An Unsupervised Hyperspectral Image Change Detection Method Using Bayesian Network to Optimize Pseudo Labels

Abstract

An unsupervised hyperspectral image change detection method using Bayesian network to optimize pseudo-labels includes two steps: initial pseudo-label generation and pseudo-label optimization. For the first step, a pseudo-label generation strategy with high confidence spatial constraints is designed to generate initial pseudo-labels. The generated pseudo-labels have both accuracy and diversity. For the second step, the initial pseudo-labels are used to train the Bayesian network. The trained network can predict the complete change map and uncertainty. Then, the pseudo-label selection mechanism based on uncertainty perception can select new pseudo-labels for the next network training. In this way, the network training, testing and pseudo-label selection are repeated in an iterative form, which enables the pseudo-labels to be continuously optimized, thereby continuously enhancing the model's ability to detect changes. The present invention can detect almost complete strong changes and weak changes, suppress false changes, and can also provide the uncertainty of the prediction results for decision-making assistance.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing hyperspectral remote sensing images, relates to hyperspectral image change detection technology, and particularly relates to an unsupervised hyperspectral image change detection method for optimizing pseudo-labels by using a Bayesian network. Background Art

[0002] Change detection refers to the process of discovering differences by comparing multiple images taken in the same area but with a certain time difference. This technology is crucial for understanding land surface changes and has currently been widely applied in fields such as ecosystem monitoring, land cover mapping, urban expansion research, resource management, etc. Therefore, change detection has always been a research hotspot in the field of remote sensing applications. The rapid development of sensor technology has promoted the formation of hyperspectral images. Hyperspectral images provide dense sampling of the spectral characteristics of targets within a relatively wide spectral wavelength range, making it possible to precisely monitor land cover at a finer spectral scale. The challenge of hyperspectral change detection is to accurately detect strong and fine changes while suppressing false changes.

[0003] Scholars at home and abroad have devoted a great deal of energy to conducting in-depth research on the change detection problem, and a large number of change detection methods and theoretical models have been proposed. Deep learning has swept through the fields of computer vision and remote sensing image interpretation. The high-dimensional characteristics of hyperspectral data have made deep learning-based methods increasingly popular. Many supervised deep learning-based hyperspectral change detection methods have been successively proposed. However, the limited data sources and the scarcity of reference change maps have led to the pixels used for training and testing coming from the same scene. These characteristics result in the inability to use large supervised networks and an overly optimistic evaluation of the performance of supervised networks. Therefore, unsupervised deep learning methods that do not rely on reference change maps are more promising. Among them, the unsupervised deep learning method based on pseudo-labels has been widely studied due to its simplicity, ease of implementation, and high accuracy. However, the quality of the pseudo-labels generated by current hyperspectral change detection methods needs to be improved. Measuring the quality of pseudo-labels requires considering accuracy and diversity. Accuracy refers to the consistency between pseudo-labels and true labels. Pseudo-labels with low accuracy will gradually transmit noise into the change map. Diversity refers to whether the pseudo-labels cover all categories, or whether the pseudo-training set can represent the entire dataset. For the change detection of hyperspectral images, the diversity of pseudo-labels requires both strong changes that are easily captured and subtle changes that are difficult to detect. A model trained with pseudo-labels that do not contain subtle changes cannot accurately distinguish subtle changes, which will lead to many missed detections. Therefore, improving the quality of pseudo-labels is crucial. Existing hyperspectral change detection methods usually obtain pseudo-labels using one or two pre-classification methods and still cannot meet the requirements of accuracy and diversity simultaneously. Because pre-classification methods tend to misclassify subtle changes and false changes. This results in insufficient diversity of pseudo-labels without subtle changes and excessive noise in pseudo-labels with subtle changes. This problem limits the detection performance and applications. Therefore, it is crucial to propose a new hyperspectral change detection method whose pseudo-label strategy takes both accuracy and diversity into account. Summary of the Invention

[0004] To overcome the above-mentioned drawbacks of the prior art, the purpose of the present invention is to provide an unsupervised hyperspectral image change detection method that optimizes pseudo-labels using a Bayesian network, so as to solve the problem that the pseudo-labels of existing methods cannot possess both accuracy and diversity simultaneously, and to obtain excellent detection performance using the optimized high-quality pseudo-labels, thereby being able to detect the change information in bi-temporal images.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] An unsupervised hyperspectral image change detection method that optimizes pseudo-labels using a Bayesian network, comprising the following steps:

[0007] Step 1: For the dual-temporal hyperspectral images, use the high-confidence spatial constraint pseudo-label generation strategy to jointly use multiple pre-classification methods and superpixel segmentation to generate initial pseudo-labels;

[0008] Step 2: Use the initial pseudo-labels to train the Bayesian deep learning classification network model, and then test all pixels to obtain the change map and uncertainty; use the pseudo-label selection mechanism based on uncertainty awareness to optimize the pseudo-labels until the model converges and outputs the final change map.

[0009] Compared with the prior art, the beneficial effects of the present invention are:

[0010] 1. When performing change detection on hyperspectral images, the designed model of the present invention uses the Bayesian network to iteratively optimize the pseudo-labels, so that the pseudo-labels with continuously improved quality promote the improvement of the model detection performance, can achieve high accuracy on hyperspectral images obtained by different hyperspectral sensors, have good generalization ability, and have high robustness to noise.

[0011] 2. The change detection model proposed by the present invention can detect relatively complete strong changes and weak changes while suppressing false changes, which gives full play to the advantages of the fine spectral features of hyperspectral images.

[0012] 3. The change detection model proposed by the present invention can provide the predicted model uncertainty, that is, output the change map and the model uncertainty corresponding to each pixel. There is a good correspondence between the uncertainty and the prediction error (the uncertainty of the pixels with prediction errors is higher, and the uncertainty of the pixels with correct predictions is lower). The prediction credibility provided by such uncertainty can assist in decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is the overall framework of the present invention.

[0014] Figure 2 is Figure 1 a schematic diagram of the generation of initial pseudo-labels in an embodiment of DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] The following describes the embodiments of the present invention in detail with reference to the drawings and embodiments.

[0016] As mentioned above, among the existing hyperspectral change detection technologies, unsupervised methods based on pseudo-labels can achieve higher accuracy without relying on ground truth. However, the existing pseudo-label-based methods rely on one or two pre-classification methods, which makes it impossible for pseudo-labels to have both accuracy and diversity, limiting the performance of the model. To this end, the hyperspectral image change detection method of the present invention designs a pseudo-label generation strategy with high-confidence spatial constraints to obtain reliable and high-quality initial pseudo-labels; then, the uncertainty-aware pseudo-label selection mechanism can continuously optimize the pseudo-labels using the uncertainty generated by the Bayesian network, so that the pseudo-labels can satisfy both accuracy and diversity.

[0017] like Figure 1 As shown, the unsupervised hyperspectral image change detection method using Bayesian network to optimize pseudo-labels mainly includes two steps: the first step is to generate initial labels (pre-classification, fusion change map, spatially constrained pseudo-label selection), that is, for dual-phase hyperspectral images, a high-confidence spatially constrained pseudo-label generation strategy is used, and multiple pre-classification methods and superpixel segmentation are used in combination to generate initial pseudo-labels. The second step is to optimize pseudo-labels (make training sets, train Bayesian networks, test, optimize pseudo-labels, and output detection results), that is, use the initial pseudo-labels to train the Bayesian deep learning classification network model, then test all pixels, obtain change maps, predict probabilities and uncertainties; optimize pseudo-labels using a pseudo-label selection mechanism based on uncertainty perception until the model converges and outputs the final change map.

[0018] In the first step of the present invention, a pseudo-label generation strategy with high confidence space constraints is designed to generate initial pseudo-labels. The generated pseudo-labels have both accuracy and diversity. In the second step of the present invention, the initial pseudo-labels are used to train the Bayesian network. The trained network can predict the complete change graph and uncertainty. Then, the pseudo-label selection mechanism based on uncertainty perception can select new pseudo-labels for the next network training. In this way, the network training, testing and pseudo-label selection are carried out in an iterative form, which enables the pseudo-labels to be continuously optimized, thereby promoting the continuous enhancement of the model's ability to detect changes. Ultimately, the present invention can detect almost complete strong changes and weak changes, suppress false changes, and can also provide the uncertainty of the prediction results for decision-making assistance.

[0019] The specific process of the present invention can be described as follows:

[0020] The first step is to generate initial pseudo labels. In one embodiment, the Hermiston hyperspectral image dataset includes dual-phase hyperspectral images T1 and T2, and a ground truth map. The generation of initial pseudo labels is as follows: Figure 2 shown.

[0021] 1) Generate high-quality pseudo-labels using three available pre-classification methods

[0022] The three pre-classification methods of the present invention are the change vector analysis method, the structural similarity method, and the spectral unmixing method. Using these three pre-classification methods to pre-classify T1 and T2 respectively, three binary change maps are obtained. In the maps, the value of 1 represents a change in the corresponding area, and the value of 0 represents no change in the corresponding area.

[0023] First, obtain the change map using the change vector analysis method. The input data are the bi-temporal hyperspectral images T1 and T2. Performing pixel-by-pixel difference on T1 and T2 gives the difference map X, expressed as:

[0024] X = abs(T1 - T2) (1)

[0025] Compressing the difference map X to 2D gives:

[0026]

[0027] where is the k-th component of X, k = 1, …, K, K is the number of spectral channels of the hyperspectral image, W is the spatial width of each hyperspectral image, H is the spatial height of each hyperspectral image. Performing K-means clustering on DI cva can obtain the result map of the change vector analysis method, that is, a binary change map, expressed as CM cva .

[0028] Next, use the structural similarity method for change detection. The input data are the bi-temporal hyperspectral images T1 and T2. Let w1 and w2 be the windows at the same spatial position in one spectral channel of T1 and T2 respectively. The calculation formula for the structural similarity index S(w1, w2) of w1 and w2 is:

[0029]

[0030] In the formula and are the mean and variance of w1, and are the mean and variance of w2, Denote the statistical covariance between w1 and w2. C1 and C2 are constants used to avoid the denominator being zero and have little impact on the accuracy of the final result. The calculated S in the above formula is the structural similarity index of the central pixel of the window in a spectral channel. Let the center of the sliding window traverse the entire image, and a structural similarity matrix of the same size as the spatial dimension of the hyperspectral image can be obtained. Then, traversing all channels in the same way can obtain a 3D structural similarity matrix of the same size as the hyperspectral image. Finally, compress the structural similarity index of the two-temporal hyperspectral images T1 and T2 to 2D, and the structural similarity index of the two-temporal hyperspectral image is obtained as:

[0031]

[0032] where S k is the structural similarity index of the k-th spectral channel. Usually, the structural similarity cannot be directly used for change detection. The value range of the structural similarity is 0 to 1. Therefore, convert S HSI to a difference:

[0033] DI ssim = 1 - S HSI (5)

[0034] Perform K-means clustering on DI ssim to obtain a binary change map, denoted as CM ssim .

[0035] Finally, perform a spectral unmixing method, and the input data is the difference map X. That is, use the linear spectral unmixing model to unmix the difference map X. The linear unmixing model is expressed as:

[0036]

[0037] where x is the spectral feature of the mixed pixel, m e is the e-th endmember in X, α e represents the corresponding abundance map, that is, the percentage of m e in the considered mixed pixel, e = 1,..., E, E represents the number of endmembers in X, and ε is the noise vector. It should be noted that for the present invention, the number of endmembers is equal to the number of change categories.

[0038] The present invention uses the fully constrained least squares method for abundance map estimation. The estimation process is as follows:

[0039]

[0040] Since the changing categories are usually associated with the intensity of change, and the unchanged categories are usually associated with low intensity of change, it is easy to rank the end members and abundances. Let m1 be the unchanged end member and α1 be the corresponding abundance map. Similarly, m2, …, m E are variable end members, α2…,α E is the corresponding abundance map.

[0041] By thresholding all abundance maps, a binary change map can be obtained. Obviously, the changes selected by the high threshold are credible. Therefore, the threshold is simply set to 0.5, which is a relatively high value because the number of end members E is generally greater than 2. The binary change map based on the unmixing method is expressed as:

[0042]

[0043] The above three pre-classifications can also be other complementary change detection methods, which can jointly determine credible pseudo labels.

[0044] 2) Fusion change graph

[0045] The change vector analysis, structural similarity-based and unmixing-based methods are used to jointly determine the credible pseudo-labels, thereby ensuring the accuracy of the pseudo-labels. The present invention formulates a fusion strategy based on the characteristics of the three pre-classification methods, and fuses the above three binary change maps into one change map, so that the fused change map has a higher credibility. In addition, spectral decomposition is specifically used to mine the subtle changes of pseudo-labels, so that it contains both strong changes and weak changes, thereby improving the diversity of pseudo-labels. Therefore, the process of fusing three binary change maps can be expressed as:

[0046]

[0047] The above formula shows that for CM cva CM ssim and CM unmix The pixels in the fusion result image are unchanged. cva and CM ssim The pixels in the fusion result image are all changed pixels, and they are judged as changed. unmix The pixels that are changed in the fusion result image are also judged as changed (this is because the change threshold based on the unmixing method is set relatively high, so the changes detected are most likely changes).

[0048] 3) Spatially constrained pseudo-label selection

[0049] Inspired by K-means clustering, superpixel segmentation can utilize spatial information and spectral information to segment all pixels into superpixels. Therefore, in the present invention, superpixel segmentation is performed on the difference map X to achieve spatial constraint. Then, according to the segmentation result, several pixels are extracted from each superpixel and their indices are recorded. Based on the indices, values are taken from the fused change map as initial pseudo-labels. The only parameter that needs to be specified in this algorithm is the number of superpixels. In the present invention, superpixel segmentation is performed on the difference map X, and the segmentation result can be expressed as:

[0050]

[0051] where sp z represents the z-th superpixel, and Z is the number of superpixels. After obtaining the superpixels, n pixels are randomly selected from each superpixel for training, which can be expressed as:

[0052]

[0053] At this time, covers Z·n difference spectra in the initial training set, and the fused result map CM p contains the required initial pseudo-labels. Each in is a difference spectrum, corresponding to a p value in CM , that is, the pseudo-label corresponding to the difference spectrum. The spatial index of is easily recorded for corresponding value. It should be noted that when the corresponding value is neither 0 nor 1, then this difference spectrum is regarded as invalid and discarded. The corresponding values of all valid ones are integrated to obtain the initial pseudo-labels:

[0054]

[0055] The pseudo-labels selected in this way can cover the entire image, including all categories, including strong change information and weak change information, and basically meet the requirements of accuracy and diversity.

[0056] Step 2: Optimization of pseudo-labels.

[0057] First, the Bayesian network is trained using the initial pseudo-labels in the first iteration. Then, the trained Bayesian network is used to make multiple predictions for all pixels in the entire hyperspectral image. The prediction results include the change map, uncertainty, and prediction probability. Among them, the uncertainty and prediction probability are used to select more reliable pseudo-labels for the next iteration. Such an iterative process continues until the number of pseudo-labels converges to a stable level. This process can be described as:

[0058] Step 2.1, make a pseudo-training set. Regard the pseudo-labels as sample labels, and according to the indexes of the pseudo-labels, obtain the differential spectra corresponding to each pseudo-label from the difference map X, that is, sample data. One pixel corresponds to one sample data, and the pseudo-training set is obtained: {sample data, sample label}; when using the initial pseudo-labels, the initial pseudo-training set is obtained. In the present invention, the sample is the pixel.

[0059] Step 2.2, build the network. Establish a Bayesian deep learning classification network model, and its network structure is sequentially set as the input layer, convolutional layer, Dropout layer, fully connected layer, and output layer. Set the dimension of the input layer as the number of channels of the difference map X, set the number of nodes of the output layer as 2 (corresponding to changed and unchanged), and set the Dropout ratio as r. In this embodiment, the Bayesian network only includes one convolutional layer, one dropout layer, and one fully connected layer. Such a network structure can ensure timeliness on the basis of meeting the accuracy requirements.

[0060] Step 2.3, train. Initialize the model parameters, and use the pseudo-training set to train the model so that it has the ability to detect change information and the ability to estimate the uncertainty of the prediction.

[0061] Step 2.4, test. Use the trained model to test the complete test set (all samples, covering all pixels of the entire data set), test B times, that is, B predictions are required, obtain the uncertainty of the pixels, and the final test result is: the change map composed of all predictions, the corresponding prediction probability, and uncertainty.

[0062] Step 2.5, pseudo-label optimization. The uncertainty-aware selection mechanism determines whether each pixel is selected as a pseudo-label according to the prediction probability and uncertainty of each pixel (two thresholds are τ σ and τ p ). The pixels with high prediction probability and low uncertainty that meet the conditions are integrated into new pseudo-labels, and a new pseudo-training set is made according to the new pseudo-labels for network training in the next iteration. Repeat steps 2.3 to 2.5, and the iterative optimization of the pseudo-labels can be realized.

[0063] Step 2.6, output the detection result. Take the number of pseudo-labels as the condition for iterative convergence. When the number of pseudo-labels no longer increases, it is regarded that the quantity and quality of the pseudo-labels reach a balance. At this time, output the current iterative change graph as the final detection result.

[0064] Specifically, during network training, in the first iteration, the training set consists of the initial pseudo-labels of Z·n samples: where is the input differential spectrum, is the corresponding pseudo-label. Taking a new differential spectrum x * as an example, the process of calculating the model uncertainty and prediction probability is as follows:

[0065] The posterior probability distribution of the new differential spectrum x * is expressed as:

[0066]

[0067] where p(y * |x * , ω) is the probability of the output y * when the input is x * and the network weight is ω. ω is the training set weight. In addition, is the probability of obtaining the network weight ω by training the training set , which can be understood as the uncertainty of the model. Therefore, the posterior probability in the above formula describes the distribution of all possible predictions given the new differential spectrum x * and the network weight ω. The width of this posterior probability distribution can reflect the confidence of the model in a specific prediction. It should be noted that cannot be calculated analytically, but can be approximated by the variational distribution q(ω). The Monte Carlo (MC) dropout method can interpret the random dropout process of neurons in the dropout layer as a variational distribution. This method alleviates the problem of representational uncertainty in deep learning without sacrificing computational complexity and model accuracy. The present invention uses a Bayesian network with a dropout layer to obtain the uncertainty of the model. By running B predictions through the Bayesian network, B random outputs can be directly obtained. Therefore, Equation (13) can be approximated as:

[0068]

[0069] The final prediction probability (prediction average value) of the new differential spectrum x * is:

[0070]

[0071] The variance of the B predictions is the uncertainty of the model's prediction of the prediction, that is, the model uncertainty:

[0072]

[0073] It can be seen that the Bayesian network with dropout not only realizes the mapping from the difference map to the change map, but also provides the uncertainty and prediction probability of each sample.

[0074] Based on the uncertainty-aware pseudo-label selection strategy, the model uncertainty σ 2 (x * ) and the prediction probability μ(x * ) are used to judge whether to select the new difference spectrum x * corresponding prediction y * as a pseudo-label; let the binary value l(x * ) indicate whether y * is selected. l(x * ) = 1 means selection, and l(x * ) = 0 means rejection. Then the discrimination criterion is:

[0075]

[0076] where τ σ and τ μ are the uncertainty threshold and the prediction probability threshold respectively. Obviously, only the samples that simultaneously meet the low uncertainty and high prediction probability will be selected as pseudo-labels. By selecting all samples, new pseudo-labels can be obtained for the network training in the next iteration. During the iteration process, the pseudo-labels are continuously optimized, the number is constantly increasing, and the diversity is gradually enriched. Until the model converges, the number and quality of the pseudo-labels reach the optimal state, and then the change map can be output as the final detection result.

[0077] In addition, the binary hyperspectral image change detection task only considers two classes, namely the changed class and the unchanged class. Therefore, this is a task with a very small number of classes. In this case, since the unchanged class is easy to detect, the classification network will seriously tend to the unchanged class, resulting in serious class imbalance. To solve this problem, the present invention balances the classes of the selected pseudo-labels. Specifically, if the number of pseudo-labels of the unchanged class is too large, only the samples with low uncertainty that are consistent with the number of the changed class are selected as the optimized pseudo-labels.

[0078] In one embodiment, the Hermiston dataset is adopted, and the number of endmembers E is set to 7. The convolution kernel size of the convolutional neural network is (3,3), and the number of feature channels is 256. The dropout parameter r in the Bayesian network is set to 0.1, and the number of runs B is 10 times. The uncertainty threshold τ σ of selecting pseudo-labels is 0.1, and the threshold τ pIt is 0.7. The number of selected initial pseudo-labels is 20% of all samples in the dataset.

[0079] Input the Hermiston dataset into the model to generate initial pseudo-labels, then continuously optimize the pseudo-labels. When the model converges, output the change diagram and evaluate the change diagram. The quantitative evaluation results of 5 evaluation indicators for different algorithms are analyzed as shown in the following table:

[0080] Table 1 Quantitative evaluation results of change detection for the USA dataset

[0081] Method OA Recall Precision F1 Kappa CVA 0.9245 0.6788 0.9804 0.8022 0.7574 SSIM 0.9040 0.5972 0.9630 0.7372 0.6824 ISFA 0.8344 0.8254 0.5957 0.6920 0.5827 IRMAD 0.9180 0.6560 0.9709 0.7830 0.7347 MaxtreeCD 0.9255 0.9097 0.7912 0.8437 0.7940 Unet 0.8920 0.7393 0.7720 0.7553 0.6861 SGCL 0.9130 0.6302 0.9747 0.7655 0.7151 SSCF 0.9383 0.8247 0.8934 0.8577 0.8184 BRPL 0.9498 0.9059 0.8757 0.8905 0.8579

[0082] The higher the values of the 5 quantitative evaluation indicators in Table 1, the better the detection performance. The higher the Precision value, the fewer false alarms. The higher the Recall value, the fewer missed detections. Usually, there is a negative correlation between the two. The F1 score is a combination of the Precision and Recall metrics. OA reveals the overall performance. The Kappa coefficient represents the consistency between the predicted change diagram and the reference change diagram. It can be seen from Table 1 that the algorithm of the present invention has obtained the highest OA, F1 and Kappa, which fully demonstrates the superior detection performance of the algorithm of the present invention.

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still modify the specific implementation manners of the present invention or make equivalent replacements. Any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention are within the protection scope of the claims of the present invention pending approval.

Claims

1. An unsupervised hyperspectral image change detection method for optimizing pseudo-labels using a Bayesian network, characterized in that, It includes the following steps: Step 1: For the dual-temporal hyperspectral image, use the high-confidence spatial-constrained pseudo-label generation strategy, and jointly use multiple pre-classification methods and superpixel segmentation to generate initial pseudo-labels; Step 2: Use the initial pseudo-labels to train a Bayesian deep learning classification network model, then test all pixels to obtain the change map and uncertainty; use the pseudo-label selection mechanism based on uncertainty awareness to optimize the pseudo-labels until the model converges and outputs the final change map; The said Step 1 includes the following steps: Step 1.1, perform pixel-by-pixel difference on the dual-temporal hyperspectral images T1 and T2 to obtain the difference map X, expressed as: X = abs(T1 - T2); Step 1.2, respectively use 3 pre-classification methods to pre-classify the dual-temporal hyperspectral images T1 and T2 to obtain 3 binary change maps. In the map, the value of 1 represents that there is a change in the corresponding area, and the value of 0 represents that there is no change in the corresponding area. The 3 pre-classification methods are the change vector analysis method, the structural similarity method, and the spectral unmixing method; Step 1.3, fuse the 3 binary change maps into one change map, so that the fused change map has higher credibility and contains both strong changes and weak changes; Step 1.4: Perform superpixel segmentation on the difference map X to achieve spatial constraints. Then, according to the segmentation results, extract several pixels from each superpixel, denoted as and record the index. Take the value from the fused change map according to the index as the initial pseudo-label, each of i.e., a difference spectrum; In the said step 2, for a new differential spectrum The process of calculating the model uncertainty is as follows: Posterior probability distribution of the new differential spectrum x * is expressed as: where p(y * |x * , ω) is the probability of output y * when the input is x * and the network weights are ω. is the probability of obtaining the network weights ω through training on the training set . The width of this posterior probability distribution can reflect the confidence of the model in a specific prediction. B random outputs are obtained through B predictions and are approximated as: The variance of B predictions is the uncertainty of the model's prediction, that is, the model uncertainty:

2. The unsupervised hyperspectral image change detection method for optimizing pseudo-labels using a Bayesian network according to claim 1, wherein In Step 1.2, the said change vector analysis method: Compress the difference map X to 2D: wherein, is the k-th component of X, k = 1, …, K, where K is the number of spectral channels of the hyperspectral image, W is the spatial width of each hyperspectral image, and H is the spatial height of each hyperspectral image. Performing K-means clustering on DI cva can obtain the result image of the change vector analysis method, that is, a binary change map, denoted as CM cva ; The said structural similarity method: Let w1 and w2 be the windows at the same spatial position in the same spectral channel of the dual-temporal hyperspectral images T1 and T2. The calculation formula of the structural similarity index S(w1, w2) of w1 and w2 is: where and are the mean and variance of w1, and are the mean and variance of w2, denotes the statistical covariance between w1 and w2, and C1 and C2 are constants used to avoid the denominator being zero; Compress the structural similarity index of the dual-temporal hyperspectral images T1 and T2 to 2D: where S k is the structural similarity index of the k-th spectral channel. Convert S HSI to a difference: DI ssim = 1 - S HSI For DI ssim Perform K-means clustering to obtain a binary change map, denoted as CM ssim ; The said spectral unmixing method: Use the linear spectral unmixing model to unmix the difference map X. The linear unmixing model is expressed as: where x is the spectral signature of the mixed pixel, m e is the e-th endmember in X, α e represents the corresponding abundance map, i.e., m e represents the percentage in the considered mixed pixel, e = 1, …, E, where E represents the number of endmembers in X, and ε is a noise vector, and the number of endmembers is equal to the number of change classes; Use the fully constrained least squares method for abundance map estimation: Let m1 be the invariant endmember, α1 be the corresponding abundance map, and m2, …, m E be the variable endmembers, and α2, …, α E be the corresponding abundance maps; threshold selection is performed on all the abundance maps to obtain a binary change map, denoted as CM unmix .

3. The unsupervised hyperspectral image change detection method for optimizing pseudo-labels using a Bayesian network according to claim 2, wherein In the said spectral unmixing method, the threshold is uniformly set to 0.

5. The binary change map based on the unmixing method is expressed as:

4. The unsupervised hyperspectral image change detection method for optimizing pseudo-labels using a Bayesian network according to claim 2, characterized in that The process of fusing the 3 binary change maps in the said Step 1.3 is expressed as: That is, CM cva , CM ssim and CM unmix are all unchanged pixels and are unchanged in the fused result image; CM cva and CM ssim are all changed pixels and are changed in the fused result image; CM unmix The changed pixels in are changed in the fused result image; The said Step 1.4, perform superpixel segmentation on the difference map X, expressed as: where sp z represents the z-th superpixel, and Z is the number of superpixels; after obtaining the superpixels, n pixels are randomly selected in each superpixel for training, which is expressed as: at this time, Covering Z n difference spectra in the initial training set, the fusion result is shown in CM p contains the required initial pseudo labels, Each of That is, a difference spectrum, corresponding to CM p one of the Value, that is, the pseudo label corresponding to the difference spectrum; The spatial index is recorded for the corresponding value; when a Corresponding When the value is neither 0 nor 1, the difference spectrum is considered invalid and discarded. All valid Corresponding The values are integrated to get the initial pseudo-label:

5. The unsupervised hyperspectral image change detection method for optimizing pseudo-labels using a Bayesian network according to claim 1 or 2 or 3 or 4, characterized in that The said Step 2 includes the following steps: Step 2.1, regard the pseudo-labels as sample labels, and according to the indexes of the pseudo-labels, obtain the difference spectra corresponding to each pseudo-label from the difference map X, that is, the sample data. One pixel corresponds to one sample data, and get the pseudo-training set: {sample data, sample label}; when using the initial pseudo-labels, get the initial pseudo-training set; Step 2.2, establish a Bayesian deep learning classification network model, whose network structure is sequentially set as the input layer, convolutional layer, Dropout layer, fully connected layer, and output layer. Set the dimension of the input layer as the number of channels of the difference map X, set the number of nodes in the output layer as 2, corresponding to change and no change, and set the Dropout ratio as r; Step 2.3, initialize the model parameters, and use the pseudo-training set to train the model so that it has the ability to detect change information and the ability to estimate the uncertainty of predictions; Step 2.4: Use the trained model to test the complete test set B times to obtain the uncertainty of pixels. The final test results are: the change map composed of all predictions, the corresponding prediction probabilities, and the uncertainties; The complete test set refers to all pixels covering the entire data set; Step 2.5: According to the prediction probability and uncertainty of each pixel, determine whether it is selected as a pseudo-label. Pixels with high prediction probabilities and low uncertainties that meet the conditions are integrated into new pseudo-labels. A new pseudo-training set is created based on the new pseudo-labels for network training in the next iteration. Repeat steps 2.3 - 2.5 to achieve iterative optimization of the pseudo-labels; Step 2.6: Use the number of pseudo-labels as the condition for iterative convergence. When the number of pseudo-labels no longer increases, it is considered that the quantity and quality of the pseudo-labels reach a balance. At this time, output the current iterative change map as the final detection result.

6. The unsupervised hyperspectral image change detection method for optimizing pseudo-labels using a Bayesian network according to claim 5, characterized in that During network training, in the first iteration, the training set consists of initial pseudo-labels: where is the input differential spectrum, is the corresponding pseudo-label; for a new differential spectrum x * , the predicted probability is the average value: An uncertainty-aware pseudo-label selection strategy that utilizes the model uncertainty σ 2 (x * ) and the predicted probability μ(x * ) to determine whether to select the new differential spectrum x * 's corresponding prediction y * as a pseudo-label; let the binary value l(x * ) represent whether y * is selected, l(x * ) = 1 means selected, l(x * ) = 0 means rejected, and the discrimination criterion is: where τ σ and τ μ are the uncertainty threshold and the prediction probability threshold, respectively.

Images