Multi-view hyperspectral image processing method capable of being used for ocean target analysis

Abstract

The invention discloses a multi-view hyperspectral image processing method capable of being used for ocean target analysis, and belongs to the technical field of hyperspectral image processing. According to the method, the spectral features and the spatial features of the hyperspectral image are extracted at the same time, and multi-dimensional information of data is effectively utilized. According to the method, AAP and SD algorithms are introduced, and dimensionality reduction is carried out on high-dimensional spectral data in a self-adaptive manner according to the characteristics of spectral bands. According to the method, an end-to-end deep learning framework is adopted, and spectral feature extraction, spatial feature fusion, unmixing and data reconstruction are integrated in the same network. The method can perform efficient unmixing and accurate classification on complex spectrum and spatial information in the multi-view hyperspectral image, is especially suitable for the fields of marine environment monitoring, resource exploration, environmental protection and the like, and has important application value in the aspects of spectrum separation and classification of complex oceans.

Description

Technical Field

[0001] The invention belongs to the technical field of hyperspectral image processing, and in particular relates to a multi-view hyperspectral image processing method that can be used for ocean target analysis. Background Art

[0002] With the rapid development of satellite remote sensing technology, hyperspectral imaging technology has been widely used in recent years, and the amount of remote sensing data has increased exponentially. Hyperspectral imaging (HSI) can capture continuous spectral information from visible light to near infrared. Due to its high spectral resolution and fine material identification capabilities, it has become a research hotspot in the fields of environmental monitoring, agricultural exploration, and ocean observation. For example, Hyperion launched by the United States and my country's hyperspectral observation satellites can provide high-precision remote sensing data covering hundreds of spectral bands. These hyperspectral data can not only provide rich spectral information, but also effectively identify the spectral characteristics of different objects.

[0003] With the rapid development of remote sensing technology in my country, the application of hyperspectral remote sensing is becoming more and more extensive. The acquired hyperspectral data is becoming more and more complex. Accurate unmixing and classification of these data has become an important direction of current research. Unmixing technology aims to separate the spectral components of different objects in mixed pixels, which is of great significance for improving the analysis accuracy of hyperspectral images. However, hyperspectral images are usually affected by noise, spatial heterogeneity and spectral redundancy. Traditional unmixing methods often perform poorly when processing these complex scenes, and it is difficult to fully extract useful information in spatial and spectral dimensions.

[0004] In recent years, research on unmixing and classification of hyperspectral images has emerged continuously, which can be mainly divided into traditional spectral-based methods and modern methods that combine spatial information. (1) Traditional spectral unmixing methods usually rely only on spectral information and perform linear or nonlinear decomposition of spectral features in hyperspectral data, but it is difficult to effectively use spatial information, resulting in limitations in the analysis of mixed pixels. (2) Modern methods combine spectral and spatial information, using deep learning autoencoder technology to extract features and perform dimensionality reduction through unsupervised learning. Autoencoders can effectively capture important information in hyperspectral data, reduce redundancy, and improve the accuracy of unmixing and robustness to noise. However, these methods also face challenges. For example, when the data dimension is extremely high, the model may have problems with low training efficiency. At the same time, the architecture design and hyperparameter adjustment of the autoencoder have a significant impact on the results, increasing the difficulty of model optimization.

[0005] In general, the current hyperspectral image unmixing and classification methods are easily affected by noise and spectral mixing when facing complex environments, and it is difficult to obtain stable and high-precision unmixing results. Therefore, developing more robust unmixing methods that can effectively utilize multi-view and multi-dimensional information has become an important direction of current hyperspectral remote sensing research. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-view hyperspectral image processing method that can be used for marine target analysis, so as to solve the technical problems faced by existing hyperspectral image processing methods, such as complex spectral mixing, insufficient unmixing accuracy and noise interference.

[0007] The present invention effectively improves the unmixing accuracy and classification ability of hyperspectral images by combining multi-view data construction with spatial spectral network. The specific technical solutions are as follows:

[0008] A multi-view hyperspectral image processing method for ocean target analysis includes the following steps:

[0009] S1: Obtain hyperspectral data HSI containing L spectral bands and n mixed pixels, expressed as HSI(L,n);

[0010] S2: Preliminary division of L spectral bands in the hyperspectral data to form a multi-view dataset;

[0011] S3: Extract features from space and spectrum simultaneously to obtain their feature matrices respectively;

[0012] S4: Fusing the features of the spectral dimension and the spatial dimension to generate a fused feature vector; restoring the input pixel vector through a decoder for the fused feature vector input;

[0013] S5: Optimize the loss function and perform network training, use the trained network model to process data, and obtain the final end-member output.

[0014] Furthermore, in S1, the hyperspectral data (HSI) is obtained by hyperspectral imaging technology and is a multidimensional data, denoted as HSI (L, n), where L represents the number of spectral bands and n represents the number of pixels; each pixel not only has spatial information, but also contains a complete spectral curve.

[0015] Furthermore, in S2, the L spectral bands in the hyperspectral data are preliminarily divided by the AAP or SD algorithm, including:

[0016] S2-1: First, perform hyperspectral data preprocessing: noise removal and normalization. After preprocessing, the processed spectral data matrix HSI′(L,n) is obtained, where HSI′ is the hyperspectral data that has been standardized and had noise removed.

[0017] S2-2: Select the AAP or SD algorithm for band division: The original L bands are divided into k band clusters, generating a new data matrix HSI(k,n), where k < L, and each band cluster represents a group of bands with similar spectral characteristics.

[0018] Furthermore, in S2-1: Noise removal: Through noise analysis, use statistical methods (such as standard deviation analysis) to remove those bands affected by atmospheric interference or with high noise; Normalization: Standardize the bands to ensure that the values of each band are at the same magnitude for more stable subsequent processing.

[0019] Furthermore, in S2-2:

[0020] AAP (Automatic Adaptive Partitioning) algorithm: The AAP algorithm is an adaptive partitioning method that dynamically partitions bands based on the local changes in the spectral curve. Assume that the change amount between bands can be represented by the gradient. Let G i =|HSI′ {i+1} –HSI′ i |. When the gradient G i exceeds the preset threshold, it is regarded as the starting point of a new band cluster.

[0021] SD (Spectral Division) algorithm: The SD algorithm divides bands through spectral clustering or similarity measurement. Common measurement methods include cosine similarity or Euclidean distance. The specific calculation formula is:

[0022]

[0023] Then, use a clustering algorithm (such as K-means or hierarchical clustering) to cluster the bands and assign bands with higher similarity to the same cluster.

[0024] Furthermore, S3 specifically includes:

[0025] S3-1: Spectral dimension feature extraction

[0026] In the spectral dimension, use a fully connected network to extract features from the input of HSI(k,n). Through dimensionality reduction technology, reduce the spectral data from k dimensions to the dimension c of the target matrix, that is, the number of pure spectral endmembers; then use the ReLU activation function to enable the network to have non-linear expression ability.

[0027] S3-2: Spatial dimension feature extraction

[0028] For each pixel point, a spatial window of size W×W is selected, and the spatial information of the surrounding pixels is extracted through the weighted averaging method to generate a spatial feature vector representing the pixel. Then, the average features of the spatial window are reduced in dimension using a fully connected network, batch normalization operation and activation function. Finally, the ReLU activation function is used to give the network nonlinear expression capabilities.

[0029] Furthermore, in S3-1: Assume that the input spectrum vector is Through the fully connected network, batch normalization operation and activation function, the feature extraction formula is:

[0030]

[0031] in, is the weight matrix of spatial features, is the bias vector, and the output is the feature representation after dimensionality reduction The scale and offset parameters γ, β are the parameters of batch normalization, respectively.

[0032] Furthermore, in S3-2, the formula for extracting the spatial feature vector is:

[0033]

[0034] Where W is the spatial window size, Represents x i and x w The spatial distance between

[0035] Then, the average feature of the spatial window is reduced in dimension using a fully connected network, batch normalization operation, and activation function. The formula is:

[0036]

[0037] in, is the weight matrix of spatial features, is the bias vector, and the scaling and offset parameters γ, β are the parameters of batch normalization, respectively.

[0038] Furthermore, in S4, the feature fusion strategy may use methods such as concatenation and element-by-element addition. Assuming that the spectral feature is f spectral (x i ), the spatial feature is f spatial (x i ), the fusion features of space and spectrum are obtained by using the product method, and the residual structure is introduced. The final f fusion (x i)The formula is:

[0039] f fusion (x i )=f spatial (x i )+f spectral (x i )×f spatial (x i )

[0040] The fused features are decoded Reconstructed into the original spectral dimension L, the formula of the reconstruction process is:

[0041]

[0042] in, is the weight matrix of the decoder, is the bias vector. The output This is the reconstructed spectral data.

[0043] Further, the S5 includes:

[0044] S5-1: Loss Function and Optimization

[0045] In order to measure the difference between the reconstructed data and the original input data, the spectral angle distance (SAD) and abundance Regularization is used as the loss function; SAD measures the similarity between the reconstructed spectrum and the original spectrum by calculating the angle between the two. The formula is:

[0046]

[0047] Among them, x i is the original spectrum vector, is the reconstructed spectrum vector;

[0048] The abundance It is expressed by the formula:

[0049]

[0050] in Considered as the abundance obtained;

[0051] So the complete loss function is considered as:

[0052]

[0053] Where λ1 and λ2 represent the weights of the two loss functions respectively;

[0054] The Adam optimization algorithm is used to optimize the network parameters. The update formula is:

[0055]

[0056] Among them, θ t represents the network parameters, α is the learning rate;

[0057] S5-2: After network training and optimization, the final output is the endmember feature, that is, the pure spectral endmember matrix extracted by the network Represents the spectral characteristics of different substances.

[0058] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0059] 1. Combine spectral and spatial features to improve unmixing accuracy:

[0060] The present invention effectively utilizes the multi-dimensional information of the data by simultaneously extracting the spectral features and spatial features of the hyperspectral image. The spectral features are used to separate the spectral components of the mixed substances, and the spatial features capture the spatial correlation of the pixels. By fusing these two features, the present invention can not only improve the spectral unmixing accuracy of complex substances, but also maintain spatial consistency and reduce the influence of noise and boundary effects during the unmixing process.

[0061] 2. Adaptive band division to reduce data redundancy:

[0062] The present invention introduces AAP (Adaptive Partitioning) and SD (Spectrum Partitioning) algorithms to adaptively reduce the dimension of high-dimensional spectral data according to the characteristics of spectral bands. This process effectively reduces redundant bands, retains key information, and improves the computational efficiency and robustness of the model. In addition, the data after band division can better perform feature extraction during the processing process, enhancing the adaptability of the demixing model to complex scenes.

[0063] 3. End-to-end deep learning framework to improve demixing efficiency and robustness:

[0064] The present invention adopts an end-to-end deep learning framework, integrating spectral feature extraction, spatial feature fusion, unmixing and data reconstruction in the same network. This architecture realizes the simultaneous processing of spectral and spatial information through fully connected layers and convolution operations, and automatically learns the optimal feature representation. By using the Adam optimization algorithm, the model can converge quickly and maintain high unmixing accuracy in complex scenarios. This end-to-end approach not only improves processing efficiency and reduces the complexity of manual parameter adjustment, but also enhances the robustness of the model in different data sets and application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of the present invention.

[0066] Figure 2 Schematic diagram of the process of the hyperspectral processing method of the present invention.

[0067] Figure 3 Schematic diagram of the module of the deep learning model of the present invention.

[0068] Figure 4 Schematic diagram of input processing of the network model of the present invention.

[0069] Figure 5 Schematic diagram of the Urban hyperspectral input data of the present invention.

[0070] Figure 6 It is a schematic diagram of the Jasper hyperspectral input data of the present invention.

[0071] Figure 7 The end member effect diagram of the implemented method on the Urban dataset.

[0072] Figure 8 The end member effect diagram of the implemented method under the Jasper dataset.

[0073] Fig. 9 Schematic diagram of ocean hyperspectral data. DETAILED DESCRIPTION

[0074] The present invention is further explained and illustrated below through specific embodiments in conjunction with the accompanying drawings.

[0075] A multi-view hyperspectral image processing method that can be used for ocean target analysis, such as Figure 1 , Figure 2 and Figure 3 As shown, the following steps are included:

[0076] Step 1: Obtain hyperspectral data and determine the number of spectral segments and hybrid pixels

[0077] Step 1.1. Hyperspectral data (HSI) is obtained through hyperspectral imaging technology. It is a multidimensional data, denoted as HSI (L, n), where L represents the number of spectral bands and n represents the number of pixels. Each pixel not only has spatial information, but also contains a complete spectral curve. Unlike traditional RGB images that only contain three bands of red, green, and blue, hyperspectral images cover a continuous spectrum from visible light to near infrared, usually including hundreds or even thousands of bands. Through this spectral information, the composition and properties of substances can be identified and analyzed. Common hyperspectral datasets include Indian Pines and Pavia University, which can be downloaded through the official links provided by the University of Pavia and Purdue University.

[0078] Step 2. Preliminary division of L spectral bands in the hyperspectral data by AAP or SD algorithm

[0079] Step 2.1. Hyperspectral data preprocessing

[0080] Before dividing the spectral bands, the hyperspectral data must first be preprocessed to ensure the integrity and quality of the data; common preprocessing steps include:

[0081] (1) Noise removal: Through noise analysis, statistical methods (such as standard deviation analysis) are used to remove those bands that are affected by atmospheric interference or have high noise.

[0082] (2) Normalization: Normalize the bands to ensure that the values ​​of each band are at the same level so that the subsequent processing process is more stable.

[0083] Let the hyperspectral data matrix be HSI(L,n), where each column represents the spectral vector of each pixel. After preprocessing, the processed spectral data matrix HSI′(L,n) is obtained, where HSI′ is the standardized and noise-removed hyperspectral data.

[0084] Step 2.2 Band division

[0085] Select AAP or SD algorithm for band division:

[0086] (1) AAP (Automatic Adaptive Partitioning) algorithm: AAP algorithm is an adaptive partitioning method that dynamically divides bands based on the local changes of the spectral curve. Assuming that the change between bands can be expressed by gradient, let G i =|HSI′ {i+1} –HSI′ i |, when the gradient Gi When it exceeds the preset threshold, it is regarded as the starting point of a new band cluster.

[0087] (2) SD (Spectral Division) algorithm: The SD algorithm divides bands through spectral clustering or similarity measurement. Common measurement methods include cosine similarity or Euclidean distance, and the specific calculation formula is:

[0088]

[0089] Cluster the bands through a clustering algorithm (such as K-means or hierarchical clustering), and assign bands with higher similarity to the same cluster.

[0090] Here, the clustering algorithm is an unsupervised learning algorithm that divides L data points into k clusters. The goal is to minimize the sum of squared errors within the clusters, that is, to assign data points to the cluster closest to the centroid of the cluster they belong to.

[0091] After being processed by the AAP or SD algorithm, the original L bands are divided into k band clusters, generating a new data matrix HSI(k,n), where k < L, and each band cluster represents a group of bands with similar spectral characteristics.

[0092] Step 3. Extract features and reconstruct data simultaneously in space and spectrum, as Figure 4 shown:

[0093] Step 3.1. Feature extraction in the spectral dimension

[0094] In the spectral dimension, use a fully connected network to extract features from the input of HSI(k,n) obtained; through dimensionality reduction technology, reduce the spectral data from k dimensions to the dimension c of the target matrix, that is, the number of pure spectral endmembers.

[0095] Assume the input spectral vector is Through the fully connected network, batch normalization operation and activation function, the feature extraction formula is:

[0096]

[0097] Among them, is the weight matrix of the spatial feature, is the bias vector, and the scaling and offset parameters γ,β are the parameters of batch normalization respectively. The output is the feature representation after dimensionality reduction

[0098] Here, batch normalization standardizes the output of the linear transformation to a mean of 0 and a variance of 1.

[0099] Finally, select the ReLU activation function to make the network have non-linear expression ability.

[0100] Step 3.2. Spatial dimension feature extraction

[0101] For each pixel, a spatial window of size W×W is selected, and the spatial information of the surrounding pixels is extracted by weighted averaging to generate a spatial feature vector representing the pixel:

[0102]

[0103] Where W is the spatial window size, Represents x i and x w The spatial distance between.

[0104] Then, the average feature of the spatial window is reduced in dimension using a fully connected network, batch normalization operation, and activation function. The formula is:

[0105]

[0106] in, is the weight matrix of spatial features, is the bias vector, and the scaling and offset parameters γ and β are the parameters of batch normalization respectively;

[0107] The batch normalization here standardizes the output of the linear transformation to have a mean of 0 and a variance of 1.

[0108] Finally, the ReLU activation function is used to give the network nonlinear expression capabilities.

[0109] Step 4. Spatial and spectral feature fusion and reconstruction

[0110] The features of the spectral dimension and the spatial dimension are fused to generate a fused feature vector. Feature fusion strategies usually use methods such as splicing and element-by-element addition. Assume that the spectral feature is f spectral (x i ), the spatial feature is f spatial (x i ), here we use the product method to obtain the fusion features of space and spectrum, and introduce the residual structure, and finally get f fusion (x i )The formula is:

[0111] f fusion (x i ) = f spatial (x i )+f spectral (x i )×f spatial (x i )

[0112] The fused features are decoded The formula for the reconstruction process of the original spectral dimension L is:

[0113]

[0114] in, is the weight matrix of the decoder, is the bias vector, the output This is the reconstructed spectral data.

[0115] Step 5. Deep learning optimization network

[0116] Step 5.1. Loss Function and Optimization

[0117] In order to measure the difference between the reconstructed data and the original input data, the spectral angle distance (SAD) and abundance Regularization is used as a loss function. SAD measures the similarity between the reconstructed spectrum and the original spectrum by calculating the angle between the two. The formula is:

[0118]

[0119] Among them, x i is the original spectrum vector, is the reconstructed spectrum vector.

[0120] The abundance It can be expressed by the formula:

[0121]

[0122] in It can be regarded as the abundance obtained;

[0123] So the complete loss function can be regarded as:

[0124]

[0125] Where λ1 and λ2 represent the weights of the two loss functions respectively.

[0126] The Adam optimization algorithm is used to optimize the network parameters. The update formula is:

[0127]

[0128] Among them, θ t represents the network parameters, and α is the learning rate.

[0129] Step 5.2. Output end member features and abundance graph

[0130] After network training and optimization, the final output is the endmember feature, that is, the pure spectral endmember matrix extracted by the network. Represents the spectral characteristics of different substances.

[0131] Embodiment 2:

[0132] This embodiment further illustrates the effect of the present invention in combination with simulation experiments.

[0133] The simulation experiment of the present invention is carried out in the hardware environment of Intel i5-12400, NVIDIAGTX 3080, memory 16GB and the software environment of Python3.7 and PyTorch 1.0. The simulation experiment data of the present invention are Urban data set and Jasper. The Urban data set is collected by HYDICE sensor and contains 307×307 pixels. After removing the low-quality bands, the remaining 162 channels are used for the experiment. The Jasper data set is collected by AVIRIS sensor and contains 100×100 pixels. After removing the low-quality bands, the remaining 198 channels are used for the experiment.

[0134] Figure 5 Shows a cube image of the Urban dataset, which contains four main materials: asphalt, grass, trees, and roofs. Figure 6 Shows a cube image of the Jasper dataset, which contains four main materials: trees, water, soil, and roads. Figure 7 The end-member features extracted by the present invention under Urban data are shown, which represent asphalt, grass, trees and roofs from left to right and from top to bottom. Figure 7 The red line in the middle represents the true value, while the blue line represents the end member spectral characteristic curve obtained by the present invention. It can be seen that the spectral characteristic curve obtained by the present invention is very close to the true value. Figure 8 The end member characteristics obtained by the present invention under the Jasper data are shown, so that the red line also represents the true value, and the blue line represents the end member curve obtained by the present invention. It can be seen from the results under the two data sets that the end member characteristic curve we obtained is very close to the true value, especially under the Jasper data set, the degree of fit is more obvious.

[0135] The comparison results of the present invention and the existing more advanced change detection methods are shown in Tables 1 and 2. Table 1 is the comparison result under the Urban data set, and Table 2 is the comparison result under the Jasper data set. The convolutional neural network autoencoder unmixing (hereinafter abbreviated as CNNAEU) method in the comparative experiment is proposed in the article "Convolutional Autoencoder for Spectral-Spatial Hyperspectral Unmixing"; the EndNet method is proposed in the article "EndNet: Sparse Autoencoder Network for Endmember Extraction and Hyperspectral Unmixing"; the TANet method is proposed in the article "TANet: An Unsupervised Two-Stream Autoencoder Network for Hyperspectral Unmixing"; the SSAE method is proposed in the article "Spatial-Spectral Autoencoder Networks for Hyperspectral Unmixing".

[0136] The present invention uses spectral angle distance (SAD) to compare with the above method in terms of objective indicators, and the calculation method is as follows:

[0137]

[0138] in,<x,y> is the inner product of vectors x and y, ||x||,||y|| represent the Euclidean norm of x and y respectively. n is the number of samples, is the predicted value of the i-th sample, y i is the true value of the i-th sample. Here, SAD is a metric that measures the angular difference between two spectral vectors, and is particularly suitable for hyperspectral data analysis. SAD measures the similarity of two spectra by calculating the angle between the spectral vectors, rather than directly comparing their size or amplitude. This means that SAD is insensitive to scale changes in the spectrum (such as brightness differences due to changes in lighting conditions), and it is more concerned with the shape characteristics of the spectrum. Therefore, SAD is a reliable metric for evaluating spectral similarity during the unmixing process.

[0139] Table 1 SAD of Urban dataset (×10 -2 ) Method Experimental Results

[0140] method asphalt grassland Trees roof average CNNAEU 5.75 3.66 3.21 3.32 3.98 EndNet 5.98 5.34 4.57 3.89 4.95 TANet 3.86 3.66 7.49 2.51 4.38 SSAE 6.02 3.59 3.18 2.76 3.89 The present invention 4.38 3.31 3.32 2.58 3.39

[0141] Table 2 SAD of Jasper dataset (×10 -2 ) Method Experimental Results

[0142]

[0143]

[0144] The present invention provides a multi-view hyperspectral image processing method that can be used for marine target analysis, which is mainly used to analyze and monitor the distribution and dynamic changes of surface materials, and classify and express the abundance of materials in remote sensing images. However, the present invention is not limited to the above-mentioned application scenarios, but is also applicable to spectral analysis of other complex surface environments, such as marine pollution assessment, coastal ecosystem monitoring, etc., and can be efficiently processed at the spatial spectrum scale, with a wide range of application scenarios and good practical effects. Hyperspectral unmixing technology is a crucial technology in marine remote sensing, which can perform fine monitoring and analysis of complex marine environments. The marine environment is highly complex, and multiple components such as seawater, phytoplankton, suspended matter, benthic organisms, and pollutants often coexist and mix together. Traditional remote sensing methods are difficult to accurately distinguish and quantify these components. Hyperspectral imaging technology can capture subtle spectral differences by acquiring continuous spectral information, which helps to deeply analyze various components in the ocean. The emergence of hyperspectral unmixing technology makes it possible to separate pure spectral information from mixed pixels, providing important support for monitoring and research of the marine environment. The application of this invention to the marine field can have a positive impact on marine observations. In terms of marine data sets, Fig. 9 (Right) The dataset shows a cropped RGB rendering of a hyperspectral image of the Trondheim Fjord taken with the HYPSO-1 satellite.

[0145] The hyperspectral image processing method based on multi-perspective and spatial-spectral combined unmixing provided by the present invention combines multi-perspective spectral and spatial information processing technology, and uses a spatial-spectral combined network to perform parallel unmixing of two types of information, thereby significantly improving the unmixing accuracy in complex scenes.

[0146] The above is a detailed description of the present invention based on specific embodiments, but it is obvious that the specific implementation of the present invention is not limited thereto. For those skilled in the art, various obvious changes made to it without departing from the scope of the claims of the present invention are within the protection scope of the present invention.

Claims

1. A multi-view hyperspectral image processing method that can be used for ocean target analysis, characterized in that: It includes the following steps: S1: Obtain hyperspectral data HSI containing L spectral bands and n mixed pixels, denoted as HSI(L,n); S2: Conduct a preliminary division of the L spectral bands in the hyperspectral data to form a multi-view dataset; S3: Extract features simultaneously in the spatial and spectral dimensions to obtain their respective feature matrices; S4: Fuse the features in the spectral dimension and the spatial dimension to generate a fused feature vector; input the fused feature vector into a decoder to restore the input pixel vector; S5: Optimize the loss function and conduct network training, and use the trained network model to process data to obtain the final endmember output.

2. The multi-view hyperspectral image processing method according to claim 1, characterized in that: In the above S1, the hyperspectral data HSI is obtained through hyperspectral imaging technology and is a multi-dimensional data, denoted as HSI(L,n), where L represents the number of spectral bands and n represents the number of pixel points; each pixel point not only has spatial information but also contains a complete spectral curve.

3. The multi-view hyperspectral image processing method according to claim 1, characterized in that: In the above S2, the L spectral bands in the hyperspectral data are preliminarily divided by the AAP or SD algorithm, including: S2-1: First, conduct hyperspectral data preprocessing: noise removal and normalization. After preprocessing, obtain the processed spectral data matrix HSI′(L,n), where HSI′ is the hyperspectral data that has been standardized and had noise removed; S2-2: Select the AAP or SD algorithm for band division: The original L bands are divided into k band clusters to generate a new data matrix HSI(k,n), where k < L, and each band cluster represents a group of bands with similar spectral characteristics.

4. The multi-view hyperspectral image processing method according to claim 3, characterized in that: In the above S2-1: Noise removal: Through noise analysis, use statistical methods to remove bands affected by atmospheric interference or noise; Normalization: Standardize the bands to ensure that the values of each band are at the same magnitude for more stable subsequent processing.

5. The multi-view hyperspectral image processing method according to claim 3, characterized in that: In the above S2-2: The AAP algorithm is described as follows: the bands are dynamically divided based on the local changes of the spectral curve. Assuming that the change between the bands is represented by a gradient, let G i =|HSI′ {i+1} –HSI′ i |, when the gradient G i When it exceeds the preset threshold, it is regarded as the starting point of a new band cluster; The specific calculation formula of the SD algorithm is: Then, use a clustering algorithm to cluster the bands and assign bands with higher similarity to the same cluster.

6. The multi-view hyperspectral image processing method according to claim 1, characterized in that: The above S3 specifically includes: S3-1: Spectral dimension feature extraction In the spectral dimension, use a fully connected network to extract features from the input of the HSI(k,n). Through dimensionality reduction technology, reduce the spectral data from k dimensions to the dimension c of the target matrix, that is, the number of pure spectral endmembers; then use the ReLU activation function to enable the network to have non-linear expression ability; S3-2: Spatial dimension feature extraction For each pixel point, select a spatial window of size W×W, and extract the spatial information of its surrounding pixels through the method of weighted average to generate a spatial feature vector representing the pixel. Then, use a fully connected network, batch normalization operation, and activation function to reduce the dimension of the average feature of this spatial window. Finally, use the ReLU activation function to enable the network to have non-linear expression ability.

7. The multi-view hyperspectral image processing method according to claim 6, characterized in that: In S3-1: Assume that the input spectrum vector is Through the fully connected network, batch normalization operation and activation function, the feature extraction formula is: in, is the weight matrix of spatial features, is the bias vector, and the output is the feature representation after dimensionality reduction The scale and offset parameters γ, β are the parameters of batch normalization, respectively.

8. The multi-view hyperspectral image processing method according to claim 6, characterized in that: In the above S3-2, the extraction formula of the spatial feature vector is: Where W is the spatial window size, Represents x i and x w The spatial distance between Then, use a fully connected network, batch normalization operation, and activation function to reduce the dimension of the average feature of this spatial window, and the formula is: in, is the weight matrix of spatial features, is the bias vector, and the scaling and offset parameters γ, β are the parameters of batch normalization, respectively.

9. The multi-view hyperspectral image processing method according to claim 1, characterized in that: In S4, the feature fusion strategy assumes that the spectral feature is f spectral (x i ), the spatial feature is f spatial (x i ), the fusion features of space and spectrum are obtained by using the product method, and the residual structure is introduced. The final f fusion (x i )The formula is: f fusion (x i )=f spatial (x i )+f spectral (x i )×f spatial (x i ) The fused features are decoded Reconstructed into the original spectral dimension L, the formula of the reconstruction process is: in, is the weight matrix of the decoder, is the bias vector. The output This is the reconstructed spectral data.

10. The multi-view hyperspectral image processing method according to claim 1, characterized in that: The above S5 includes: S5-1: Loss function and optimization Using spectral angle distance SAD and abundance Regularization is used as the loss function; SAD is calculated by calculating the angle between the reconstructed spectrum and the original spectrum, and its formula is: Among them, x i is the original spectrum vector, is the reconstructed spectrum vector; Abundance It is expressed by the formula: in Considered as the abundance obtained; So the complete loss function is regarded as: Where λ1 and λ2 represent the weights of the two loss functions respectively; The Adam optimization algorithm is used to optimize the network parameters. The update formula is: Among them, θ t represents the network parameters, α is the learning rate; S5-2: After network training and optimization, the final output is the endmember feature, that is, the pure spectral endmember matrix extracted by the network Represents the spectral characteristics of different substances.

Images