A hyperspectral remote sensing image change detection method
By combining the ConvLSTM model with CNN and LSTM, and utilizing Siamese convolutional neural networks and attention mechanisms, the problems of high dimensionality and information redundancy in hyperspectral remote sensing image change detection are solved, and high-precision change detection is achieved.
Patent Information
- Application Number
- CN202210026016.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2021-12-31
- Filing Date
- 2022-01-11
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-01-11
AI Technical Summary
Existing methods for detecting changes in hyperspectral remote sensing images face challenges in terms of high-dimensional features and redundancy of information in adjacent bands. Traditional methods have limited effectiveness in detecting changes in hyperspectral images and struggle to effectively identify subtle changes.
We employ a ConvLSTM model that combines CNN and LSTM, using a Siamese convolutional neural network for feature extraction and dimensionality reduction, and combining an attention mechanism for feature fusion to construct a multi-scale temporal-spatial-spectral feature representation. This allows us to adaptively learn changes in land cover types and spurious changes under imaging conditions.
It improves the accuracy of hyperspectral image change detection, enhances robustness to real change pixels, reduces the impact of spurious change pixels, and improves classification accuracy.
Smart Images

Figure CN114359735B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a hyperspectral remote sensing image change detection method, in particular to a hyperspectral remote sensing image change detection method which introduces time dependency and space-spectral features. Background Art
[0002] With the rapid development of remote sensing technology, more and more remote sensing sensors are used for earth observation. Dynamic observation of the earth's surface has produced a large number of available remote sensing images, which can be used to monitor land cover and land use changes. At present, remote sensing change detection has been widely used in urban expansion monitoring, disaster monitoring and other aspects.
[0003] Hyperspectral images can provide rich spectral and spatial information, and multi-temporal hyperspectral remote sensing images are of great significance for revealing subtle changes in surface objects. However, most of the current change detection methods for hyperspectral remote sensing images are derived from classic single-band or multi-spectral change detection methods, and their performance is limited in hyperspectral remote sensing images with high-dimensional features. In recent years, deep neural networks have achieved great success in various computer vision and remote sensing applications, and they also have good application prospects for change detection. At present, multispectral / hyperspectral change detection based on deep learning mainly uses two neural network structures, CNN and RNN, or combines CNN with RNN to extract deep features for change detection.
[0004] Most change detection methods are proposed for multispectral and synthetic aperture radar images. Hyperspectral images are not adaptable enough. The spatial resolution of hyperspectral images is usually low, but the spectral resolution is very high. Although hyperspectral images with rich spectral information are considered to be able to identify subtle changes. Traditional change detection methods for single-band or multispectral images are rarely used in hyperspectral image change detection tasks, mainly due to their high-dimensional characteristics. The information redundancy between adjacent bands and the high dimensionality of images bring additional difficulties to the development of change detection algorithms for hyperspectral images. For the classification of hyperspectral images, the classification accuracy usually increases and then decreases with the increase of the number of participating bands. This is the so-called "Hughes" effect, which cannot be ignored in change detection of hyperspectral images. Summary of the invention
[0005] In order to solve the above technical problems, the present invention takes into account that there is a temporal correlation between the hyperspectral remote sensing images of the two phases, and each pixel corresponds to a longer spectral sequence. Therefore, for hyperspectral image change detection, the Long Short-Term Memory (LSTM) model has greater advantages than other deep learning models, and the ConvLSTM improved on the basis of LSTM can not only establish a temporal relationship like LSTM, but also can characterize local spatial features like CNN. The present invention proposes a hyperspectral remote sensing image change detection method based on ConvLSTM.
[0006] LSTM is a recurrent neural network suitable for processing time series. It has a strong modeling ability for time series data and can adaptively learn the temporal dependencies between multi-phase images. However, LSTM lacks the ability of CNN to extract spatial features and ignores the neighborhood spatial information of pixels in the image. The ConvLSTM model that combines LSTM and CNN in the present invention combines the advantages of both and can learn the joint spectral-spatial-temporal feature representation of two-phase hyperspectral images.
[0007] The technical solution of the present invention is: a hyperspectral remote sensing image change detection method, which specifically includes the following steps:
[0008] Step 1: input the first phase hyperspectral image and the second phase hyperspectral image, and remove abnormal bands of the hyperspectral images respectively;
[0009] Step 2: Generate high confidence samples using the change vector analysis method as neural network training samples;
[0010] Step 3: Perform PCA dimensionality reduction on the hyperspectral image;
[0011] Step 4: Train the twin convolutional neural network. The trained twin convolutional neural network is used to reduce the dimension and extract features of the input hyperspectral images of different phases.
[0012] Step 5: Input the feature tensors of the two time phases and output the multi-scale temporal-spatial-spectral variation feature representation;
[0013] Step 6: Combine the attention mechanism to fuse the multi-scale temporal-spatial-spectral change features of the output to obtain the final detection result.
[0014] Furthermore, the step 2, using the change vector analysis method to generate high confidence samples as neural network training samples, specifically includes:
[0015] The hyperspectral image of the previous phase is subtracted from the hyperspectral image of the latter phase to generate a difference image. The initial difference image is obtained through Otsu threshold segmentation. The 10% pixels with the largest difference are selected as training labels, and then sample pixels are randomly selected. A three-dimensional image patch is constructed with the selected sample pixels as the center as the training set based on the convolutional long short-term memory neural network model.
[0016] Furthermore, the step 3 specifically includes:
[0017] The following operations are performed on the hyperspectral images of the first phase and the second phase respectively: the average value of each band of each pixel in the hyperspectral image is calculated to obtain the band mean image, the original image is subtracted from the band mean image to obtain the average deviation image, the covariance matrix of the average deviation image is calculated, the covariance matrix is eigenvalue decomposed to obtain the eigenvalues and corresponding eigenvectors, the eigenvectors with larger eigenvalues are extracted and then combined as the image after dimensionality reduction, thereby obtaining the hyperspectral images of the two phases after dimensionality reduction.
[0018] Furthermore, the step 4 specifically includes:
[0019] The hyperspectral images of the two phases after dimensionality reduction are respectively input into the two convolutional neural network branches of the twin neural network. The weight values of the two branches are shared, each branch includes only one convolution layer, and the number of convolution kernels is less than the number of channels after dimensionality reduction in the previous step, so that the feature dimension is further reduced, and the features corresponding to the hyperspectral images of the first phase and the second phase are output.
[0020] Furthermore, the step 5 specifically includes:
[0021] The pair of feature tensors output in the previous step are input into the parallel convolutional long short-term memory neural network module to extract the features of the spatial, temporal, and spectral dimensions from the features of the two phases. Due to the parallel structure, a multi-scale time-space-spectral feature output consisting of low-level and high-level features is obtained.
[0022] Beneficial effects:
[0023] The change detection method proposed in the present invention uses a twin convolutional neural network to reduce the dimension of the image, thereby alleviating the impact of the high dimension of the hyperspectral image and the redundancy of adjacent band information on change detection and highlighting the change characteristics. Through a parallel structure, a hierarchical convolutional long short-term memory network is constructed to form different gradient propagation paths to output hidden states of different scales, adaptively learn the dependencies and spectral change characteristics related to real changes, and enhance the robustness to pseudo-change pixels. The hidden states output by the convolutional long short-term memory units of different paths are fused using the attention mechanism, and the multi-level time dependency between the hyperspectral images of the two phases is modeled. The training set is constructed using the traditional change vector analysis method. A small number of training samples can learn the multi-scale time-space-spectral feature representation between the two phase images. The trained neural network model can effectively distinguish between false change pixels caused by imaging conditions and noise and real change pixels caused by changes in ground object types. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 : ConvLSTM neural network structure;
[0025] Figure 2 :Flowchart of change detection based on ConvLSTM neural network;
[0026] Figure 3 :Schematic diagram of the MP-ConvLSTM module structure. DETAILED DESCRIPTION
[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.
[0028] ConvLSTM determines the state of a pixel at a future moment through the input and past state of a pixel's local neighborhood. Compared with LSTM, it can not only obtain time dependency but also extract spatial features through convolution. Based on the above characteristics of the convolutional long short-term memory neural network, the neural network can be used to learn the spectral-spatial-temporal features of hyperspectral remote sensing images of different phases for land use / cover change detection.
[0029] According to an embodiment of the present invention, the present invention proposes a hyperspectral remote sensing image change detection method, which can be implemented using Python programming language and specifically includes the following steps:
[0030] Step 1: input the first phase hyperspectral image and the second phase hyperspectral image, and remove abnormal bands of the hyperspectral images respectively;
[0031] Water vapor in the atmosphere has a greater impact on electromagnetic waves in the spectral range of 1363-1440nm, 1810-1957nm and greater than 2395nm. In electromagnetic waves of these wavelengths, ground information is greatly interfered, so the bands in these wavelength ranges need to be removed. For the EO-1 satellite Hyperion sensor, these are the 107th-116th, 152th-170th, and 214th-224th bands.
[0032] Step 2: Generate high confidence samples using the change vector analysis method as neural network training samples. The training samples here serve as the training samples of step 4, and the output obtained after processing in step 4 serves as the input of step 5.
[0033] The classic change vector analysis method is used to generate samples for network training. In order to identify changed and unchanged pixels, the hyperspectral image of the previous phase is subtracted from the hyperspectral image of the latter phase to generate a difference image. The initial difference image is obtained by Otsu threshold segmentation. The 10% pixels with the largest difference are selected as training labels, and then sample pixels are randomly selected. A three-dimensional image patch is constructed with the selected sample pixels as the center as the training set based on the convolutional long short-term memory neural network model.
[0034] Step 3: Perform PCA dimensionality reduction on the hyperspectral image (the dimension is c);
[0035] The central idea of PCA is to reduce the dimensionality of hyperspectral images composed of a large number of related bands while retaining as many variation features as possible in the data set. Principal component analysis is an orthogonal linear transformation that transforms the data into a new coordinate system to obtain the same number of principal components as the original data features. These principal components are uncorrelated, thereby achieving the purpose of reducing the redundancy of band information. In the principal component analysis of hyperspectral images, the principal components are the eigenvectors of the covariance matrix of the spectral vectors corresponding to each pixel of the hyperspectral image, and the spectral vector of each pixel is transformed into an eigenvector in the new space. The pixel spectral vector is defined as:
[0036] x idx =[x1, x2, ...x N ] T
[0037] Each element represents the pixel value of a pixel in a certain band;
[0038] The dimension of the vector is the number of bands N of the initial hyperspectral image. For a hyperspectral image with h rows and w columns, there will be M = h*w such vectors, i.e., idx = 1, ..., M. The average vector of all image vectors is represented and calculated as:
[0039]
[0040] The covariance matrix Cov(x) is defined as:
[0041] Cov(x)=E{(xE(x))(xE(x))} T
[0042] Where E is the expectation, T represents the transpose, and the covariance matrix CV of the pixel feature vector in the hyperspectral image is x The calculation is as follows:
[0043]
[0044] In principal component analysis, the eigenvalue decomposition of the covariance matrix is:
[0045] CV x =ADA T
[0046] where D=diag(λ1,λ2,...,λ N ) is the covariance matrix CV x Eigenvalues (λ1, λ2,...,λ N ), A is a diagonal matrix composed of N-dimensional eigenvectors a k (k=1,2,...a N ) is an orthogonal matrix:
[0047] A=(a1,a2,..a N )
[0048] The linear transformation of principal component analysis is defined as:
[0049]
[0050] Then y idx is the pixel spectrum vector x idx The principal component vector of each pixel in the present invention only retains the first c principal components.
[0051] Step 4: Dimensionality reduction and feature extraction based on the twin convolutional neural network (i.e., the twin "convolutional long short-term memory neural network model"); the twin convolutional neural network is the preprocessing module of the convolutional long short-term memory neural network, and the output of the twin convolutional neural network is used as the output of the convolutional long short-term memory neural network.
[0052] The method proposed by the present invention includes a training phase and a testing phase.
[0053] I t ∈{x(i, j)|1≤i≤h, 1≤j≤w} represents the hyperspectral image at time t, where h, w, and c represent the number of columns, rows, and bands of HSI, respectively. The number of principal components is equal to the number of image bands after dimensionality reduction, both of which are c, and CV represents the covariance matrix.
[0054] 3D Image Patch By extracting I t The pixel at position (m, n) and the surrounding pixels have a size of s×s×c, where s represents the number of columns and rows, and c represents the number of bands. Replace m in the principal component with ma. A patch refers to a local small image composed of the central pixel and the neighboring pixels. The columns (width) and rows (height) of the small image are equal, both are s. The number of principal components is equal to the number of image bands after dimensionality reduction, both are c. CV represents the covariance matrix.
[0055] In the training phase, the twin convolutional neural network can learn a trainable nonlinear function to transform the two-phase image patches P t , P t-1 Mapped to the tensor X in feature space t , X t-1 , the tensor size is s×s×k, so that if the two image patches are similar, the distance (e.g., Euclidean distance) between the two tensors is small, and if they change, the distance is large. The automatic learning process of the distance is achieved by training two convolutional neural network branches that share the same weights. In the present invention, a twin convolutional neural network (the number of convolution kernels is less than the number of input image channels) comprising two two-dimensional convolutional layer branches is used to extract preliminary features and reduce the spectral dimension of the hyperspectral image. After input into the twin convolutional neural network, the spectral dimension of the patch t is reduced from c to k, while the spatial dimension remains unchanged.
[0056] Step 5: Input the feature tensors of the two time phases and output the multi-scale temporal-spatial-spectral variation feature representation
[0057] According to an embodiment of the present invention, a parallel convolutional long short-term memory network module (MP-ConvLSTM) is proposed, which inputs a feature tensor of two time phases and outputs a multi-scale temporal-spatial-spectral change feature representation. In the present invention, the parallel convolutional long short-term memory network module (MP-ConvLSTM) and the twin neural network have different functions. The twin neural network performs preliminary feature extraction, and the parallel convolutional long short-term memory network is used for the extraction of multi-scale temporal-spatial-spectral features.
[0058] Traditional neural networks, including convolutional neural networks, fully connected neural networks and other forward propagation neural networks, assume that the input data are independent of each other, and do not consider the time dependency between inputs when extracting features. In the change detection task, due to the influence of imaging conditions and noise, the change in reflectivity of certain bands is not caused by the change in the type of land object. In order to strengthen the band reflectivity characteristics caused by the change in the type of land object, the band change characteristics (pseudo-change characteristics) that are not related to the change in the type of land object are suppressed. Recurrent neural networks have the ability to model temporal dependencies. Long short-term memory networks are a classic recurrent network, but they are not suitable for two-dimensional image processing. The convolutional long short-term memory neural network (ConvLSTM) with the convolution module can model the time dependency of two-phase hyperspectral images. Through the convolution operation, all input X t (X t In the change detection task, it is represented as an image patch P t Here X t Should be changed to P t Each band), cell state C t , hidden state H t AND Gate (I t , F t , O t )Cell State C t Combined with the current X t The cell state C at the previous moment t-1 The gate is used to control the on and off of information. t , F t , O t They are input gate, forget gate, output gate, and input gate control X t The information is integrated into the cell state C t , the forget gate controls the cell state C at the previous moment t-1 Integration into cell state C t , output goalkeeper C t-1 The output is the hidden state H t-1 They are all three-dimensional tensors The first dimension encodes temporal dependencies, and the second two dimensions learn spatial dependencies (e.g. Figure 2 shown).
[0059] From X t The input and past states of the center pixel value (and neighboring pixels) of each band are the inputs of ConvLSTM to determine the future state of the center pixel. The key equation of ConvLSTM is as follows:
[0060] I t =σ(W XI *X t +W HI *H t-1 +WCI ⊙C t-1 +b I )
[0061] F t =σ(W XF *X t +W HF *H t-1 +W CF ⊙C t-1 +b F )
[0062] C t =F t ⊙C t-1 +I t ⊙tanh(W XC *X t +W HC *H t-1 +b C )
[0063] O t =σ(W XO *X t +W HO *H t-1 +W CO ⊙C t +b O )
[0064] H t =O t ⊙tanh(C t-1 )
[0065] W represents the calculation weight between feature tensors, b represents the bias tensor, and W is the weight between features, for example, W XI represents the calculation weight between the input feature X and the input gate I, and σ represents the sigmoid activation function.
[0066] Where '*' represents the convolution operator, '⊙' represents the Hadamard product, and the convolutional long short-term memory network reduces the number of parameters by weight sharing, and on the other hand, effectively utilizes the spatial domain information by considering the spatial neighborhood pixels. Hidden state H t is considered as short-term memory in the time domain, while the cell state C t is considered as long-term memory. For the change detection task, the input data is a hyperspectral image obtained at two different times, which focuses more on short-term temporal dependencies. Therefore, the hidden state H output by ConvLSTM is tAs a time-space-spectral variation feature. In order to extract short-term dependencies at multiple levels (channel dimensions), the present invention does not simply stack ConvLSTM layers in series, but connects ConvLSTM units in parallel into MP-ConvLSTM modules (see Figure 3 ) The low-level hidden state H t Then send it to the ConvLSTM unit to get the high-level hidden state H′ t . Thus, multi-level short-term dependencies are extracted. High-level dependencies capture changes in spectral trends, and low-level dependencies are more sensitive to local changes in the spectrum. At the same time, in order to alleviate the possible gradient disappearance caused by deepening the layer, the hidden state output H of the front end is t-1 With the hidden state H of the backend t The modules are connected as the final output of the whole module. This module realizes multi-level modeling of short-term dependencies through its parallel branches and strengthens the learning of local discriminative temporal-spatial-spectral features.
[0067] Step 6: Feature fusion with attention mechanism
[0068] Although the feature map size of MP-ConvLSTM output is the same, the scale and time dependence of the spatial-spectral features are different to some extent. When fusing these features, an automatic channel selection mechanism is needed to suppress the information redundancy between channels. The attention mechanism proposed by Squeeze and Excite Network (SENet) is used to recalibrate the feature response between channels to establish a model of inter-channel dependency, so that the model pays more attention to the channel features with the largest amount of information and suppresses those unimportant features. The basic equation is defined as:
[0069] F = [H0, H'0, H1, H'1], H0 represents the hidden state of the low level at the first moment, H1 represents the hidden state of the low level at the second moment, H'0 represents the hidden state of the low level at the first moment, and H'1 represents the hidden state of the high level at the second moment;
[0070] M se =σ(MLP(AvgPool(F))+MLP(MaxPool(F)))
[0071] where F is the merged tensor of different semantic-level features (H0, H′0, H1, H′1) from MP-ConvLSTM, M se represents the attention tensor, the dimension of the tensor is 1×1×4k, and σ represents the sigmoid function. MLP, AvgPool, and MaxPool represent multi-layer perception, average pooling, and maximum pooling operations, respectively.
[0072]
[0073] in Represents the point-by-point product of the elements, the fused feature F se The data is sequentially input to the convolution layer and average pooling layer at the back end of the network to obtain the change probability. Pixels with high change probability are identified as changed pixels, and pixels with low change probability are identified as unchanged pixels.
[0074] According to one embodiment of the present invention, 2% of the labeled pixels of the two data sets are randomly selected as training sets to train the model, and the remaining labeled pixels are used as test sets to verify the accuracy of the model.
[0075] This paper takes two-view hyperspectral data obtained by the Hyperion sensor of the EO-1 satellite as a case to evaluate the change detection performance and compares it with CNN, RNN and traditional change detection methods.
[0076] Experimental Dataset
[0077] It includes two-phase hyperspectral images from two different scenes obtained by the Airborne Visible Infrared Imaging Spectrometer (AVIRIS), with 224 spectral bands, a spectral range of 200-2400nm, a spectral resolution of 10nm, and a spatial resolution of 20m. Both scenes include three types of pixels: changed pixels, unchanged pixels, and unlabeled pixels. Geometric registration has been completed. The Bay Area dataset was taken in 2013 and 2015 in Paterson, California, USA, with a size of 600x500 pixels.
[0078] The Santa Barbara dataset was taken in Santa Barbara, California, USA in 2013 and 2014, with a size of 984x740 pixels.
[0079] Table 1 AVIRIS hyperspectral remote sensing change detection dataset
[0080]
[0081] Accuracy evaluation indicators:
[0082] Change and no change can be regarded as two categories, so change detection can be regarded as a binary classification problem. Common accuracy evaluation indicators include overall accuracy (OA), user accuracy (UA), producer accuracy (PA) and Kappa coefficient. The present invention uses these indicators to evaluate the accuracy of change detection.
[0083] The standard change result is used as the reference value, and the model output result is used as the predicted value. By comparing the predicted value with the reference value, the confusion matrix is obtained, as shown in Table 2.
[0084] Table 2 Change detection confusion matrix
[0085]
[0086] N is the total number of pixels
[0087] Overall classification accuracy:
[0088]
[0089] User Precision:
[0090]
[0091] Producer precision:
[0092]
[0093] Kappa coefficient:
[0094]
[0095] (1) Abnormal band elimination
[0096] Water vapor in the atmosphere has a greater impact on electromagnetic waves in the spectral range of 1363-1440nm, 1810-1957nm and greater than 2395nm. In electromagnetic waves of these wavelengths, ground information is greatly interfered with, so the 107-116, 152-170, and 214-224 bands need to be eliminated.
[0097] (2) Hyperspectral Dimensionality Reduction
[0098] Although hyperspectral remote sensing data has a large number of bands, adjacent bands often have a strong correlation. When all bands are used for change detection, the efficiency is low and the computational complexity is large. Therefore, in this study, PCA dimensionality reduction is required for hyperspectral remote sensing images of two phases.
[0099] The principal component analysis takes the first 5 principal components to reduce the number of spectra of the hyperspectral image to 5, the window size is set to 5*5, 2% of the labeled pixels are selected as the training set for model training, the training batch size is 128, and all samples in the training set participate in the training times (Epoch) for 20 times. The remaining labeled pixels are used as the test set to verify the accuracy of the model. The experiment is repeated 10 times. The accuracy evaluation is shown in Table 3:
[0100] Table 3 Change detection accuracy evaluation table
[0101]
[0102] In order to verify the effectiveness of the change detection method proposed in the present invention, a comparative analysis was performed with the traditional change vector analysis method and the deep learning change detection method based on CNN and RNN.
[0103] Change vector analysis uses the magnitude of the vector between two time phases to represent the degree of change. After obtaining the change vector image, the change vector is classified by threshold value to divide the image into unchanged area and changed area. According to the change measurement, it can be divided into change vector analysis method based on Euclidean distance and Otsu threshold (EUC+OTSU), change vector analysis method based on spectral angle distance and Otsu threshold (SAM+OTSU), and SAM+WAT+OTSU change vector analysis method based on SAM+OTSU method and combined with watershed segmentation method to consider spatial neighborhood information. These three methods are used for change detection experiments respectively.
[0104] Convolutional neural network (CNN), recurrent neural network (RNN), and convolutional long short-term memory neural network (ConvLSTM) are used respectively, and the input samples are used to train the three neural networks. The training sample is set to 2%, and the parameters such as batch size and Epoch are kept consistent. The whole image is input into the trained neural network, and the change detection accuracy evaluation indicators are shown in Table 4.
[0105] Table 4 Santa Barbara change detection accuracy evaluation indicators
[0106]
[0107] The ConvLSTM-based method achieves higher change detection accuracy than traditional methods, CNN, and RNN. For the Bay Area dataset, the OA and Kappa coefficients of change detection based on ConvLSTM can reach 0.9909 and 0.9818, which are 2.2% and 4.6% higher than the change vector analysis method, and 2.8% and 5.9% higher than RNN and CNN, respectively. For the Santa Barbara dataset, the OA and Kappa coefficients of change detection based on ConvLSTM are 0.9969 and 0.9936, which are 3.6% and 8.0% higher than the change vector analysis method, and 2.1% and 4.6% higher than RNN and CNN, respectively. 0.2% and 0.3% higher.
[0108] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, and it should be clear that the present invention is not limited to the scope of the specific embodiments, for those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.
Claims
1. A hyperspectral remote sensing image change detection method, characterized in that: The specific steps include: Step 1, input the first phase hyperspectral image and the second phase hyperspectral image, and remove abnormal bands of the hyperspectral images respectively; Step 2: Generate high confidence samples using the change vector analysis method as neural network training samples; Step 3: Perform PCA dimensionality reduction on the hyperspectral image; Step 4: Train the twin convolutional neural network. The trained twin convolutional neural network is used to reduce the dimension and extract features of the input hyperspectral images of different phases. Step 5: Input the feature tensors of the two time phases and output the multi-scale temporal-spatial-spectral variation feature representation; Step 6: Combine the attention mechanism to perform feature fusion on the output multi-scale temporal-spatial-spectral change features to obtain the final detection result; The step 4 specifically includes: The hyperspectral images of the two phases after dimensionality reduction are respectively input into the two convolutional neural network branches of the twin neural network. The weight values of the two branches are shared. Each branch includes only one convolution layer. The number of convolution kernels is less than the number of channels after dimensionality reduction in the previous step, so that the feature dimension is further reduced, and the features corresponding to the hyperspectral images of the first phase and the second phase are output; The step 5 specifically includes: The pair of feature tensors output in the previous step are input into the parallel convolutional long short-term memory neural network module to extract the features of the spatial, temporal, and spectral dimensions from the features of the two phases. Due to the parallel structure, a multi-scale time-space-spectral feature output consisting of low-level and high-level features is obtained.
2. The hyperspectral remote sensing image change detection method according to claim 1, characterized in that: The step 2, using the change vector analysis method to generate high confidence samples as neural network training samples, specifically includes: The hyperspectral image of the latter phase is subtracted from the hyperspectral image of the previous phase to generate a difference image. The initial difference image is obtained by Otsu threshold segmentation, and the 10% pixels with the largest difference are selected as training labels. Then, sample pixels are randomly selected, and a three-dimensional image patch is constructed with the selected sample pixels as the center as the training set based on the convolutional long short-term memory neural network model.
3. The hyperspectral remote sensing image change detection method according to claim 1, characterized in that: The step 3 specifically includes: The following operations are performed on the hyperspectral images of the first phase and the second phase respectively: the average value of each band of each pixel in the hyperspectral image is calculated to obtain the band mean image, the original image is subtracted from the band mean image to obtain the average deviation image, the covariance matrix of the average deviation image is calculated, the covariance matrix is eigenvalue decomposed to obtain the eigenvalues and corresponding eigenvectors, the eigenvectors with larger eigenvalues are extracted and then combined as the image after dimensionality reduction, thereby obtaining the hyperspectral images of the two phases after dimensionality reduction.
