Hyperspectral remote sensing image unmixing method and system

Abstract

The application provides a hyperspectral remote sensing image unmixing method and system, which can effectively improve the unmixing efficiency and effect of the hyperspectral remote sensing image, and the method comprises the following steps: step 1, obtaining a training sample; step 2, first, constructing a Ghost module, decomposing a 3D convolution layer into three-stage convolution operations, adopting ordinary convolution in the first stage to generate an intermediate output feature map; performing inexpensive operation on each intrinsic feature map of the intermediate output feature map in the second stage to obtain a ghost feature map; and combining the intermediate output feature map and the ghost feature map in the third stage; then, constructing a Ghost attention module, and further building a Ghost attention unmixing network; step 3, training the Ghost attention unmixing network by using the training sample; step 4, obtaining a hyperspectral image to be unixed; and step 5, performing unmixing by using the trained Ghost attention unmixing network to obtain a hyperspectral remote sensing image unmixing result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing image processing, and particularly relates to a hyperspectral remote sensing image unmixing method and system. BACKGROUND

[0002] Hyperspectral remote sensing images have a wide range of applications in geology, agriculture, and the environment. They have the characteristics of high dimensionality, high spectral resolution, and high spatial resolution, and can extract rich information of target objects (Document 1). However, in practical applications, due to the mutual interference and overlap between spectral data, it is difficult to accurately reflect the composition and content of the surface material of the obtained remote sensing images. Therefore, unmixing technology has always been an important problem in the processing of hyperspectral remote sensing images (Document 2).

[0003] Deep learning methods are widely used in multiple fields due to their strong learning feature capabilities, such as object detection and image classification (Document 3). In particular, convolutional neural networks are one of the most popular learning tools in the field of deep learning, which can extract feature information of target objects in actual remote sensing images (Document 4). However, current hyperspectral unmixing methods based on convolutional neural networks usually require a large number of training parameters and have high hardware requirements, lacking unmixing efficiency.

[0004] Zhao J, Hu L, Huang L, et al. MSRA-G: Combination of multi-scale residual attention network and generative adversarial networks for hyperspectral image classification [J]. Engineering Applications of Artificial Intelligence, 2023, 121: 106017.

[0005] Heylen R, Parente M, Gader P. A review of nonlinear hyperspectral unmixing methods [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2014, 7(6): 1844-1868.

[0006] Zhang X, Sun Y, Zhang J, et al. Hyperspectral unmixing via deep convolutional neural networks [J]. IEEE Geoscience and Remote Sensing Letters, 2018, 15(11): 1755-1759.

[0007] Zhao J, Hu L, Dong Y, et al. A combination method of stacked autoencoder and 3D deep residual network for hyperspectral image classification [J]. International Journal of Applied Earth Observation and Geoinformation, 2021, 102: 102459. SUMMARY

[0008] The present application is to solve the above problems, and aims to provide a hyperspectral remote sensing image unmixing method and system, which can effectively improve the unmixing efficiency and effect of hyperspectral remote sensing image.

[0009] In order to achieve the above purpose, the present application adopts the following scheme:

[0010] <Method>

[0011] The present application provides a hyperspectral remote sensing image unmixing method, characterized in that it comprises the following steps:

[0012] Step 1, obtaining hyperspectral image as training sample;

[0013] Step 2, building Ghost attention unmixing network:

[0014] Step 2-1, constructing Ghost module: decomposing 3D convolution layer into three-stage convolution operation, first stage, using ordinary convolution kernel size 1 f'∈ k×k×B×n′ to generate intermediate output feature map Y'∈ H′×W′×B×n′ , H' and W' represent the height and width of the intermediate output data respectively, B represents the number of spectral bands, the kernel size of the three-dimensional convolution filter f is k x k, and n' represents the number of channels of the intermediate output feature map; the second stage, using convolution kernel size 3, cheap operation is performed on each intrinsic feature map in Y' to obtain g ghost feature maps Y GThe third stage generates intermediate output feature maps Y' and ghost feature maps Y G The joint operation is performed, and the number of channels is n = n' + g.

[0015] Y = X * f + b

[0016] Y' = X * f' + b'

[0017]

[0018]

[0019] In the formula, cat[·] represents a joint operation, Y' represents that the output feature map has n channels, and y' represents the i-th intrinsic feature map in Y' i Φ represents the i-th intrinsic feature map in Y', and Φ i,j represents that the j-th linear operation is performed on the i-th intrinsic feature map to obtain the j-th ghost feature map; y' represents the i-th ghost feature map in Y' i has at least one ghost feature map.

[0020] Step 2-2, constructing a Ghost attention unmixing network: a Ghost module is connected with a normalized attention module to form a Ghost attention module, and then a Ghost attention unmixing network is built based on the Ghost attention module; the Ghost attention unmixing network includes three convolution modules C1-C3, two Ghost attention modules, two residual modules, and a full connection module.

[0021] Step 3, training the Ghost attention unmixing network by using training samples.

[0022] Step 4, obtaining a hyperspectral image to be unmixed.

[0023] Step 5, using the trained Ghost attention unmixing network to unmix the hyperspectral image to be unmixed, to obtain a hyperspectral remote sensing image unmixing result.

[0024] Preferably, in step 2-2 of the hyperspectral remote sensing image unmixing method, the Ghost attention unmixing network is set as follows:

[0025] The input of the first convolution module C1 is the preprocessed hyperspectral image, the convolution kernel size is 3, and the step size is 1; the input of the first Ghost attention module is the input feature map F1 of C1.

[0026] The input of the first residual module is the feature map F1, which is used to connect the output of C1 and the output of the first Ghost attention module.

[0027] The second convolution module C2 includes one average pooling layer and one convolution layer, and the input is the result F3 of the residual connection of the output feature map F1 of the convolution module C1 and the output F2 of the first Ghost attention module, the kernel size of the one average pooling layer and the one convolution layer is 2 and 1 respectively, and the step size is 1;

[0028] The input of the second Ghost attention module is the input feature map F4 of the second convolution module C2;

[0029] The input of the second residual module is the feature map F4, which is used to connect the output of the second convolution module C2 and the output of the second Ghost attention module;

[0030] The third convolution module C3 is composed of one average pooling layer, the kernel size is 2, the step size is 1, and the input is the result F6 of the residual connection of the output feature map F4 of the second convolution module C2 and the output F5 of the second Ghost attention module;

[0031] The full connection module includes an unfolding layer and three full connection layers, and the input is the output feature map F7 of the third convolution module C3;

[0032] After convolution, a normalization layer and an activation layer are sequentially added; in the full connection module, a Dropout layer is connected after the first two full connection layers, and the last full connection layer is provided with a Softmax function.

[0033] Preferably, the hyperspectral remote sensing image unmixing method provided by the application, in step 2-2, the output feature M of the normalization attention module is:

[0034]

[0035] In the formula, sigmoid() represents the sigmoid function operation, W γ = γ i / ∑ j＝0 γ j represents the obtained weight, μ p and σ p are the average value and the standard deviation of the small batch p respectively, γ and β represent the c scale and the transfer factor respectively, which are trainable transformation parameters; B in represents the input feature.

[0036] Preferably, the hyperspectral remote sensing image unmixing method provided by the application, in step 1, samples are obtained in proportion from each type of ground object of the original hyperspectral remote sensing image; the samples are preprocessed by normalization operation to obtain the preprocessed hyperspectral image X H×W×B As a training sample, H, W and B represent height, width and spectral band number respectively.

[0037] Preferably, the hyperspectral remote sensing image unmixing method provided by the present application updates the network parameters through a back propagation mechanism in step 3, adopts a cross-entropy loss function, uses Adam as a model training optimizer, sets the initial value of the learning rate to 0.001, and sets the number of iterations to 500; the root mean square error and the average root mean square error are used as model performance evaluation indexes.

[0038] Preferably, the hyperspectral remote sensing image unmixing method provided by the present application, in step 4, obtains the hyperspectral image to be unmixing and performs the same preprocessing as step 1; in step 5, the preprocessed hyperspectral image to be unmixing is input into the trained Ghost attention unmixing network to obtain the hyperspectral remote sensing image unmixing result.

[0039] <system>

[0040] Further, the present application also provides a hyperspectral remote sensing image unmixing system capable of automatically implementing the above <method>, characterized in that it comprises:

[0041] a sample acquisition unit for acquiring a hyperspectral image as a training sample;

[0042] a unmixing network building unit for first building a Ghost module and then building a Ghost attention unmixing network based on the Ghost module; the Ghost module is built by decomposing a 3D convolutional layer into three-stage convolutional operations, in the first stage, a normal convolution f'∈ k×k×B×n′ to generate an intermediate output feature map Y'∈ H′×W′×B×n′ , H' and W' represent the height and width of the intermediate output data respectively, B represents the number of spectral bands, the kernel size of the three-dimensional convolutional filter f is k×k, and n' represents the number of channels of the intermediate output feature map; in the second stage, a cheap operation is performed on each intrinsic feature map in Y' to obtain g ghost feature maps Y G ; in the third stage, the intermediate output feature map Y' generated by the normal convolution and the ghost feature map Y G are combined, and the number of channels is n=n'+g;

[0043] Y'=X*f'+b'

[0044]

[0045]

[0046] In the formula, cat[·] is a joint operation, represents that the output feature map has n channels, y' i is the i-th intrinsic feature map in Y', Φ i,j represents that the j-th linear operation is performed on the i-th intrinsic feature map to obtain the j-th ghost feature map; y'i has at least one ghost feature map; a ghost attention unmixing network is built: a ghost module is connected with a normalized attention module to form a ghost attention module, and then a ghost attention unmixing network is built based on the ghost attention module; the ghost attention unmixing network includes three convolution modules C1-C3, two ghost attention modules, two residual modules and a full connection module;

[0047] a training unit that trains the ghost attention unmixing network by using training samples;

[0048] a to-be-solved image acquisition unit that acquires hyperspectral images to be unmixing;

[0049] an unmixing unit that unmixes the hyperspectral images to be unmixing by using the trained ghost attention unmixing network to obtain a hyperspectral remote sensing image unmixing result;

[0050] a control unit that is in communication with the sample acquisition unit, the unmixing network building unit, the training unit, the to-be-solved image acquisition unit and the unmixing unit, and controls the operation of the units.

[0051] Preferably, the hyperspectral remote sensing image unmixing system provided by the application can further include an input display unit that is in communication with the control unit, allows an operator to input operation instructions, and displays corresponding information according to the control instructions.

[0052] Preferably, in the unmixing network building unit, the ghost attention unmixing network is:

[0053] the first convolution module C1 has a pre-processed hyperspectral image as an input, a convolution kernel size of 3 and a step size of 1; and the first ghost attention module has an input feature map F1 of C1 as an input;

[0054] the first residual module has the feature map F1 as an input, and is used to connect the output of C1 with the output of the first ghost attention module;

[0055] the second convolution module C2 includes one layer of average pooling and one layer of convolution, has the output feature map F1 of the convolution module C1 and the output F2 of the first ghost attention module as an input, and has the result F3 of residual connection, a kernel size of one layer of average pooling and one layer of convolution is 2 and 1 respectively, and a step size is 1;

[0056] the second ghost attention module has an input feature map F4 of the second convolution module C2 as an input;

[0057] The input of the second residual module is the feature map F4, which is used to connect the output of the second convolution module C2 and the output of the second Ghost attention module;

[0058] The third convolution module C3 is composed of one layer of average pooling, the kernel size is 2, the step size is 1, and the input is the output feature map F4 of the second convolution module C2 and the output F5 of the second Ghost attention module, which are connected in residual connection to obtain the result F6;

[0059] The full connection module includes an unfolding layer and three full connection layers, and the input is the output feature map F7 of the third convolution module C3;

[0060] After convolution, a normalization layer and an activation layer are sequentially added; in the full connection module, a Dropout layer is connected after the first two full connection layers, and the last full connection layer is provided with a Softmax function.

[0061] Preferably, the hyperspectral remote sensing image unmixing system provided by the application, in the unmixing network building part, the output feature M of the normalization attention module is:

[0062]

[0063] In the formula, sigmoid() represents the sigmoid function operation, W γ = gamma i / sum j＝0 gamma j represents the obtained weight, mu p and sigma p are the average value and standard deviation of the small batch p respectively, gamma and beta represent the c scale and the transfer factor respectively, which are the trainable transformation parameters; B in represents the input feature.

[0064] Effects of the application

[0065] The application relates to a hyperspectral remote sensing image unmixing method and system, which comprises the following steps: firstly, a Ghost module is constructed, a 3D convolution layer is decomposed into three-stage convolution operations, in the first stage, a normal convolution with a convolution kernel size of 1 is adopted to generate an intermediate output feature map, in the second stage, a convolution kernel size of 3 is adopted to perform inexpensive operations on each intrinsic feature map in the intermediate output feature map to obtain g ghost feature maps, further, the Ghost module and a normalization attention module are connected in series to form a Ghost attention module, then a special Ghost attention unmixing network is built based on the Ghost attention module, the Ghost attention unmixing network comprises three convolution modules C1-C3, two Ghost attention modules, two residual modules and a full connection module, so that a special unmixing network architecture is established, the hyperspectral image to be unmixing is unmixing by using the unmixing network architecture, the learning performance of the convolution in the prior art can be bettered by using more inexpensive operations, the training parameter amount is greatly reduced, key features are highlighted, the values of meaningless channels are suppressed, the information flow in the network model is fully utilized, the defects of low unmixing efficiency and unsatisfactory effect of the hyperspectral remote sensing image unmixing under the condition of ground object spectral mixing are solved, and the application can be widely applied to the hyperspectral data mixed pixel decomposition application in the fields of environmental monitoring, precision agriculture and food safety.

[0066] In addition, in view of the complex data structure of the hyperspectral image, the balance between the model performance and the unmixing efficiency is comprehensively considered, a normalization BN layer, a Dropout layer and a pooling layer are introduced, and ReLU is used as the activation function of the neural network, so that the model converges faster. BRIEF DESCRIPTION OF DRAWINGS

[0067] Fig. 1 A flowchart of the hyperspectral remote sensing image unmixing method according to the embodiment of the application is shown in the figure.

[0068] Fig. 2 A Ghost attention unmixing network structure diagram according to the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0069] The specific implementation of the hyperspectral remote sensing image unmixing method and system according to the application will be described in detail below with reference to the accompanying drawings.

[0070] <Embodiment I>

[0071] As shown in the figure, the hyperspectral remote sensing image unmixing method according to the embodiment I comprises the following steps: Figs. 1-2

[0072] Step 1: Obtain a training sample

[0073] ​Training samples are obtained proportionally from each land cover class in the original hyperspectral remote sensing image; the training samples are preprocessed by normalization to shorten the model processing time, resulting in the preprocessed hyperspectral image X∈ H×W×B , where H, W and B represent height, width and number of spectral bands, respectively.

[0074] Step 2, build the Ghost attention unmixing network:

[0075] Step 2-1, Build the Ghost module:

[0076] The process of processing raw hyperspectral data X using a traditional 3D convolutional layer can be simply represented as:

[0077] Y = X * f + b

[0078] In the formula, f∈ k×k×B×n This represents a three-dimensional convolutional filter, where b is the bias.

[0079] This invention decomposes the original 3D convolutional layer into three stages of convolutional operations. Specifically:

[0080] The 3D convolutional layer is decomposed into three stages of convolution operations. In the first stage, a regular convolution f′∈ is used with a kernel size of 1. k×k×B×n′ Generate intermediate output feature map Y′∈ H′×W′×B×n′ In the second stage, n′≤n, the total number of channels is strictly controlled; a convolution kernel of size 3 is used to perform inexpensive operations on each intrinsic feature map in Y′ to obtain g ghost feature maps Y. G The third stage attempts to extract enough feature maps; the intermediate output feature map Y′ generated by ordinary convolution and the ghost feature map Y′ are then combined. G The number of channels is n = n′ + g;

[0081] Y′=X*f′+b′

[0082]

[0083]

[0084] In the formula, cat[·] represents a union operation. The output feature map represents n channels, y′ i Φ is the i-th intrinsic feature map in Y′. i,j This indicates that the j-th linear operation is performed on the i-th intrinsic feature map to obtain the j-th ghost feature map; y′ i It has one or more ghost feature maps; Y′∈ H′×W′×B×nThe output feature map representing the three-dimensional convolutional layer has n channels, in addition, H' and W' represent the height and width of the output data, respectively, and the kernel size of the three-dimensional convolutional filter f is k x k.

[0085] Step 2-2, constructing a Ghost attention unmixing network: connecting a Ghost module and a normalized attention module at the head and tail to form a Ghost attention module, and then building a Ghost attention unmixing network based on the Ghost attention module; the Ghost attention unmixing network is composed of three convolutional modules C1-C3, two Ghost attention modules, two residual modules, and a fully connected module.

[0086] Specifically, the input of the first convolutional module C1 of the Ghost attention unmixing network is the preprocessed hyperspectral image, the convolution kernel size is 3, and the step size is 1; the input of the first Ghost attention module of the Ghost attention unmixing network is the input feature map F1 of C1; the input of the first residual module of the Ghost attention unmixing network is the feature map F1, which is used to connect the output of C1 and the output of the first Ghost attention module, to ensure the transmission of the information flow of the encoding module; the second convolutional module C2 of the Ghost attention unmixing network is composed of an average pooling layer and a convolutional layer, the input is the result F3 of the residual connection of the output feature map F1 of the convolutional module C1 and the output F2 of the first Ghost attention module, the kernel size of the average pooling layer and the convolutional layer is 2 and 1 respectively, and the step size is 1; the input of the second Ghost attention module of the Ghost attention unmixing network is the input feature map F4 of C2; the input of the second residual module of the Ghost attention unmixing network is the feature map F4, which is used to connect the output of C2 and the output of the second Ghost attention module; the third convolutional module C3 of the Ghost attention unmixing network is composed of an average pooling layer, the kernel size is 2, the step size is 1, and the input is the result F6 of the residual connection of the output feature map F4 of the convolutional module C2 and the output F5 of the second Ghost attention module; the fully connected module of the Ghost attention unmixing network is composed of an unfolding layer and three fully connected layers, and the input is the output feature map F7 of the convolutional module C3; a normalization layer and an activation layer are added after convolution in sequence; a Dropout layer is connected after the first two fully connected layers in the fully connected module, and the last fully connected layer is attached with a Softmax function to obtain the abundance estimation.

[0087] The output feature M of the normalized attention module is:

[0088]

[0089] In the formula, sigmoid() represents the sigmoid function operation, W γ = γi / ∑ j＝0 γ j represent the obtained weights, μ p and σ p are the mean and standard deviation of the mini-batch p, respectively, and γ and β represent the c-scale and shift factor, respectively, are trainable transformation parameters; B in denotes the input features.

[0090] Step 3, training the Ghost attention unmixing network with the training samples:

[0091] The network parameters are updated through the back propagation mechanism. Specifically, the cross-entropy loss function is used, Adam is used as the model training optimizer, the initial value of the learning rate is set to 0.001, and the iteration number is 500. In addition, the root mean square error and the average root mean square error are used as the model performance evaluation indexes.

[0092] Step 4, obtaining the hyperspectral image to be unmixing:

[0093] Obtain the hyperspectral image to be unmixing and perform the same preprocessing as in step 1.

[0094] Step 5, using the trained Ghost attention unmixing network to unmix the hyperspectral image to be unmixing, to obtain the hyperspectral remote sensing image unmixing result.

[0095] Input the preprocessed hyperspectral image to be unmixing into the trained Ghost attention unmixing network to obtain the hyperspectral remote sensing image unmixing result.

[0096] <Embodiment Two>

[0097] The embodiment two provides a system for unmixing hyperspectral remote sensing images based on the above method of the application. The system comprises a sample acquisition unit, an unmixing network building unit, a training unit, an image to be solved acquisition unit, an unmixing unit, an input display unit and a control unit.

[0098] The sample acquisition unit can perform the content described in step 1 above to acquire the training samples.

[0099] The unmixing network building unit can perform the content described in step 2 above to build the Ghost attention unmixing network; and construct the Ghost module.

[0100] The training unit can perform the content described in step 3 above to train the Ghost attention unmixing network with the training samples.

[0101] The image to be solved acquisition unit can perform the content described in step 4 above to acquire the hyperspectral image to be unmixing.

[0102] The unmixing unit can perform the content described in step 5 above, and uses the trained Ghost attention unmixing network to unmix the hyperspectral image to be unmixed to obtain a hyperspectral remote sensing image unmixing result.

[0103] The input display unit can allow an operator to input an operation instruction, and display corresponding information according to the control instruction. Specifically, the input display unit can statically or dynamically display the input, output data and processing process of the sample acquisition unit, the unmixing network building unit, the training unit, the image to be solved acquisition unit and the unmixing unit in the form of a data table or an image according to the control instruction.

[0104] The control unit is in communication with the sample acquisition unit, the unmixing network building unit, the training unit, the image to be solved acquisition unit, the unmixing unit and the input display unit, and controls the operation of them.

[0105] The above embodiments are only illustrative of the technical solutions of the present application. The hyperspectral remote sensing image unmixing method and system involved in the present application are not limited to the content described in the above embodiments, but are subject to the scope defined in the claims. Any modification, supplement or equivalent replacement made by a person skilled in the art on the basis of the embodiments is within the scope claimed by the claims of the present application.

Claims

1. A method for unmixing hyperspectral remote sensing imagery, characterized in that, The method comprises the following steps: Step 1, obtaining a hyperspectral image as a training sample; Step 2, building a Ghost attention unmixing network: Step 2-1, constructing Ghost module: decompose the 3D convolutional layer into three-stage convolutional operations, the first stage, using ordinary convolution with kernel size 1 produce intermediate output feature maps , and represent the height and width of the intermediate output data, respectively, B represent the number of spectral bands, the kernel size of the three-dimensional convolutional filter f , n’ represent the number of channels of the intermediate output feature maps; the second stage, using convolution with kernel size 3, cheaply operates on each intrinsic feature map in to obtain g ghost feature maps ;​ The third stage combines the intermediate output feature maps generated by the ordinary convolution and the ghost feature maps with a channel number of ; wherein is jointly operated, represents an output feature map with n a number of channels, is an intrinsic feature map at the i th out of i the j th j ghost feature map; has at least one ghost feature map; b is a bias; Step 2-2, constructing the Ghost attention unmixing network: a Ghost module is connected with a normalization attention module to form a Ghost attention module, and then the Ghost attention unmixing network is built based on the Ghost attention module; the Ghost attention unmixing network comprises three convolution modules C1-C3, two Ghost attention modules, two residual modules and a full connection module; Step 3, training the Ghost attention unmixing network by using the training sample; Step 4, obtaining a hyperspectral image to be unmixed; Step 5, using the trained Ghost attention unmixing network to unmix the hyperspectral image to be unmixed, and obtaining a hyperspectral remote sensing image unmixing result; In step 2-2, the Ghost attention unmixing network is set as follows: The first convolution module C1 takes the preprocessed hyperspectral image as input, and the convolution kernel size is 3 and the step size is 1; the first Ghost attention module takes the input feature map F1 of C1 as input; The first residual module takes the feature map F1 as input, and is used to connect the output of C1 with the output of the first Ghost attention module; The second convolution module C2 comprises an average pooling layer and a convolution layer, and takes the result F3 of the residual connection between the output feature map F1 of the convolution module C1 and the output F2 of the first Ghost attention module as input; the kernel size of the average pooling layer and the convolution layer is 2 and 1 respectively, and the step size is 1; The second Ghost attention module takes the input feature map F4 of the second convolution module C2 as input; The second residual module takes the feature map F4 as input, and is used to connect the output of the second convolution module C2 with the output of the second Ghost attention module; The third convolution module C3 is composed of an average pooling layer, and takes the result F6 of the residual connection between the output feature map F4 of the second convolution module C2 and the output F5 of the second Ghost attention module as input; the kernel size is 2 and the step size is 1; The full connection module comprises an unfolding layer and three full connection layers, and takes the output feature map F7 of the third convolution module C3 as input; Normalization layers and activation layers are sequentially added after convolution; in the full connection module, the last full connection layer is provided with a Softmax function.

2. The hyperspectral remote sensing image unmixing method according to claim 1, wherein in step 2-2, the output feature M of the normalization attention module is: wherein 3. The hyperspectral remote sensing image unmixing method according to claim 1, wherein in step 2-2, the output feature M of the normalization attention module is: where sigmoid() denotes the sigmoid function operation, represent the obtained weights, and are the mean and standard deviation of the mini-batch p respectively, and denote the c-scale and the shift factor respectively, are trainable transformation parameters; B in denotes the input features.

4. The hyperspectral remote sensing image unmixing method according to claim 1, wherein in step 2-2, the output feature M of the normalization attention module is: wherein, In step 1, samples are proportionally obtained from each type of ground object in the original hyperspectral remote sensing image; the samples are preprocessed by normalization operation to obtain the preprocessed hyperspectral image As training samples, H 、 W and B respectively represent the height, width and spectral band number. In step 3, the network parameters are updated through a back propagation mechanism, a cross-entropy loss function is used, Adam is used as a model training optimizer, the initial value of the learning rate is set to 0.001, the iteration number is 500, and the root mean square error and the average root mean square error are used as model performance evaluation indexes. wherein ​ 5. The hyperspectral remote sensing image unmixing method according to claim 3, characterized in that: wherein In step 4, the hyperspectral image to be unmixing is obtained and preprocessed in the same way as step 1; in step 5, the preprocessed hyperspectral image to be unmixing is input into the trained Ghost attention unmixing network to obtain the hyperspectral remote sensing image unmixing result.

6. A hyperspectral remote sensing image unmixing system, characterized in that, It comprises: A sample acquisition unit acquires a hyperspectral image as a training sample; The unmixing network building unit first constructs a Ghost module, and then constructs a Ghost attention unmixing network based on the Ghost module; the Ghost module is constructed: the 3D convolutional layer is decomposed into three-stage convolutional operations, in the first stage, an ordinary convolutional kernel with a size of 1 is adopted to generate an intermediate output feature map , , respectively representing the height and width of the intermediate output data, B , the kernel size of the three-dimensional convolutional filter f is , n’ , representing the number of channels of the intermediate output feature map; in the second stage, a convolutional kernel with a size of 3 is adopted to perform inexpensive operations on each intrinsic feature map in to obtain g ghost feature maps ; in the third stage, the intermediate output feature map generated by the ordinary convolution and the ghost feature map are combined, and the number of channels is ;​ In the formula, For joint operations, The output feature map represents... n Number of channels, Is The Middle i The inherent feature map of each, Indicates the first i On the first inherent feature map, perform the... j The nth linear operation is obtained to obtain the nth j One ghost feature map; It has at least one ghost feature map; constructs a ghost attention demixing network: the ghost module and the normalized attention module are concatenated to form a ghost attention module, and then the ghost attention demixing network is built based on the ghost attention module; the ghost attention demixing network includes three convolutional modules C1~C3, two ghost attention modules, two residual modules and one fully connected module; A training unit trains the Ghost attention unmixing network using the training sample; An image to be solved acquisition unit acquires a hyperspectral image to be unmixing; An unmixing unit uses the trained Ghost attention unmixing network to unmix the hyperspectral image to be unmixing, and obtains a hyperspectral remote sensing image unmixing result; A control unit is in communication with the sample acquisition unit, the unmixing network building unit, the training unit, the image to be solved acquisition unit, and the unmixing unit, and controls their operation; In the unmixing network building unit, the Ghost attention unmixing network is set as: The input of the first convolution module C1 is the preprocessed hyperspectral image, the convolution kernel size is 3, and the step size is 1; the input of the first Ghost attention module is the input feature map F1 of C1; The input of the first residual module is the feature map F1, which is used to connect the output of C1 and the output of the first Ghost attention module; The second convolution module C2 includes one layer of average pooling and one layer of convolution layer, the input is the result F3 of the residual connection of the output feature map F1 of the convolution module C1 and the output F2 of the first Ghost attention module, and the kernel size of one layer of average pooling and one layer of convolution layer is 2 and 1 respectively, and the step size is 1; The input of the second Ghost attention module is the input feature map F4 of the second convolution module C2; The input of the second residual module is the feature map F4, which is used to connect the output of the second convolution module C2 and the output of the second Ghost attention module; The third convolution module C3 is composed of one layer of average pooling, the kernel size is 2, the step size is 1, and the input is the result F6 of the residual connection of the output feature map F4 of the second convolution module C2 and the output F5 of the second Ghost attention module; The full connection module includes an unfolding layer and three full connection layers, and the input is the output feature map F7 of the third convolution module C3; After convolution, a normalization layer and an activation layer are added in sequence; in the full connection module, a Dropout layer is connected after the first two full connections, and the last full connection layer is provided with a Softmax function.

7. The hyperspectral remote sensing imagery unmixing system of claim 6, wherein, It also includes: An input display unit is in communication with the control unit, allowing an operator to input operation instructions and displaying corresponding information according to the control instructions.

8. The hyperspectral remote sensing image unmixing system according to claim 6, characterized in that: wherein In the unmixing network building unit, the output feature M of the normalization attention module is: where sigmoid() denotes a sigmoid function operation, denotes the obtained weights, and are the mean and standard deviation of the mini-batch p respectively, and denote the c-scale and the shift factor, respectively, are trainable transformation parameters; B in denotes the input features.

Images