A spectral feature extraction method for dual-channel hyperspectral images

Abstract

The present invention discloses a method for extracting spectral features from dual-path hyperspectral images, belonging to the field of remote sensing image processing. The method comprises: obtaining a hyperspectral data block from the local neighborhood of a pixel to be classified; performing a band delocalization operation on the hyperspectral data block to obtain a processed hyperspectral data block; constructing a dual-branch hyperspectral image spectral feature extraction network to extract spectral features from the hyperspectral data blocks before and after processing, respectively; and fusing the spectral features extracted from the two branches. The present invention enables the simultaneous extraction and utilization of correlated features between closely adjacent bands and between long-range non-adjacent bands in high-dimensional spectral features using a convolution kernel of limited size.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a method for extracting spectral features of dual-channel hyperspectral images. Background Art

[0002] Remote sensing image processing technology, with its advantages of high flexibility and good reproducibility, is playing an increasingly important role in production and daily life. Hyperspectral image processing is a key branch of this field. Among the many hyperspectral image processing technologies, hyperspectral image classification has been a research focus and hotspot. Hyperspectral image classification can be used to determine the distribution and growth of crops, enabling effective and scientific agricultural management. It can also accurately identify and classify urban buildings and road surfaces, thus assisting in urban planning and other aspects.

[0003] The effective utilization of hyperspectral remote sensing imagery is inseparable from the extraction of its features. Efficiently extracting the spectral features of hyperspectral imagery is currently a key research topic in hyperspectral imagery feature extraction. Currently, the main methods for extracting spectral features include 1D convolution, LSTM, and RNN. However, unlike traditional image data, hyperspectral imagery has a higher spectral dimension (hundreds of dimensions), making it difficult to globally consider this high-dimensional data using convolution kernels of limited size. Simply put, existing methods only extract correlation features between locally adjacent bands while neglecting the extraction of features between non-adjacent, long-range bands. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of existing spectral feature extraction technology in hyperspectral images, a dual-path hyperspectral image spectral feature extraction method is provided, which can effectively solve the shortcomings of insufficient spectral feature extraction in existing hyperspectral images and thus improve the performance of subsequent tasks.

[0005] Technical solution: To achieve the above-mentioned purpose, the present invention provides a method for extracting spectral features from dual-channel hyperspectral images, comprising the following steps:

[0006] S1: Obtain the hyperspectral data block of the local neighborhood of the pixel to be classified;

[0007] S2: performing a band non-localization operation on the hyperspectral data block to obtain a processed hyperspectral data block;

[0008] S3: Construct a dual-branch hyperspectral image spectral feature extraction network to extract spectral features of hyperspectral data blocks before and after processing respectively;

[0009] S4: Splice and fuse the spectral features extracted from the two branches.

[0010] Furthermore, in step S1, the method for obtaining the hyperspectral data block of the local neighborhood of the pixel to be classified is:

[0011] A1: Take each pixel to be classified as the center of the original hyperspectral data and obtain the cube neighborhood block set A = {α1, α2, ..., α n},in Where W×W represents the size of the neighborhood block space, and B represents the number of bands.

[0012] 7. Furthermore, the method of performing a band non-localization operation on the hyperspectral data block in step S2 to obtain a processed hyperspectral data block includes:

[0013] B1: The obtained cubic neighborhood block α i The jth band in and the B-j+1th band Splicing along the spectral dimension.

[0014] Furthermore, the processed hyperspectral data block obtained in step S2 is described by the formula:

[0015]

[0016] Among them, [] represents the splicing operation, α′ i Represents a neighborhood block of hyperspectral data after delocalization of spectral bands.

[0017] Furthermore, in step S3, a dual-path hyperspectral image spectral feature extraction network is constructed to extract spectral features from the hyperspectral data blocks before and after processing, respectively. The specific construction and feature extraction methods are as follows:

[0018] C1: Use a 3D convolution kernel with a structure of 1×1×n to construct a convolutional neural network to extract spectral features;

[0019] C2: The two branches use the same network structure to extract characteristic spectral features from the neighborhood blocks before processing and after band delocalization respectively;

[0020] C3: The two branches obtain the short-range inter-band relationship feature F1 and the long-range inter-band relationship feature F2 respectively.

[0021] Furthermore, in step S4, the spectral features extracted from the two branches are spliced ​​and fused. The specific fusion method is expressed by the formula:

[0022] F=[F1,F2]

[0023] Where F represents the fused features. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0025] Figure 1 A schematic diagram of a method flow chart provided for an example of the present invention;

[0026] Figure 2 Schematic diagram of the spectral feature extraction network structure provided by the example of the present invention;

[0027] Figure 3 This is the Pavia University hyperspectral remote sensing image data map used in the simulation experiment of the present invention;

[0028] Figure 4 This is a pseudo-color labeled image of the hyperspectral remote sensing image data of Pavia University in the simulation experiment of the present invention;

[0029] Figure 5 This is a classification result diagram of the hyperspectral remote sensing image data of Pavia University in the simulation experiment of the present invention without using the present invention;

[0030] Figure 6 This is a diagram showing the classification results of the hyperspectral remote sensing image data of Pavia University used in the simulation experiment of the present invention, which is based on the extraction of spectral inter-band relationship features between short-range bands and long-range bands; DETAILED DESCRIPTION

[0031] The present invention is further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0032] like Figure 1 As shown, the present invention provides a method for extracting spectral features of a dual-path hyperspectral image, comprising the following steps:

[0033] S1: Obtain the hyperspectral data block of the local neighborhood of the pixel to be classified;

[0034] S2: performing a band non-localization operation on the hyperspectral data block to obtain a processed hyperspectral data block;

[0035] S3: Construct a dual-branch hyperspectral image spectral feature extraction network to extract spectral features of hyperspectral data blocks before and after processing respectively;

[0036] S4: Splice and fuse the spectral features extracted from the two branches.

[0037] In this embodiment, the method for obtaining the hyperspectral data block of the local neighborhood of the pixel to be classified in step S1 is:

[0038] A1: Take each pixel to be classified as the center of the original hyperspectral data and obtain the cube neighborhood block set A = {α1, α2, ..., α n},in Where W×W represents the size of the neighborhood block space, and B represents the number of bands.

[0039] In this embodiment, the processed hyperspectral data block obtained in step S2 is described by the formula:

[0040]

[0041] Among them, [] represents the splicing operation, α′ i Represents a neighborhood block of hyperspectral data after delocalization of spectral bands.

[0042] In this embodiment, in step S3, a dual-path hyperspectral image spectral feature extraction network is constructed to extract spectral features from the hyperspectral data blocks before and after processing. The specific construction and feature extraction methods are as follows:

[0043] C1: Use a 3D convolution kernel with a structure of 1×1×n to construct a convolutional neural network to extract spectral features;

[0044] C2: The two branches use the same network structure to extract characteristic spectral features from the neighborhood blocks before processing and after band delocalization respectively;

[0045] C3: The two branches obtain the short-range inter-band relationship feature F1 and the long-range inter-band relationship feature F2 respectively.

[0046] In step S4 of this embodiment, the spectral features extracted from the two branches are spliced ​​and fused. The specific fusion method is expressed by the formula:

[0047] F=[F1,F2]

[0048] Where F represents the fused features.

[0049] Based on the above scheme, in order to verify the effect of the method of the present invention, a two-way spectral feature extraction network was constructed to verify the performance of feature extraction using a hyperspectral image classification task. 1% of the samples in the Pavia University hyperspectral image data set were used as the training set, and 99% of the samples were used as the test set for a simulation experiment. No training strategy was adopted. The TensorFlow deep learning framework was used, and the training parameter settings were shown in Table 1. The model parameter settings are shown in Table 2.

[0050] Table 1 Training parameter settings

[0051] Input Size 3×3×103 Decay Rate 0.01 Epochs 200 Batch Size 64 Learning Rate 0.01

[0052] Table 2 Model parameter settings

[0053]

[0054] Implementation Results

[0055] In this embodiment, the overall classification accuracy (OA), average classification accuracy (AA), and Kappa coefficient are used. Table 3 shows the classification accuracy comparison between the spectral feature extraction model without the spectral band delocalization operation proposed in the present invention and the spectral feature extraction model with the spectral band delocalization operation proposed in the present invention. Figure 3-4 They are the Pavia University hyperspectral remote sensing image data used in the experiment and the corresponding pseudo-color labeled image. Figure 3 Applying the above method, we get Figure 5 and Figure 6 The diagram shows the classification results of the spectral feature extraction model obtained by Pavia University without using the spectral band non-localization operation proposed by the present invention and using the spectral feature extraction model with the spectral band non-localization operation proposed by the present invention.

[0056] Table 3 Classification accuracy of each category and overall classification evaluation

[0057]

[0058] According to the data in Table 3, the classification effect of the present invention is better than that of the spectral feature extraction model that does not use the spectral band non-localization operation proposed by the present invention. When subdivided into OA, the average classification AA and Kappa coefficient are also higher than those of the spectral feature extraction model that does not use the spectral band non-localization operation proposed by the present invention. Figure 5 and Figure 6 Compared with the classification truth map, Figure 6The classification performance is closer to the true value, thereby verifying the effectiveness of the method of the present invention.

Claims

1. A method for extracting spectral features from dual-channel hyperspectral images, characterized in that: The steps include: S1: Obtain the hyperspectral data blocks of the local neighborhood of the pixel to be classified. Specifically, the original hyperspectral data is centered on each pixel to be classified, and a cube neighborhood block set A = {α1, α2, ..., α n },in Where W×W represents the size of the neighborhood block space, and B represents the number of bands; S2: performing a band non-localization operation on the hyperspectral data block to obtain a processed hyperspectral data block, specifically comprising: performing a band non-localization operation on the hyperspectral data block to obtain a processed hyperspectral data block. i The jth band in and the B-j+1th band Splice along the spectral dimension to obtain the processed hyperspectral data block Among them, [] represents the splicing operation, α i Represents the neighborhood block of hyperspectral data after delocalization of spectral bands; S3: Construct a dual-branch hyperspectral image spectral feature extraction network to extract spectral features of hyperspectral data blocks before and after processing respectively; S4: Splice and fuse the spectral features extracted from the two branches.

2. A dual-path hyperspectral image spectral feature extraction method according to claim 1, characterized in that: In step S3, a dual-path hyperspectral image spectral feature extraction network is constructed to extract spectral features from the hyperspectral data blocks before and after processing. The specific construction and feature extraction methods are as follows: C1: Use a 3D convolution kernel with a structure of 1×1×n to construct a convolutional neural network to extract spectral features; C2: The two branches use the same network structure to extract characteristic spectral features from the neighborhood blocks before processing and after spectral band delocalization respectively; C3: The two branches obtain the short-range inter-band relationship feature F1 and the long-range inter-band relationship feature F2 respectively.

3. The method for extracting spectral features from dual-path hyperspectral images according to claim 2, wherein: Based on the two branch spectral features obtained in step C3, they are spliced ​​and fused. The specific fusion method is expressed by the formula: F=[F1,F2] Where F represents the fused features.

Images