Hyperspectral unmixing method of cascaded auto-encoder based on spatial constraint

Abstract

The invention discloses a spatial constraint-based hyperspectral unmixing method of a cascade auto-encoder. The method comprises the following steps of S1, acquiring a ground object hyperspectral image; s2, constructing a cascading auto-encoder unmixing model of spatial constraint; s3, performing unmixing on the obtained surface feature hyperspectral image by using the constructed spatial constraint cascade auto-encoder unmixing model to obtain surface feature abundance corresponding to each surface feature; the cascade auto-encoder unmixing model adopts a layer-by-layer training mode; the self-encoder comprises at least one encoder, a fusion layer and a decoder. The encoder comprises a multi-scale feature extraction module, a BN layer and an activation function layer. The multi-scale feature extraction module comprises a small convolution kernel for focusing on local feature extraction and a large convolution kernel for capturing global context information. According to the invention, precise identification and classification of ground feature components are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperspectral image processing in the field of remote sensing, and in particular relates to a hyperspectral unmixing method of a cascaded autoencoder based on spatial constraints. Background Art

[0002] Imaging spectroscopy technology organically combines the spectrum of ground objects determined by material composition with the spatial distribution image reflecting the existence pattern of ground objects, providing an innovative method so that each pixel on the remote sensing image has both spectral information representing its physical properties and spatial information reflecting its geometric characteristics. In the process of ground object identification and classification, hyperspectral imaging technology is particularly important. By obtaining hyperspectral images of surface cover, researchers can deeply analyze the spectral characteristics of various ground objects. After interpreting these spectral sample data, the composition, chemical composition and structural characteristics of the ground objects in the detection data can be accurately determined, thereby identifying potential ground object resources. Furthermore, by comparing with the spectral database of known ground object resources, rapid identification and quantitative analysis of potential ground object resources can be achieved. This process not only improves the efficiency of data processing, but also provides important experimental data support for the development and utilization of ground object resources. This data support is of great significance to fields such as environmental monitoring, land management and resource assessment. Therefore, hyperspectral technology shows broad application prospects in ground object analysis. With the continuous advancement of technology and the improvement of data processing capabilities, it is expected that hyperspectral imaging technology will play a more critical role in the exploration and development of land resources in the future, promoting sustainable development and ecological protection.

[0003] The difficulties in object recognition are mainly reflected in the following aspects. First, the spectral characteristics of different objects may be similar, leading to confusion and misidentification. Second, the spectral information of objects is affected by environmental factors such as climate, season and soil conditions, which makes the stability and reliability of spectral data challenging. In addition, the amount of data is huge and complex, requiring efficient algorithms and models for processing and analysis, which places higher demands on computing power and technical level. Finally, the lack of a comprehensive spectral database also limits the accurate identification and analysis of potential resources.

[0004] Spectral variability is caused by the interaction of multiple factors such as surface material, lighting conditions, and atmospheric effects, which leads to differences in spectral response of the same object under different conditions. This characteristic makes it difficult for the linear mixed model to accurately capture the true spectral characteristics during the unmixing process, thus affecting the accuracy and robustness of the results.

[0005] Therefore, it is necessary to design a new hyperspectral unmixing method. Summary of the invention

[0006] The purpose of the present invention is to provide a hyperspectral unmixing method based on a spatially constrained cascaded autoencoder to solve the problems faced by hyperspectral image unmixing, such as spatial resolution limitation, complex and diverse distribution of objects, low data signal-to-noise ratio, and spectral variability.

[0007] To achieve the above object, the present invention provides a hyperspectral unmixing method of a cascaded autoencoder based on spatial constraints, comprising the following steps: S1, obtaining a hyperspectral image of a ground object;

[0008] S2, construct a spatially constrained cascaded autoencoder unmixing model;

[0009] S3, using the constructed spatially constrained cascade autoencoder unmixing model to unmix the acquired hyperspectral image of the ground objects, and obtain the ground object abundance corresponding to each ground object;

[0010] The spatially constrained cascaded autoencoder unmixing model adopts a layer-by-layer training method, that is, the first layer of autoencoder is trained first, and then the output of the trained first layer of autoencoder is used as the input of the second layer of autoencoder, and so on, iteratively performed until the termination condition of the iteration is reached;

[0011] The autoencoder specifically includes a preprocessing stage module and a spectral unmixing stage module; the preprocessing stage module includes at least one encoder, and the spectral unmixing stage module includes a fusion layer and a decoder;

[0012] The encoder includes a multi-scale feature extraction module, a BN layer and an activation function layer. The multi-scale feature extraction module is a Conv layer, and the activation function layer adopts a LeakyReLU activation function; the multi-scale feature extraction module includes a small convolution kernel for focusing on local feature extraction and a large convolution kernel for capturing global context information.

[0013] In a specific implementation, a spatial constraint term is introduced into the autoencoder, and the spatial constraint term is used to take information of adjacent pixels into consideration to optimize the parameters of the autoencoder.

[0014] In a specific implementation, when the spatially constrained cascaded autoencoder unmixing model uses a deep learning-based spatial regularization unmixing algorithm to process a hyperspectral image of a ground object:

[0015] First, quantify the similarity between adjacent pixels; use Euclidean distance as an indicator to measure the similarity of adjacent pixels, and describe the similarity between pixels by calculating the Euclidean distance between pixels;

[0016] Then, based on the Euclidean distance information between pixels, the neighborhood weight ∈ ij , neighborhood weight ∈ ij It is expressed as:

[0017]

[0018] Among them, im(·) represents the pixel x i and x j The Euclidean distance between pixels x i and pixel x j The closer the position, the smaller the Euclidean distance, and the weight ∈ ij The bigger;

[0019] (a, b) and (c, d) are pixel x respectively. i and pixel x j The spatial coordinates of , then the Euclidean distance is expressed as

[0020] Use neighborhood weights to mine spatial correlations as follows:

[0021] In the formula, represents the abundance element neighborhood x ij gather;

[0022] Construct spatial weighting factors:

[0023] Introduce the constructed spatial weights into the model:

[0024]

[0025] In a specific embodiment, a loss function is set for the spatially constrained cascaded autoencoder unmixing model, and the loss function is used to evaluate the accuracy of spectral unmixing and the effective utilization of spatial information; an optimization algorithm is set to minimize the loss function, and the model parameters of the spatially constrained cascaded autoencoder unmixing model are iteratively updated until convergence to an optimal solution, thereby ensuring that the spatially constrained cascaded autoencoder unmixing model can optimally fuse spectral and spatial information.

[0026] In a specific implementation, the loss function is constructed as follows:

[0027]

[0028] Among them, L rec Represents the true pixel vector Y and the predicted pixel vector and The average value of the sum of squares of the differences between , Y represents the true pixel vector, and denotes the predicted pixel vector, and N represents the total number of pixels.

[0029] In a specific implementation, the autoencoder also introduces a structured sparsity loss function:

[0030] L sparse =||Y|| 2,1 .

[0031] In a specific embodiment, the fused layer is a fusion layer.

[0032] In a specific embodiment, the training process of the spatially constrained cascaded autoencoder unmixing model uses six encoders to work together. When the encoders perform multi-scale feature extraction, two encoders are selected from three encoders with convolution kernels of different scales. One of the encoders uses a small 1x1 convolution kernel to focus on local feature extraction, and the other encoder uses a large 5x5 or 9x9 convolution kernel to capture global context information.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention solves the problems of spatial resolution limitation, complex and diverse distribution of objects, low data signal-to-noise ratio and spectral variability faced by hyperspectral image unmixing, thereby achieving accurate identification and classification of object components.

[0035] The present invention innovatively adds spatial weighting factors to the framework of deep learning, constructs a spatially constrained cascade autoencoder that fuses multiple encoders, and constructs a new multi-scale feature extraction mechanism to utilize the spatial weighting factors to fully mine the spatial information in the hyperspectral image. At the same time, a noise sparse regularization term is introduced to resist noise interference to the maximum extent, effectively improving the classification and recognition of objects. The experimental results on simulated hyperspectral data and real hyperspectral data sets show that the present invention has achieved good recognition and classification effects.

[0036] In addition to the above-described objects, features and advantages, the present invention has other objects, features and advantages. The present invention is further described in detail below. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0038] Figure 1 It is a schematic diagram of a flow chart of an embodiment of the present invention;

[0039] Figure 2 is a true abundance diagram of each end member in a simulated data embodiment of the present invention;

[0040] Figure 3 is an unmixed abundance map of end member 2 in a simulated data embodiment of the present invention;

[0041] Figure 4 It is a color image of the end member simulated by the model in one embodiment of the present invention. DETAILED DESCRIPTION

[0042] The embodiments of the present invention are described in detail below. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0043] The present invention provides a hyperspectral unmixing method of a cascaded autoencoder based on spatial constraints, comprising the following steps: S1, obtaining a hyperspectral image of a ground object;

[0044] S2, construct a spatially constrained cascaded autoencoder unmixing model;

[0045] S3, using the constructed spatially constrained cascade autoencoder unmixing model to unmix the acquired hyperspectral image of the ground objects, and obtain the ground object abundance corresponding to each ground object;

[0046] The spatially constrained cascaded autoencoder unmixing model adopts a layer-by-layer training method, that is, the first layer of autoencoder is trained first, and then the output of the trained first layer of autoencoder is used as the input of the second layer of autoencoder, and so on, iteratively performed until the termination condition of the iteration is reached;

[0047] The autoencoder specifically includes a preprocessing stage module and a spectral unmixing stage module; the preprocessing stage module includes at least one encoder, and the spectral unmixing stage module includes a fusion layer and a decoder;

[0048] The encoder includes a multi-scale feature extraction module, a BN layer and an activation function layer. The multi-scale feature extraction module is a Conv layer, and the activation function layer adopts a LeakyReLU activation function; the multi-scale feature extraction module includes a small convolution kernel for focusing on local feature extraction and a large convolution kernel for capturing global context information.

[0049] The Conv layer performs linear transformation, laying the foundation for data flow. The conv layer is trained layer by layer.

[0050] The BN layer improves model stability and accelerates convergence. The fusion layer can integrate features from different encoders or different modalities to form a more comprehensive feature representation. The use of the Leaky ReLU activation function solves the gradient vanishing problem of ReLU and further enhances the nonlinear processing capability of the model. The decoder reconstructs the high-dimensional output from the fused low-dimensional features, compares it with the original pixel data, and calculates the reconstruction error to ensure the stability and accuracy of the output.

[0051] The present invention utilizes a spatially constrained cascaded autoencoder unmixing model to accurately estimate each ground feature component contained in a mixed pixel; based on the solved ground feature abundance, the corresponding proportion of each ground feature is quantified to achieve accurate identification and classification.

[0052] A spatial constraint term is introduced into the autoencoder, and the spatial constraint term is used to take the information of adjacent pixels into consideration to optimize the parameters of the autoencoder.

[0053] When the spatially constrained cascaded autoencoder unmixing model uses a deep learning-based spatial regularization unmixing algorithm to process hyperspectral images of ground objects:

[0054] In order to deeply explore the rich spatial information contained in the image, the similarity between adjacent pixels is quantified first; the Euclidean distance is used as an indicator to measure the similarity of adjacent pixels, and the Euclidean distance between pixels is calculated to describe the similarity between pixels;

[0055] Then, based on the Euclidean distance information between pixels, the neighborhood weight ∈ ij , so as to make full use of these spatial correlations in subsequent processing, thereby improving the unmixing effect and accuracy. Neighborhood weight ∈ ij It is expressed as:

[0056]

[0057] Among them, im(·) represents the pixel x i and x j The Euclidean distance between pixels x i and pixel x j The closer the position, the smaller the Euclidean distance, and the weight ∈ ij The bigger;

[0058] (a, b) and (c, d) are pixel x respectively. i and pixel x j The spatial coordinates of , then the Euclidean distance is expressed as

[0059] Use neighborhood weights to mine spatial correlations as follows:

[0060] In the formula, represents the abundance element neighborhood x ij gather;

[0061] Construct spatial weighting factors:

[0062] Introduce the constructed spatial weights into the model:

[0063]

[0064] The process of mining the spatial information contained in the image is to perform multi-scale feature extraction by two dual-path parallel encoders to mine the spatial information in the image, and fuse the outputs of the two encoders through the fusion layer.

[0065] The spatially constrained cascaded autoencoder unmixing model sets a loss function, which is used to evaluate the accuracy of spectral unmixing and the effective utilization of spatial information; an optimization algorithm is set to minimize the loss function, and the model parameters of the spatially constrained cascaded autoencoder unmixing model are iteratively updated until convergence to an optimal solution, thereby ensuring that the spatially constrained cascaded autoencoder unmixing model can optimally fuse spectral and spatial information.

[0066] The loss function is used to evaluate model performance and provides a comprehensive quantitative standard for model performance.

[0067] The preferred optimization algorithm is gradient descent.

[0068] The loss function is constructed as:

[0069] Among them, L rec Represents the true pixel vector Y and the predicted pixel vector and The average value of the sum of squares of the differences between , Y represents the true pixel vector, and denotes the predicted pixel vector, and N represents the total number of pixels.

[0070] In order to improve the accuracy of unmixing, reduce the computational complexity or optimize the model performance, it can promote sparsity, prevent overfitting and utilize the role of data group structure, which helps to improve the performance and stability of the model. The following constraints are also introduced, which can also be called structured sparsity loss function:

[0071] L sparse =||Y|| 2，1 .

[0072] In the formula, ||·|| 2,1 Indicates l 2,1 Norm, l 2,1 The norm plays a role in promoting sparsity, preventing overfitting, and exploiting the structure of data sets in hyperspectral image processing, which helps to improve the performance and stability of the model.

[0073] The added structured sparsity loss function also performs sparse processing on the reconstruction results. Structured sparsity takes into account the contextual information of the image and the correlation between pixels. By introducing structured sparsity loss, the sparsity and detail retention capabilities of the image can be further improved.

[0074] In addition, a SAD constraint is introduced, which can improve the accuracy and reliability of spectral data correction while ensuring that the corrected spectral vector is consistent with the true spectrum in shape and direction. The SAD constraint is defined as:

[0075]

[0076] Based on the above introduction, the overall constraint can be written as:

[0077] L=L rec +L spa +L sparse +L SAD

[0078] The fusion layer is a fusion layer.

[0079] The training process of the spatially constrained cascaded autoencoder unmixing model uses six encoders to work together. When the encoders perform multi-scale feature extraction, two encoders are selected from three encoders with convolution kernels of different scales. One of the encoders uses a small 1x1 convolution kernel to focus on local feature extraction, and the other encoder uses a large 5x5 or 9x9 convolution kernel to capture global context information.

[0080] A multi-scale feature extraction mechanism is introduced to capture local and global features through different convolution kernels, such as 1x1, 5x5, and 9x9, to enhance the model's understanding of complex input data.

[0081] The decoder reconstructs high-dimensional output data. In the output layer of the decoder, according to the specific task requirements, the model can flexibly select Sigmoid or Softmax function for probability output or normalization processing to meet the needs of different classification scenarios.

[0082] Based on deep learning, the hyperspectral image is unmixed to obtain the abundance of each feature, and the features are identified based on the abundance obtained. As we all know, the abundance of features is presented in the model as a matrix composed of a set of numbers. The method will automatically match the final unmixed abundance to complete the identification and classification of the abundance.

[0083] Example 1

[0084] A hyperspectral unmixing method based on a cascaded autoencoder with spatial constraints, the specific steps are as follows:

[0085] Put in the hyperspectral image that needs to be identified. The algorithm starts a multi-scale feature extraction mechanism using different convolution kernels. This mechanism works together through three independent encoders. The entire spatially constrained cascaded autoencoder unmixing model sets up a total of six encoders to work together to obtain the trained spatial features. A comprehensive capture of the input data features is achieved. One of the encoders uses a small convolution kernel (1x1) to focus on fine local feature extraction. One of the other two encoders uses a large convolution kernel (such as 9x9 or 5x5) to capture more extensive and global spatial context information. The outputs of the two encoders are then fused through the fusion layer to construct a spatial weighting factor:

[0086] W spa The i-th row and j-th column element at the t+1-th iteration in (i=1,...,m,j=1,...,n).

[0087] Among them, f(x ij ) represents spatial correlation, which can be expressed as:

[0088]

[0089] Where N(j) represents the element x ij The neighborhood set of , f(·) represents the function of mining spatial correlation through the neighborhood system.

[0090] Further define the neighborhood weight ∈ ij , so as to make full use of these spatial correlations in subsequent processing, thereby improving the unmixing effect and accuracy. Neighborhood weight ∈ ij It is expressed as:

[0091] Among them, im(·) represents the pixel x i and x j The Euclidean distance between pixels x i and pixel x j The closer the position, the smaller the Euclidean distance, and the weight ∈ ij The bigger;

[0092] (a, b) and (c, d) are pixel x respectively. i and pixel x j The spatial coordinates of , then the Euclidean distance is expressed as

[0093] The model integrates 6 encoders, 1 fusion layer and 1 decoder to build an efficient spectral data processing system. Among them, the Conv layer, as a fully connected (linear) layer, is responsible for the linear transformation of data and lays the foundation for data flow. The BN layer (batch normalization layer) effectively improves the stability and convergence speed of the model with its unique batch normalization operation, ensuring a smoother training process.

[0094] LReLU (Leaky Rectified Linear Unit) is used as the activation function, which cleverly solves the gradient vanishing problem that ReLU may encounter in negative value areas, and further enhances the nonlinear processing capability of the model.

[0095] In order to accurately evaluate and quantify the difference between the network output and the target output, the mean square error is used as the core measurement indicator. The REC constraints are as follows:

[0096]

[0097] At the same time, l 2,1 Norm constraints, the constraints are as follows:

[0098] L sparse =||Y|| 2，1

[0099] In addition, the introduction of SAD constraints can improve the accuracy and reliability of spectral data correction, while ensuring that the corrected spectral vector is consistent with the true spectrum in shape and direction. The SAD constraint is defined as:

[0100]

[0101] Based on the above introduction, the overall constraint can be written as:

[0102] L=L rec +L spa +L sparse +L SAD

[0103] The decoder reconstructs high-dimensional output data. In the output layer of the decoder, according to the specific task requirements, the model can flexibly select Sigmoid or Softmax function for probability output or normalization processing to meet the needs of different classification scenarios.

[0104] Based on deep learning, hyperspectral image unmixing is completed to obtain the abundance of each land feature.

[0105] Example 2

[0106] This embodiment constructs a set of simulated data, which consists of 5 end members. In order to fully cover the frequency range of spectral data, this embodiment sets 162 bands, and each band is regarded as a unique data dimension to capture the spectral response of substances at different frequencies. It is particularly critical that a 162×162 pixel matrix is ​​constructed in each band. This matrix structure not only reflects the fine distribution of spectral data in the spatial dimension, but also ensures the integrity and analyzability of the data set. In order to simulate the complexity and uncertainty that spectral data may encounter in real scenes, this embodiment deliberately introduces 30dB of noise into this set of data to enhance the impact of spectral variability.

[0107] In this dataset, the encoder learning rate of the model is set to 4.5×(10)^(-3), and the decoder learning rate is 1×(10)^(-3); the number of iterations is 300; for the loss function, its parameters are beta1, beta2, beta3, beta4, and beta5, among which beta1 is set to 8×(10)^(-7); beta2 is set to 4.4×(10)^(-6); beta3 is set to 9.5×(10)^(-2); beta4 is set to 3×(10)^(-1); and beta5 is set to 5×(10)^(-2).

[0108] The fully constrained least squares algorithm FCLS (Fully Constrained Least Squares), the sparse unmixing algorithm SUnSAL (Sparse Unmixing via Variable Splitting and Augmented Lagrangian), the linear augmented mixture algorithm ALMM (Augmented Linear Mixed Model) and the convolutional neural network automatic endmember extraction and unmixing algorithm CNNAEU (Convolutional Neural Network for Automatic Endmember Extraction and Unmixing) were used for comparative experiments to verify the effectiveness of the model.

[0109] SAD, aSAD and aRMSE are used to quantitatively measure the quality of the measurement results, and their definitions are:

[0110]

[0111] SAD is the abbreviation of Sum of Absolute Difference, which is used to measure the difference in pixel values ​​between two images. It is obtained by calculating the sum of the absolute differences of each corresponding pixel between two image blocks of the same size. Mathematically, for two image blocks A and B, their size is h×w.

[0112] Where A(i,j) and B(i,j) represent the pixel values ​​of image blocks A and B at position (i,j) respectively.

[0113]

[0114] aSAD is a variation of SAD, which is obtained by dividing the SAD value by the total number of pixels in the image block, thus obtaining an average difference value.

[0115] aRMSE is short for "average root mean square error", which measures the global error between two images. First, the sum of the squares of each corresponding pixel difference between the two image patches is calculated, and then the square root of their average is taken. Mathematically, for two image patches A and B, whose size is h×w, RMSE is defined as:

[0116]

[0117] The smaller the SAD, the better the result. aSAD and aRMSE can reflect the overall performance of the algorithm.

[0118] Table 1. Comparison results of different algorithms (×10 -2 )

[0119]

[0120] In order to more intuitively display the unmixing effect, the abundance of the end members estimated by each algorithm and the real abundance map are displayed. Compared with other comparative methods, the abundance image obtained in this embodiment better fits the real situation, which proves the effectiveness of this embodiment for abundance estimation.

[0121] Example 3

[0122] This embodiment uses a real data set named Samson to further verify the beneficial effects of the present invention.

[0123] The Samson dataset occupies a prominent position in the field of hyperspectral image unmixing and is one of the widely used benchmark datasets. The original image of the dataset consists of 952×952 pixels and covers a wide wavelength range from 401 nanometers to 889 nanometers, with a total of 156 bands, and the spectral resolution between each band is fine 3.13 nanometers.

[0124] This embodiment uses a sub-image area of ​​95×95 pixels cut out from the complete image.

[0125] In this dataset, this embodiment focuses on three basic and representative substances, which are often regarded as pure endmembers in remote sensing and spectral analysis, namely trees, soil, and water.

[0126] These three pure endmembers were chosen not only because of their widespread distribution on the Earth's surface, but also because of their significant differences in spectral characteristics, which provide a natural distinction for the unmixing and analysis of spectral data.

[0127] In this dataset, the encoder learning rate of the model is set to 9.5×(10)^(-2), and the decoder learning rate is 4.5×(10)^(-2); the number of iterations is 300; for the loss function, beta1 is set to 7×(10)^(-7); beta2 is set to 1×(10)^(-5); beta3 is set to 1×(10)^(-1); beta4 is set to 2×(10)^(-1); beta5 is set to 1×(10)^(-5); the spatial weight factor incorporates the information in the 7x7 neighborhood around each pixel during evaluation, and imposes corresponding constraints accordingly.

[0128] Table 2. Comparison results of different algorithms on the Samson dataset (×10 -2 )

[0129]

[0130] Through experimental verification, the present invention has shown outstanding advantages in two key evaluation indicators, aSAD (average spectral angular distance) and aRMSE (average root mean square error).

[0131] This achievement not only demonstrates the superiority of the algorithm itself, but also strongly proves the key role played by the spatial weighting factor in improving the unmixing accuracy.

[0132] As can be seen from Table 2, in the extraction of the three end members, the present invention has achieved the optimal values ​​among all the methods. In particular, in the extraction of the soil end member, the value is far less than the suboptimal value.

[0133] aSAD and aRMSE also achieved the best state among these methods, and the numerical gap with the suboptimal methods was obviously widened.

[0134] The reconstructed predicted pixel data is analyzed by this method to complete the recognition and classification.

[0135] In summary, the accuracy of the present invention in demixing can be proved, and it can also be shown that the spatial weighting factor has an obvious effect in end member extraction, which greatly improves the accuracy.

[0136] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions and substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A hyperspectral unmixing method based on a cascaded autoencoder with spatial constraints, characterized in that: The following steps are involved: S1. Obtaining hyperspectral images of ground objects; S2, construct a spatially constrained cascaded autoencoder unmixing model; S3, using the constructed spatially constrained cascade autoencoder unmixing model to unmix the acquired hyperspectral image of the ground objects, and obtain the ground object abundance corresponding to each ground object; The spatially constrained cascaded autoencoder unmixing model adopts a layer-by-layer training method, that is, the first layer of autoencoder is trained first, and then the output of the trained first layer of autoencoder is used as the input of the second layer of autoencoder, and so on, iteratively performed until the termination condition of the iteration is reached; The autoencoder specifically includes a preprocessing stage module and a spectral unmixing stage module; The preprocessing stage module includes at least one encoder, and the spectral unmixing stage module includes a fusion layer and a decoder; The encoder includes a multi-scale feature extraction module, a BN layer and an activation function layer. The multi-scale feature extraction module is a Conv layer, and the activation function layer adopts a LeakyReLU activation function; the multi-scale feature extraction module includes a small convolution kernel for focusing on local feature extraction and a large convolution kernel for capturing global context information.

2. The hyperspectral unmixing method based on the spatially constrained cascaded autoencoder according to claim 1, characterized in that: A spatial constraint term is introduced into the autoencoder, and the spatial constraint term is used to take the information of adjacent pixels into consideration to optimize the parameters of the autoencoder.

3. The hyperspectral unmixing method based on the spatially constrained cascaded autoencoder according to claim 2, characterized in that: When the spatially constrained cascaded autoencoder unmixing model uses a deep learning-based spatial regularization unmixing algorithm to process hyperspectral images of ground objects: First, quantify the similarity between adjacent pixels; use Euclidean distance as an indicator to measure the similarity of adjacent pixels, and describe the similarity between pixels by calculating the Euclidean distance between pixels; Then, based on the Euclidean distance information between pixels, the neighborhood weight ∈ ij , neighborhood weight ∈ ij It is expressed as: Among them, im(·) represents the pixel x i and x j The Euclidean distance between pixels x i and pixel x j The closer the position, the smaller the Euclidean distance, and the weight ∈ ij The bigger; (a, b) and (c, d) are pixel x respectively. i and pixel x j The spatial coordinates of , then the Euclidean distance is expressed as Use neighborhood weights to mine spatial correlations as follows: In the formula, represents the abundance element neighborhood x ij gather; Construct spatial weighting factors: Introduce the constructed spatial weights into the model:

4. The hyperspectral unmixing method based on the spatially constrained cascaded autoencoder according to claim 1, characterized in that: The spatially constrained cascaded autoencoder unmixing model sets a loss function, which is used to evaluate the accuracy of spectral unmixing and the effective utilization of spatial information; an optimization algorithm is set to minimize the loss function, and the model parameters of the spatially constrained cascaded autoencoder unmixing model are iteratively updated until convergence to an optimal solution, thereby ensuring that the spatially constrained cascaded autoencoder unmixing model can optimally fuse spectral and spatial information.

5. The hyperspectral unmixing method based on the spatially constrained cascaded autoencoder according to claim 4 is characterized in that: The loss function is constructed as: Among them, L rec Represents the true pixel vector Y and the predicted pixel vector and The average value of the sum of squares of the differences between , Y represents the true pixel vector, and denotes the predicted pixel vector, and N represents the total number of pixels.

6. The hyperspectral unmixing method based on spatially constrained cascaded autoencoders according to claim 1, characterized in that: The autoencoder also introduces a structured sparsity loss function: L sparse =||Y|| 2，1 .

7. The hyperspectral unmixing method based on the spatially constrained cascaded autoencoder according to claim 1, characterized in that: The fusion layer is a fusion layer.

8. The hyperspectral unmixing method based on spatially constrained cascaded autoencoders according to claim 1, characterized in that: The training process of the spatially constrained cascaded autoencoder unmixing model uses six encoders to work together. When the encoders perform multi-scale feature extraction, two encoders are selected from three encoders with convolution kernels of different scales. One of the encoders uses a small 1x1 convolution kernel to focus on local feature extraction, and the other encoder uses a large 5x5 or 9x9 convolution kernel to capture global context information.

Images