Polarimetric SAR image classification method based on complex NSCT neural network and statistical features

Abstract

The present invention discloses a polarimetric SAR image classification method based on a complex NSCT neural network and statistical features, comprising: preprocessing the polarimetric SAR images to be classified and dividing them into a training set and a test set; constructing a complex NSCT neural network and setting network parameters; the complex NSCT neural network includes a complex spatial domain feature learning module, a complex frequency domain NSCT feature learning module, a statistical feature integration module, and a classification module; inputting the training set into the complex NSCT neural network to train the network; and classifying the test set using the trained complex NSCT neural network to obtain polarimetric SAR image classification results. This method alleviates the interference of speckle noise on texture information extraction, enhances the sparse representation of the directionality, singularity, and regularity of image features, and improves the network's classification capability and approximation performance, thereby improving polarimetric SAR image classification results.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a polarimetric SAR image classification method based on a complex NSCT neural network and statistical features. Background Art

[0002] Polarimetric Synthetic Aperture Radar (PolSAR) is a multi-channel coherent microwave imaging system that represents the polarimetric scattering characteristics of ground objects using a polarimetric scattering matrix. Polarimetric SAR images, due to their all-day, all-weather, and high-resolution capabilities, play an important role in tasks such as land cover classification. Polarimetric SAR image classification is a hot research topic in remote sensing.

[0003] Traditional polarimetric SAR image classification methods include those based on polarimetric scattering characteristics and those based on statistical characteristics. The former extracts the polarimetric scattering characteristics of the target through polarimetric target decomposition, such as Pauli decomposition and Freeman decomposition. The obtained features are then fed into a classifier for classification. Common classifiers include the Wishart classifier and support vector machine. The latter classifies images based on the statistical distribution characteristics of polarimetric SAR. For example, in 1988, Kong et al. proposed a maximum likelihood classifier based on complex Gaussian distribution for single-look polarimetric SAR (Kong JA, Swartz AA, Yueh HA, et al. Identification of terrain cover using the optimum polarimetric classifier [J]. Journal of Electromagnetic Waves and Applications, 1988, 2(2):171-94). In 1994, Lee et al. established a maximum likelihood classifier based on the Wishart distribution for multi-look polarimetric SAR data (Lee JS, Grunes MR, Kwok R. Classification of multi-look polarimetric SAR imagery based on complex Wishart distribution [J]. International Journal of Remote Sensing, 1994, 15(11): 2299-311.). Subsequent research has been developed under these two frameworks. However, traditional polarimetric SAR image classification methods are based on shallow features designed based on artificial experience and cannot fully characterize the ground object attributes and neighborhood correlation of target pixels. Moreover, as the spatial resolution of polarimetric SAR increases, its computational cost also increases accordingly.

[0004] In recent years, convolutional neural networks (CNN) have achieved remarkable classification performance in remote sensing image classification tasks due to their powerful feature extraction and abstraction capabilities.

[0005] However, existing remote sensing image classification methods based on convolutional neural networks still have the following limitations: 1. In order to increase the learning ability of the network model, the number of network layers and neurons needs to be adjusted, which introduces a large number of parameters for learning and adjustment, while increasing the time cost; and insufficient number of samples will lead to overfitting of the network, reducing the classification accuracy. 2. The parameters of CNN are randomly initialized, and when the background of the input image is relatively complex, it is easy to fall into local optimality. 3. CNN has certain limitations in representing data structure information. This is especially true for polarization SAR images with complex phase information. 4. CNN lacks the ability to mine the frequency domain information of the image. Since CNN learns the pixel matrix of the image in the spatial domain, a single feature space cannot fully represent the image features, ignoring the multi-scale and multi-directional characteristics in the frequency domain.

[0006] Therefore, how to fully exploit the effective representation of images based on the characteristics of polarimetric SAR images by using CNN in the spatial domain and multi-scale and multi-resolution analysis in the frequency domain has become the key to improving the performance of polarimetric SAR image classification models. Summary of the Invention

[0007] To address the above-mentioned problems in the prior art, the present invention provides a polarimetric SAR image classification method based on a complex NSCT neural network and statistical features. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0008] A polarimetric SAR image classification method based on complex NSCT neural network and statistical features, comprising:

[0009] Step 1: Preprocess the polarimetric SAR images to be classified and divide them into training and test sets;

[0010] Step 2: Construct a complex NSCT neural network and set network parameters; the complex NSCT neural network includes a complex spatial domain feature learning module, a complex frequency domain NSCT feature learning module, a statistical feature integration module and a classification module; wherein,

[0011] The complex spatial domain feature learning module is used to extract features from the input image to obtain a spatial domain feature map;

[0012] The complex frequency domain NSCT feature learning module is used to perform non-subsampled Contourlet transform on the input image and perform feature extraction on the obtained NSCT decomposition coefficient map to obtain a high-dimensional representation of the frequency domain feature map;

[0013] The statistical feature integration module is used to analyze the NSCT decomposition coefficient graph to obtain statistical features of the NSCT decomposition coefficients;

[0014] The classification module is used to cascade the spatial domain feature map and the frequency domain feature map, and combine the cascade result with the statistical characteristics of the NSCT decomposition coefficient to perform classification;

[0015] Step 3: inputting the training set into the complex NSCT neural network to train the network;

[0016] Step 4: Classify the test set using the trained complex NSCT neural network to obtain polarimetric SAR image classification results.

[0017] Beneficial effects of the present invention:

[0018] 1. The complex-valued NSCT neural network constructed by the present invention, which has spatial and frequency domain branches, integrates the frequency domain branch based on the multi-resolution non-subsampled Contourlet transform into the complex-valued convolutional neural network in the spatial domain. This combines the advantages of the complex-valued convolutional network in the spatial domain and the NSCT feature learning in the frequency domain, and can fully represent images at multiple scales, multiple directions, and multiple resolutions. It enhances the sparse representation of the directionality, singularity, and regularity of image features, realizes a shallow network structure with a small number of trainable parameters, improves the classification ability and approximation performance of the network, and thus obtains better classification results than a single domain.

[0019] 2. The network structure designed in this paper can fully utilize the phase information of polarimetric SAR images, extending the convolution, activation function, pooling, BP algorithm, and NSCT feature decomposition in the network model to the complex domain. At the same time, the proposed statistical feature integration module can further integrate the statistical feature vectors of the NSCT decomposition coefficients into the network to enhance the distinguishability of feature learning and the classification accuracy of polarimetric SAR images.

[0020] 3. The present invention designs a complex NSCT neural network based on NSCT technology. NSCT approximates singular curves with "long strip" basis functions. With its overcomplete transformation and compact expression ability, it well preserves the edge and texture details of the image, thereby increasing the geometric transformation and feature perception capabilities of the network model and improving the approximation performance of the model.

[0021] 4. The present invention uses NSCT filters to initialize the frequency domain branch of the network. Compared with traditional random initialization, it avoids falling into local optimality when the background of the input image is relatively complex. At the same time, the NSCT kernel increases the interpretability of the network model.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 11 is a flow chart of a polarimetric SAR image classification method based on a complex NSCT neural network and statistical features provided by an embodiment of the present invention;

[0024] Figure 2 1 is a schematic diagram of the structure of a complex NSCT neural network provided by an embodiment of the present invention;

[0025] Figure 3 It is the RGB pseudo-color image and the corresponding labeled image of the Flevoland dataset to be classified in the simulation experiment;

[0026] Figure 4 It is the RGB pseudo-color image and the corresponding labeled image of the dataset Xi'an to be classified in the simulation experiment;

[0027] Figure 5 It is the RGB pseudo-color image and the corresponding labeled image of the dataset Germany to be classified in the simulation experiment;

[0028] Figure 6 This is a diagram showing the classification results of polarimetric SAR images of Flevoland, an agricultural area in the Netherlands, using the present invention and other existing methods;

[0029] Figure 7 This is a diagram showing the classification results of polarimetric SAR images of the Weihe River area in Xi'an using the present invention and other existing methods;

[0030] Figure 8 This is a diagram showing the classification results of the polarimetric SAR image of Oberpfaffenhofen in Germany using the present invention and other existing methods. DETAILED DESCRIPTION

[0031] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0032] Example 1

[0033] See Figure 1 , Figure 1 1 is a flow chart of a polarimetric SAR image classification method based on a complex NSCT neural network and statistical features provided by an embodiment of the present invention, which includes:

[0034] Step 1: Preprocess the polarimetric SAR images to be classified and divide them into training and test sets.

[0035] 11) Obtain the polarimetric SAR image to be classified and the corresponding label matrix, extract the upper triangular elements of the coherence matrix of each pixel to form a 6-dimensional complex-valued vector, and perform Z-score normalization on each vector.

[0036] Specifically, let the coherence matrix T of each pixel i in the input polarimetric SAR image be T. In the upper triangular matrix of T, the diagonal elements are real numbers and the rest are complex numbers. To keep the numerical format of the input network consistent, the diagonal elements T are 11 、T 22 、T 33 Expand to a complex number with an imaginary part of 0. Then normalize the expanded coherence matrix.

[0037] Common normalization methods include minimum-maximum normalization, zero-mean normalization (Z-score normalization), and nonlinear normalization. This example uses, but is not limited to, Z-score normalization, and the processed data conforms to a standard normal distribution.

[0038] 12) Perform a sliding window operation on the normalized polarimetric SAR image to obtain an image block, and use the label corresponding to the target pixel as the label of the corresponding image block.

[0039] In this embodiment, a sliding window operation is performed on the normalized polarimetric SAR image using a window of size 12×12.

[0040] 13) The obtained dataset is randomly shuffled and divided into training set and test set in proportion.

[0041] Specifically, the obtained data set is randomly shuffled, and the first 1% of the data is selected as the training set, and the rest is used as the test set. Assume that the training set is X = {x m |m=1,2,…,M}, M is the number of samples, And the corresponding labels i=1,…,C, where C is the number of sample types.

[0042] Step 2: Construct a complex NSCT neural network and set network parameters; the complex NSCT neural network includes a complex spatial domain feature learning module, a complex frequency domain NSCT feature learning module, a statistical feature integration module, and a classification module.

[0043] See Figure 2 , Figure 2 The diagram is a schematic diagram of the structure of the complex NSCT neural network provided by the embodiment of the present invention. Specifically, the network model has a dual-branch structure for spatial and frequency domain feature learning, in which all operations such as convolution, activation function, pooling, BP algorithm, etc. are performed in the complex domain. The complex NSCT neural network includes a spatial branch and frequency domain branch It is followed by a feature concatenation layer and a fully connected layer [f stack , f c2 , f c ].

[0044] The following is a detailed introduction to the construction process of each module.

[0045] 21) Construct a complex spatial domain feature learning module.

[0046] Specifically, a first convolutional layer and a first pooling layer are set to form a complex spatial feature learning module to extract spatial features of the input image and obtain a spatial feature map.

[0047] Among them, the first convolutional layer (conv1_spa) is used to convolutionally transform the input image. Perform feature extraction to obtain feature maps The first pooling layer (pooling1_spa) is used to pool the feature maps Perform dimensionality reduction to obtain the spatial feature map

[0048] Optionally, for the specific network parameters of the complex spatial domain feature learning module, the convolution kernel size of the first convolution layer can be set to 3×3×6, and the step size can be set to 1; the kernel size of the first pooling layer can be set to 2×2, and the step size can be set to 1.

[0049] 22) Construct a complex frequency domain NSCT feature learning module.

[0050] First, the types of scaling filters and directional filters and the number of decomposed subbands are set to form an NSCT unit, which is used to perform non-subsampled Contourlet transform on the input image and obtain several NSCT decomposition coefficient maps.

[0051] Specifically, the essence of NSCT (Nonsubsampled Contourlet Transform) is the process of iteratively convolving a signal with a nonsubsampled pyramid (NSP) and a nonsubsampled directional filter bank (NSDFB).

[0052] If the input is X, let X l,i and X h,i represents the low-pass component and high-pass component of the i-th level, and X l,0 =X, then the i-th level output of NSCT decomposition can be written as:

[0053]

[0054] Among them, F NSP = "maxflat" and F NSDFB= "dmaxflat7" are non-subsampled pyramid filters and non-subsampled directional filter banks, respectively. The subscripts l and h of X represent the low-pass and high-pass components, respectively. The subscript h_bds represents the bandpass directional subband in the frequency domain.

[0055] In this embodiment, one scale and eight directional sub-bands may be used but are not limited thereto. Then, the input six-channel data is processed by the NSCT unit to obtain 54 NSCT decomposition coefficient maps.

[0056] At the same time, X h_bds,i+1 Adjust to M 2 Size, among which

[0057] M=max(height(X h_bds,i+1 ),width(X h_bds,i+1 )).

[0058] Among them, M represents the width and height of the adjusted feature map.

[0059] Then, the second convolutional layer and the second pooling layer are sequentially set after the NSCT unit to perform feature extraction and dimensionality reduction on the NSCT decomposition coefficient map to obtain a high-dimensional representation of the frequency domain feature map.

[0060] Specifically, use Represents the NSCT decomposition coefficient map, the second convolutional layer (conv2_spe) is used to Perform feature extraction to obtain feature maps The second pooling layer (pooling2_spe) is used to pool the feature maps Perform dimensionality reduction to obtain the frequency domain feature map

[0061] Optionally, for the specific network parameters of the complex frequency domain NSCT feature learning module, the convolution kernel size of the second convolution layer can be set to 3×3×9, and the step size can be set to 1; the kernel size of the second pooling layer can be set to 2×2, and the step size can be set to 1.

[0062] The NSCT technology used in the present invention approximates singular curves with "long strip" basis functions. With its overcomplete transformation and compact expression capabilities, it well preserves the edge and texture detail information of the image, thereby increasing the geometric transformation and feature perception capabilities of the network model and improving the approximation performance of the model.

[0063] 23) Build a statistical feature integration module.

[0064] like Figure 2 As shown in Figure 2, the process of performing statistics on the 54 NSCT decomposition coefficient graphs obtained by the NSCT unit and obtaining the corresponding statistical feature vectors can be expressed as:

[0065]

[0066] Among them, f sfim Represents the statistical characteristics of the NSCT decomposition coefficient, μ n It represents the statistical characteristics of the nth decomposition coefficient graph, n∈[1,N], N is the total number of NSCT decomposition coefficient graphs. In this embodiment, N is 54, and W and H are the coefficient graphs C, n The width and height of (x,y).

[0067] 24) Build a classification module.

[0068] First, a feature cascade layer is set to cascade the spatial domain feature map and the frequency domain feature map to obtain the cascade feature. The cascade feature can be expressed as f stack .

[0069] Then, a third convolutional layer is set to concatenate the features f stack Perform feature extraction and obtain feature map f c2 .

[0070] Finally, a fully connected layer is set up and connected to the statistical feature integration module to integrate the features f extracted by the third convolutional layer c2 and the statistical characteristics of the NSCT decomposition coefficient f sfim Splice and get feature f c , for f c Classification can be performed to obtain classification results.

[0071] Among them, the convolution kernel size of the third convolutional layer (conv3) is 3×3×6 and the stride is 12.

[0072] At this point, the design of each module of the complex NSCT neural network is completed, and combining them can be obtained Figure 2 The complex NSCT neural network shown.

[0073] The structural parameters of the complex NSCT neural network can be designed according to Table 1 below.

[0074] Table 1 Structural parameters of complex NSCT neural network

[0075]

[0076] Step 3: Input the training set into the complex NSCT neural network to train the network.

[0077] Specifically, the training set constructed in step 1 is input into the complex NSCT neural network, and the model is trained using stochastic gradient descent until convergence, resulting in a trained network. The frequency domain branch of the network can be directly initialized using NSCT filters. Compared to traditional random initialization, this avoids falling into local optimality when the input image background is complex. The NSCT kernel also increases the interpretability of the network model.

[0078] Step 4: Use the trained complex NSCT neural network to classify the test set to obtain the polarimetric SAR image classification results.

[0079] Specifically, the network model trained in step 3 is used to classify the test set constructed in step 1 to obtain pixel-by-pixel classification results. The resulting pixel-by-pixel classification result vector is then rearranged using the reshape function in Matlab. While maintaining the total number of elements in the result matrix, it is rearranged into a two-dimensional matrix of the same size as the input remote sensing image, thereby obtaining the final polarimetric SAR image classification results.

[0080] The complex-valued NSCT neural network with spatial and frequency domain branches constructed by the present invention integrates the frequency domain branch based on the multi-resolution non-subsampled Contourlet transform into the complex-valued convolutional neural network in the spatial domain, combining the advantages of the complex-valued convolutional network in the spatial domain and the NSCT feature learning in the frequency domain, and can fully perform multi-scale, multi-directional, and multi-resolution representation of images, enhance the sparse representation of the directionality, singularity, and regularity of image features, realize a shallow network structure with a small number of trainable parameters, improve the classification ability and approximation performance of the network, and thus obtain better classification results than a single domain.

[0081] In addition, the network structure designed in the present invention can make full use of the phase information of polarimetric SAR images and extend the convolution, activation function, pooling, BP algorithm and eigendecomposition of NSCT in the network model to the complex domain. At the same time, the proposed statistical feature integration module can further integrate the statistical eigenvectors of the NSCT decomposition coefficients into the network to enhance the distinguishability of feature learning and the classification accuracy of polarimetric SAR images.

[0082] Example 2

[0083] The beneficial effects of the present invention are verified and explained through simulation experiments below.

[0084] 1. Simulation conditions

[0085] The simulation test platform was run on an HP-Z840 high-performance graphics workstation running Ubuntu 16.04 LTS. It was equipped with two NVIDIA GeForce GTX 1080 graphics cards and an Intel Xeon E5 processor with 64GB of video memory. The computer software was Matlab R2021b.

[0086] The data used in this simulation are three real polarimetric SAR scene datasets with different view counts, bands, and imaging systems. The RGB pseudo-color images after Pauli decomposition and the corresponding annotated images of the Flevoland dataset in the Netherlands are shown in the figure. Figure 3 RGB pseudo-color images after Pauli decomposition of Xi'an dataset and corresponding labeled images, see Figure 4 RGB pseudo-color images after Pauli decomposition of the Germany dataset and the corresponding labeled images, see Figure 5 .

[0087] 2. Simulation content and result analysis

[0088] Simulation 1: Simulation experiment on the classification of the Flevoland dataset using the present invention and the existing methods SVM, SAE, DBN, and CV-CNN. The results are shown in Figure 6 As shown. Among them, Figure 6 Figure (a) shows the classification result of the Flevoland dataset using the existing SVM method; Figure (b) shows the classification result of the Flevoland dataset using the existing SAE method; Figure (c) shows the classification result of the Flevoland dataset using the existing DBN method; Figure (d) shows the classification result of the Flevoland dataset using the existing CV-CNN method; Figure (e) shows the classification result of the Flevoland dataset using the method of the present invention.

[0089] contrast Figure 6 From Figure (a) to Figure (d), it can be seen that the compared methods cannot distinguish wheat, wheat 2 and wheat 3 well. Among them, SAE misclassifies most pixels into rapeseed and sugar beet. DBN misclassifies a large number of pixels into wheat 3. In addition, SVM and CV-CNN perform poorly in most land cover types. This may be due to insufficient training samples (1%), which affects the classification performance of the compared methods. The model of the present invention achieves good classification results and has good local consistency. As shown Figure 6Our method, shown in the white ovals and rectangular areas labeled 1-5, reduces the number of isolated pixels and achieves better visual continuity. This is due to two factors: (1) The complex non-subsampled contourlet neural network designed in this paper has multi-scale and multi-directional representation capabilities, which can capture detailed information about ground object coverage. (2) The statistical feature integration module proposed in this paper increases the separability between classes, thereby improving the classification performance of the model.

[0090] Simulation 2: Simulation experiment of object classification on Xi'an dataset using the present invention and existing methods SVM, SAE, DBN and CV-CNN. The results are shown in Figure 7 .in, Figure 7 Figure (a) shows the classification result of the dataset Xi'an using the existing SVM method; Figure (b) shows the classification result of the dataset Xi'an using the existing SAE method; Figure (c) shows the classification result of the dataset Xi'an using the existing DBN method; Figure (d) shows the classification result of the dataset Xi'an using the existing CV-CNN method; Figure (e) shows the classification result of the dataset Xi'an using the method of the present invention.

[0091] from Figure 7 It can be seen that the existing SVM and SAE have a very low recognition rate for water and mistakenly classify water areas as grass, such as Figure 7 The white circle marked with the number 2 in (a) and (b) shows this. DBN performs poorly in classifying urban and grassland areas, misclassifying grass as water and urban areas as grass. Compared to other methods, the proposed model achieves better classification results (highlighted areas are marked with white ovals), with clearer boundaries and better regional consistency.

[0092] Simulation 3: Simulation experiment of object classification on the dataset Germany using the present invention and existing methods SVM, SAE, DBN, CV-CNN. The results are shown in Figure 8 .in, Figure 8 Figure (a) shows the classification result of the dataset Germany using the existing SVM method; Figure (b) shows the classification result of the dataset Germany using the existing SAE method; Figure (c) shows the classification result of the dataset Germany using the existing DBN method; Figure (d) shows the classification result of the dataset Germany using the existing CV-CNN method; Figure (e) shows the classification result of the dataset Germany using the method of the present invention.

[0093] and Figure 8 Compared with (a)-(c), Figure 8 The number of isolated points in (d)-(e) is significantly reduced. This shows that the proposed model performs better than CV-CNN in forest and building areas. Figure 8 Shown as a medium blue rectangle.

[0094] 3. Evaluation of land feature classification results

[0095] In the classification simulation experiment evaluation of polarimetric SAR images, the overall accuracy OA and Kappa coefficient are often used for evaluation. The calculation formula of OA is as follows:

[0096]

[0097] Among them, TP(k) and TN(k) represent the number of true positives and true negatives of the kth class, respectively, and FN(k) and FP(k) represent the number of false negatives and false positives.

[0098] The Kappa coefficient is calculated based on the confusion matrix. It measures the consistency between the model prediction results and the actual labels and can be defined as:

[0099]

[0100] Where N is the total number of test samples, x k: and x :k denote the sum of the k-th row and k-th column of the confusion matrix respectively.

[0101] The accuracy of each class, overall accuracy OA and Kappa coefficient of the above simulation on the Flevoland dataset are calculated and compared with the classic SVM, SAE, DBN and CV-CNN. The results are shown in Table 2.

[0102] Table 2 Results of the present invention and other comparative methods on the Flevoland dataset (%)

[0103]

[0104]

[0105] As shown in Table 2, the proposed method outperforms the other four compared methods, achieving OA and Kappa coefficients that are 29.4% and 0.3165 higher than those of CV-CNN. Specifically, the proposed method not only achieves the highest classification accuracy in large categories such as forest, wheat, and potato, but also performs well in small categories such as stem beans, bare soil, and barley, demonstrating the effectiveness of phase information in the complex domain and spectral information in multiscale analysis on classification results.

[0106] The accuracy of each category, overall accuracy OA and Kappa coefficient of the above simulation on the Xi'an dataset are calculated and compared with the classic SVM, SAE, DBN and CV-CNN. The results are shown in Table 3.

[0107] Table 3 Results of the present invention and other comparative methods on the Xi'an dataset (%)

[0108] category Support Vector Machine SAE DBN CV-CNN The present invention Grass 91.36 78.39 13.03 82.58 78.29 City 52.73 67.47 2.54 86.65 90.77 waters 17.78 30.23 99.00 65.95 90.03 OA 66.77 67.20 22.76 80.61 88.53 Kappa 39.43 42.19 0.87 68.48 82.80

[0109] As shown in Table 3, the proposed method outperforms other methods in terms of OA and Kappa coefficient evaluation criteria. In terms of overall accuracy, the proposed method achieves improvements of 21.76% (SVM), 21.33% (SAE), 65.77% (DBN), and 7.92% (CV-CNN), respectively. The improvements in Kappa coefficient are 0.4337 (SVM), 0.4061 (SAE), 0.8193 (DBN), and 0.1432 (CV-CNN), respectively. Although the sampling rate of our training set is only 1%, the significant performance on the Xi'an dataset validates the effectiveness of the proposed model.

[0110] The accuracy of each category, overall accuracy OA and Kappa coefficient of the above simulation on the Germany dataset are calculated and compared with the classic SVM, SAE, DBN and CV-CNN. The results are shown in Table 4.

[0111] Table 4 Results of the present invention and other comparative methods on the Germany dataset (%)

[0112] category Support Vector Machine SAE DBN CV-CNN The present invention building 10.70 18.09 64.86 61.70 68.37 woodland 89.32 62.72 36.31 77.25 82.58 Open area 24.53 22.96 3.63 95.23 93.72 OA 33.25 29.22 25.09 83.56 85.36 Kappa 0.59 -10.53 -0.18 75.33 78.03

[0113] As can be seen from Table 4, the present invention is superior to other comparison methods. Among them, OA is 52.11%, 56.14%, 60.27% and 1.8% higher than SVM, SAE, DBN and CV-CNN respectively. For the Kappa coefficient, the present invention is 0.7744 (SVM), 0.8856 (SAE), 0.7821 (DBN) and 0.027 (CV-CNN) higher than other methods respectively. Especially in the classification of buildings, the OA of the present invention is improved by 6.67% on the basis of CV-CNN. Because the high-frequency component of the contour wave has certain advantages in describing edge and texture information, and the building area in the Pauli decomposition diagram of the Germany dataset has snowflake-like texture and rich details, the added frequency domain branch improves the representation and classification ability of the model.

[0114] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A polarimetric SAR image classification method based on complex NSCT neural network and statistical features, characterized by: include: Step 1: Preprocess the polarimetric SAR images to be classified and divide them into training and test sets; Step 2: Construct a complex NSCT neural network and set network parameters; the complex NSCT neural network includes a complex spatial domain feature learning module, a complex frequency domain NSCT feature learning module, a statistical feature integration module and a classification module; wherein, The complex spatial domain feature learning module is used to extract features from the input image to obtain a spatial domain feature map; The complex frequency domain NSCT feature learning module is used to perform non-subsampled Contourlet transform on the input image and perform feature extraction on the obtained NSCT decomposition coefficient map to obtain a high-dimensional representation of the frequency domain feature map; The statistical feature integration module is used to analyze the NSCT decomposition coefficient graph to obtain statistical features of the NSCT decomposition coefficients; The classification module is used to cascade the spatial domain feature map and the frequency domain feature map, and combine the cascade result with the statistical characteristics of the NSCT decomposition coefficient to perform classification; Step 3: inputting the training set into the complex NSCT neural network to train the network; Step 4: Classify the test set using the trained complex NSCT neural network to obtain polarimetric SAR image classification results.

2. The polarimetric SAR image classification method based on complex NSCT neural network and statistical features according to claim 1, characterized in that: Step 1 includes: 11) Obtain the polarimetric SAR image to be classified and the corresponding label matrix, extract the upper triangular elements of the coherence matrix of each pixel to form a 6-dimensional complex-valued vector, and perform Z-score normalization on each vector; 12) Perform a sliding window operation on the normalized polarimetric SAR image to obtain an image block, and use the label corresponding to the target pixel as the label of the corresponding image block; 13) The obtained dataset is randomly shuffled and divided into training set and test set in proportion.

3. The polarimetric SAR image classification method based on complex NSCT neural network and statistical features according to claim 1, characterized in that: In step 2, constructing a complex spatial domain feature learning module includes: A first convolution layer and a first pooling layer are set to extract features of the input image and obtain a spatial feature map.

4. The polarimetric SAR image classification method based on complex NSCT neural network and statistical features according to claim 3, characterized in that: In the complex spatial domain feature learning module, the convolution kernel size of the first convolution layer is 3×3×6, and the step size is 1; the kernel size of the first pooling layer is 2×2, and the step size is 1.

5. The polarimetric SAR image classification method based on complex NSCT neural network and statistical features according to claim 1, characterized in that: In step 2, constructing a complex frequency domain NSCT feature learning module includes: Set the type of scale filter and directional filter and the number of decomposed subbands to form an NSCT unit, which is used to perform non-subsampled Contourlet transform on the input image and obtain several NSCT decomposition coefficient maps; A second convolution layer and a second pooling layer are sequentially arranged after the NSCT unit to perform feature extraction and dimensionality reduction processing on the NSCT decomposition coefficient map to obtain a high-dimensional representation of the frequency domain feature map.

6. The polarimetric SAR image classification method based on complex NSCT neural network and statistical features according to claim 5, characterized in that: The setting of the types of the scale filter and the directional filter and the number of decomposed subbands includes: Setting F NSP ="maxflat" and F NSDFB ="dmaxflat7" are non-subsampled pyramid filter and non-subsampled directional filter bank respectively; Set the number of decomposition subbands to 1 scale and 8 directional subbands.

7. The polarimetric SAR image classification method based on complex NSCT neural network and statistical features according to claim 5, characterized in that: In the complex frequency domain NSCT feature learning module, the convolution kernel size of the second convolution layer is 3×3×9, and the step size is 1; the kernel size of the second pooling layer is 2×2, and the step size is 1.

8. The polarimetric SAR image classification method based on complex NSCT neural network and statistical features according to claim 1, characterized in that: In step 2, the statistical feature integration module analyzes the NSCT decomposition coefficient graph to obtain the statistical features of the NSCT decomposition coefficients. The process is expressed as follows: f sfim =(μ1,μ2,…,μ N ) Among them, f sfim Represents the statistical characteristics of the NSCT decomposition coefficient, μ n Represents the statistical characteristics of the nth decomposition coefficient graph, n∈[1,N], N is the total number of NSCT decomposition coefficient graphs, W and H are the coefficient graphs C n The width and height of (x,y).

9. The polarimetric SAR image classification method based on complex NSCT neural network and statistical features according to claim 1, characterized in that: In step 2, building the classification module includes: Set a feature cascade layer to cascade the spatial domain feature map and the frequency domain feature map to obtain cascade features; A third convolutional layer is set to extract features from the cascaded features; A fully connected layer is set and connected to a statistical feature integration module to combine the features extracted by the third convolutional layer with the statistical features of the NSCT decomposition coefficients for classification.

10. The polarimetric SAR image classification method based on complex NSCT neural network and statistical features according to claim 9, characterized in that: The convolution kernel size of the third convolutional layer is 3×3×6, and the stride is 12.

Images