CN114764884A - End-to-end polarization SAR image classification method based on superpixel and graph convolution - Google Patents

Abstract

The invention discloses an end-to-end polarization SAR image classification method based on superpixels and graph convolution, which comprises the following steps: step 1, inputting a polarized SAR image to be classified and cutting the image into uniform size; step 2, dividing the cut pictures into a training set and a test set according to the proportion; step 3, decomposing the complex scattering matrix of each pixel point of each image of the test set training set to generate a polarization coherent matrix and converting the polarization coherent matrix into a row vector serving as the polarization characteristic of the pixel point; step 4, splicing the polarization characteristics to the horizontal and vertical coordinates of the pixel point, and splicing the row vectors as the characteristics of the pixel point; step 5, building an end-to-end network based on the full convolution network, the graph convolution network and the convolution neural network; and 6, sending the training set into an end-to-end network for joint training, and sending the test set into the trained end-to-end network to obtain a result. The method can further improve the classification precision of the polarized SAR image.

Description

An end-to-end polarimetric SAR image classification method based on superpixels and graph convolution

technical field

The invention belongs to the technical field of image processing and remote sensing, and relates to an end-to-end polarimetric SAR image classification method based on superpixels and graph convolution.

Background technique

Polarimetric synthetic aperture radar (PolSAR) image classification is one of the most popular research directions in the PolSAR field, which can provide basic support for many fields such as land use surveys, geographic national conditions monitoring, and urban and rural planning. The so-called classification problem of polarimetric SAR images is to determine the category of each pixel in the image through an algorithm, so as to judge the corresponding ground objects.

In the traditional method of classifying polarimetric SAR images based on superpixel segmentation, superpixel segmentation is to preprocess the image as a separate task, and then the preprocessed image results of superpixel segmentation are input to polarimetric SAR classification The classification results of polarimetric SAR images are obtained in the network.

Because the PolSAR image itself has more random coherent speckle noise and low resolution, and different ground object types tend to show relatively similar characteristics, the traditional superpixel segmentation method will be greatly disturbed, and the superpixel segmentation results will be directly Influence the output of the downstream polarimetric SAR classification network.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an end-to-end polarimetric SAR image classification method based on superpixels and graph convolution, which can further improve the classification accuracy of polarimetric SAR images.

The technical scheme adopted in the present invention is:

An end-to-end polarimetric SAR image classification method based on superpixels and graph convolution, including the following steps:

Step 1, input the polarimetric SAR image to be classified and crop it into a uniform size;

Step 2: Divide the cropped image into a training set and a test set in proportion;

Step 3, decompose the complex scattering matrix of each pixel of each image of the test set training set, generate a polarization coherence matrix and convert it into a row vector as the polarization feature of the pixel;

Step 4, splicing the polarization feature on the horizontal and vertical coordinates of the pixel, and splicing the row vector as the pixel feature;

Step 5, build an end-to-end network based on a fully convolutional network, a graph convolutional network, and a convolutional neural network;

Step 6: The training set is sent to the end-to-end network for joint training, and the test set is sent to the trained end-to-end network to obtain the result.

The specific steps of step 3 are:

Decompose the complex scattering matrix of each pixel of each image cropped in step 2 to generate a polarization coherence matrix and convert it into a 1×9 row vector as the polarization feature of the pixel. The expression of the generated polarization coherence matrix T is as follows:

然后将极化相干矩阵转换为行向量得到特征矩阵：T′＝[T₁₁，T₁₂，T₁₃，T₂₁，T₂₂，T₂₃，T₃₁，T₃₂，T₃₃]，并且该极化相干矩阵为复共轭矩阵，对此复共轭矩阵进行预处理：得到该像素点极化特征向量F：Denote the polarization coherence matrix T as

Then convert the polarization coherence matrix to row vector to get the characteristic matrix: T′=[T ₁₁ , T ₁₂ , T ₁₃ , T ₂₁ , T ₂₂ , T ₂₃ , T ₃₁ , T ₃₂ , T ₃₃ ], and the polarization The coherence matrix is a complex conjugate matrix, and the complex conjugate matrix is preprocessed to obtain the pixel polarization eigenvector F:

Among them, Re represents the real part of the complex number, Im represents the imaginary part of the complex number, and T _ij represents the data of the i-th row and the j-th column of the polarization coherence matrix.

Step 4 is specifically: splicing the polarization feature vector of each pixel point generated in step 3 with the horizontal and vertical coordinates of the pixel point to obtain the pixel point feature.

The structure of the fully convolutional network in step 5 is a connected input layer, a first downsampling layer, a second downsampling layer, a third downsampling layer, a fourth downsampling layer, a fifth downsampling layer, and a first upsampling layer. layer, the second upsampling layer, the third upsampling layer, the fourth upsampling layer and the softMax output layer, the loss function of the fully convolutional network is:

代表更新后像素点的特征，

代表求两者之间的交叉熵损失函数；where Φ represents the pixel feature before updating,

represents the feature of the updated pixel point,

Represents the cross entropy loss function between the two;

The structure of the graph convolutional neural network is the input layer, the first graph convolutional layer, the second graph convolutional layer, the third graph convolutional layer and the softmax output layer connected in sequence, and the activation function of each graph convolutional layer is tanh function;

The convolutional neural network structure is sequentially connected input layer, first convolutional layer, first pooling layer, second convolutional layer, second pooling layer and softMax output layer. The activation function of each convolutional layer is LeakyRelu function.

The specific steps of step 6 are:

Step 6.1, initialize the training set and test set as superpixel blocks.

Step 6.2, send the pixel point feature obtained in step 3 and step 4 of the training set into the fully convolutional network to obtain the output result, and the output matrix Q is the soft correlation matrix of the superpixel and the pixel;

Step 6.3, obtain the adjacency matrix A of superpixel block, feature matrix B, conversion matrix C between superpixel and pixel by superpixel and pixel soft correlation matrix;

According to the superpixel and pixel soft correlation matrix Q obtained in step 6.2, each pixel is assigned to the surrounding superpixel block with the highest probability to obtain the superpixel segmentation result of the entire image, and each pixel is obtained according to the superpixel segmentation result. The adjacency matrix A and feature matrix B of the image;

Among them, A _i,j in the adjacency matrix A represents the elements of the i-th row and the j-th column of the adjacency matrix A:

表示超像素块i中第j个像素点特征，n表示第i个超像素块中像素点个数。B ⁱ in the feature matrix B represents the feature of the i-th superpixel block;

represents the feature of the j-th pixel in the superpixel block i, and n represents the number of pixels in the i-th superpixel block.

The transformation matrix of superpixels and pixels is transformation matrix C

Step 6.4, input the adjacency matrix and feature matrix of each image into the graph convolution network, the input is the adjacency matrix and feature matrix of the image, and the output is a two-dimensional tensor H as the superpixel feature of the image;

Step 6.5, convert the superpixel features of the image to pixel features:

_Hgcn = C·H (14)

where C represents the transformation matrix and H represents the output of the graph convolutional network. H _gcn is the pixel feature obtained from the superpixel feature output by the graph convolution network through the transformation matrix of superpixel and pixel

Step 6.6, the convolutional neural network obtains the pixel-level features of the image;

The polarization features obtained in step 3 of the training set are sent to the convolutional neural network as input to obtain pixel-level features;

Step 6.7, classify the image pixel-level features obtained in step 6.6 and the pixel-level features obtained in step 6.5 through the transformation matrix of superpixels and pixels;

Step 6.8, compute the total loss function and pass back iteratively update the network until convergence. The total loss function loss of the end-to-end network is:

loss=loss1+loss2 (15)

where loss1 is the fully convolutional network loss function for superpixel segmentation; loss2 is the classification loss function. Finally, the network is iteratively updated until the network converges, and the end-to-end network model training is completed;

Step 6.9, send the test set into the trained end-to-end network model to obtain the classification result.

The beneficial effects of the present invention are:

The present invention uses a fully convolutional network to perform superpixel segmentation on the image and combines the downstream classification network. The superpixel segmentation network and the downstream classification network perform end-to-end joint training. The downstream classification network is composed of a graph convolutional network and a convolutional network. The neural network consists of these two networks. The results of the superpixel segmentation network can affect the output results of the downstream classification network, and the downstream classification network results can in turn affect the superpixel segmentation results. By iteratively training the network model to complete the image classification work and achieve better classification results.

On the other hand, the downstream classification network uses the graph convolutional network and the convolutional neural network to extract image features at the same time, and fuse the features extracted by the two, so that the graph convolution can learn the global information of the polarimetric SAR image, and the convolutional neural network The pixel-level features of polarimetric SAR images can be learned, and the features extracted from the two can be fused, which can further improve the classification accuracy of polarimetric SAR images.

Description of drawings

Figure 1 is a flow chart of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

The present invention's end-to-end polarimetric SAR image classification method based on superpixels and graph convolution, as shown in Figure 1, includes the following steps:

Step 1, input the polarimetric SAR image to be classified, and crop the polarimetric SAR image into 500 images with a width and height of 512 pixels;

Step 2: Divide the cropped images proportionally, 20% as the training set and 80% as the test set;

Step 3: Decompose the complex scattering matrix of each pixel of each image in the training set and the test set to generate a polarization coherence matrix and convert it into a 1×9 row vector as the polarization feature of the pixel ;

Step 4, splicing the polarization feature on the horizontal and vertical coordinates of the pixel point, and using the spliced row vector with a size of 1×11 as the pixel point feature.

Step 5, build an end-to-end network based on fully convolutional network, graph convolutional network and convolutional neural network

Step 6: The training set is sent to the end-to-end network for joint training, and the test set is sent to the trained end-to-end network to obtain the classification result.

The specific steps of step 3 are:

The complex scattering matrix of each pixel of each image in the training set and the test set is decomposed to generate a polarization coherence matrix and convert it into a 1×9 row vector as the polarization feature of the pixel. The expression of the generated polarization coherence matrix T is as follows:

然后将极化相干矩阵转换为行向量得到特征矩阵：T′＝[T₁₁，T₁₂，T₁₃，T₂₁，T₂₂，T₂₃，T₃₁，T₃₂，T₃₃]，并且该极化相干矩阵为复共轭矩阵，对此复共轭矩阵进行预处理：得到该像素点极化特征向量F：：Denote the polarization coherence matrix T as

Then convert the polarization coherence matrix to row vector to get the characteristic matrix: T′=[T ₁₁ , T ₁₂ , T ₁₃ , T ₂₁ , T ₂₂ , T ₂₃ , T ₃₁ , T ₃₂ , T ₃₃ ], and the polarization The coherence matrix is a complex conjugate matrix, and the complex conjugate matrix is preprocessed to obtain the pixel polarization eigenvector F:::

Among them, Re represents the real part of the complex number, Im represents the imaginary part of the complex number, and T _ij represents the data of the i-th row and the j-th column of the polarization coherence matrix.

The specific steps of step 4 are:

The 9-latitude polarization feature vector of each pixel point generated in step 3 is spliced with the 2-dimensional position feature vector of the pixel point, that is, the horizontal and vertical coordinates of the pixel point, and the spliced 11-latitude vector S is used as the pixel point feature:

Where x, y represent the horizontal and vertical coordinates of the pixel.

In the specific steps of step 5, the structures of the fully convolutional network, the graph convolutional network and the convolutional neural network are:

Build a fully convolutional network:

a) Build a fully convolutional network with the following structure: input layer (1,11,512,512)→first downsampling layer (1,16,512,512)→second downsampling layer (1,32,512,512)→third downsampling layer (1) ,64,512,512)→the fourth downsampling layer (1,128,512,512)→the fifth downsampling layer (1,256,512,512)→the first upsampling layer (1,128,512,512)→the second upsampling layer (1,64,512,512)→the third upsampling layer (1, 32, 512, 512) → the fourth upsampling layer (1, 16, 512, 512) → softMax output layer.

The input data size of the input layer is: (1, 11, 512, 512), where 1 represents the batch size, 11 represents the 1×11 row vector of each pixel, and 512 is the width and height of the image.

The output size of the softMax output layer is: (1, 9, 512, 512). We denote the output data as Q, the soft correlation matrix of superpixels and pixels. Among them, 1 is the batchsize size, 9 means that the pixel belongs to the superpixel block where the pixel is located, and the superpixel block where the pixel is located and its adjacent eight superpixel blocks (sit, top, top right, left, right , lower left, right lower, lower right) probability values. The sum of these 9 probability values is 1.

b) Construct the loss function of the fully convolutional network:

Obtain the polarization features of a superpixel block:

为超像素m的极化特征，f_i,j为像素点(i,j)的极化特征，其中q为超像素和像素的软关联矩阵，

表示像素点(i,j)属于超像素块m的概率值。in,

is the polarization feature of superpixel m, f _i,j is the polarization feature of pixel point (i,j), where q is the soft correlation matrix of superpixel and pixel,

Indicates the probability value that the pixel point (i, j) belongs to the superpixel block m.

Get the coordinate features of the superpixel block:

According to formula (4) and formula (5), the superpixel update process can be written as:

为超像素m的极化特征；

为超像素m的位置特征，p_i，j为像素点(i，j)的位置特征，Φ代表更新前的像素点特征，

表示像素点(i，j)属于超像素块m的概率值。in

is the polarization characteristic of superpixel m;

is the position feature of the superpixel m, p _{i, j} is the position feature of the pixel point (i, j), Φ represents the pixel point feature before updating,

Indicates the probability value that the pixel point (i, j) belongs to the superpixel block m.

Combining the superpixel block feature and the soft correlation matrix of superpixels and pixels to obtain the polarization feature of each pixel:

为像素点(i，j)更新后的极化特征，

为超像素m的极化特征，

表示像素点(i，j)属于超像素块m的概率值。in

is the updated polarization feature of the pixel point (i, j),

is the polarization characteristic of superpixel m,

Indicates the probability value that the pixel point (i, j) belongs to the superpixel block m.

Combine the superpixel block features and the soft correlation matrix of superpixels and pixels to get the position feature of each pixel:

According to formula (7) and formula (8), the pixel feature update process can be written as:

为像素点(i，j)更新后的位置特征，

为超像素m的位置特征，

表示像素点(i，j)属于超像素块m的概率值。

代表更新后的像素点特征。in

is the updated position feature of the pixel point (i, j),

is the location feature of superpixel m,

Indicates the probability value that the pixel point (i, j) belongs to the superpixel block m.

Represents the updated pixel feature.

The fully convolutional network loss function for superpixel segmentation is:

代表更新后像素点的特征，

代表求两者之间的交叉熵损失函数；where Φ represents the pixel feature before updating,

represents the feature of the updated pixel point,

Represents the cross entropy loss function between the two;

代表求两者之间的交叉熵损失函数。in

Represents the cross-entropy loss function between the two.

Build a graph convolutional neural network: its structure is: input layer→first graph convolutional layer→second graph convolutional layer→third graph convolutional layer→softmax output layer, and the activation function of each graph convolutional layer is tanh function.

Build a convolutional neural network: its structure is: input layer→first convolutional layer→first pooling layer→second convolutional layer→second pooling layer→softMax output layer. The activation function of each convolutional layer is the LeakyRelu function.

The specific steps of step 6 are:

The training set is sent to the end-to-end network for training, and the test set is sent to the trained end-to-end network to obtain the classification accuracy of the model.

Step 6.1, initialize superpixels;

Initialize the superpixel, the initial size of the superpixel is 16 pixels in height and width, then the image with a length and width of 512 pixels is divided into 1024 superpixel blocks.

Step 6.2, obtain the soft correlation matrix Q between superpixels and pixels;

The pixel point features obtained in step 3 and step 4 of the training set are sent to the fully convolutional network to obtain the output result. The output result size is: (1, 9, 512, 512). We denote the output matrix as Q, which is the soft-association matrix of superpixels and pixels. Among them, 1 is the batchsize size, 9 means that the pixel belongs to the superpixel block where the pixel is located, and the superpixel block where the pixel is located and its adjacent eight superpixel blocks (sit, top, top right, left, right , lower left, right lower, lower right) probability values. The sum of these 9 probability values is 1.

Step 6.3, obtain the adjacency matrix A of superpixel block, feature matrix B, conversion matrix C between superpixel and pixel by superpixel and pixel soft correlation matrix;

According to the superpixel and pixel soft correlation matrix Q obtained in step 6.2, each pixel is assigned to the surrounding superpixel block with the highest probability to obtain the superpixel segmentation result of the entire image. The adjacency matrix A and feature matrix B of each image are obtained according to the superpixel segmentation results.

The size of the adjacency matrix A is (1024, 1024), and A _i,j represents the elements of the i-th row and the j-th column of the adjacency matrix A.

表示超像素块i中第j个像素点特征，n表示第i个超像素块中像素点个数。The size of the feature matrix B is (1024, 9), and B ⁱ represents the feature of the i-th superpixel block;

represents the feature of the j-th pixel in the superpixel block i, and n represents the number of pixels in the i-th superpixel block.

The size of the transformation matrix C of superpixels and pixels is (512×512,1024), and C _i,j represent the elements of the ith row and jth column of the superpixel and pixel transformation matrix C.

where i represents the pixel subscript of the image, and j represents the superpixel block subscript of the image. i∈[1,512×512]; j∈[1,1024].

Step 6.4, graph convolution to obtain superpixel features of the image;

The adjacency matrix and feature matrix of each image are input into the graph convolutional network, and the input is the adjacency matrix and feature matrix of the image. The output is a two-dimensional tensor H, which is the superpixel feature of the image; the size of the tensor is (1024,64), where the first latitude 1024 represents that the image has 1024 superpixel blocks, and the second latitude 64 represents each superpixel block. Features of the pixel block.

Step 6.5, convert the superpixel features of the image into pixel features;

_Hgcn = C·H (14)

Among them, C represents the transformation matrix between superpixels and pixels obtained in step 6.3, and H represents the output of the graph convolution network, and the size is: (1024,64). The superpixel feature output by the _Hgcn graph convolutional network is the pixel feature obtained by the transformation matrix of superpixel and pixel, and the size is: (512×512,64).

Step 6.6, the convolutional neural network obtains the pixel-level features of the image;

The polarization features obtained in step 3 of the training set are sent to the convolutional neural network as input to obtain pixel-level features;

The input to the convolutional neural network is a four-dimensional tensor of size: (1,512,512,9). The first latitude 1 of the input data represents batchsizoe; the second latitude and the third latitude represent the width and height of the image; the fourth latitude 9 represents the polarization characteristics of each pixel in the image.

The output of the convolutional neural network is a four-dimensional tensor of size: (1,256,256,64). The fourth latitude 64 represents the features extracted by each pixel.

Step 6.7, classify the image pixel-level features obtained in step 6.6 and the pixel-level features obtained in step 6.5 through the transformation matrix of superpixels and pixels;

The superpixel-level feature size obtained in step 6.5 is: (512×512,64); the pixel-level feature size obtained in step 6.6 is: (512×512,64). The pixel-level features and super-pixel-level features are fused to obtain the final extracted feature size of the image: (512×512,128), and finally the softmax classification is performed. Get the classification result and calculate the classification loss loss2. The classification loss here uses the cross-entropy loss function.

Step 6.8, compute the total loss function and pass back iteratively update the network until convergence.

The total loss function loss of the end-to-end network is:

loss=loss1+loss2 (15)

where loss1 is the fully convolutional network loss function for superpixel segmentation; loss2 is the classification loss function. Finally, the network is iteratively updated until the network converges.

Step 6.9, send the test set into the trained end-to-end network model to obtain the classification result.

Claims (5)

1. The end-to-end polarization SAR image classification method based on the superpixel and graph convolution is characterized by comprising the following steps of:

step 1, inputting polarized SAR images to be classified and cutting the images into uniform sizes

Step 2, dividing the cut pictures into a training set and a test set according to the proportion;

step 3, decomposing the complex scattering matrix of each pixel point of each image of the test set training set to generate a polarization coherent matrix and converting the polarization coherent matrix into a row vector serving as the polarization characteristic of the pixel point;

step 4, splicing the polarization characteristics to the horizontal and vertical coordinates of the pixel point, and splicing the row vectors as the characteristics of the pixel point;

step 5, building an end-to-end network based on the full convolution network, the graph convolution network and the convolution neural network;

and 6, sending the training set into an end-to-end network for joint training, and sending the test set into the trained end-to-end network to obtain a result.

2. The end-to-end polarization SAR image classification method based on superpixel and graph convolution according to claim 1, characterized in that the specific steps of step 3 are:

decomposing the complex scattering matrix of each pixel point of each image cut in the step 2 to generate a polarization coherent matrix, and converting the polarization coherent matrix into a row vector with the size of 1 multiplied by 9 as the polarization characteristic of the pixel point; the expression of the generated polarization coherence matrix T is as follows:

let the polarization coherence matrix T be

Then, converting the polarization coherent matrix into a row vector to obtain a characteristic matrix: t ═ T₁₁,T₁₂,T₁₃,T₂₁,T₂₂,T₂₃,T₃₁,T₃₂,T₃₃]And the polarization coherent matrix is a complex conjugate matrix, and the complex conjugate matrix is preprocessed: obtaining a polarization characteristic vector F of the pixel point:

where Re represents the real part of the complex number, Im represents the imaginary part of the complex number, and T_ijRepresents the ith row and jth column data of the polarized coherent matrix.

3. The end-to-end polarization SAR image classification method based on superpixel and graph convolution according to claim 1, characterized in that said step 4 specifically is: and (4) splicing the polarization characteristic vector of each pixel point generated in the step (3) with the horizontal and vertical coordinates of the pixel point to obtain the pixel point characteristics.

4. The method for classifying the SAR image with end-to-end polarization based on the super pixel and graph convolution as claimed in claim 1, wherein the structure of the full convolution network in step 5 is an input layer, a first down-sampling layer, a second down-sampling layer, a third down-sampling layer, a fourth down-sampling layer, a fifth down-sampling layer, a first up-sampling layer, a second up-sampling layer, a third up-sampling layer, a fourth up-sampling layer and a softMax output layer which are connected once, and the loss function of the full convolution network is as follows:

wherein phi represents the pixel point characteristics before updating,

representing the characteristics of the pixel points after the update,

representing the cross entropy loss function between the two;

the graph convolution neural network is structurally characterized by comprising an input layer, a first graph convolution layer, a second graph convolution layer, a third graph convolution layer and a softmax output layer which are sequentially connected, wherein the activation function of each graph convolution layer is a tanh function;

the convolutional neural network structure comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a softMax output layer which are sequentially connected, and the activation function of each convolutional layer is a LeakyRelu function.

5. The end-to-end polarization SAR image classification method based on superpixel and graph convolution according to claim 1, characterized in that the specific steps of step 6 are:

step 6.1, initializing the training set and the test set into a superpixel block;

step 6.2, the pixel point characteristics obtained by the training set through the step 3 and the step 4 are sent to a full convolution network to obtain an output result, and an output matrix Q is a soft correlation matrix of super pixels and pixels;

6.3, acquiring an adjacent matrix A and a characteristic matrix B of the super pixel block and a conversion matrix C between the super pixels and the pixels through the super pixels and the pixel soft association matrix;

according to the superpixel and pixel soft association matrix Q obtained in the step 6.2, distributing each pixel point to the peripheral superpixel blocks with the highest probability to obtain a superpixel segmentation result of the whole image, and obtaining an adjacent matrix A and a feature matrix B of each image according to the superpixel segmentation result;

wherein A in the adjacency matrix A_i,jElements representing the ith row and jth column of the adjacency matrix a:

b in the feature matrix BⁱFeatures representing the ith superpixel block;

representing the characteristics of the jth pixel point in the superpixel block i, wherein n represents the number of the pixel points in the ith superpixel block;

conversion matrix of super-pixel and pixel into conversion matrix C

Step 6.4, inputting the adjacent matrix and the characteristic matrix of each image into a graph convolution network, wherein the input is the adjacent matrix and the characteristic matrix of the image, and the output is the superpixel characteristic of which the two-dimensional tensor H is the image;

step 6.5, converting the super pixel characteristics of the image into pixel characteristics:

H_gcn＝C·H (14)

wherein C represents a transformation matrix and H represents the output of the graph convolution network; h_gcnPixel characteristics obtained by converting super-pixel characteristics output by the image convolution network into a super-pixel and pixel conversion matrix

6.6, obtaining pixel level characteristics of the image by the convolutional neural network;

the polarization characteristics obtained by the training set through the step 3 are used as input and sent to a convolutional neural network to obtain pixel level characteristics;

step 6.7, fusing and classifying the image pixel level characteristics obtained in the step 6.6 and the pixel level characteristics obtained in the step 6.5 through the conversion matrix of the super pixels and the pixels;

6.8, calculating a total loss function and reversely transmitting an iterative updating network until convergence;

the total loss function loss of the end-to-end network is:

loss＝loss1+loss2 (15)

wherein loss1 is the super-pixel-partitioned full convolution network loss function; loss2 is a classification loss function. Finally, the network is updated iteratively until the network is converged, and the end-to-end network model training is completed;

and 6.9, sending the test set into the trained end-to-end network model to obtain a classification result.

Images