same backscattering coefficient in the same image, which may cause misinterpretation during sample interpretation. The purpose of introducing a corrected backscattering coefficient feature is to eliminate the brightness difference caused by different radar beam incident angles, so that similar ground features have similar brightness values ​​at any location in the image, thereby mitigating the incident angle effect in HH and HV polarized images.

[0041] Sentinel-1 EW images range from approximately 19° in angle of incidence. ° up to 47 ° The incident angles of all images will eventually approach the average incident angle of 33°. ° .

[0042] Specifically, the morphological feature set is obtained through the following method: based on the GEE platform, the erosion features, dilation features, opening operation features, closing operation features, top cap transformation features, and bottom cap transformation features of the preprocessed image are calculated. The contribution rate of each morphological feature of the preprocessed image is calculated using principal component analysis (PCA). The five morphological features with the highest contribution rates are selected to form the morphological feature set. The five morphological features with the highest contribution rates are: erosion features, dilation features, opening operation features, closing operation features, and bottom cap transformation features. The morphological feature set can enhance the contrast of sea ice edges and suppress noise.

[0043] The contribution rates of each feature calculated by Principal Component Analysis (PCA) are as follows: the contribution rate of corrosion feature is 0.306, the contribution rate of expansion feature is 0.154, the contribution rate of opening operation feature is 0.267, the contribution rate of closing operation feature is 0.230, the contribution rate of bottom-hat transformation feature is 0.201, and the contribution rate of top-hat transformation feature is 0.091. The top-hat transformation feature was removed to construct the morphological feature set.

[0044] Specifically, the OTSU feature is a binary mask of the preprocessed image based on the OTSU algorithm. The OTSU algorithm is suitable for image classification where there is a significant grayscale difference between the target and the background. Adaptive thresholding is performed on the preprocessed image based on the OTSU algorithm to generate a binary mask representing the potential sea ice distribution, providing a spatial constraint benchmark for subsequent feature space construction. The specific implementation of the OTSU algorithm is prior art and will not be elaborated upon in this invention.

[0045] Step 3: Use the SNIC superpixel segmentation method to segment the multidimensional data feature set to obtain a set of superpixel blocks;

[0046] The SNIC superpixel segmentation method identifies pixels with similar colors, textures, and other features, clustering pixels with common characteristics to form many superpixel blocks. The advantage of the SNIC algorithm lies in its ability to fully consider the neighborhood information of pixels, preserving the spatial structure information of the image, and exhibiting stronger noise suppression capabilities in sea ice classification. The parameter settings for the SNIC superpixel segmentation method include seed size, compactness, connectivity, and neighborhood size. In the embodiments of this invention, since the classification task is based on multi-dimensional data features with a spatial resolution of 40 meters, the number of seeds is set to 36. The seed size affects the size and number of generated superpixels. The compactness parameter is set to 0.1, controlling the smoothness and shape regularity of the superpixel boundaries. The connectivity parameter is set to 8, indicating that the neighborhood relationship between adjacent pixels and diagonal pixels is considered simultaneously. The neighborhood size is set to 256, determining the sliding window size used in the clustering process.

[0047] Step 4: Input the superpixel block set into a supervised-trained machine learning classifier to classify seawater and sea ice, thereby extracting sea ice. The machine learning classifier can be a support vector machine classifier, a random forest classifier, a gradient tree classifier, etc. In this invention, a random forest classifier is preferred.

[0048] Steps 1 to 4 are the sea ice extraction process in actual application. Before the machine learning classifier is used for actual classification, it is pre-trained using labeled sample data. The labeled sample data is randomly collected from dual-polarization SAR images that have completed visual interpretation of sea ice, and the visual interpretation results are used as ground truth labels.

[0049] In addition, different classification models need to be supervised training according to their respective training requirements. Specifically, the parameter settings of the machine learning classifier during supervised training are dynamically adjusted through preset parameter combinations. The adjusted parameters include classifier type selection strategy, model training iteration control method and accuracy optimization mechanism, and select the best parameter combination that can achieve the best classification performance.

[0050] The present invention will further demonstrate the detailed implementation process and technical effects of the multi-feature superpixel segmentation sea ice extraction method shown in steps 1 to 4 above on a specific dataset through specific embodiments, so as to facilitate understanding of the essence of the present invention.

[0051] Example 1

[0052] To systematically verify that the present invention can effectively improve the accuracy of sea ice classification, an experimental group and three control groups were set up to classify sea ice respectively.

[0053] In this embodiment, Sentinel-1 dual-polarization SAR images of Hokkaido from the past 8 years were collected and preprocessed. The Sentinel-1 imaging mode was selected as EW, and the dual-polarization bands were HH and VV. One winter image and one summer image were selected for sea ice identification and extraction. An experimental area was determined in both the winter and summer images. An experimental group and three control groups were established based on both the winter and summer images.

[0054] Experimental group: The multidimensional data feature set is obtained by fusing backscattering coefficient features, normalized features, morphological feature set, OTSU features and texture features. The multidimensional data feature set is used to obtain a set of superpixel blocks by the SNIC superpixel segmentation method. The set of superpixel blocks is input into a supervised machine learning classifier to classify seawater and sea ice and extract sea ice.

[0055] Control group 1: The extracted features are backscattered features, which are input into a supervised machine learning classifier to classify seawater and sea ice and extract sea ice.

[0056] Control group 2: The multidimensional data feature set is obtained by fusing backscattering coefficient features and texture features. The multidimensional data feature set is input into a supervised machine learning classifier to classify seawater and sea ice and extract sea ice.

[0057] Control group 3: The multidimensional data feature set is obtained by fusing backscattering coefficient features, normalized features, morphological feature set, OTSU features and texture features. The multidimensional data feature set is input into a supervised machine learning classifier to classify seawater and sea ice and extract sea ice.

[0058] from Figure 2 and Figure 3It can be seen that the classification results of the single backscatter feature (control group 1) are poor, with not all sea ice being identified, and only the general outline of the sea ice being visible. This is mainly due to the influence of SAR image noise. Control group 2 incorporates texture features, which enhances the image representation difference between sea ice and open water, resulting in improved sea ice classification accuracy compared to control group 1. However, some sea ice is still not identified, and the classification results for sea ice edges are blurry. The sea ice classification results of control group 3 and the experimental group are better, with all sea ice being identified. However, the sea ice classification results of control group 3 contain misidentifications caused by image noise, while the sea ice classification results of the experimental group do not have misidentifications caused by noise interference, and the sea ice edges are uniform and natural.

[0059] Example 2

[0060] This embodiment demonstrates, by comparing the experimental group and the corresponding sea ice concentration data, that the present invention can effectively improve the accuracy and robustness of sea ice region classification.

[0061] The experimental group in this embodiment is: the winter and summer images selected in Example 1 are used to classify sea ice using the experimental group method of Example 1.

[0062] from Figure 4 As can be seen, even with a spatial resolution of 1km for sea ice concentration data, the sea ice classification results still lose a lot of sea ice information. In contrast, the sea ice classification map of the experimental group, with a spatial resolution of 40m, can still fully show the distribution of sea ice, and has higher sea ice classification accuracy and robustness.

[0063] In summary, this invention extracts and fuses backscattering coefficient features, normalized features, morphological feature sets, OTSU features, and texture features to obtain a multidimensional data feature set. This multidimensional data feature set is then used to obtain a set of superpixel blocks through the SNIC superpixel segmentation method. The set of superpixel blocks is then input into a supervised machine learning classifier to classify seawater and sea ice, which can effectively improve the accuracy and robustness of sea ice classification.

[0064] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions and substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A sea ice extraction method based on multi-feature superpixel segmentation, characterized in that, The application relates to a method for extracting sea ice from a target sea area, comprising the following steps: Step 1, acquiring a dual-polarization SAR image of a target sea area and performing image preprocessing to obtain a preprocessed image; Step 2, extracting backscattering coefficient features, normalized features, a morphological feature set, OTSU features and texture features from the preprocessed image and fusing the features to obtain a multi-dimensional data feature set; Step 3, using an SNIC superpixel segmentation method to segment the multi-dimensional data feature set to obtain a superpixel block set; Step 4, inputting the superpixel block set into a supervised machine learning classifier to classify seawater and sea ice and extract sea ice. The normalized features are corrected backscattering coefficient features, which are obtained by correcting the backscattering coefficient features of the preprocessed image by using incident angle normalization, and the correction is specifically as follows: A plurality of pixel points are randomly selected in the preprocessed image, and the following steps are performed on each randomly selected pixel point to obtain a corresponding corrected backscattering coefficient value, and the corrected backscattering coefficient values of all the randomly selected pixel points form the corrected backscattering coefficient feature of the preprocessed image. Step a, extract the backscattering coefficient value and the incident angle value of the pixel point, construct a linear fitting function, and the linear fitting function is: ; Step b, constructing the inverse function of the linear fitting function using the average value of the incident angle, the inverse function of the linear fitting function is: ; Step c, constructing the backscattering coefficient equation normalized by incident angle to obtain the corrected backscattering coefficient value at incident angle θ The backscattering coefficient equation normalized by incident angle is: ; where a and b are the slope and intercept of the linear fit function, respectively, denotes the backscatter coefficient value at an incidence angle of θ, is the inverse function of , θ ref is the average value of the incidence angle, θ ref is 33 ° , is the corrected backscatter coefficient value.

2. The multi-feature superpixel segmentation based sea ice extraction method of claim 1, wherein, The dual-polarization SAR image is a dual-polarization SAR image based on HH and HV dual-polarization bands obtained by Sentinel-1 in an EW imaging mode, and HH and HV are both used as backscattering feature bands.

3. The multi-feature superpixel segmentation based sea ice extraction method of claim 2, wherein, The texture features are extracted by using principal component analysis (PCA) to reduce the dimensionality of the GLCM texture features of the preprocessed image, and three texture features with the highest contribution rates are selected, namely contrast, correlation and variance.

4. The multi-feature superpixel segmentation based sea ice extraction method of claim 3, wherein, The morphological feature set is obtained by the following method: the erosion feature, the dilation feature, the opening operation feature, the closing operation feature, the top-hat transformation feature and the bottom-hat transformation feature of the preprocessed image are calculated based on the GEE platform, the contribution rates of each morphological feature of the preprocessed image are calculated by using the principal component analysis (PCA) method, the five morphological features with the highest contribution rates are selected to form the morphological feature set, and the five morphological features with the highest contribution rates are respectively the erosion feature, the dilation feature, the opening operation feature, the closing operation feature and the bottom-hat transformation feature.

5. The multi-feature superpixel segmentation based sea ice extraction method of claim 4, wherein, The OTSU features are binary masks of the preprocessed image based on the OTSU algorithm.

6. The multi-feature superpixel segmentation based sea ice extraction method of claim 5, wherein, The machine learning classifier is a random forest classifier.

7. The multi-feature superpixel segmentation based sea ice extraction method of claim 6, wherein The machine learning classifier is supervised trained by using labeled sample data before being used for actual classification, the labeled sample data is randomly collected from a dual-polarization SAR image with completed visual interpretation of sea ice, and the visual interpretation result is used as a true value label.

8. The multi-feature superpixel segmentation based sea ice extraction method of claim 7, wherein The parameter settings of the machine learning classifier during the supervised training are dynamically adjusted by using a preset parameter combination, the adjustment parameters include a classifier type selection strategy, a model training iteration control mode and a precision optimization mechanism, and the best parameter combination achieving the optimal classification performance is selected.

Images