Classification method and system for polsar data combining deep learning model and traditional classifier

Abstract

The application discloses a kind of classification method and system for PolSAR data combining deep learning model and traditional classifier, wherein the method adopts random stratified sampling method to select training sample, then trains CNN using training sample, and then divides entire PolSAR data into general pixel and key pixel, then uses AdaBoost algorithm to combine SVM classifier, wishart classifier and decision tree classifier into a strong classifier to classify key pixel again.Finally, the class of important pixel and the class of general pixel are combined as the final result.On the one hand, multiple traditional classifiers are fused, and on the other hand, the deep learning model is combined with the fused traditional classifier to classify the key pixel again, which can greatly improve the classification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data classification, and particularly relates to a classification method and system for PolSAR data combining a deep learning model and a traditional classifier. BACKGROUND

[0002] Synthetic aperture radar (SAR) obtains ground information by actively emitting electromagnetic waves and receiving echoes, and can work all day and all weather without being blocked by light and cloud. Polarimetric synthetic aperture radar (PolSAR) has four polarization channels, and the electromagnetic wave is horizontally transmitted and horizontally received and vertically transmitted and vertically received, so that the full polarization backscatter information of the ground object is obtained, and PolSAR has more advantages in characterizing the polarization characteristics of the ground object. Therefore, using PolSAR data to classify ground objects is one of the important applications of PolSAR.

[0003] The traditional classification method for PolSAR data can be divided into supervised polarization classification and unsupervised polarization classification according to whether a training sample is needed.

[0004] Supervised classification is a process of first training a classifier with training samples whose attribute categories are known, so as to master the statistical characteristics of each category, and then classifying and identifying according to the classification decision rule. The supervised polarization classification for PolSAR data mainly includes maximum likelihood-based supervised classification, support vector machines (SVM)-based supervised classification and decision tree-based supervised classification.

[0005] The maximum likelihood-based supervised classification is based on the assumption that the Bayesian criterion and the probability distribution of each class are Gaussian distribution, and is a risk minimum decision analysis. For example, Cheng et al. proposed a maximum likelihood classifier by analyzing the general characteristics of surface scattering, double bounce scattering and volume scattering, and divided the PolSAR pixels into 9 categories according to 3 PolSAR indexes. Lee et al. extended the maximum likelihood rule to the SAR multi-view case, and developed a polarization covariance matrix supervised classification algorithm based on complex Wishart distribution. However, due to the high complexity and randomness of the statistical distribution of remote sensing information, when the distribution of classes in the feature space is relatively discrete, the classification result obtained will deviate from the actual situation.

[0006] SVM-based supervised classification, based on statistical learning theory, has good generalization ability even with limited training samples. For example, Aghababaee et al. constructed a new fractal feature vector using Pauli decomposition vector and fractal dimension, and used SVM to deal with the problem of nonlinear classifier. However, when only considering the elements of the polarimetric coherence matrix, the SVM results in the L-band are similar to those based on the maximum likelihood criterion. But SVM is highly sensitive to its kernel function and parameters, and its efficiency may decrease in the presence of outliers.

[0007] Decision tree-based supervised classification is a non-parametric algorithm, independent of data distribution, good at mining the correlation between features, and usually uses various polarimetric SAR target decomposition to decompose the scattering matrix into multiple components for classification. For example, Deng et al. proposed a polarimetric SAR data classification method combining polarimetric decomposition and decision tree algorithm. However, when the number of classes is too large, the classification errors of this method may increase rapidly.

[0008] Unsupervised classification refers to a method based on cluster theory, which uses computer to perform statistical analysis on images in the absence of prior knowledge, and establishes decision rules for classification according to the statistical characteristics of the feature parameters of the samples to be classified. For example, Van Zyl proposed a method of unsupervised classification using the relationship between the phase and rotation of incident and scattered waves, which classifies targets into single scattering targets, double scattering targets, mixed targets and unclassifiable targets. This method has limited use of polarimetric information. K-means algorithm converges quickly and efficiently, and can be used to classify the coherence matrix T or the covariance matrix C by calculating the maximum likelihood estimate of each cluster to initialize the cluster center. However, K-means algorithm is sensitive to the selection of initial points, and can only guarantee local optimal classification results. H / α unsupervised classification classifies targets according to the two parameters of scattering entropy H and average scattering angle obtained by Cloude decomposition. This method estimates the type of observed scattering medium from the physical interpretation of the basic scattering mechanism, but any fixed boundary in the H / α plane may not conform to the distribution of the data, and the clustering of similar natural objects may cross the boundary of the decision plane.

[0009] With the development of machine learning, researchers found that the above-mentioned shallow structure classification method has obvious shortcomings in feature extraction and generalization ability when the target object has rich meaning. In recent years, due to the strong learning ability of deep learning model, it can directly learn the rich features of the image from the image data, can solve very complex problems, has good portability, greatly improves the precision of image classification, and shows good performance in PolSAR data classification. At present, deep belief network (deep Boltzmann machine, DBN), deep automatic coding (stacked auto-encoder, SAE), convolutional neural network (, CNN), capsule network and other deep learning models have been applied in PolSAR data classification. For example, Hua et al. proposed a multi-layer Wishart restricted Boltzmann machine (MWRBM) model for PolSAR image classification based on the consideration of PolSAR data obeying Wishart distribution, to improve the classification results. Shang et al. added the relationship between local pixels as a classification feature, and proposed a new method of combining scattering power and stacked sparse auto-encoder (scattering SSAE) for PolSAR classification. Bi et al. proposed an active deep learning method of combining active learning and convolutional neural network (Convolutional Neural Networks, CNN) for minimum supervised PolSAR image classification in order to greatly reduce the labeling cost while improving the classification performance. Cheng et al. believe that a single neuron in CNN cannot represent multiple polarization attributes of land cover, and that a capsule network (CapsNet) can use a vector instead of a single neuron to represent polarization attributes, and proposed a hierarchical capsule network (HCapsNet) for land cover classification of PolSAR images. Although the deep learning method has high classification accuracy, it has large calculation amount, high requirement for hardware, complex model design, dependence on data, low interpretability, and precision decline in the case of unbalanced training data.

[0010] The above classification methods, whether traditional shallow classification methods or deep learning classification methods, do not focus on pixels that are easily mislabeled, affecting the final classification results. SUMMARY

[0011] The present application proposes a classification method and system for PolSAR data combining deep learning model and traditional classifier, to solve or at least partially solve the technical problem of low classification accuracy in the prior art.

[0012] To solve the above technical problems, the technical scheme of the present application is as follows:

[0013] The first aspect provides a classification method for PolSAR data combining a deep learning model and a traditional classifier, comprising:

[0014] classifying the PolSAR data by using the trained CNN model;

[0015] taking the pixels in the classification result that are easily mislabeled as key pixels, and taking other pixels as general pixels, and retaining the class labels of the general pixels;

[0016] using an AdaBoost algorithm to combine an SVM classifier, a wishart classifier and a decision tree classifier into a strong classifier to reclassify the key pixels, and obtaining the classes of the key pixels;

[0017] combining the classes of the key pixels and the classes of the general pixels as the final classification result of the PolSAR data.

[0018] In an embodiment, the classification of the PolSAR data by using the trained CNN model comprises:

[0019] a stratified random sampling method is used to select training samples for each type of ground object;

[0020] the PolSAR data is padded with 0 around the four sides, the padding size is 14, the PolSAR data is cut into different patches by window sliding, the sliding step is 1, the patch size is 15 pixels x 15 pixels x 9 channels, and the class of each patch corresponds to the class of the center pixel;

[0021] the selected patches corresponding to the training samples are used to train the CNN model to obtain the trained CNN model, wherein the trained CNN model comprises an input layer, a hidden layer and an output layer, the hidden layer comprises a plurality of activation functions, a plurality of 3D pooling layers, a plurality of 3D convolution layers and a plurality of fully connected layers;

[0022] the PolSAR data is input into the trained CNN model to obtain the output of the last fully connected layer of the hidden layer and the classification result.

[0023] In an embodiment, the last fully connected layer of the hidden layer outputs an NxC matrix according to the input data, wherein N is the number of pixels of the PolSAR data, C is the total number of classes, the row of the NxC matrix represents the C class prediction probabilities of a certain pixel; the maximum prediction probability and the second largest prediction probability of each pixel are found, and the non-negative difference between the two is taken as a probability difference, if the probability difference is less than a given threshold, the pixel is taken as a key pixel, and other pixels are taken as general pixels.

[0024] In an embodiment, when the PolSAR data is classified by using the trained CNN model, the selected classification features include 9 elements of the coherence matrix T storing the backscattering information of the ground objects: T11, T12_real, T12_imaginary, T13_real, T13_imaginary, T22, T23_real, T23_imaginary, and T33.

[0025] In an embodiment, the key pixels are reclassified by using an AdaBoost algorithm to combine an SVM classifier, a wishart classifier, and a decision tree classifier into a strong classifier, including:

[0026] Different weak classifiers are trained by using the same training set;

[0027] The trained weak classifiers are combined to construct a strong classifier, in the process, the weight of the sample misclassified by the previous weak classifier is increased, the weight of the correctly classified sample is reduced, and is used again to train the next weak classifier, at the same time, a new weak classifier is added in each iteration until the preset error rate or the maximum number of iterations is reached to determine the final strong classifier, and in each iteration process, a class participating in the classification is regarded as 1, and other classes not participating in the classification are regarded as -1.

[0028] The key pixels are reclassified by using the constructed strong classifier.

[0029] Based on the same inventive concept, the second aspect of the present application provides a classification system for PolSAR data combining a deep learning model and a traditional classifier, including:

[0030] An initial classification module is configured to classify the PolSAR data by using a trained CNN model;

[0031] A key pixel screening module is configured to regard the pixels easily mislabeled in the classification result as key pixels, and regard other pixels as general pixels, and retain the class labels of the general pixels;

[0032] A reclassification module is configured to reclassify the key pixels by using an AdaBoost algorithm to combine an SVM classifier, a wishart classifier, and a decision tree classifier into a strong classifier, and obtain the classes of the key pixels;

[0033] A final classification result acquisition module is configured to combine the classes of the key pixels and the classes of the general pixels as the final classification result of the PolSAR data.

[0034] Based on the same inventive concept, the third aspect of the present application provides a computer readable storage medium, having stored thereon a computer program which, when executed, implements the method of the first aspect.

[0035] Based on the same inventive concept, the fourth aspect of the present application provides a computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of the first aspect when executing the program.

[0036] Compared with the prior art, the technical solution provided by the present application has at least the following technical effects:

[0037] Firstly, the present application uses CNN to classify PolSAR data, locates pixels that are easily mislabeled as key pixels, and other pixels as general pixels, and retains the class labels of general pixels; then uses the AdaBoost algorithm to combine the SVM classifier, the wishart classifier and the decision tree classifier into a strong classifier to classify the key pixels again. Finally, the class of the important pixels and the class of the general pixels are combined as the final result, first using a deep learning model for preliminary classification, and using the AdaBoost algorithm to combine the SVM classifier, the wishart classifier and the decision tree classifier into a strong classifier to classify the key pixels again. Since the key pixels that are easily mislabeled are considered, and the deep learning model and the traditional classifier are combined for classification, the accuracy and precision of the classification can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings described below are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0039] Figure 1 is the overall framework diagram of the classification method for PolSAR data combining deep learning model and traditional classifier provided by the embodiment of the present application;

[0040] Figure 2 is the flowchart of locating key pixels in the embodiment of the present application;

[0041] Figure 3 is the schematic diagram of the key pixels found out when the probability difference is 0.1 to 0.8 in the experimental data in the embodiment of the present application;

[0042] Figure 4is the overall accuracy of the PolSAR image (experimental data) in different probability differences in the embodiment of the application;

[0043] Figure 5 is the label of the key pixel in the experimental data when the probability difference is 0.7 in the embodiment of the application, wherein part (a) is before reclassification, and part (b) is after reclassification;

[0044] Figure 6 is a classification result schematic diagram of the experimental data using different methods respectively in the embodiment of the application, wherein part (a) adopts the method of CNN, part (b) adopts the method of SVM, part (c) adopts the method of Wishart, part (d) adopts the method of decision tree, part (e) adopts the method of AdaBoost, and part (f) adopts the method of the application. DETAILED DESCRIPTION

[0045] The present application inventors find through a large amount of research and practice that any classifier has its own advantages and disadvantages, and the advantages of multiple classifiers can be obtained by fusing multiple classifiers. The classification results of multiple classifiers are fused to obtain better classification results than using a single classifier.

[0046] However, in the existing classification methods, whether it is a traditional shallow classification method, a deep learning classification method, or a classification method of fusing multiple classifiers, the pixels that are easily mislabeled are not considered, which affects the final classification result. To solve this problem, the present application reclassifies the pixels that are easily mislabeled to obtain the final classification result. The present application first classifies the PolSAR data using CNN, locates the pixels that are easily mislabeled as key pixels, and the others as general pixels, and retains the class labels of the general pixels; then uses the AdaBoost algorithm to combine the SVM classifier, the wishart classifier and the decision tree classifier into a strong classifier to reclassify the key pixels. Finally, the class of the important pixels and the class of the general pixels are combined as the final result. The present application fuses multiple traditional classifiers on the one hand, and combines the deep learning model with the fused traditional classifiers to reclassify the key pixels on the other hand, which can greatly improve the classification accuracy.

[0047] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0048] Embodiment one

[0049] The application provides a classification method for PolSAR data combining a deep learning model and a traditional classifier, comprising:

[0050] Classifying the PolSAR data by using the trained CNN model;

[0051] Taking the pixels in the classification result that are easily mislabeled as key pixels, taking other pixels as general pixels, and retaining the class labels of the general pixels;

[0052] Using an AdaBoost algorithm to combine an SVM classifier, a wishart classifier and a decision tree classifier into a strong classifier to reclassify the key pixels, and obtaining the classes of the key pixels;

[0053] Combining the classes of the key pixels and the classes of the general pixels as the final classification result of the PolSAR data.

[0054] Specifically, referring to Figure 1 is the overall framework diagram of the classification method for PolSAR data combining a deep learning model and a traditional classifier provided by the embodiment of the application. Firstly, the PolSAR data is preliminarily classified by using a trained convolutional network model, and key pixels are located, then a strong classifier is obtained by integrating multiple classifiers by using an AdaBoost algorithm, the located key pixels are reclassified by using the strong classifier, and finally the classification result of the general pixels and the classification result of the key pixels are combined as the final classification result.

[0055] In one embodiment, the PolSAR data is classified by using the trained CNN model, comprising:

[0056] A stratified random sampling method is used to select training samples for each type of surface feature;

[0057] The PolSAR data is filled with 0 around the four sides, the filling size is 14, the PolSAR data is cut into different patches by window sliding, the sliding step is 1, the patch size is 15 pixels x 15 pixels x 9 channels, and the class of each patch corresponds to the class of the center pixel;

[0058] The selected training samples are used to train the CNN model, and a trained CNN model is obtained, wherein the trained CNN model comprises an input layer, a hidden layer and an output layer, the hidden layer comprises multiple activation functions, multiple 3D pooling layers, multiple 3D convolution layers and multiple fully connected layers;

[0059] The PolSAR data is input into the trained CNN model to obtain the output of the last full connection layer of the hidden layer and the classification result.

[0060] Specifically, the training data for training the CNN model is from a ground object real distribution reference map corresponding to each PolSAR data, and the number of pixels of each ground object is different. In order to avoid the adverse effects of the imbalance of the number of training samples of each ground object on the final result, the present application adopts a stratified random sampling method to select training samples for each ground object. Stratified random sampling is suitable for survey objects with a large number of units and large internal differences, and has a small sampling error.

[0061] The PolSAR data is filled with 0 around the four sides, and the filling size is 14. The PolSAR data is cut into patches by window sliding, the sliding step is 1, and the patch size is 15 pixels x 15 pixels x 9 channels. The category of each patch corresponds to the category of the center pixel. Then, a CNN classification model is obtained by training the CNN using the patch corresponding to the training sample. The structure of the CNN used in the present embodiment is shown in Table 1.

[0062] Table 1 CNN structure

[0063]

[0064] wherein, Input Layer, Hidden Layers, Output Layer are input layer, hidden layer and output layer respectively, 3D-convolution Layer, Activation Function, 3D-Pooling Layer, Fully Connected Layer, Classification Layer represent 3D convolution layer, activation function, 3D pooling layer, full connection layer and classification layer respectively.

[0065] In one embodiment, the last full connection layer of the hidden layer outputs an NxC matrix according to the input data, wherein N is the number of pixels of the PolSAR data, and C is the total number of categories, the row of the NxC matrix represents the C category prediction probability of a certain pixel; the maximum prediction probability and the second largest prediction probability of each pixel are found, and the non-negative difference between the two is taken as the probability difference. If the probability difference is less than a given threshold, the pixel is taken as a key pixel, and other pixels are taken as general pixels. The category information of the general pixels is stored.

[0066] Specifically, the PolSAR image data is input into the trained CNN classification model to obtain the output of the last fully connected layer and the final classification result. The fully connected layer (FC) can further fuse the features, so that the neural network finally sees the global features.

[0067] The positioning process of the key pixels is as shown in Figure 2

[0068] In an embodiment, when the PolSAR data is classified by using the trained CNN model, the selected classification features include 9 elements of the coherence matrix T storing the backscattering information of the ground object: T11, T12_real, T12_imaginary, T13_real, T13_imaginary, T22, T23_real, T23_imaginary, T33, wherein T11 represents the element of the first row and the first column of the coherence matrix, T12_real represents the real part of the element of the first row and the first column of the coherence matrix, T12_imaginary represents the imaginary part of the element of the first row and the first column of the coherence matrix, and the others are similar.

[0069] Specifically, the PolSAR obtains the polarization scattering matrix by measuring the scattering echo in each resolution unit of the ground:

[0070]

[0071] The scattering matrix [S] can only describe the so-called coherent or pure scatterers, and cannot be used to describe the so-called distributed scatterers. Due to the existence of speckle noise, such scatterers can only be statistically described. In order to reduce the influence of speckle noise, only the second-order polarization expression can be used to analyze the distributed scatterers. The coherence matrix is one of the second-order description factors.

[0072]

[0073] It has been proved by existing research work that the backscattering information of the ground object is mainly concentrated in the coherence matrix T, and therefore the 9 elements T11, T12_real, T12_imaginary, T13_real, T13_imaginary, T22, T23_real, T23_imaginary, T33 of the coherence matrix T are selected as the classification features of the present application.

[0074] In an embodiment, the SVM classifier, the wishart classifier and the decision tree classifier are combined into a strong classifier by using the AdaBoost algorithm to re-classify the key pixels, including:

[0075] ​Different weak classifiers are trained using the same training set;

[0076] The weak classifiers trained are assembled to build a strong classifier, in which the weight of the sample misclassified by the previous weak classifier is increased, and the weight of the sample correctly classified is decreased, and is used to train the next weak classifier again, while in each iteration, a new weak classifier is added until the preset error rate or the maximum number of iterations is reached to determine the final strong classifier, and in each iteration process, one of the classes participating in the classification is regarded as 1, and the other classes not participating in the classification are regarded as -1.

[0077] The strong classifier built is used for re-classification of the key pixels.

[0078] Specifically, in this step, the SVM classifier, the Wishart classifier and the decision tree classifier are used as weak classifiers, and the three are combined into a strong classifier by using the AdaBoost algorithm to re-classify the key pixels.

[0079] Firstly, the SVM classifier, the Wishart classifier and the decision tree classifier are trained using the training samples to obtain three classification models.

[0080] The SVM is a binary classification model. The basic idea of SVM learning is to solve the separating hyperplane w T x+b=0, wherein x is a value of the hyper space, w is a normal vector of the hyperplane, and b is the distance of the hyperplane from the origin. The libsvm open source library is used in the application to realize the SVM multi-class classification.

[0081] Lee et al. extended the maximum likelihood rule to the SAR multi-view case, and developed a supervised classification algorithm based on the complex Wishart distribution of the polarization covariance matrix. Since the polarization covariance matrix can be transformed into the polarization coherence matrix [T] by linear transformation, the polarization coherence matrix [T] also obeys the complex Wishart distribution. The maximum likelihood criterion is applied to the complex Wishart distribution to obtain the decision rule wherein ω i is the class; Argmaxf(x) returns the x value corresponding to the maximum value of f(x); is the maximum likelihood estimation of the coherence matrix. Taking the negative of the above formula and removing the term irrelevant to the study of clustering, a new decision rule can be obtained wherein, Tr is the trace of the matrix. The pixel is assigned to the class ω i with the smallest distance.

[0082] The decision tree algorithm recursively selects the optimal feature and splits the training data according to the feature, so that each sub-data set has the best classification. The decision tree algorithm is to measure the attribute selection by information gain, and select the attribute with the maximum information gain after splitting. Assuming that the proportion of the kth class of samples in the current sample set X is P k , then the information entropy H(X) of X is defined as H(X) = -∑ k＝1 P k log2P k , the smaller the value of H(x), the higher the purity of X. H(X|Y) = ∑ k＝1 P(y k )H(X=x k |Y) is the conditional entropy, which represents the information entropy of a single feature, and P(y k ) represents the proportion of a certain feature within its feature. Information gain I(X,Y) = H(X)-H(X|Y) refers to the degree of reduction in overall feature uncertainty of the sample after knowing a certain feature. Finally, pruning the generated tree obtains a decision tree classification model.

[0083] AdaBoost is an iterative algorithm, and its core idea is to use the same training set to train different weak classifiers, and then combine these weak classifiers to form a strong classifier. In this process, the weight of the sample misclassified by the previous weak classifier will increase, and the weight of the correctly classified sample will decrease, and will be used again to train the next weak classifier. At the same time, a new weak classifier is added in each iteration until a predetermined sufficiently small error rate or a predetermined maximum number of iterations is reached to determine the final strong classifier.

[0084] Given a training data set: (x1,y1),…,(x N ,y N ), where y i ∈{1,-1}, i = 1,…,N, indicating the class label of the training sample. Initialize the weight distribution D1 of the training sample set. Each training sample is assigned the same weight W i = 1 / N, i = 1,…,N, and the initial weight distribution of the training set is:

[0085] D1 = (w1,…,w N ) = (1 / N,…,1 / N) (3)

[0086] where w i is the weight of each training sample, and N is the number of training samples.

[0087] Then the following tth iteration is performed, t = 1,2,3.

[0088] Select the weak classifier h with the lowest current error rate as the Hth classifier. t Calculate the distribution D using a base classifier. t Error on

[0089]

[0090]

[0091] Among them, D t =(w1,…, N ): The weight distribution of the training sample set at the t-th iteration; e t I is the error rate, and I is the indicator function.

[0092] Calculate the weight of the base classifier in the final strong classifier:

[0093]

[0094] Where, α t : Weights of the weak classifier.

[0095] Update the weight distribution D of the training samples t+1 :

[0096]

[0097] Z t Normalization constant

[0098] Finally, the weak classifiers are combined according to their weights.

[0099]

[0100] By applying the sign function, a strong classifier is obtained as follows:

[0101]

[0102] The AdaBoost algorithm described above is for binary classification problems. Since the PolSAR data used in the specific experiments contains multiple categories, in this invention, each iteration treats a category that has not yet been classified as 1, and the other categories that have not yet been classified as -1. Finally, a C-fold strong classifier is constructed to reclassify the key pixels, obtaining new categories for the key pixels.

[0103] Finally, the classification results of general pixels and the classification results of key pixels are combined to form the PolSAR classification result.

[0104] The method proposed in the application is verified and explained by specific experimental data.

[0105] 1. Experimental data

[0106] The first PolSAR data is L-band full polarimetric 4-look data of Flevoland region in central Netherlands acquired by NASA / JPL AIRSAR system on August 16, 1989. The image size is 750 pixels x 1024 pixels, the azimuth resolution is 12.1 m, the range resolution is 6.7 m, and 11 types of ground objects are contained, namely, Bareland, Beet, Grass, Lucerne, Pea, Potato, Rape, Soybean, Water, Wheat, and Wood (representing bare land, beet, grass, lucerne, pea, potato, rape, soybean, water, wheat, and wood, respectively). A detailed survey of the region was conducted by the JPL laboratory at the same time of imaging, and a reference map of the real ground object distribution was obtained, which provides a basis for evaluating the segmentation accuracy.

[0107] 2. Parameter setting

[0108] The structure of the CNN in the embodiment is shown in Table 1. The input window size of the CNN is 15x15x9. The first layer of the convolution kernel is 3x3x5, the number of filters of the convolution layer is 8, the excitation function is Rectified Linear Unit (ReLU), and the pooling layer is maximum pooling with a size of 2x2x1 and a step of 1. The second layer of the convolution kernel is 3x3x4, the number of filters of the convolution layer is 4, the excitation function is Rectified Linear Unit (ReLU), and the pooling layer is maximum pooling with a size of 2x2x1 and a step of 1. The first full connection layer has 256 neurons, and dropout is used to prevent overfitting during training, and the probability of dropout is 0.4. The second full connection layer has 128 neurons, and dropout is used to prevent overfitting during training, and the probability of dropout is 0.4. The number of neurons in the last full connection layer is the number of categories. Finally, a softmax classifier with cross-entropy loss is used for classification. The learning rate is 0.001, and the number of iterations is 100.

[0109] 3. Training samples

[0110] In the specific implementation process, the training samples are selected from the PolSAR data by stratified random sampling. Table 2 shows the number of samples of each type of ground object for each data. The total number of training samples is determined according to the number of categories of ground objects, and the PolSAR data used in the experiment has 11 types of ground objects, so the total number of training samples is 11000.

[0111] Table 2 Number of samples of each type of ground object

[0112]

[0113] 4. Experimental results and analysis

[0114] 4.1 Precision comparison before and after re-classification of key pixels

[0115] Figure 3 Figures 4a-4h show the key pixels found in the experimental data when the probability difference is 0.1-0.8, respectively. In Figure 4a, the probability difference is 0.1; in Figure 4b, the probability difference is 0.2; in Figure 4c, the probability difference is 0.3; in Figure 4d, the probability difference is 0.4; in Figure 4e, the probability difference is 0.5; in Figure 4f, the probability difference is 0.6; in Figure 4g, the probability difference is 0.7; and in Figure 4h, the probability difference is 0.8. It can be found that most of these points are located at the edge of each type of ground object, i.e., in the area where the backscattering information is relatively complex. As the value of the probability difference increases, the coverage of the key pixels gradually expands from the edge of the ground object to the interior of the ground object, and some even cover the entire block.

[0116] In order to obtain the best probability difference by comparison, classification was performed using the present application under each probability difference. The overall classification accuracy of the classification results of the experimental data is shown in Figure 5. It can be found that when the probability difference is 0.7, the overall classification accuracy reaches a maximum value of 92.10%. Figure 4

[0117] Figure 6 shows the category diagrams of the key pixels in the experimental data before and after re-classification, and it can be found that the category of some pixels has changed. Figure 5

[0118] According to the reference diagram of the real ground object distribution of the experimental data, the classification correctness of each type of ground object in the key pixels before and after re-classification was counted, and the results are shown in Table 3. Among the 11 types of ground objects, the classification correctness of 10 types of ground objects was improved. Among them, the correctness of wood was improved by the smallest amount of 0.3%, the correctness of soybean was improved by the largest amount of 38.29%, and the correctness of water was reduced from 88.89% to 59.26%. From Table 3, it can be found that the overall classification accuracy of the key pixels in the experimental data was improved by 20.77%, which indicates that the present application is effective in improving the classification accuracy of key pixels.

[0119] Table 3 Correctness of each type of ground object in the key pixels in the experimental data before and after re-classification when the probability difference is 0.7

[0120]

[0121] ​​

[0122] Note that the average improvement accuracy in Table 3 is an average of the improvement accuracy of 11 ground objects, which is taken as the overall accuracy improvement degree.

[0123] 4.2 Comparison of visual effects of various methods

[0124] SVM, Wishart and Decision tree in the shallow classifier, CNN in deep learning, and AdaBoost algorithm using SVM, Wishart and Decision tree as weak classifiers are trained respectively using the same training sample, and then the corresponding classification results are obtained. The classification results of the present application are compared in visual effect.

[0125] As shown in Figure 6 , the classification results of CNN, SVM, Wishart, decision tree, AdaBoost and the present application are shown respectively. Referring to the real ground object distribution map, 4 oval regions are selected on each classification result for comparison. It can be found that the classification result of CNN is more pure than that of SVM, Wishart, decision tree and AdaBoost, and the spot condition is less, but the classification result is more blurred and the texture is lost, while the texture of the result of the present application is clearer than that of the deep learning model CNN, and the spot-like classification condition is greatly improved than that of the classification results of the above-mentioned shallow classifiers.

[0126] 4.3 Comparison of classification accuracy of various methods

[0127] Using the confusion matrix, the classification results of two experimental data are analyzed from the aspects of overall accuracy, Kappa coefficient, producer accuracy and user accuracy with the real ground object distribution map as a reference. The classification results of the present application are compared with the classification results of CNN, SVM, Wishart, decision tree and AdaBoost.

[0128] The overall accuracy and Kappa coefficient of the classification results of the experimental data are shown in Table 4. In terms of accuracy, the overall classification accuracy of the present application is improved by 2.22 to 20.38 percentage points compared with the other five classification methods, and the Kappa coefficient is improved by 0.03 to 0.24.

[0129] Table 4 Overall accuracy and Kappa coefficient of experimental data

[0130]

[0131] The producer's accuracy of each ground object in the experimental data under different methods is shown in Table 5. The last row of the table is the standardized difference (STD) of the producer's accuracy of all ground objects under each method. The producer's accuracy of each ground object under the present application is greater than the minimum value of the producer's accuracy of various classification methods and close to or greater than the maximum value. The producer's accuracy of bareland under the Decision tree is the minimum value, 0.00%, and the producer's accuracy of bareland under the Wishart is the maximum value, 97.64%, and the producer's accuracy of bareland under the present application is 84.66%. Similar to bareland, the producer's accuracy of beet, grass, pea, potato, rape and water under the present application is greater than the minimum value and close to the maximum value. The producer's accuracy of lucerne, soybean, wheat and wood under the present application is greater than the producer's accuracy under other methods, for example, the producer's accuracy of lucerne under the Decision tree is the minimum value, 41.58%, and the producer's accuracy of lucerne under the present application is the maximum value, 94.74%. The standardized difference is used to judge the balance of the producer's accuracy of all ground objects under each method, and the smaller the value, the more balanced the data. From the data in Table 5, the standard deviation of the six classification methods is the maximum value of 0.3112 for the Decision tree and the minimum value of 0.0557 for the AdaBoost. The standard deviation of the present application is 0.0569, which is 0.0012 greater than the minimum value. However, compared with the AdaBoost, the producer's accuracy of 10 of the 11 ground objects under the present application is improved, and only the producer's accuracy of bareland is reduced by 8.69%. This shows that the producer's accuracy of the present application is relatively balanced. Therefore, whether locally or globally, the present application is superior to the other five classification methods in terms of the producer's accuracy.

[0132] Table 5 Producer's accuracy of experimental data (PolSAR image)

[0133]

[0134]

[0135] As shown in Table 6, the user accuracy of each ground object in the experimental data under different methods, the last row of the table is the standardized difference of the user accuracy of all ground objects under each method, and the proposed method is the method of the application. The user accuracy of each ground object under the application is the minimum value of the user accuracy of various classification methods, close to the maximum value or even greater than the maximum value. The user accuracy of lucerne under CNN is the minimum value, which is 87.95%, the user accuracy of lucerne under Decision tree is the maximum value, which is 98.29%, and the user accuracy of lucerne under the classification method of the application is 95.08%. Similarly, in the application, the user accuracy of pea, rape, water, wheat and wood is greater than the minimum value and close to the maximum value. The user accuracy of beet, grass, potato and soybean under the application is greater than the user accuracy under other methods, for example, the user accuracy of beet under Wishart is the minimum value, which is 78.17%, and the user accuracy of beet under the application is the maximum value, which is 92.95%. The standardized difference is used to judge the balance of the user accuracy of all ground objects under each method, and the smaller the value, the more balanced the data. From the data in Table 6, the standard deviation of the six classification methods, the maximum value is 0.2852 of Decision tree, and the minimum value is 0.0435 of SVM. The standard deviation of the application is 0.0482, which is 0.0047 greater than the minimum value and very close to the minimum value. Compared with SVM, the user accuracy of 9 of the 11 ground objects under the application is improved, and the user accuracy of pea and rape is reduced by 1.47% and 0.07% respectively. Therefore, in terms of user accuracy, the application is better than the other five classification methods.

[0136] Table 6 User accuracy of the first PolSAR image

[0137]

[0138]

[0139] Embodiment two

[0140] Based on the same inventive concept, the embodiment discloses a classification system for PolSAR data combining a deep learning model and a traditional classifier, comprising:

[0141] An initial classification module is configured to classify the PolSAR data by using the trained CNN model;

[0142] A key pixel screening module is configured to screen the pixels that are easily mislabeled in the classification result as key pixels, and other pixels as general pixels, and retain the class labels of the general pixels;

[0143] a re-classification module, configured to re-classify the key pixels by using an AdaBoost algorithm to combine the SVM classifier, the wishart classifier and the decision tree classifier into a strong classifier, to obtain the categories of the key pixels;

[0144] a final classification result acquisition module, configured to combine the categories of the key pixels and the categories of the general pixels as the final classification result of the PolSAR data.

[0145] Since the system introduced in the embodiment two is the system used in the method for classifying the PolSAR data by combining the deep learning model and the traditional classifier in the embodiment one, the specific structure and the deformation of the system can be understood by those skilled in the art based on the method introduced in the embodiment one, and thus will not be repeated here. Any system used in the method in the embodiment one belongs to the range of the present application.

[0146] Embodiment three

[0147] Based on the same inventive concept, the present application further provides a computer readable storage medium having a computer program stored thereon, the program being executed to implement the method described in the embodiment one.

[0148] Since the computer readable storage medium introduced in the embodiment three is the computer readable storage medium used in the method for classifying the PolSAR data by combining the deep learning model and the traditional classifier in the embodiment one, the specific structure and the deformation of the computer readable storage medium can be understood by those skilled in the art based on the method introduced in the embodiment one, and thus will not be repeated here. Any computer readable storage medium used in the method in the embodiment one belongs to the range of the present application.

[0149] Embodiment four

[0150] Based on the same inventive concept, the present application further provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor executing the above-mentioned program to implement the method in the embodiment one.

[0151] Since the computer device introduced in the embodiment four is the computer device used in the method for classifying the PolSAR data by combining the deep learning model and the traditional classifier in the embodiment one, the specific structure and the deformation of the computer device can be understood by those skilled in the art based on the method introduced in the embodiment one, and thus will not be repeated here. Any computer device used in the method in the embodiment one belongs to the range of the present application.

[0152] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer-readable program code.

[0153] The present application is described in reference to the flowchart and / or block diagrams of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing element or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified by one or more blocks Figure 1 one or more functions specified by one or more blocks

[0154] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such modifications and variations as fall within the true spirit and scope of the application. It is evident that those skilled in the art can make various changes and modifications of the embodiments without departing from the spirit and scope of the application. Therefore, the application is not intended to be limited to the embodiments described herein but is to be accorded the full scope of the claims and their equivalents.

Claims

1. A classification method for PolSAR data combining a deep learning model and a traditional classifier, characterized in that, The application relates to a method for classifying PolSAR data, comprising: classifying PolSAR data by using a trained CNN model; taking pixels in a classification result that are easily mislabeled as key pixels, and taking other pixels as general pixels, and retaining the class labels of the general pixels; using an AdaBoost algorithm to combine an SVM classifier, a wishart classifier and a decision tree classifier into a strong classifier to reclassify the key pixels, and obtaining the classes of the key pixels; combining the classes of the key pixels and the classes of the general pixels as the final classification result of the PolSAR data.

2. The method for classification of PolSAR data combining deep learning model and traditional classifier according to claim 1, wherein, The method for classifying PolSAR data by using a trained CNN model comprises: adopting a stratified random sampling method to select training samples for each type of ground object; filling the periphery of PolSAR data with 0, and filling the size to be 14; cutting the PolSAR data into different patches by window sliding, and taking the sliding step to be 1; the size of each patch is 15 pixels*15 pixels*9 channels, and the class of each patch corresponds to the class of the center pixel; training the CNN model by using the patches corresponding to the selected training samples, and obtaining the trained CNN model, wherein the trained CNN model comprises an input layer, a hidden layer and an output layer; the hidden layer comprises multiple activation functions, multiple 3D pooling layers, multiple 3D convolution layers and multiple fully connected layers; inputting the PolSAR data into the trained CNN model, and obtaining the output of the last fully connected layer of the hidden layer and a classification result.

3. The method for classification of PolSAR data combining deep learning model and traditional classifier according to claim 2, characterized in that, The last fully connected layer of the hidden layer outputs an NxC matrix according to the input data, wherein N is the number of pixels of the PolSAR data, C is the total number of classes, the row of the NxC matrix represents the C-class prediction probability of a certain pixel; the maximum prediction probability and the second maximum prediction probability of each pixel are found, and the non-negative difference between the two is taken as a probability difference; if the probability difference is less than a given threshold, the pixel is taken as a key pixel, and other pixels are taken as general pixels.

4. The method for classification of PolSAR data combining deep learning model and traditional classifier according to claim 1, wherein, When the PolSAR data is classified by using the trained CNN model, the selected classification features comprise nine elements of a coherence matrix T that stores the backscattering information of a ground object: T11, T12_real, T12_imaginary, T13_real, T13_imaginary, T22, T23_real, T23_imaginary and T33.

5. The method for classification of PolSAR data combining deep learning model and traditional classifier according to claim 1, characterized in that, The method for reclassifying the key pixels by using an AdaBoost algorithm to combine an SVM classifier, a wishart classifier and a decision tree classifier into a strong classifier comprises: training different weak classifiers by using the same training set; The trained weak classifiers are assembled to construct a strong classifier, in the process, the weight of the sample misclassified by the previous weak classifier is increased, the weight of the correctly classified sample is reduced, and is used again to train the next weak classifier, at the same time, a new weak classifier is added in each iteration until the preset error rate or the maximum number of iterations is reached to determine the final strong classifier, and in each iteration process, one of the categories involved in the classification is regarded as 1, and the other categories not involved in the classification are regarded as-1; The constructed strong classifier is used for re-classification of the key pixels.

6. A classification system for PolSAR data combining a deep learning model and a traditional classifier, characterized in that, The program is executed to implement the method in any one of claims 1 to 5. The processor implements the method in any one of claims 1 to 5 when executing the program. The processor implements the method in any one of claims 1 to 5 when executing the program. ​ ​ 7. A computer-readable storage medium having stored thereon a computer program, characterized in that, ​ 8. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, ​

Images