A multi-view hyperspectral image processing method for marine target analysis

Abstract

The application discloses a multi-view hyperspectral image processing method which can be used for marine target analysis and belongs to the technical field of hyperspectral image processing.The application effectively utilizes multi-dimensional information of data by simultaneously extracting spectral features and spatial features of a hyperspectral image.The application introduces an AAP and an SD algorithm, and adaptively reduces high-dimensional spectral data according to the characteristics of spectral bands.The application adopts an end-to-end deep learning framework, integrates spectral feature extraction, spatial feature fusion, unmixing and data reconstruction in the same network, and can efficiently unmix and accurately classify complex spectral and spatial information in multi-view hyperspectral images, and is especially suitable for marine environment monitoring, resource exploration and environmental protection and the like, and has important application value in spectral separation and classification of complex marine environments.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral image processing technology, and in particular relates to a multi-view hyperspectral image processing method that can be used for marine target analysis. Background Technology

[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 acquired has grown 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 fields such as environmental monitoring, agricultural exploration, and ocean observation. For example, the Hyperion satellite launched by the United States and my country's hyperspectral observation satellites can provide high-precision remote sensing data covering hundreds of spectral bands. This hyperspectral data not only provides rich spectral information but also effectively identifies the spectral characteristics of different ground features.

[0003] With the rapid development of remote sensing technology in my country, hyperspectral remote sensing applications are becoming increasingly widespread, resulting in a growing volume and complexity of acquired hyperspectral data. Accurate unmixing and classification of this data has become an important research direction. Unmixing techniques aim to separate the spectral components of different land features within mixed pixels, which is crucial for improving the analysis accuracy of hyperspectral images. However, hyperspectral images are often affected by noise, spatial heterogeneity, and spectral redundancy. Traditional unmixing methods often perform poorly when dealing with these complex scenarios, making it difficult to fully extract useful information from both spatial and spectral dimensions.

[0004] In recent years, research on the unmixing and classification of hyperspectral images has emerged continuously, which can be mainly divided into traditional methods based on spectral information and modern methods that combine spatial information. (1) Traditional spectral unmixing methods usually rely only on spectral information to perform linear or nonlinear decomposition of spectral features in hyperspectral data, but it is difficult to effectively utilize spatial information, resulting in limitations in the analysis of mixed pixels. (2) Modern methods combine spectral and spatial information and use 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, such as the problem of low training efficiency of the model when the data dimension is extremely high. 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 summary, current hyperspectral image unmixing and classification methods are easily affected by noise and spectral mixing in complex environments, making it 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 in current hyperspectral remote sensing research. Summary of the Invention

[0006] The purpose of this 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] This invention effectively improves the unmixing accuracy and classification ability of hyperspectral images by combining multi-view data construction with spatial spectral networks. The specific technical solution is as follows:

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

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

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

[0011] S3: Simultaneously extract features from both spatial and spectral data to obtain their respective feature matrices;

[0012] S4: Fuse the features of the spectral dimension and the spatial dimension to generate a fused feature vector; use the decoder to restore the input pixel vector from the input fused feature vector.

[0013] S5: Optimize the loss function and train the network. Use the trained network model to process the data and obtain the final endmember output.

[0014] Furthermore, in S1, the hyperspectral data (HSI) is acquired through hyperspectral imaging technology and is a type of 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 step S2, the L spectral bands in the hyperspectral data are initially divided using the AAP or SD algorithm, including:

[0016] S2-1: Hyperspectral data preprocessing is first performed: noise removal and normalization. After preprocessing, a processed spectral data matrix HSI′(L,n) is obtained, wherein HSI′ is hyperspectral data that has been standardized and denoised;

[0017] S2-2: Select AAP or SD algorithm for band partitioning: 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.

[0018] Further, in said S2-1: Noise removal: through noise analysis, a statistical method (e.g., standard deviation analysis) is used to remove those bands that are affected by atmospheric interference or have relatively high noise; Normalization: standardization processing is performed on the bands to ensure that the values of each band are on the same order of magnitude, so that the subsequent processing process is more stable.

[0019] Further, in said S2-2:

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

[0021] SD (Spectral Division) algorithm: the SD algorithm partitions bands through spectral clustering or similarity measurement. Commonly used measurement methods include cosine similarity or Euclidean distance, and the specific calculation formula is:

[0022]

[0023] Then the bands are clustered by a clustering algorithm (such as K-means or hierarchical clustering), and bands with high similarity are assigned to the same cluster.

[0024] Further, said S3 specifically comprises:

[0025] S3-1: Spectral dimension feature extraction

[0026] In the spectral dimension, a fully connected network is used to extract features from the input of said HSI(k,n), and through dimension reduction technology, the spectral data is reduced from k dimensions to the dimension c of the target matrix, that is, the number of pure spectral endmembers; then the ReLU activation function is used to enable the network to have nonlinear expression capability;

[0027] S3-2: Spatial Dimension Feature Extraction

[0028] For each pixel, a spatial window of size W×W is selected. The spatial information of the surrounding pixels is extracted by weighted averaging to generate a spatial feature vector representing the pixel. Then, the average feature of the spatial window is reduced in dimensionality using a fully connected network, batch normalization, and activation function. Finally, the ReLU activation function is used to give the network non-linear expressive power.

[0029] Furthermore, in S3-1: assuming the input spectral vector is... The feature extraction formula, derived using a fully connected network, batch normalization, and activation functions, is as follows:

[0030]

[0031] in, It is the weight matrix of spatial features. It is a bias vector, and the output is the dimensionality-reduced feature representation. The scaling and offset parameters γ and β are the parameters for batch normalization.

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

[0033]

[0034] Where W is the size of the spatial window, x represents i and x w Spatial distance between them;

[0035] Then, the dimensionality of the average features of the spatial window is reduced using a fully connected network, batch normalization, and activation function, as shown in the formula:

[0036]

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

[0038] Furthermore, in S4, the feature fusion strategy can use methods such as splicing and element-wise addition; assuming the spectral feature is f spectral (x i ), spatial features are f spatial (x i The spatial and spectral fusion features are obtained using a product method, and a residual structure is introduced to finally obtain f. fusion (x iThe formula is:

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

[0040] The fused features are processed by the decoder. The original spectral dimension L is reconstructed, and the formula for the reconstruction process is:

[0041]

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

[0043] Furthermore, S5 includes:

[0044] S5-1: Loss Function and Optimization

[0045] To measure the difference between the reconstructed data and the original input data, spectral angular distance (SAD) and abundance were used. Regularization is used as the loss function; SAD measures the similarity between the reconstructed spectrum and the original spectrum by calculating the angle between them, and the formula is as follows:

[0046]

[0047] Where, x i The original spectral vector, The reconstructed spectral vector;

[0048] And abundance This can be expressed by the formula:

[0049]

[0050] in Considered as the obtained abundance;

[0051] Therefore, the complete loss function can be considered as:

[0052]

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

[0054] The network parameters are optimized using the Adam optimization algorithm, and the update formula is as follows:

[0055]

[0056] Where, θ t This represents the network parameters, where α is the learning rate;

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

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

[0059] 1. By combining spectral and spatial characteristics, the unmixing accuracy is improved:

[0060] This invention effectively utilizes the multidimensional information of data by simultaneously extracting spectral and spatial features from hyperspectral images. Spectral features are used to separate the spectral components of mixed substances, while spatial features capture the spatial correlation of pixels. By fusing these two features, this invention not only improves the accuracy of spectral unmixing for complex substances but also maintains spatial consistency and reduces the impact of noise and boundary effects during the unmixing process.

[0061] 2. Adaptive band division reduces data redundancy:

[0062] This invention introduces AAP (Adaptive Partitioning) and SD (Spectrum Partitioning) algorithms to adaptively reduce the dimensionality of high-dimensional spectral data based on the characteristics of spectral bands. This process effectively reduces redundant bands, preserves key information, and improves the computational efficiency and robustness of the model. Furthermore, the data after band partitioning allows for better feature extraction during processing, enhancing the adaptability of the demixing model to complex scenarios.

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

[0064] This invention employs an end-to-end deep learning framework, integrating spectral feature extraction, spatial feature fusion, unmixing, and data reconstruction into a single network. This architecture, through fully connected layers and convolutional operations, enables simultaneous processing of spectral and spatial information and automatically learns the optimal feature representation. By using the Adam optimization algorithm, the model converges quickly and maintains high unmixing accuracy in complex scenarios. This end-to-end approach not only improves processing efficiency and reduces the complexity of manual parameter tuning but also enhances the model's robustness across different datasets and application scenarios. Attached Figure Description

[0065] Figure 1 This is a flowchart of the present invention.

[0066] Figure 2 This is a schematic flowchart of the hyperspectral processing method of the present invention.

[0067] Figure 3 This is a schematic diagram of the modules of the deep learning model of the present invention.

[0068] Figure 4 This is a schematic diagram of the input processing of the network model of the present invention.

[0069] Figure 5 This is a schematic diagram of the Urban hyperspectral input data of the present invention.

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

[0071] Figure 7 The implementation method is shown in the endmember effect diagram on the Urban dataset.

[0072] Figure 8 The implementation of the method is shown in the endmember effect diagram on the Jasper dataset.

[0073] Figure 9 This is a schematic diagram of marine hyperspectral data. Detailed Implementation

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

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

[0076] Step 1: Acquire hyperspectral data and determine the number of spectral bands and mixed pixels.

[0077] Step 1.1. Hyperspectral Data (HSI) is acquired through hyperspectral imaging technology. It is a multidimensional dataset, denoted as HSI(L,n), where L represents the number of spectral bands and n represents the number of pixels. Each pixel not only contains spatial information but also a complete spectral curve. Unlike traditional RGB images, which only contain red, green, and blue bands, hyperspectral images cover a continuous spectrum from visible light to near-infrared, typically including hundreds or even thousands of bands. This spectral information allows for the identification and analysis of the composition and properties of substances. Common hyperspectral datasets include those from Indian Pines and Pavia University, which can be downloaded via official links provided by the University of Pavia and Purdue University.

[0078] Step 2. Perform preliminary division of the L spectral bands in the hyperspectral data using the AAP or SD algorithm.

[0079] Step 2.1. Hyperspectral data preprocessing

[0080] Before dividing the spectral bands, necessary preprocessing of the hyperspectral data is performed to ensure data integrity and quality. 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 levels.

[0082] (2) Normalization: Standardize the bands to ensure that the values ​​of each band are on the same order of magnitude so that the subsequent processing 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, we obtain the processed spectral data matrix HSI′(L,n), where HSI′ is the normalized and noise-removed hyperspectral data.

[0084] Step 2.2 Band Division

[0085] Choose either the AAP or SD algorithm for band division:

[0086] (1) AAP (Automatic Adaptive Partitioning) Algorithm: The AAP algorithm is an adaptive partitioning method that dynamically divides bands based on local changes in the spectral curve. It assumes that the changes between bands can be represented by gradients, and lets G... i =|HSI′ {i+1} –HSI′ i |, when gradient Gi When the value exceeds a 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. Commonly used measurement methods include cosine similarity or Euclidean distance, and the specific calculation formula is:

[0088]

[0089] Bands are clustered by a clustering algorithm (such as K-means or hierarchical clustering), and bands with high similarity are assigned to the same cluster.

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

[0091] After processing by the AAP or SD algorithm, 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 set of bands with similar spectral characteristics.

[0092] Step 3. Performing feature extraction simultaneously in spatial and spectral domains and reconstructing data, as shown in Figure 4 below:

[0093] Step 3.1. Spectral dimension feature extraction

[0094] In the spectral dimension, a fully connected network is used to perform feature extraction on the obtained input of HSI(k,n); through dimension reduction technology, the spectral data is reduced from k-dimension to c-dimension of the target matrix, where c is the number of pure spectral endmembers.

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

[0096]

[0097] Wherein, is a weight matrix of spatial features, is a bias vector, the scaling and offset parameters γ and β are respectively parameters of batch normalization, and the output is the feature representation after dimension reduction

[0098] The batch normalization herein standardizes the output of linear transformation to adjust the output to have a mean of 0 and a variance of 1.

[0099] Finally, the ReLU activation function is selected to endow the network with nonlinear expression capability.

[0100] Step 3.2. Spatial Dimension Feature Extraction

[0101] For each pixel, a spatial window of size W×W is selected, and spatial information of its surrounding pixels is extracted using a weighted average method to generate a spatial feature vector representing the pixel.

[0102]

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

[0104] Then, the dimensionality of the average features of the spatial window is reduced using a fully connected network, batch normalization, and activation function, as shown in the formula:

[0105]

[0106] in, It is the weight matrix of spatial features. It 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, adjusting it to have a mean of 0 and a variance of 1.

[0108] Finally, the ReLU activation function is used to give the network non-linear expressive capabilities.

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

[0110] Features from the spectral and spatial dimensions are fused to generate a fused feature vector. Common feature fusion strategies include concatenation and element-wise addition. Assume the spectral features are f... spectral (x i ), spatial features are f spatial (x i Here we use the product method to obtain the fused spatial and spectral features, and introduce a residual structure to finally obtain 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 processed by the decoder. The formula for reconstructing the original spectral dimension L is:

[0113]

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

[0115] Step 5. Optimize the network using deep learning

[0116] Step 5.1. Loss Function and Optimization

[0117] To measure the difference between the reconstructed data and the original input data, spectral angular distance (SAD) and abundance were used. Regularization is used as the loss function. SAD measures the similarity between the reconstructed spectrum and the original spectrum by calculating the angle between them, and the formula is as follows:

[0118]

[0119] Where, x i The original spectral vector, This is the reconstructed spectral vector.

[0120] And abundance This can be expressed by the formula:

[0121]

[0122] in This can be considered as the obtained abundance;

[0123] Therefore, 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 network parameters are optimized using the Adam optimization algorithm, and the update formula is as follows:

[0127]

[0128] Where, θ t This represents the network parameters, where α is the learning rate.

[0129] Step 5.2. Output endmember features and abundance maps

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

[0131] Example 2:

[0132] This embodiment, combined with simulation experiments, further illustrates the effectiveness of the present invention.

[0133] The simulation experiments of this invention were conducted in a hardware environment with an Intel i5-12400 processor, an NVIDIA GTX 3080 graphics card, and 16GB of RAM, and a software environment using Python 3.7 and PyTorch 1.0. The simulation experimental data used in this invention are the Urban dataset and the Jasper dataset. The Urban dataset was acquired by the HYDICE sensor and contains 307×307 pixels. After removing low-quality bands, the remaining 162 channels were used in the experiments. The Jasper dataset was acquired by the AVIRIS sensor and contains 100×100 pixels. After removing low-quality bands, the remaining 198 channels were used in the experiments.

[0134] Figure 5 The image shows a cube of the Urban dataset containing four main materials: asphalt, grass, trees, and rooftops. Figure 6 The image shows cube images from the Jasper dataset, which contain four main materials: trees, water, soil, and roads. Figure 7 This demonstrates the endmember features extracted from Urban data using this invention. From left to right and top to bottom, they represent asphalt, grass, trees, and rooftops, respectively. Figure 7 The red line represents the true value, while the blue line represents the endmember spectral characteristic curve obtained by this invention. It can be seen that the spectral characteristic curve obtained by this invention is very close to the true value. Figure 8 The diagram illustrates the endmember features obtained by this invention on the Jasper dataset. The red line represents the ground truth, and the blue line represents the endmember curve obtained by this invention. The results from both datasets show that the endmember feature curves we obtained are very close to the ground truth, especially on the Jasper dataset, where this fit is even more pronounced.

[0135] The comparison results between this invention and existing advanced change detection methods are shown in Tables 1 and 2. Table 1 shows the comparison results on the Urban dataset, and Table 2 shows the comparison results on the Jasper dataset. The convolutional neural network autoencoder unmixing method (hereinafter abbreviated as CNNAEU) used in the comparative experiments is proposed in the paper "Convolutional Autoencoder for Spectral–Spatial Hyperspectral Unmixing"; the EndNet method is proposed in the paper "EndNet: Sparse AutoEncoder Network for EndmemberExtraction and Hyperspectral Unmixing"; the TANet method is proposed in the paper "TANet: An Unsupervised Two-Stream Autoencoder Network for Hyperspectral Unmixing"; and the SSAE method is proposed in the paper "Spatial-Spectral Autoencoder Networks for Hyperspectral Unmixing".

[0136] This invention uses spectral angular distance (SAD) to compare with the above methods in terms of objective indicators. The calculation method is as follows:

[0137]

[0138] in,<x,y> It is the dot product of vectors x and y, where ||x|| and ||y|| represent the Euclidean normal forms of x and y, respectively. n is the number of samples. y is the predicted value of the i-th sample. i This represents the true value of the i-th sample. Here, SAD is a metric that measures the angular difference between two spectral vectors, 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 magnitudes or amplitudes. This means that SAD is insensitive to scale changes in the spectrum (e.g., brightness differences due to variations in illumination conditions), and is more concerned with the shape characteristics of the spectrum. Therefore, SAD is a reliable metric for evaluating spectral similarity during unmixing.

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

[0140] 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 This invention 4.38 3.31 3.32 2.58 3.39

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

[0142]

[0143]

[0144] This invention provides a multi-view hyperspectral image processing method for marine target analysis, primarily used to analyze and monitor the distribution and dynamic changes of surface materials, classifying and representing the abundance of substances present in remote sensing images. However, this invention is not limited to the aforementioned application scenarios; it is also applicable to spectral analysis of other complex surface environments, such as marine pollution assessment and coastal ecosystem monitoring, and can perform efficient processing at the spatial-spectral scale, demonstrating broad application scenarios and good practical results. Hyperspectral unmixing technology is a crucial technique in marine remote sensing, enabling precise monitoring and analysis of complex marine environments. The marine environment is highly complex, with various components such as seawater, phytoplankton, suspended matter, benthic organisms, and pollutants often coexisting and mixing together. Traditional remote sensing methods struggle to accurately distinguish and quantify these components. Hyperspectral imaging technology, by acquiring continuous spectral information, can capture subtle spectral differences, aiding in in-depth analysis of various marine components. The emergence of hyperspectral unmixing technology makes it possible to separate pure spectral information from mixed pixels, providing crucial support for marine environmental monitoring and research. Applying this invention to the marine field can have a positive impact on marine observation. Regarding marine datasets… Figure 9 (Right) The dataset shows a cropped RGB rendering of a hyperspectral image of Trondheim Fjord taken by the HYPSO-1 satellite.

[0145] The hyperspectral image processing method based on multi-view and spatial spectral combination provided by this invention combines multi-view spectral and spatial information processing techniques, and uses a spatial spectral combination network to perform parallel demixing of the two types of information, which significantly improves the demixing accuracy in complex scenes.

[0146] The foregoing detailed description of the present invention based on specific embodiments is obviously not limited thereto. For those skilled in the art, various obvious modifications made to the invention without departing from the scope of the claims are within the protection scope of the present invention.

Claims

1. A multi-view hyperspectral image processing method for marine target analysis, characterized in that, Includes the following steps: S1: Obtain hyperspectral data HSI containing L spectral bands and n mixed pixels, denoted as HSI(L, n); S2: Perform preliminary division of L spectral bands in the hyperspectral data to form a multi-view dataset; In step S2, the L spectral bands in the hyperspectral data are initially divided using the AAP or SD algorithm, including: S2-1: First, the hyperspectral data is pre-processed: noise removal and normalization, after pre-processing, the processed spectral data matrix is obtained wherein is the normalized and noise-removed hyperspectral data; S2-2: Select either 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; in S2-2: The AAP algorithm: dynamically divides wave bands based on local changes of spectral curves, assumes that the change amount between wave bands is represented by a gradient, and lets When the gradient exceeds a preset threshold, it is considered as the starting point of a new wave band cluster. The specific calculation formula for the SD algorithm is as follows: ； Then, a clustering algorithm is used to cluster the bands, assigning bands with high similarity to the same cluster; S3: Simultaneously extract features from both spatial and spectral data to obtain their respective feature matrices; S3 specifically includes: S3-1: Spectral Dimension Feature Extraction In the spectral dimension, a fully connected network is used to extract features from the input of HSI(k, n). Dimensionality reduction techniques are used to reduce the spectral data from k dimensions to the dimension c of the target matrix, i.e., the number of pure spectral endmembers. Then, the ReLU activation function is used to enable the network to have non-linear expressive capabilities. S3-2: Spatial Dimension Feature Extraction For each pixel, a spatial window with a size of is selected, the spatial information of the surrounding pixels is extracted by the weighted average method to generate a spatial feature vector representing the pixel, then the average features of the spatial window are reduced by using a fully connected network, batch normalization operation and activation function, and finally the ReLU activation function is used to make the network have nonlinear expression ability; S4: Fuse the features of the spectral dimension and the spatial dimension to generate a fused feature vector; use the decoder to restore the input pixel vector from the input fused feature vector. S5: Optimize the loss function and train the network. Use the trained network model to process the data and obtain the final endmember output.

2. The multi-view hyperspectral image processing method as described in claim 1, characterized in that, In S1, the hyperspectral data HSI is acquired through hyperspectral imaging technology and is a type of 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.

3. The multi-view hyperspectral image processing method as described in claim 1, characterized in that, In S2-1: Noise removal: Through noise analysis, statistical methods are used to remove bands affected by atmospheric interference or noise; Normalization: Standardize the bands to ensure that the values ​​of each band are on the same order of magnitude, so that subsequent processing is more stable.

4. The multi-view hyperspectral image processing method as described in claim 1, characterized in that, In S3-1: assuming the input spectral vector is Through a fully connected network, batch normalization, and activation functions, the feature extraction formula is as follows: in, It is the weight matrix of spatial features. It is a bias vector, and the output is the dimensionality-reduced feature representation. Scaling and offset parameters These are the parameters for batch normalization.

5. The multi-view hyperspectral image processing method as described in claim 1, characterized in that, In S3-2, the formula for extracting spatial feature vectors is: Where W is the size of the spatial window, express and Spatial distance between them; Then, the dimensionality of the average features of the spatial window is reduced using a fully connected network, batch normalization, and activation function, as shown in the formula: in, It is the weight matrix of spatial features. It is the bias vector, scaling, and offset parameters. These are the parameters for batch normalization.

6. The multi-view hyperspectral image processing method as described in claim 1, characterized in that, In S4, the feature fusion strategy assumes that the spectral features are Spatial features are The spatial and spectral fusion features are obtained using a product method, and a residual structure is introduced to finally obtain... The formula is: The fused features are obtained through the decoder. The original spectral dimension L is reconstructed, and the formula for the reconstruction process is: in, It is the weight matrix of the decoder. It is the bias vector; the output This refers to the reconstructed spectral data.

7. The multi-view hyperspectral image processing method as described in claim 1, characterized in that, S5 includes: S5-1: Loss Function and Optimization Using spectral angular distance SAD and abundance Regularization is used as the loss function; SAD calculates the angle between the reconstructed spectrum and the original spectrum, using the following formula: in, The original spectral vector, The reconstructed spectral vector; Abundance This can be expressed by the formula: in Considered as the obtained abundance; Therefore, the complete loss function can be considered as: in These represent the weights of the two loss functions, respectively. The network parameters are optimized using the Adam optimization algorithm, and the update formula is as follows: in, Represents network parameters, The learning rate; S5-2: After network training and optimization, the final output is the endmember feature, which is the pure spectral endmember matrix extracted by the network. , representing the spectral characteristics of different substances.

Images