SAR image sea ice detection method

Abstract

The invention discloses an SAR image sea ice detection method, and the method comprises the steps: obtaining an SAR image of a preset target region in an SAR image library, so as to obtain an SAR image data set for sea ice detection; preprocessing the SAR image data set to obtain a first sea ice mask image set and a sea ice feature vector; training a pre-created support vector machine, and performing classification prediction on the SAR image data set according to the trained support vector machine to obtain a second sea ice mask image; randomly dividing the second sea ice mask image into a training set and a verification set according to a preset proportion; inputting the training set into a pre-created HRU-Net model for model training, and determining whether the trained HRU-Net model converges according to the verification set based on the constructed model loss function so as to obtain a support vector machine-HRU-Net detection model for SAR image sea ice detection; and sea ice detection of the SAR image to be detected is realized according to the support vector machine-HRU-Net detection model. The problems that an existing SAR sea ice detection method is high in dependence on labeled samples, low in sea ice detection efficiency, low in precision and the like are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to a method for detecting sea ice using SAR images. Background Art

[0002] In recent years, the global climate has been abnormal, and sea ice has appeared more and more frequently. The appearance of sea ice will cause huge losses to marine transportation, marine fisheries, and marine oil and gas resource development. Many ports around the world may freeze in the nearby waters during winter or cold seasons, especially ports located in cold regions or high latitudes. Therefore, sea ice monitoring is of great significance for ensuring shipping safety, maintaining port navigation, ensuring marine operations, and improving emergency response capabilities.

[0003] Traditional ground measurement methods can usually only cover limited areas. In remote areas or high-latitude areas, weather conditions such as bad weather and low visibility may make monitoring tasks impossible. In addition, ground measurement requires a lot of manpower, material resources and time costs, especially for large-scale sea ice monitoring tasks, where data acquisition is difficult and analysis is slow. Therefore, using SAR images for sea ice monitoring can make up for the limitations of ground measurement, provide all-weather and all-day sea ice monitoring data, and provide a basis for sea ice early warning and response.

[0004] Currently, the use of SAR satellite images for accurate sea ice detection still faces huge challenges. First, with the rapid development of space-borne SAR technology at home and abroad, there are currently a variety of sensors, and the multi-resolution and multi-scale changes of images are important challenges currently faced. The existing methods only use data from the same satellite, resulting in the model not having the ability to generalize multi-scale images; second, the existing models have poor mobility, and models trained using samples from a certain region and a certain year are usually difficult to apply to other regions or other years. This leads to low sea ice detection accuracy and low detection efficiency; third, in the development process of existing neural network models for sea ice detection, training the network model usually requires a large number of training samples, but this requires the participation of personnel with professional knowledge and experience, which increases the complexity and cost of data processing.

[0005] Therefore, in order to solve the problems of high dependence on labeled samples, low efficiency and low accuracy of sea ice detection in the current SAR sea ice detection methods, it is urgent to construct an efficient and accurate method for sea ice detection using SAR images. Summary of the invention

[0006] The present invention provides a method for sea ice detection using SAR images to overcome the above technical problems.

[0007] In order to achieve the above object, the technical solution of the present invention is:

[0008] A method for sea ice detection using SAR images, comprising the following steps:

[0009] S1: Obtain SAR images of the preset target area in the SAR image library to obtain a SAR image dataset for sea ice detection;

[0010] S2: Preprocess the SAR image dataset;

[0011] The pre-processing comprises:

[0012] Marking points of sea ice in SAR images to obtain the first sea ice mask image set;

[0013] Extract the sea ice features in SAR images to obtain sea ice feature vectors;

[0014] The sea ice characteristics include sea ice scattering characteristics, sea ice texture characteristics and sea ice space characteristics;

[0015] S3: training the pre-created support vector machine according to the first mask image set and the sea ice feature vector, and classifying and predicting the SAR image data set according to the trained support vector machine to obtain the second sea ice mask image;

[0016] The second sea ice mask image includes sea ice features and corresponding sea ice mask labels;

[0017] S4: randomly dividing the second sea ice mask image into a training set and a validation set according to a preset ratio;

[0018] S5: inputting the training set into the pre-created HRU-Net model for model training, and confirming whether the trained HRU-Net model converges based on the constructed model loss function and the validation set;

[0019] If the output of the trained HRU-Net model converges, the trained HRU-Net model is the optimal HRU-Net model, and step S6 is executed;

[0020] Otherwise, the weight parameters of the HRU-Net model are adaptively adjusted based on the back propagation method, and step S4 is repeated;

[0021] S6: Evaluate the model accuracy of the optimal HRU-Net model after training through the validation set;

[0022] If the model accuracy of the trained HRU-Net model meets the preset accuracy threshold, a support vector machine-HRU-Net detection model for SAR image sea ice detection is obtained based on the trained HRU-Net model;

[0023] Otherwise, reacquire the SAR image data in the SAR image library, and repeat steps S2 to S5 until the preset accuracy threshold is met;

[0024] S7: Implement sea ice detection on the SAR images to be detected based on the support vector machine-HRU-Net detection model.

[0025] Furthermore, the method for marking points of sea ice in SAR images in S2 is as follows:

[0026] Get the coordinates of each sea ice pixel in the SAR image;

[0027] And the sea ice pixel coordinate point includes a pixel abscissa and a pixel ordinate;

[0028] The average pixel abscissa and the average pixel ordinate of each sea ice pixel coordinate point are taken as the location of the sea ice in the SAR image.

[0029] Taking the location of the sea ice as the center point, obtain the corresponding square area of ​​interest;

[0030] And the vertex coordinates of the square region of interest are marked as

[0031] (x c -r,y c -r),(x c +r,y c -r),(x c +r,y c +r),(x c -r,y c +r)

[0032] where x c The horizontal coordinate of the center point is the location of the sea ice; c represents the vertical coordinate of the center point where the sea ice is located; r represents the preset parameter.

[0033] Furthermore, S2 extracts the sea ice scattering features, sea ice texture features, and sea ice spatial features in the SAR image, and fuses and obtains the sea ice feature vector, specifically:

[0034] S21: obtaining a cumulative distribution function value corresponding to each gray level according to the pixels of the SAR image to obtain a pixel contrast of the SAR image;

[0035] S22: Based on the exhaustive method, confirm the image segmentation parameters and merging parameters in the image segmentation process of the watershed transform algorithm according to the pixel contrast;

[0036] The SAR image is segmented according to the image segmentation parameters and the merging parameters to obtain a number of sub-images;

[0037] S23: Acquire sea ice scattering features, sea ice texture features, and sea ice spatial features in each sub-image, and acquire a sea ice feature vector based on the corresponding images of sea ice scattering features, sea ice texture features, and sea ice spatial features.

[0038] Furthermore, the sea ice scattering characteristics in S23 include the mean of sea ice scattering intensity, the variance of sea ice scattering intensity and the information entropy of sea ice scattering histogram;

[0039] The mean sea ice scattering intensity W 1 for

[0040]

[0041] Where: f(i,j) represents the scattering intensity at the pixel point (i,j) in the sub-image; V represents the area of ​​the sub-image; S represents the width of the image; M represents the height of the image;

[0042] The sea ice scattering intensity variance W 2 for

[0043]

[0044] The sea ice scattering histogram information entropy W 3 for

[0045]

[0046] Where: h[g] represents the pixel value of the sub-image histogram; h[g](v) represents the percentage of pixels at a certain gray level (v) in the entire image; U represents the maximum gray level;

[0047] The sea ice texture characteristics include sea ice texture contrast, sea ice texture correlation, sea ice texture second-order moment and sea ice texture image entropy;

[0048] The sea ice texture contrast W 4 for

[0049]

[0050] Where: p(i,j) represents the sea ice texture contrast at the pixel point (i,j) in the sub-image;

[0051] The sea ice texture correlation W 5 for

[0052]

[0053] Where: μ and σ 2 All represent intermediate parameter variables;

[0054] The second-order moment of sea ice texture W6 for

[0055]

[0056] The image entropy W of the sea ice texture 7 for

[0057]

[0058] The sea ice spatial characteristics include the roundness of the sea ice image and the compactness of the sea ice image;

[0059] The circularity W of the sea ice image 8 for

[0060]

[0061] Where: A represents the area of ​​sea ice in the sub-image; P represents the circumference of sea ice in the sub-image;

[0062] The compactness of the sea ice image W 9 for

[0063]

[0064] Where: P 1 Represents the outer contour length of the sea ice in the sub-image;

[0065] The obtained sea ice characteristic vector θ is recorded as

[0066] Θ=[W 1 ,W 2 ,W 3 ,W 4 ,W 5 ,W 6 ,W 7 ,W 8 ,W 9 ].

[0067] Furthermore, the kernel function used in S3 to train the pre-created support vector machine is a polynomial kernel function;

[0068] The expression of the polynomial kernel function is

[0069] K(x i ,x j )=(γx i x j +r) d

[0070] In the formula: γ represents an adjustable parameter; r represents a constant term; d represents the order of the polynomial; x i ,x jis the input of the support vector machine, representing the first sea ice mask image set and the sample K(x i ,x j ) is the output of the kernel function, representing the input vector x i ,x j Similarity in feature space.

[0071] Furthermore, the model loss function constructed in S5 is

[0072] L total =αL ce +βL dice

[0073]

[0074] Where: y i represents the true sea ice mask label of the i-th sample; p i represents the predicted label probability of the i-th sample; N represents the total number of samples, which is the number of pixels in the image in the training or validation set, and i∈N; L total Represents the model loss function; L dice represents Dice loss; L ce represents the cross entropy loss; α, β represent the weight parameters.

[0075] Beneficial effects: The present invention provides a method for sea ice detection in SAR images, which preprocesses the SAR image data set, that is, performs point annotation on the sea ice in the SAR image to obtain a first sea ice mask image set; performs feature extraction on the sea ice features in the SAR image to obtain a sea ice feature vector; uses point annotation and then uses a support vector machine to preliminarily form a sea ice mask, thereby greatly reducing the production cost of samples required for a deep learning model; inputs the training set into a pre-created HRU-Net model for model training, the HRU-Net network model can not only extract multi-scale features, but also gradually restore spatial information in its decoder, so that the mask finally generated has both a global field of view and retains image details, and is suitable for complex sea ice scene segmentation tasks; the support vector machine-HRU-Net detection model for sea ice detection in SAR images is suitable for multi-source SAR satellites in the sea ice detection process, supports multiple polarization modes including HH, HV and VV, supports images in multiple imaging modes, and has strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0077] Figure 1 A flow chart of the method for sea ice detection using SAR images of the present invention;

[0078] Figure 2 is the SAR image to be detected in this embodiment (FSII mode);

[0079] Figure 3 In this embodiment, the corresponding Figure 2 Detected sea ice mask (FSII mode);

[0080] Figure 4 is the SAR image to be detected in this embodiment (QPSI mode);

[0081] Figure 5 In this embodiment, the corresponding Figure 3 Detected sea ice mask (QPSI mode). DETAILED DESCRIPTION

[0082] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0083] This embodiment provides a method for detecting sea ice in SAR images. Figure 1 As shown, the following steps are included:

[0084] S1: Obtain SAR images of the preset target area in the SAR image library to obtain a SAR image dataset for sea ice detection;

[0085] Specifically, in this embodiment, 200 GF-3 satellite images of the Bohai Bay area from December to February in winter are selected from the SAR image database, and the FSII mode and the QPSI mode are included;

[0086] S2: Preprocess the SAR image dataset;

[0087] The pre-processing comprises:

[0088] Marking points of sea ice in SAR images to obtain the first sea ice mask image set;

[0089] Specifically, the method for marking points on sea ice in SAR images is as follows:

[0090] Get the coordinates of each sea ice pixel in the SAR image;

[0091] And the sea ice pixel coordinate point includes a pixel abscissa and a pixel ordinate;

[0092] The average pixel abscissa and the average pixel ordinate of each sea ice pixel coordinate point are taken as the location of the sea ice in the SAR image.

[0093] Taking the point where the sea ice is located as the center point, obtain the corresponding square area of ​​interest; that is, the square area centered on the point, whose side length is 2r;

[0094] Then the vertex coordinates of the square region of interest are

[0095] (x c -r,y c -r),(x c +r,y c -r),(x c +r,y c +r),(x c -r,y c +r);

[0096] where x c The horizontal coordinate of the center point is the location of the sea ice; c represents the vertical coordinate of the center point where the sea ice is located; r represents the preset parameter.

[0097] Extract the sea ice features in SAR images to obtain sea ice feature vectors;

[0098] The sea ice characteristics include sea ice scattering characteristics, sea ice texture characteristics and sea ice space characteristics;

[0099] Specifically, the sea ice features in the SAR image are extracted, and the sea ice feature vector is obtained by fusion, which is specifically:

[0100] S21: obtaining a cumulative distribution function value corresponding to each gray level according to the pixels of the SAR image to obtain a pixel contrast of the SAR image;

[0101] And the expression for obtaining the cumulative distribution function value corresponding to each gray level is

[0102]

[0103] Where: h(j) represents the frequency of pixels with gray value j; N represents the total number of pixels in the image; F(i) represents the proportion of pixels with pixel value less than or equal to i;

[0104] S22: Based on the exhaustive method, confirm the image segmentation parameters and merging parameters in the image segmentation process of the watershed transform algorithm according to the pixel contrast;

[0105] Specifically, several combinations of segmentation parameters and merging parameters are randomly defined as parameter combinations to be selected, and experimental tests are performed based on expert experience. By comparing the experimental results, it is found that the segmentation parameter is set to 30 and the merging parameter is set to 95 to achieve the best segmentation effect. The experimental process of selecting image segmentation parameters and merging parameters is an existing well-known technical means, which will not be described in detail here.

[0106] The SAR image is segmented according to the image segmentation parameters and the merging parameters to obtain a number of sub-images;

[0107] S23: Acquire sea ice scattering features, sea ice texture features, and sea ice spatial features in each sub-image, and acquire a sea ice feature vector according to the corresponding images of sea ice scattering features, sea ice texture features, and sea ice spatial features;

[0108] Specifically, the sea ice scattering characteristics include the mean of sea ice scattering intensity, the variance of sea ice scattering intensity and the information entropy of sea ice scattering histogram;

[0109] The mean sea ice scattering intensity W 1 for

[0110]

[0111] Where: f(i,j) represents the scattering intensity at the pixel point (i,j) in the sub-image; V represents the area of ​​the sub-image; S represents the width of the image; M represents the height of the image;

[0112] The sea ice scattering intensity variance W 2 for

[0113]

[0114] The sea ice scattering histogram information entropy W 3 for

[0115]

[0116] Where: h[g] represents the pixel value of the sub-image histogram; h[g](v) represents the percentage of pixels at a certain gray level (v) in the entire image; U represents the maximum gray level;

[0117] The sea ice texture characteristics include sea ice texture contrast, sea ice texture correlation, sea ice texture second-order moment and sea ice texture image entropy;

[0118] The sea ice texture contrast W 4 for

[0119]

[0120] Where: p(i,j) represents the sea ice texture contrast at the pixel point (i,j) in the sub-image;

[0121] The sea ice texture correlation W 5 for

[0122]

[0123] Where: μ and σ 2 All represent intermediate parameter variables;

[0124] The second-order moment of sea ice texture W 6 for

[0125]

[0126] The image entropy W of the sea ice texture 7 for

[0127]

[0128] The sea ice spatial characteristics include the roundness of the sea ice image and the compactness of the sea ice image;

[0129] The circularity W of the sea ice image 8 for

[0130]

[0131] Where: A represents the area of ​​sea ice in the sub-image; P represents the circumference of sea ice in the sub-image;

[0132] The compactness of the sea ice image W 9 for

[0133]

[0134] Where: P 1 Represents the outer contour length of the sea ice in the sub-image;

[0135] The obtained sea ice characteristic vector θ is recorded as

[0136] Θ=[W 1 ,W 2 ,W 3 ,W 4 ,W5 ,W 6 ,W 7 ,W 8 ,W 9 ].

[0137] Where W 1 ,W 2 ,W 3 ,W 4 ,W 5 ,W 6 ,W 7 ,W 8 ,W 9 Each feature of can be regarded as a single-band image of the same size as the image. Here, the 9 features calculated are simply stacked together to form a 9-band image, which is the sea ice feature vector.

[0138] S3: training the pre-created support vector machine according to the first mask image set and the sea ice feature vector, and classifying and predicting the SAR image data set according to the trained support vector machine to obtain the second sea ice mask image;

[0139] The second sea ice mask image includes sea ice features and corresponding sea ice mask labels;

[0140] Specifically, the support vector machine SVM classifier is a multi-layer perceptron containing one hidden layer, and the number of hidden layer nodes is automatically determined by the algorithm through the training process;

[0141] The nonlinear prediction model for classification prediction of SAR image data sets based on support vector machine SVM is:

[0142]

[0143] Among them, f(x) represents the sea ice mask label value of the SAR image dataset; x represents the characteristic data related to the sea ice mask label, including the mean of sea ice scattering intensity, the variance of sea ice scattering intensity and the information entropy of sea ice scattering histogram; sea ice texture contrast, sea ice texture correlation, sea ice texture second-order moment and sea ice texture image entropy; the roundness of sea ice image and the compactness of sea ice image; b represents the prediction residual term; represents the nonlinear mapping from the input space to the high-dimensional space, i.e., the kernel function; w represents the model parameters of the support vector machine SVM;

[0144] The penalty function of the support vector machine SVM is defined as

[0145]

[0146] Where: e i represents the error term; r represents the regularization parameter; N represents the number of samples;

[0147] After introducing the Lagrange multiplier λ, the nonlinear prediction model expression based on the support vector machine SVM classifier is transformed into

[0148]

[0149] Where K(x i ,x j ) represents the kernel function for training or predicting the support vector machine SVM, that is, a polynomial kernel function, and the expression of the polynomial kernel function is

[0150] K(x i ,x j )=(γx i x j +r) d

[0151] In the formula: γ represents an adjustable parameter; r represents a constant term; d represents the order of the polynomial; x i ,x j is the input of the support vector machine, representing the first sea ice mask image set and the sample K(x i ,x j ) is the output of the kernel function, representing the input vector x i ,x j Similarity in feature space;

[0152] The polynomial kernel function is used as the kernel function of the support vector machine SVM. The support vector machine SVM is trained through the first mask image set and the sea ice feature vector, wherein the first mask image set can be regarded as a mask image with only 0 and 1; the sea ice feature vector can be regarded as a 9-band image. After inputting into the support vector machine SVM, the feature vector value of the corresponding position is read according to the mask, the position of 0 is not read, and the position of 1 is read. After reading the value, the support vector machine starts training, and the penalty factor, polynomial kernel function parameters and weight factor are iterated during the training to optimize the model, so as to obtain the sea ice mask prediction model of the support vector machine SVM, and the SAR image data set is classified and predicted according to the trained support vector machine to obtain the second sea ice mask image;

[0153] S4: randomly dividing the second sea ice mask image into a training set and a validation set according to a preset ratio; specifically, the preset ratio of randomly dividing the second sea ice mask image into the training set and the validation set is 8:2;

[0154] S5: inputting the training set into the pre-created HRU-Net model for model training, and confirming whether the trained HRU-Net model converges based on the constructed model loss function and the validation set;

[0155] Specifically, the constructed model loss function is

[0156] L total =αL ce +βL dice

[0157]

[0158] Where: y i represents the true sea ice mask label of the i-th sample; p i represents the predicted label probability of the i-th sample; N represents the total number of samples, usually the number of pixels in the image in the training or validation set, and i∈N; L total Represents the model loss function; L dice represents Dice loss; L ce represents the cross entropy loss; α, β represent weight parameters;

[0159] If the output of the trained HRU-Net model converges, the trained HRU-Net model is the optimal HRU-Net model, and step S6 is executed;

[0160] Otherwise, the weight parameters of the HRU-Net model are adaptively adjusted based on the back propagation method, and step S4 is repeated;

[0161] The HRU-Net model described in this embodiment is an existing network model structure, and can integrate multi-level feature maps layer by layer to ensure that when generating the final mask, not only the global semantic features of the image are captured, but also the fine local details are retained. This integration process is achieved by fusing the high-resolution and low-resolution features of HRNet in the decoder part of U-Net;

[0162] And the global semantic features refer to the macroscopic, high-level semantic information in the image. These features are usually extracted through low-resolution feature maps, capturing the overall structure and contextual relationships. For example, in SAR images, global features can reflect the overall shape and distribution of sea ice areas, helping the model understand the semantic categories of these areas (sea ice or non-sea ice); the local details refer to small features in the image, which are usually extracted through high-resolution feature maps, mainly describing information such as the fine boundaries and textures of objects. Local features can more accurately capture the precise boundaries between sea ice and other areas (such as sea water). This is very important for segmentation tasks; for sea ice in SAR images, there may be slight brightness changes or texture features, which can be effectively captured through high-resolution feature maps;

[0163] Specifically, feature maps from different layers are combined and restored to a higher resolution through the upsampling process of U-Net. The integration of features in each layer not only relies on its own convolutional features, but also relies on high-resolution jump connections with adjacent layers to achieve cross-layer information sharing between features. During the upsampling process, low-resolution features are gradually restored through methods such as bilinear interpolation, and fused with high-resolution local detail features. In this way, the features output by each layer not only contain their own global information, but also carry the local details of the previous level. The expression of this feature fusion process is

[0164]

[0165] Among them, X 1 ,X 2 ,...,X N Represents input features of different resolutions; H i Represents the convolution and fusion operations on these features; Y final is the final mask output.

[0166] S6: Evaluate the model accuracy of the optimal HRU-Net model after training through the validation set;

[0167] If the model accuracy of the trained HRU-Net model meets the preset accuracy threshold, a support vector machine-HRU-Net detection model for sea ice detection in SAR images is obtained based on the trained HRU-Net model;

[0168] Otherwise, reacquire the SAR image data in the SAR image library, and repeat steps S2 to S5 until the preset accuracy threshold is met;

[0169] In this embodiment, the performance indicators for evaluating the accuracy of the model using the test set are the model recall rate and the model accuracy rate; and if the recall rate is greater than 90% and the accuracy rate is greater than 90%, step S7 is executed. If the requirements are not met, S2 to S51 are executed, that is, samples are supplemented (sample supplementation means reselecting SAR images from the SAR image library to obtain a SAR image data set for sea ice detection) and parameters are adjusted until the preset accuracy threshold is reached. The expressions of the model recall rate Recall and the model accuracy Precision are:

[0170]

[0171] Among them, TP represents the number of pixels correctly classified as sea ice; FP represents the number of pixels incorrectly classified as sea ice; FN represents the number of sea ice pixels incorrectly classified as background;

[0172] S7: Implement sea ice detection on the SAR images to be detected based on the support vector machine-HRU-Net detection model.

[0173] The experimental results of this embodiment are as follows Figures 2 to 5 As shown in the figure, they are the SAR images to be detected, Figure 3 and Figure 5 The corresponding detected sea ice mask is shown in FIG. 1 . It can be seen that the SAR image sea ice detection method provided by this embodiment can achieve higher detection accuracy and efficiency for SAR images in different imaging modes.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for sea ice detection using SAR images, characterized in that: The following steps are involved: S1: Obtain SAR images of the preset target area in the SAR image library to obtain a SAR image dataset for sea ice detection; S2: Preprocess the SAR image dataset; The pre-processing comprises: Marking points of sea ice in SAR images to obtain the first sea ice mask image set; Extract the sea ice features in SAR images to obtain sea ice feature vectors; The sea ice characteristics include sea ice scattering characteristics, sea ice texture characteristics and sea ice space characteristics; S3: training the pre-created support vector machine according to the first mask image set and the sea ice feature vector, and classifying and predicting the SAR image data set according to the trained support vector machine to obtain the second sea ice mask image; The second sea ice mask image includes sea ice features and corresponding sea ice mask labels; S4: randomly dividing the second sea ice mask image into a training set and a validation set according to a preset ratio; S5: inputting the training set into the pre-created HRU-Net model for model training, and confirming whether the trained HRU-Net model converges based on the constructed model loss function and the validation set; If the output of the trained HRU-Net model converges, the trained HRU-Net model is the optimal HRU-Net model, and step S6 is executed; Otherwise, the weight parameters of the HRU-Net model are adaptively adjusted based on the back propagation method, and step S4 is repeated; S6: Evaluate the model accuracy of the optimal HRU-Net model after training through the validation set; If the model accuracy of the trained HRU-Net model meets the preset accuracy threshold, a support vector machine-HRU-Net detection model for SAR image sea ice detection is obtained based on the trained HRU-Net model; Otherwise, reacquire the SAR image data in the SAR image library, and repeat steps S2 to S5 until the preset accuracy threshold is met; S7: Implement sea ice detection on the SAR images to be detected based on the support vector machine-HRU-Net detection model.

2. The method for sea ice detection using SAR images according to claim 1, characterized in that: The method for marking points of sea ice in SAR images in S2 is: Get the coordinates of each sea ice pixel in the SAR image; And the sea ice pixel coordinate point includes a pixel abscissa and a pixel ordinate; The average pixel abscissa and the average pixel ordinate of each sea ice pixel coordinate point are taken as the location of the sea ice in the SAR image. Taking the location of the sea ice as the center point, obtain the corresponding square area of ​​interest; And the vertex coordinates of the square region of interest are marked as (x c -r,y c -r),(x c +r,y c -r),(x c +r,y c +r),(x c -r,y c +r) where x c The horizontal coordinate of the center point is the location of the sea ice; c represents the vertical coordinate of the center point where the sea ice is located; r represents the preset parameter.

3. The method for sea ice detection using SAR images according to claim 1, characterized in that: In S2, the sea ice scattering features, sea ice texture features, and sea ice spatial features in the SAR image are extracted to obtain the sea ice feature vector, specifically: S21: obtaining a cumulative distribution function value corresponding to each gray level according to the pixels of the SAR image to obtain a pixel contrast of the SAR image; S22: Based on the exhaustive method, confirm the image segmentation parameters and merging parameters in the image segmentation process of the watershed transform algorithm according to the pixel contrast; The SAR image is segmented according to the image segmentation parameters and the merging parameters to obtain a number of sub-images; S23: Acquire sea ice scattering features, sea ice texture features, and sea ice spatial features in each sub-image, and acquire a sea ice feature vector based on the corresponding images of sea ice scattering features, sea ice texture features, and sea ice spatial features.

4. The method for sea ice detection using SAR images according to claim 3, characterized in that: The sea ice scattering characteristics in S23 include the mean sea ice scattering intensity, the variance of the sea ice scattering intensity, and the information entropy of the sea ice scattering histogram; The mean sea ice scattering intensity W1 is Where: f(i,j) represents the scattering intensity at the pixel point (i,j) in the sub-image; V represents the area of ​​the sub-image; S represents the width of the image; M represents the height of the image; The sea ice scattering intensity variance W2 is The sea ice scattering histogram information entropy W3 is Where: h[g] represents the pixel value of the sub-image histogram; h[g](v) represents the percentage of pixels at a certain gray level (v) in the entire image; U represents the maximum gray level; The sea ice texture characteristics include sea ice texture contrast, sea ice texture correlation, sea ice texture second-order moment and sea ice texture image entropy; The sea ice texture contrast W4 is Where: p(i,j) represents the sea ice texture contrast at the pixel point (i,j) in the sub-image; The sea ice texture correlation W5 is Where: μ and σ 2 All represent intermediate parameter variables; The second-order moment W6 of the sea ice texture is The image entropy W7 of the sea ice texture is The sea ice spatial characteristics include the roundness of the sea ice image and the compactness of the sea ice image; The circularity W8 of the sea ice image is Where: A represents the area of ​​sea ice in the sub-image; P represents the circumference of sea ice in the sub-image; The compactness of the sea ice image W9 is Where: P1 represents the outer contour length of the sea ice in the sub-image; The obtained sea ice characteristic vector θ is recorded as Θ=[W1, W2, W3, W4, W5, W6, W7, W8, W9].

5. The method for sea ice detection using SAR images according to claim 4, characterized in that: The kernel function used in S3 to train the pre-created support vector machine is a polynomial kernel function; The expression of the polynomial kernel function is K(x i ,x j )=(γx i x j +r) d In the formula: γ represents an adjustable parameter; r represents a constant term; d represents the order of the polynomial; x i ,x j is the input of the support vector machine, representing the first sea ice mask image set and the sample K(x i ,x j ) is the output of the kernel function, representing the input vector x i ,x j Similarity in feature space.

6. The method for sea ice detection using SAR images according to claim 4, characterized in that: The model loss function constructed in S5 is L total =αL ce +βL dice Where: y i represents the true sea ice mask label of the i-th sample; p i represents the predicted label probability of the i-th sample; N represents the total number of samples, which is the number of pixels in the image in the training or validation set, and i∈N; L total Represents the model loss function; L dice represents Dice loss; L ce represents the cross entropy loss; α, β represent the weight parameters.

Images