Change Detection Method and Device Based on a Single Remote Sensing Image

Abstract

A method and apparatus for change detection based on a single remote sensing image are disclosed. In the method, a remote sensing image before a change event is obtained; an autoencoder network is trained on the remote sensing image before the change event, and the K-means algorithm is applied to cluster the intermediate layer features of the autoencoder network, thereby generating a clustering map of the remote sensing image before the change event; in each category of the clustering map, multivariate Gaussian modeling is performed according to the intermediate layer features of the autoencoder network to obtain the feature distribution information of each category of the remote sensing image before the change event; a remote sensing image after the change event is obtained and input into the trained autoencoder network to extract the intermediate layer features; the intermediate layer features are combined with the clustering map before the change event, and the Mahalanobis distance from the feature distribution information before the change event is calculated in each category; threshold segmentation is performed on the full-image Mahalanobis distance after the change event to obtain a binary change map of "changed" and "unchanged".

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing satellites, and particularly relates to a change detection method and device based on a single remote sensing image. Background Technique

[0002] Natural disasters such as floods, landslides, wildfires, and deforestation caused by both natural forces and human activities pose a major threat to the Earth's environment and human society. Accurately and timely detecting these disasters is crucial for post-disaster assessment and recovery work. With the development of Earth observation technology and the increasing abundance of high spatio-temporal resolution multi-source image data, it provides the necessary data support for extensive and timely monitoring of disaster events. Therefore, remote sensing change detection technology has become a widely adopted method for identifying the areas where disaster events occur by comparing images before and after the occurrence of disaster events.

[0003] In the past few decades, scholars have developed a variety of change detection methods. Among them, deep learning-based methods have shown the best performance, which benefits from the powerful feature extraction and transformation capabilities of deep learning models, especially when dealing with high-resolution images with high spectral variability. However, existing deep learning-based change detection methods are mainly trained in a supervised manner, so the quantity and representativeness of samples will significantly affect their performance and transferability. However, rare disaster change events increase the difficulty of obtaining sufficient labeled samples. At the same time, different natural backgrounds also pose challenges to model migration. Therefore, unsupervised change detection is a more robust solution. In addition, current remote sensing change detection methods usually execute by storing multi-temporal image data from satellites. The whole process takes an average of 1 to 3 days, which may not meet the timeliness requirements of ground disaster emergency response. With the development of on-orbit image processing technology, on-orbit change detection methods can meet the high timeliness requirements in specific scenarios (such as disaster monitoring). However, it is unrealistic to store a large amount of high-resolution remote sensing images on orbit. Therefore, developing a change detection method that only relies on a single remote sensing image and a small amount of prior knowledge has great research and application value.

[0004] The above information disclosed in the background technique section is only used to enhance the understanding of the background of the present invention, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] Aiming at the problems existing in the prior art, the present invention proposes a change detection method and device based on a single remote sensing image, which is used to obtain the "changed" and "unchanged" position information of the image relative to the previous moment from a single-temporal remote sensing image.

[0006] The object of the present invention is achieved through the following technical solutions. A change detection method based on a single remote sensing image includes:

[0007] Step 1, obtain the remote sensing image before the change event;

[0008] Step 2, perform autoencoder network training on the remote sensing image before the change event. Among them, the autoencoder network consists of 7 fully connected layers, and the number of neurons in each layer is [256, 128, 64, 32, 64, 128, 256]. During training, the remote sensing image is segmented into non-overlapping image patches, and the loss function is set as the mean square error MSE between the input image patch and the network output image patch. At the 5th round of training, the intermediate layer features of the 4th layer of the network are output, and then the K-means algorithm is applied to cluster the features to generate the clustering map of the remote sensing image before the change event;

[0009] Step 3, since the input of the autoencoder network training is non-overlapping image patches, the spatial resolution of the obtained clustering map is smaller than that of the remote sensing image. Based on the dense connection Markov conditional random field algorithm, refine the spatial resolution of the clustering map to obtain the clustering map with the spatial resolution of the remote sensing image;

[0010] Step 4, in each category of the clustering map, perform multivariate Gaussian modeling on the intermediate layer features extracted from the remote sensing image through the autoencoder network, that is, calculate the mean vector and covariance matrix of the features in each category to obtain the feature distribution information of each category of the remote sensing image before the change event;

[0011] Step 5, obtain the remote sensing image after the change event and input it into the trained autoencoder network to extract intermediate layer features;

[0012] Step 6, based on the intermediate layer features of the image after the change event and the clustering map before the event, calculate the Mahalanobis distance between the intermediate layer features after the change event and the feature distribution information before the change event in each category, and obtain the full-image Mahalanobis distance;

[0013] Step 7, perform threshold segmentation on the full-image Mahalanobis distance after the change event. The pixels greater than the threshold are assigned 1 to indicate change, and the pixels less than the threshold are assigned 0 to indicate non-change, so as to obtain the binary classification map of change and non-change.

[0014] In the described change detection method based on a single remote sensing image, in Step 2, the optimization objective of the autoencoder network includes two stages. The first stage is to reconstruct the input remote sensing image, and the second stage includes reconstructing and optimizing the objective function of the K-means algorithm. Therefore, the loss function in the first stage of the autoencoder network training is The loss function in the second stage is where, x i is the input image slice, is the image slice reconstructed by the network, and N represents the number of image slices; Indicates the category c assigned to the image slice by the K-means algorithm, μ c Represents the feature center of category c.

[0015] In the described change detection method based on a single remote sensing image, in step 3, the dense-connected Markov conditional random field algorithm is used to refine the spatial resolution of the clustering map. The dense-connected Markov conditional random field takes into account the spectral information and spatial context relationship of the remote sensing image before the change event, and its energy function is defined as E(x) = ∑ i θ u (x i ) + ∑ ij θ p (x i , x j ) and Among them, θ u (x i ) and θ p (x i , x j ) are unary and binary potentials; θ u (x i ) represents the clustering map before refinement, μ(x i , x j ) represents the tolerance function, p i and p j represent pixels at different positions respectively, I i and I j represent the spectral information at positions i and j, θ α , θ β , and θ γ are algorithm hyperparameters. After refinement, a clustering map of the spatial resolution of the remote sensing image is obtained.

[0016] In the described change detection method based on a single remote sensing image, in step 6, the calculation method of the Mahalanobis distance Md for each category of the remote sensing image after the change event is Among them, i and c represent the position and clustering category of the pixel, The intermediate layer features of the i-th pixel belonging to category c, Σ c and μ c represent the covariance matrix and mean vector of the feature distribution of category c before the change event respectively. The Mahalanobis distance between categories is normalized by , where represents the average Mahalanobis distance of category c, and M represents the number of categories.

[0017] In the described change detection method based on a single remote sensing image, after obtaining the normalized global Mahalanobis distance, calculate its mean and standard deviation, and use the mean plus twice the standard deviation and a threshold to segment the Mahalanobis distance map. Among them, pixels greater than the threshold are assigned a value of 1, and pixels less than the threshold are assigned a value of 0, and finally a binary classification map of "changed" and "unchanged" is obtained.

[0018] A change detection device based on a single remote sensing image for implementing the described change detection method based on a single remote sensing image includes:

[0019] A first unit for obtaining a remote sensing image before a change event;

[0020] A second unit for training the remote sensing image. Among them, perform autoencoder network training on the remote sensing image before the change event, and apply the K-means algorithm to cluster the intermediate layer features of the autoencoder network, so as to generate a clustering map of the remote sensing image before the change event;

[0021] A third unit for refining the clustering map;

[0022] A fourth unit for performing multivariate Gaussian modeling on the intermediate layer features to obtain the feature distribution information of each category of the remote sensing image before the change event;

[0023] A fifth unit for obtaining the intermediate layer features of the remote sensing image after the change event;

[0024] A sixth unit for obtaining the Mahalanobis distance and the global Mahalanobis distance between the intermediate layer features of the remote sensing image after the change event and the feature distribution information before the change event,

[0025] A seventh unit for performing threshold segmentation on the global Mahalanobis distance to obtain a final binary classification map of "changed" and "unchanged".

[0026] In the described change detection device based on a single remote sensing image, the first unit is a remote sensing image acquisition device.

[0027] In the described change detection device based on a single remote sensing image, the second unit is an autoencoder network processing unit.

[0028] Beneficial effects

[0029] The change detection method and device based on a single remote sensing image can obtain a binary classification map of "changed" and "unchanged" only relying on a single remote sensing image and a small amount of prior knowledge, providing necessary data support for widely and timely monitoring disaster events. Description of the drawings

[0030] Upon reading the detailed descriptions in the following preferred specific embodiments, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The accompanying drawings of the specification are only for the purpose of showing the preferred embodiments and are not considered as a limitation of the present invention. Obviously, the following described drawings are only some embodiments of the present invention, and for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Moreover, throughout the drawings, the same reference numerals are used to represent the same components.

[0031] In the drawings:

[0032] Figure 1 A schematic diagram showing the steps of the present invention;

[0033] Figure 2 A result diagram showing an embodiment of the present invention;

[0034] Figure 3 A schematic flowchart showing the present invention.

[0035] The present invention will be further explained below in conjunction with the drawings and embodiments. Specific Embodiments

[0036] The specific embodiments of the present invention will be described in more detail below with reference to the drawings. Although the specific embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0037] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different terms to refer to the same component. The specification and claims do not use the difference in terms as a way to distinguish components, but use the difference in the functions of components as the criterion for distinction. As mentioned throughout the specification and claims, "comprising" or "including" is an open-ended term and should be interpreted as "including but not limited to". The subsequent description of the specification is for the purpose of implementing the preferred embodiments of the present invention, but the description is for the general principle of the specification and is not used to limit the scope of the present invention. The scope of protection of the present invention shall be determined by the scope defined by the appended claims.

[0038] For the convenience of understanding the embodiments of the present invention, the following will further explain with specific embodiments as examples in conjunction with the drawings, and each drawing does not constitute a limitation to the embodiments of the present invention.

[0039] In one embodiment, as Figures 1 to 2As shown in the figure, a change detection method based on a single remote sensing image includes:

[0040] Step 1: Obtain the remote sensing image before the change event;

[0041] Step 2: Train an autoencoder network on the remote sensing image before the change event. The autoencoder network consists of 7 fully connected layers, and the number of neurons in each layer is [256, 128, 64, 32, 64, 128, 256]. During training, the original remote sensing image is segmented into non-overlapping image patches (9×9), the training batch size is set to 512, the learning rate is set to 1e-4, and iterative training is performed for 10 rounds to ensure model convergence. The loss function is set to the mean square error (MSE) between the input image patch and the network output image patch. When training reaches the 5th round, the intermediate layer features of the network (the 4th layer, with a dimension of 32) are output, and then the K-means algorithm is applied to cluster the features, thereby generating a clustering map of the remote sensing image before the change event;

[0042] Step 3: Since the input of the autoencoder network training is non-overlapping image patches, the spatial resolution of the obtained clustering map is smaller than that of the original image (reduced by 9 times). At this time, based on the dense connection Markov conditional random field algorithm, the spatial resolution of the clustering map is refined to obtain a clustering map with the spatial resolution of the original image;

[0043] Step 4: In each category of the clustering map, perform multivariate Gaussian modeling on the intermediate layer features extracted from the remote sensing image by the autoencoder network, that is, calculate the mean vector and covariance matrix of the features of each category to obtain the feature distribution information of each category of the remote sensing image before the change event;

[0044] Step 5: Obtain the remote sensing image after the change event and input it into the trained autoencoder network to extract intermediate layer features;

[0045] Step 6: Based on the intermediate layer features of the image after the change event and the clustering map before the event, calculate the Mahalanobis distance between the intermediate layer features after the change event and the feature distribution information before the change event in each category;

[0046] Step 7: Perform threshold segmentation on the Mahalanobis distance of the entire image after the change event. Among them, the pixels greater than the threshold are assigned a value of 1 (changed), and the pixels less than the threshold are assigned a value of 0 (unchanged). To obtain a binary classification map of "changed" and "unchanged".

[0047] In the described change detection method based on a single remote sensing image, in Step 2, the optimization objective of the autoencoder network includes two stages. The first stage is to reconstruct the input remote sensing image, and the second stage includes reconstructing and optimizing the objective function of the K-means algorithm. Therefore, the loss function in the first stage of the autoencoder network training is The loss function in the second stage is Among them, x i is the input image slice, is the image slice after network reconstruction, and N represents the number of image slices; represents the category c assigned to the image slice by the K-means algorithm, and μ c represents the feature center of category c.

[0048] In the described change detection method based on a single remote sensing image, in step 3, the dense-connected Markov conditional random field algorithm is used to refine the clustering map. The dense-connected Markov conditional random field takes into account the spectral information and spatial context relationship of the remote sensing image before the change event, and its energy function is defined as E(x) = ∑ i θ u (x i ) + ∑ ij θ p (x i , x j ), Among them, θ u (x i ) and θ p (x i , x j ) are unary and binary potentials; θ u (x i ) represents the clustering map before refinement, and μ(x i , x j ) represents the tolerance function. p i and p j represent pixels at different positions respectively, and I i and I j represent the spectral information at positions i and j. θ α , θ β , and θ γ are algorithm hyperparameters.

[0049] In the described change detection method based on a single remote sensing image, in step 6, the calculation method of the Mahalanobis distance (Md) for each category of the remote sensing image after the change event is Among them, i and c represent the position and clustering category of the pixel. The intermediate layer feature of the i-th pixel belonging to category c, Σ c and μ c represent the covariance matrix and mean vector of the feature distribution of category c before the change event respectively. Then, the Mahalanobis distance between categories is normalized by . Among them, Denote the average Mahalanobis distance of class c, and M represents the number of classes.

[0050] In one embodiment, a change detection method based on a single remote sensing image includes, first, using high-resolution remote sensing data of remote sensing images to obtain input data, including but not limited to Planet, domestic high-resolution satellites, etc. Then, using the obtained input remote sensing image; thereafter, using the change monitoring method designed by the present invention to identify potential change regions.

[0051] In one embodiment, a change detection method based on a single remote sensing image includes,

[0052] S1. Data input and model initialization

[0053] Data input: Select a high-resolution remote sensing image.

[0054] Model initialization: This algorithm is implemented by Python code and can be directly called through the host command program in any computer.

[0055] S2. Obtain prior information before the change event

[0056] S2-1. Obtain the pre-event image clustering map

[0057] Data input: Remote sensing image data

[0058] Data output: The autoencoder network model optimizes the network through reconstruction and clustering tasks, and outputs the intermediate features of the image. The intermediate features and the K-means algorithm are used for clustering, and finally the image clustering map is output (as Figure 1 shown).

[0059] S2-2. Refine the spatial resolution of the clustering map

[0060] Data input: The image clustering map output by S2-1 and the remote sensing image before the change event.

[0061] Data output: Refine the clustering map through a densely connected Markov conditional random field to obtain a clustering map with refined spatial resolution.

[0062] S2-3. Feature distribution modeling

[0063] Data input: The intermediate layer features of the network and the clustering map output by S2-1.

[0064] Data output: Model the intermediate layer features in each clustering category through a multivariate Gaussian model, and output the feature distribution information of each category (including covariance matrix and mean vector information).

[0065] S3. Change detection module

[0066] Data input: Remote sensing images after change events, refined clustering maps output by S2-2, and feature distribution information of S2-3.

[0067] Data output: A change map containing the location information of "changed" (pixel value is 1) and "unchanged" (pixel value is 0).

[0068] Image processing process: In each clustering map category, calculate the Mahalanobis distance between each pixel of the remote sensing image after the change event and the feature distribution information of this category before the change event where i and c represent the position of the pixel and the clustering category. The intermediate layer features of the i-th pixel belonging to category c, Σ c and μ c represent the covariance and mean matrix of the feature distribution of category c before the change event respectively. Then, normalize the Mahalanobis distance between categories through where, represents the average Mahalanobis distance of category c, and M represents the number of categories. Finally, binarize by using the mean of the full-image Mahalanobis distance plus twice the variance as the threshold to obtain the change map.

[0069] In one embodiment, according to the images and true change map information of two study areas (Hebei Province, China and Hawaii, USA) of the present invention, through the extraction method provided by the above method of the present invention, the result of comparing the extracted change information with the distribution of the true value is as Figure 2 shown, indicating that the change detection established by the present invention has relatively reliable accuracy. As shown in Table 1 below, when the true change area of this area is known, the predicted change result by the application method is very close to the actual situation. Among them, Mean F1 are 81.03 and 94.84 respectively, and are significantly better than other change detection methods.

[0070] Table 1 shows the change detection accuracy table of the method of the present invention and other methods in two study areas (the bold indicates the optimal score).

[0071]

[0072] Figure 2 Shows the visualization results of the detection results of the method of the present invention and the comparative method in two study areas. The method of the present invention can accurately detect the change area and greatly reduce the false detection area.

[0073] In addition, in one embodiment, a change detection device based on a single remote sensing image includes:

[0074] A first unit for obtaining a remote sensing image before a change event;

[0075] The second unit is used to train the remote sensing image. Specifically, the remote sensing image before the change event is trained by an autoencoder network, and the K-means algorithm is applied to cluster the features of the middle layer of the autoencoder network, so as to generate a clustering map of the remote sensing image before the change event.

[0076] The third unit is used to refine the spatial resolution of the clustering map.

[0077] The fourth unit is used to perform multivariate Gaussian modeling on the middle layer features to obtain the feature distribution information (including covariance matrix and mean vector information) of each category of the remote sensing image before the change event.

[0078] The fifth unit is used to obtain the middle layer features of the image after the change event.

[0079] The sixth unit is used to obtain the Mahalanobis distance between the middle layer features of the image after the change event and the feature distribution information before the change event.

[0080] The seventh unit is used to perform threshold segmentation on the Mahalanobis distance map. Specifically, the pixels greater than the threshold are assigned a value of 1 (changed), and the pixels less than the threshold are assigned a value of 0 (unchanged). To obtain the final binary classification map of "changed" and "unchanged".

[0081] In the preferred implementation manner of the change detection device based on a single remote sensing image, for the second unit, the optimization objective of the autoencoder network includes two stages. The first stage is to reconstruct the input remote sensing image, and the second stage includes reconstructing and optimizing the objective function of the K-means algorithm. Therefore, the loss function of the first stage of the autoencoder network training is The loss function of the second stage is where x i is the input image slice, is the image slice after network reconstruction, and N represents the number of image slices; represents the category c assigned to this image slice by the K-means algorithm, and μ c represents the feature center of category c.

[0082] In the preferred implementation manner of the change detection device based on a single remote sensing image, for the third unit, the dense connected Markov conditional random field algorithm is used to refine the clustering map. The dense connected Markov conditional random field takes into account the spectral information and spatial context relationship of the remote sensing image before the change event, and its energy function is defined as E(x) = ∑ i θ u (x i ) + ∑ ij θ p (x i , x j ) and Among them, θ u (x i ) and θ p (x i , x j ) are unary and binary potentials; θ u (x i ) represents the clustering map before refinement, and μ(x i , x j ) represents the tolerance function. p i and p j respectively represent pixels at different positions, and I i and I j represent the spectral information at positions i and j. θ α , θ β , and θ γ are algorithm hyperparameters.

[0083] In the preferred implementation manner of the change detection device based on a single remote sensing image described above, the sixth unit calculates the Mahalanobis distance (Md) of each category in the remote sensing image after the change event as Among them, i and c represent the position and clustering category of the pixel. The intermediate layer feature of the i-th pixel belonging to category c, Σ c and μ c respectively represent the covariance and mean matrix of the feature distribution of category c before the change event. Then, the Mahalanobis distance between categories is normalized through . Among them, represents the average Mahalanobis distance of category c, and M represents the number of categories.

[0084] Those skilled in the art should understand that although the present invention is described in the manner of multiple embodiments, not every embodiment contains only one independent technical solution. Such a narration in the specification is only for clarity. Those skilled in the art should understand the specification as a whole and consider the technical solutions involved in each embodiment as ways that can be combined with each other to form different embodiments to understand the protection scope of the present invention.

[0085] Although the embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields. The above specific embodiments are merely illustrative and guiding, rather than restrictive. Those of ordinary skill in the art can also make many forms under the inspiration of this specification and without departing from the scope protected by the claims of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A change detection method based on a single remote sensing image, characterized in that: It includes the following steps: Step 1, obtaining remote sensing images before the change event; Step 2, the autoencoder network is trained on the remote sensing image before the change event, where the autoencoder network consists of 7 fully connected layers, and the number of neurons in each layer is [256, 128, 64, 32, 64, 128, 256]. During training, the remote sensing image is divided into non-overlapping image blocks, and the loss function is set to the mean square error (MSE) between the input image block and the network output image block. When the training reaches the fifth round, the intermediate layer features of the fourth layer of the network are output, and then the K-means algorithm is used to cluster the features, thereby generating a clustering diagram of the remote sensing image before the change event; Step 3: Since the input of the autoencoder network training is non-overlapping image blocks, the spatial resolution of the obtained clustering graph is smaller than that of the remote sensing image. The spatial resolution of the clustering graph is refined based on the densely connected Markov conditional random field algorithm to obtain a clustering graph with the spatial resolution of the remote sensing image. Step 4: In each category of the cluster diagram, multivariate Gaussian modeling is performed on the intermediate layer features extracted from the remote sensing image through the autoencoder network, that is, the mean vector and covariance matrix of each category feature are calculated to obtain the feature distribution information of each category of the remote sensing image before the change event; Step 5, obtain the remote sensing image after the change event and input it into the trained autoencoder network to extract the intermediate layer features; Step 6, based on the image intermediate layer features after the change event and the clustering graph before the event, the Mahalanobis distance between the feature distribution information of the intermediate layer features after the change event and before the change event is calculated in each category, and the Mahalanobis distance of the whole image is obtained; Step 7, perform threshold segmentation on the Mahalanobis distance of the entire image after the change event, assign a value of 1 to pixels greater than the threshold to indicate a change, and assign a value of 0 to pixels less than the threshold to indicate no change, so as to obtain a binary classification map of change and no change.

2. The change detection method based on a single remote sensing image according to claim 1, characterized in that: In step 2, the optimization goal of the autoencoder network includes two stages. The first stage is to reconstruct the input remote sensing image, and the second stage includes reconstruction and optimization of the objective function of the K-means algorithm. Therefore, the loss function of the first stage of the autoencoder network training is The loss function of the second stage is Among them, x i is the input image slice, is the image slice reconstructed by the network, and N represents the number of image slices; Indicates the category c assigned to the image slice by the K-means algorithm, μ c represents the feature center of category c.

3. The change detection method based on a single remote sensing image according to claim 2, characterized in that: In step 3, the densely connected Markov conditional random field algorithm is used to refine the spatial resolution of the cluster graph. The densely connected Markov conditional random field takes into account the spectral information and spatial context of the remote sensing image before the change event, and its energy function is defined as E(x) = ∑ i θ u (x i )+∑ ij θ p (x i ,x j )and Among them, θ u (x i ) and θ p (x i ,x j ) are the unary and binary potential energies; θ u (x i ) represents the clustering graph before refinement, μ(x i ,x j ) represents the tolerance function, p i and p j The subtables represent pixels at different positions, I i and I j represents the spectral information of positions i and j, θ α ,θ β , and θ γ is the algorithm hyperparameter. After refinement, the clustering diagram of the spatial resolution of the remote sensing image is obtained.

4. The change detection method based on a single remote sensing image according to claim 1, characterized in that: In step 6, the Mahalanobis distance Md of each category of the remote sensing image after the change event is calculated as follows: Among them, i and c represent the location and cluster category of the pixel, The intermediate layer feature of the i-th pixel belonging to category c, Σ c and μ c Represent the covariance matrix and mean vector of the feature distribution of category c before and after the change event, respectively. The Mahalanobis distance between categories is calculated by Normalize, where represents the average Mahalanobis distance of category c, and M represents the number of categories.

5. The change detection method based on a single remote sensing image according to claim 4, characterized in that: After obtaining the normalized Mahalanobis distance of the entire image, its mean and standard deviation are calculated, and the Mahalanobis distance map is segmented using the mean plus twice the standard deviation and the threshold. Pixels greater than the threshold are assigned a value of 1, and pixels less than the threshold are assigned a value of 0, ultimately obtaining a binary classification map of "change" and "non-change".

6. A change detection device based on a single remote sensing image for implementing the change detection method based on a single remote sensing image as claimed in any one of claims 1 to 5, characterized in that: It includes: The first unit is used to obtain remote sensing images before the change event; The second unit is used to train the remote sensing image, wherein an autoencoder network is trained on the remote sensing image before the change event, and a K-means algorithm is applied to the intermediate layer features of the autoencoder network for clustering, thereby generating a clustering graph of the remote sensing image before the change event; The third unit is used to refine the clustering graph; The fourth unit is used to perform multivariate Gaussian modeling on the intermediate layer features to obtain the feature distribution information of each category of the remote sensing image before the change event; The fifth unit is used to obtain the intermediate layer features of the remote sensing image after the change event; Unit 6 is used to obtain the Mahalanobis distance between the middle layer features of the remote sensing image after the change event and the feature distribution information before the change event and the Mahalanobis distance of the whole image. The seventh unit is used to perform threshold segmentation on the Mahalanobis distance of the entire image to obtain the final binary classification image of "change" and "non-change".

7. The change detection device based on a single remote sensing image according to claim 6, characterized in that: The first unit is the remote sensing image acquisition device.

8. The change detection device based on a single remote sensing image according to claim 6, characterized in that: The second unit is a self-encoding network processing unit.

Images