Multi-class change detection method and device for remote sensing images suitable for imbalanced categories
Through the twin dual-branch network with weight sharing and the global hierarchical sampling mechanism, combined with the mask constraints of the second-class change detection results, the problem of training samples imbalance in the multi-class change detection of remote sensing images is solved, and the detection accuracy and robustness are improved. , realizes multi-class variation detection of high-spatial resolution images.
Patent Information
- Application Number
- CN202111219302.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-20
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-10-20
AI Technical Summary
The existing multi-category change detection methods for remote sensing images cannot guarantee high accuracy when training samples are unbalanced and there are few samples. The traditional methods ignore the spectral information and multi-scale geometric features in high-spatial resolution images, resulting in poor detection results. precise.
Using a twin dual-branch network structure with weight sharing, the balanced training samples are obtained through the global hierarchical sampling mechanism, combined with the second-class change detection results as mask constraints, reweight the classification probability, extract more distinctive features, and reduce the background without change the noise effect.
The robustness of the model and the accuracy of the detection results are improved, especially in the case of uneven categories, which can better identify multiple types of change areas and subtle changes and obtain better detection results.
Smart Images

Figure CN113963269B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the combination of deep learning and remote sensing image processing, and mainly solves the problem of remote sensing image change detection. Specifically, it relates to a multi-category change detection method and device for remote sensing images suitable for imbalanced categories, which realizes multi-category change detection of high spatial resolution remote sensing images. Background Art
[0002] The surface of the earth is constantly changing due to human acquisition and the influence of natural forces. In order to record and monitor the changes in the earth's environment and keep it in a sustainable state, it is necessary to use change detection technology. Change detection in remote sensing aims to identify changes in the earth's surface by analyzing multi-temporal images obtained from the same geographical area on different dates. At present, remote sensing image change detection is widely used in various fields, such as urban planning and natural resource management. Due to the improvement of spatial resolution, high spatial resolution remote sensing images can provide more detailed surface information for change detection tasks. For example, subtle changes in land cover types can be detected by fine geometric information. However, most of the existing change detection methods are two-class change detection, which only focuses on changed and unchanged areas, resulting in application limitations. Therefore, multi-class change detection methods that can simultaneously identify changed areas and change categories have become a topic of increasing concern in recent years.
[0003] Among the traditional methods for multi-class change detection in remote sensing images, the most widely used method is the classification-based method, which can be divided into the post-classification comparison method and the direct classification method. The post-classification comparison method first classifies the multi-temporal images separately, and then obtains the multi-class change map by comparing the two classification maps. Although the post-classification comparison method is simple and intuitive, the final accuracy of the detection result is affected by the accuracy of the classification map. Unlike the post-classification comparison method, the direct classification method can directly obtain the results of multi-class change detection from the original multi-temporal image. Therefore, the direct classification method will not have error accumulation like the post-classification comparison method. In recent years, many direct classification methods have been proposed. However, due to the rich spectral information and multi-scale geometric characteristics of high spatial resolution remote sensing images, the traditional direct classification method ignores the diversity characteristics of land cover change.
[0004] In recent years, the rapid development of deep learning technology has promoted the progress of multi-class change detection in remote sensing images. At present, multi-class change detection methods for remote sensing images based on deep learning mainly use deep learning-based segmentation networks, such as U-Net, DeepLabv3 and FarSeg, which can be regarded as classifiers to obtain multi-class changes. Despite this, there are still few studies on multi-class change detection in remote sensing images based on deep learning. In addition, most of the current multi-class change detection methods based on deep learning use a single segmentation network branch as the backbone, which weakens the robustness of the model, and cannot guarantee the accuracy of multi-class change detection in remote sensing images when the training samples are unbalanced or small. Summary of the invention
[0005] The main technical problem to be solved by the present invention is to provide a method and device for multi-category change detection of remote sensing images suitable for imbalanced categories, which can ensure the accuracy of multi-category change detection of remote sensing images when the training samples are unbalanced or the training samples are small.
[0006] In order to achieve the above object, the present invention provides a multi-class change detection method for remote sensing images applicable to imbalanced categories, comprising the following steps:
[0007] Constructing a change detection framework, wherein the change detection framework includes an encoder and a decoder, wherein the encoder includes two weight-sharing network branches;
[0008] Obtain dual-phase high spatial resolution remote sensing images before and after the change;
[0009] Inputting the dual-phase high spatial resolution remote sensing images before and after the change into the two weight-sharing network branches respectively, and extracting initial features of the image pairs respectively;
[0010] Connecting the initial features of the image pair through a connection block to obtain a feature map;
[0011] Performing mask processing on the feature map to obtain a masked feature map;
[0012] Inputting the masked feature map into the decoder to calculate the probability of change categories, thereby obtaining a change probability map;
[0013] Obtaining a stratified sequence of training samples through a global stratified sampling mechanism, training the change detection framework through the stratified sequence of training samples, and obtaining a trained change detection framework after the training is completed;
[0014] The change probability map is discriminated by the trained change detection framework, and multiple types of change detection result maps are output.
[0015] Furthermore, the step of obtaining a stratified sequence of training samples through a global stratified sampling mechanism includes:
[0016] Acquire a sample data set corresponding to dual-phase high spatial resolution remote sensing images before and after the change;
[0017] A global stratified sampling mechanism is adopted for the sample data set to generate a stratified sequence of training samples;
[0018] The layered sequence of training samples is used to train and optimize the change detection framework.
[0019] Furthermore, the step of adopting a global stratified sampling mechanism for the sample data set to generate a stratified sequence of training samples further includes:
[0020] Dividing the sample data set into training data and test data, wherein the number of each category in the training data is divided according to a ratio;
[0021] The training data is used to generate a set of random sequence training samples. A threshold is set for the number of samples in each category in each batch of training samples so that the number of samples in each category is within the range of 5-10.
[0022] Furthermore, the step of inputting the dual-phase high spatial resolution remote sensing images before and after the change into the two weight-sharing network branches respectively and extracting the initial features of the image pairs respectively includes:
[0023] Initializing the change detection framework and setting hyperparameters to obtain an initialized model framework;
[0024] The dual-phase high spatial resolution remote sensing images before and after the change are respectively input into two weight-shared network branches of the encoder in the initialization model framework, and convolution and pooling processing are performed on the dual-phase high spatial resolution remote sensing images before and after the change to obtain the initial features of the image pair.
[0025] Furthermore, the step of connecting the initial features of the image pair through the connection block to obtain the feature map includes:
[0026] The initial features of the image pair are input into the connection block, and the feature maps of the two images are connected by a cat function to obtain a feature map.
[0027] Furthermore, the step of performing mask processing on the feature map to obtain the masked feature map includes:
[0028] Applying a change mask obtained by two-class change detection to the feature map;
[0029] The feature map and the change mask are subjected to matrix multiplication processing, and the unchanged background area value is replaced with 0 to obtain the masked feature map.
[0030] Furthermore, the step of inputting the masked feature map into a decoder to calculate the probability of change categories and obtain a change probability map includes:
[0031] The decoder upsamples the masked feature map, connects the absolute values of the differences of the two branches in the encoder through jump connections to supplement the more abstract information after encoding, and restores the object details and spatial dimensions of the masked feature map through deconvolution to obtain the change probability map after the decoder.
[0032] Furthermore, the step of obtaining a hierarchical sequence of training samples and training the change detection framework by using the hierarchical sequence of training samples includes:
[0033] Obtain a stratified sequence of training samples;
[0034] Calculate the cross loss function;
[0035] Optimizing model parameters of the change detection framework according to the cross loss function and the back propagation process;
[0036] After the training is completed, the trained change detection framework is obtained.
[0037] Furthermore, the step of discriminating the change probability map by using the trained change detection framework and outputting multiple types of change detection result maps includes:
[0038] The change probability map is discriminated by using a trained change detection framework to obtain a category label of each pixel in the image and obtain a change category feature map;
[0039] The change category feature map is cropped and restored to the spatial size of the original image to obtain the multi-category change detection result map.
[0040] In addition, in order to achieve the above-mentioned purpose, the present invention also provides a remote sensing image multi-category change detection device suitable for imbalanced categories, comprising the following modules:
[0041] A construction module is used to construct a change detection framework, wherein the change detection framework includes an encoder and a decoder, wherein the encoder includes two weight-sharing network branches;
[0042] The acquisition module is used to obtain dual-phase high spatial resolution remote sensing images before and after the change;
[0043] An extraction module, used to input the dual-phase high spatial resolution remote sensing images before and after the change into the two weight-sharing network branches respectively, and extract the initial features of the image pairs respectively;
[0044] A connection module, used for connecting the initial features of the image pair through a connection block to obtain a feature map;
[0045] A mask module, used for performing mask processing on the feature map to obtain a masked feature map;
[0046] A calculation module, used for inputting the masked feature map into the decoder to calculate the probability of change category and obtain a change probability map;
[0047] A training module, used for obtaining training samples through a global stratified sampling mechanism, training the change detection framework through the training samples, and obtaining a trained change detection framework after the training is completed;
[0048] A discrimination module is used to discriminate the change probability map through the trained change detection framework and output multiple types of change detection result maps
[0049] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:
[0050] The features of the dual-phase images are extracted through a weight-sharing twin dual-branch network structure to retain more distinguishing features to improve the robustness of the model. Relatively balanced training sample data is obtained through a global stratified sampling mechanism, ensuring that the global learning framework can still be applied under unbalanced categories. The classification probability is re-weighted by using the second-class change detection results as a change mask constraint to filter out the background noise of the unchanged area. This minimizes the impact of the complex unchanged background when performing multi-class change detection, thereby obtaining better detection results. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:
[0052] Figure 1 It is a flowchart of a method for detecting changes in multiple categories of imbalanced categories in a specific embodiment of the present invention;
[0053] Figure 2 The overall structure of the weight-sharing twin dual-branch network in the specific implementation manner of the present invention;
[0054] Figure 3 The number of training data and test data for each variation category of the Shiyan dataset in the specific implementation manner of the present invention;
[0055] Figure 4This is a comparison chart of the results obtained by using different multi-class change detection methods for the Shiyan dataset in a specific embodiment of the present invention;
[0056] Figure 5 It is a structural diagram of a multi-category change detection device for remote sensing images suitable for imbalanced categories in a specific implementation of the present invention. DETAILED DESCRIPTION
[0057] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.
[0058] DETAILED DESCRIPTION OF THE EMBODIMENTS A multi-class change detection method for remote sensing images applicable to imbalanced classes is disclosed. Figure 1 , the specific steps are as follows:
[0059] S1. Construct a change detection framework, wherein the change detection framework includes an encoder and a decoder, wherein the encoder includes two weight-sharing network branches;
[0060] The overall structure of the weight-sharing twin dual-branch network is shown in Figure 2 As shown. The structure of the twin network is designed based on the following considerations. First, in traditional change detection, a single-branch network needs to convert two input images into one input, which will result in information loss. Therefore, a dual-branch network is used to input two images at the same time, which has faster convergence speed and higher stability than a single-branch network. Secondly, in order to make model training easier, the kernel size used in the model is the most common 3*3. This is because as the kernel size becomes smaller, the number of model parameters will also decrease. The twin network structure consists of two network branches, each of which can be regarded as an image feature extractor. Each feature extraction network includes four 3*3 convolutional layers and 1 2*2 maximum pooling layer. Assume and They are high spatial resolution remote sensing images in phase 1 and phase 2 respectively. The feature map can be expressed as:
[0061]
[0062] Among them, t1, t2 represent phase 1 and phase 2, are the feature maps of phase 1 and phase 2 images respectively, l = {1, ..., L} represents the feature map of the lth layer in the feature extraction network, R is the set of feature maps, w, d, h are the length, width and height of the feature map respectively.
[0063] S2, obtain dual-phase high spatial resolution remote sensing images before and after the change;
[0064] This experiment uses data sets of three typical cities, but only the Shiyan data set is analyzed and discussed in this example. The Shiyan data set contains two images taken by the QuickBird satellite in 2002 and 2004, with a spatial resolution of 2.4 meters, four bands (red, green, blue and near infrared), and a data size of 250*250. In this data set, the main land cover categories are concentrated on the transition between bare land, vegetation, roads and water bodies.
[0065] S3, inputting the dual-phase high spatial resolution remote sensing images before and after the change into the two weight-sharing network branches respectively, and extracting initial features of the image pairs respectively;
[0066] S4, connecting the initial features of the image pair through a connection block to obtain a feature map;
[0067] The feature maps of the image pairs are concatenated using the cat function. For multi-class change detection of high spatial resolution remote sensing images, the use of a differential strategy may result in information loss in the final detection results due to high intra-class or low inter-class variability. Moreover, for different spectral feature conversions, the differential strategy can obtain non-unique differential values, while the concatenation operation can retain uniqueness (all feature values). Therefore, for multi-class change detection, the concatenation operation is a better choice.
[0068] S5, performing mask processing on the feature map to obtain a masked feature map;
[0069] The feature map obtained after connection is matrix multiplied with the two-class change mask, and the value of the unchanged background area is replaced by 0 to obtain the masked feature map. Unlike the two-class change detection, the multi-class change detection task aims to obtain more information about the changed classes, which is not an easy task. It is not difficult to find from several existing multi-class change detection datasets that the proportion of unchanged classes in the image is always greater than the proportion of changed classes. In order to suppress the influence of the unchanged class, after the encoder obtains the feature map, the decoder first generates a two-class change mask to constrain the results of the multi-class change detection, thereby obtaining more accurate results. The generation of the two-class change mask can be expressed by the following formula:
[0070]
[0071]
[0072] Among them, C i,j ∈{0,1} represents the value of the change graph at (i,j), c i,j =1 means that the pixel at (i, j) has changed, p represents the probability of change at (i, j), Θ is the model parameter set, Θ *is the updated set of model parameters, which can be solved by maximization.
[0073] S6, inputting the masked feature map into the decoder to calculate the change category probability, and obtaining a change probability map;
[0074] The decoder in the change detection framework upsamples the masked feature map, connects the absolute values of the differences of the two branches in the encoder through jump connections to supplement the more abstract information after encoding, and restores the object details and spatial dimensions of the feature map through deconvolution to obtain the change probability map after the decoder.
[0075] S7, obtaining a stratified sequence of training samples through a global stratified sampling mechanism, training the change detection framework through the stratified sequence of training samples, and obtaining a trained change detection framework after the training is completed;
[0076] In this example, the global stratified sampling mechanism is used to divide the Shiyan dataset. The first sampling is to select 5% of the samples from all the labeled samples as training data. If the number of training samples of a certain category is less than 5, 5 samples of the category are randomly selected from the labeled samples as the training data of the category. The remaining 95% of the samples are used as test data. The division results are as follows: Figure 3 The second sampling is to randomly distribute the 5% training data to each batch of training samples. The number of each category in each training sample is set between 5 and 10, and a total of 20 batches are divided and used as a set of random training sample sequences.
[0077] The cross entropy loss function is used to calculate the training loss, and the weight parameters in the model are updated by stochastic gradient descent, where the momentum is set to 0.9 and the weight decay rate is set to 0.001. At the same time, the learning rate decay method is adopted and the initial learning rate is set to 0.001 and the decay rate is set to 0.9. In order to ensure the convergence of the model, the number of iterations is set to 400.
[0078] S8. Using the trained change detection framework, the change probability map is judged, and multiple types of change detection result maps are output.
[0079] The change probability map is discriminated through the trained change detection framework, the category label of each pixel in the image is obtained, and the change category feature map is obtained;
[0080] The change category feature map is cropped and restored to the spatial size of the original image to obtain the multi-category change detection result map.
[0081] In order to verify the effectiveness of the multi-class change detection method for remote sensing images applicable to imbalanced categories of the present invention, in this embodiment, the classification results are also analyzed and evaluated. Figure 4 The following figure shows the results of multi-class change detection using different methods on the Shiyan dataset. Figure 4 (a) is the real surface change label used for reference (the black area is the unchanged area, and the other colors represent different change categories) Figure 4 (b)- Figure 4 (g) Classification diagrams of PCA, MCVA, PCC_SVM, PCC_NNet, CRF, and DSCN algorithms, respectively. Figure 4 (h) is the classification diagram of the Siam-LG (Siamese global learning framework) method proposed in this experiment. By comparison, it can be seen that the results of the method proposed in this invention are closest to the real surface change labels, and the boundaries of small change categories and change areas can be well identified. Table 1 shows the comparison of the accuracy evaluation results of different methods. Overall, it can be seen that the method proposed in this experiment has the best detection accuracy, and the average accuracy of each change category is also better than other methods, which illustrates the effectiveness of the method of the present invention.
[0082] Table 1 is a comparison of the accuracy evaluation results of each classification method
[0083]
[0084] As an optional implementation, this embodiment also provides a remote sensing image multi-class change detection device applicable to imbalanced categories, referring to Figure 5 , including the following modules:
[0085] Building module 1, for building a change detection framework, wherein the change detection framework includes an encoder and a decoder, wherein the encoder includes two weight-sharing network branches;
[0086] Acquisition module 2, used to acquire dual-phase high spatial resolution remote sensing images before and after the change;
[0087] Extraction module 3, used for inputting the dual-phase high spatial resolution remote sensing images before and after the change into the two weight-sharing network branches respectively, and extracting the initial features of the image pairs respectively;
[0088] A connection module 4, used for connecting the initial features of the image pair through a connection block to obtain a feature map;
[0089] A mask module 5, used for performing mask processing on the feature map to obtain a masked feature map;
[0090] A calculation module 6, used for inputting the masked feature map into the decoder to calculate the change category probability, and obtaining a change probability map;
[0091] A training module 7 is used to obtain training samples through a global stratified sampling mechanism, train the change detection framework through the training samples, and obtain a trained change detection framework after the training is completed;
[0092] The discrimination module 8 is used to discriminate the change probability map through the trained change detection framework and output multiple types of change detection result maps.
[0093] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
[0094] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.
[0095] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A multi-class change detection method for remote sensing images suitable for imbalanced categories, characterized in that: The following steps are involved: Constructing a change detection framework, wherein the change detection framework includes an encoder and a decoder, wherein the encoder includes two weight-sharing network branches; Obtain dual-phase high spatial resolution remote sensing images before and after the change; Inputting the dual-phase high spatial resolution remote sensing images before and after the change into the two weight-sharing network branches respectively, and extracting initial features of the image pairs respectively; Connecting the initial features of the image pair through a connection block to obtain a feature map; Performing mask processing on the feature map to obtain a masked feature map; Inputting the masked feature map into the decoder to calculate the probability of change categories, thereby obtaining a change probability map; Obtaining a stratified sequence of training samples through a global stratified sampling mechanism, training the change detection framework through the stratified sequence of training samples, and obtaining a trained change detection framework after the training is completed; The change probability map is discriminated by the trained change detection framework, and multiple types of change detection result maps are output.
2. The multi-class change detection method for remote sensing images applicable to imbalanced categories as claimed in claim 1, characterized in that: The step of obtaining a stratified sequence of training samples through a global stratified sampling mechanism includes: Acquire a sample data set corresponding to dual-phase high spatial resolution remote sensing images before and after the change; A global stratified sampling mechanism is adopted for the sample data set to generate a stratified sequence of training samples; The layered sequence of training samples is used to train and optimize the change detection framework.
3. The multi-class change detection method for remote sensing images applicable to imbalanced categories as claimed in claim 2, characterized in that: The step of adopting a global stratified sampling mechanism to the sample data set to generate a stratified sequence of training samples also includes: Dividing the sample data set into training data and test data, wherein the number of each category in the training data is divided according to a ratio; The training data is used to generate a group of random sequence training samples, and a threshold is set for the number of samples in each category in each batch of training samples so that the number of samples in each category is within a certain range.
4. The multi-class change detection method for remote sensing images applicable to imbalanced categories as claimed in claim 1, characterized in that: The step of inputting the dual-phase high spatial resolution remote sensing images before and after the change into the two weight-sharing network branches respectively and extracting the initial features of the image pairs respectively comprises: Initializing the change detection framework and setting hyperparameters to obtain an initialized model framework; The dual-phase high spatial resolution remote sensing images before and after the change are respectively input into two weight-shared network branches of the encoder in the initialization model framework, and convolution and pooling processing are performed on the dual-phase high spatial resolution remote sensing images before and after the change to obtain the initial features of the image pair.
5. The multi-class change detection method for remote sensing images applicable to imbalanced categories as claimed in claim 1, characterized in that: The step of connecting the initial features of the image pair through the connection block to obtain the feature map includes: The initial features of the image pair are input into the connection block, and the feature maps of the two images are connected by a cat function to obtain a feature map.
6. The multi-class change detection method for remote sensing images applicable to imbalanced categories as claimed in claim 1, characterized in that: The step of performing mask processing on the feature map to obtain the masked feature map includes: Applying a change mask obtained by two-class change detection to the feature map; The feature map and the change mask are subjected to matrix multiplication processing, and the unchanged background area value is replaced with 0 to obtain the masked feature map.
7. The multi-class change detection method for remote sensing images applicable to imbalanced categories as claimed in claim 1, characterized in that: The step of inputting the masked feature map into a decoder to calculate the probability of change categories and obtain a change probability map comprises: The decoder upsamples the masked feature map, connects the absolute values of the differences of the two branches in the encoder through jump connections to supplement the more abstract information after encoding, and restores the object details and spatial dimensions of the masked feature map through deconvolution to obtain the change probability map after the decoder.
8. The remote sensing image multi-class change detection method applicable to imbalanced classes as claimed in claim 2, characterized in that: The step of obtaining a hierarchical sequence of training samples and training the change detection framework by using the hierarchical sequence of training samples comprises: Obtain a stratified sequence of training samples; Calculate the cross loss function; Optimizing model parameters of the change detection framework according to the cross loss function and the back propagation process; After the training is completed, the trained change detection framework is obtained.
9. The remote sensing image multi-class change detection method applicable to imbalanced classes as claimed in claim 2, characterized in that: The step of discriminating the change probability map by using the trained change detection framework and outputting multiple types of change detection result maps comprises: The change probability map is discriminated by using a trained change detection framework to obtain a category label of each pixel in the image and obtain a change category feature map; The change category feature map is cropped and restored to the spatial size of the original image to obtain the multi-category change detection result map.
10. A remote sensing image multi-class change detection device suitable for imbalanced categories, characterized in that: Includes the following modules: A construction module is used to construct a change detection framework, wherein the change detection framework includes an encoder and a decoder, wherein the encoder includes two weight-sharing network branches; The acquisition module is used to obtain dual-phase high spatial resolution remote sensing images before and after the change; An extraction module, used to input the dual-phase high spatial resolution remote sensing images before and after the change into the two weight-sharing network branches respectively, and extract the initial features of the image pairs respectively; A connection module, used for connecting the initial features of the image pair through a connection block to obtain a feature map; A mask module, used for performing mask processing on the feature map to obtain a masked feature map; A calculation module, used for inputting the masked feature map into the decoder to calculate the probability of change category and obtain a change probability map; A training module, used for obtaining training samples through a global stratified sampling mechanism, training the change detection framework through the training samples, and obtaining a trained change detection framework after the training is completed; The discrimination module is used to discriminate the change probability map through the trained change detection framework and output multiple types of change detection result maps.
Citation Information
Patent Citations
Remote sensing image change detection method based on dense connection and geometric structure constraint
CN111222453A
Small sample change detection method based on multi-scale feature extraction
CN112668494A
Remote sensing image change detection method based on twinborn multi-scale difference feature fusion
CN113420662A
