A semi-supervised high-resolution remote sensing image change detection method based on label propagation

Abstract

The application discloses a semi-supervised high-resolution remote sensing image change detection method based on label expansion, comprising the following steps: setting an input data set; constructing an encoder-decoder change detection model; constraining the consistency of the prediction results between weak enhancement and strong enhancement in the construction of the encoder-decoder change detection model; expanding pseudo labels through a position interaction graph; optimizing data through supervised loss and unsupervised loss; extracting features of the input data set through an encoder to obtain differential features; and obtaining a change detection probability graph through decoding of the differential features by a decoder. The application proposes a consistency regularization framework of one weak and two strong, and constrains the consistency of the prediction results between weak enhancement and strong enhancement and between the two strong enhancements. The position interaction graph is introduced, the global-local relationship between pixels is utilized, and the internal consistency of pseudo labels is mined, so that the model precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of image detection technology, and in particular to a semi-supervised high-resolution remote sensing image change detection method based on label expansion. Background Art

[0002] Remote sensing image change detection involves analyzing temporal differences in remote sensing images of the same geographic area acquired at different times to identify changes within the region and the type and extent of these changes. This task primarily relies on comparing dual-temporal or long-term remote sensing imagery to dynamically monitor and assess the target area over a specific time period. High-resolution remote sensing image change detection methods based on semi-supervised learning introduce a semi-supervised learning strategy to the change detection task. By collaboratively training a small number of labeled samples with a large number of unlabeled samples, the model's generalization and change detection performance are enhanced in the absence of sufficient labeling. The resulting model can perform a binary classification of the study area as either changed or unchanged, enabling both qualitative judgment and quantitative analysis. This technology has broad application value in a variety of practical scenarios, including urban planning, dynamic land use monitoring, natural disaster assessment, and forest resource change analysis, providing strong data support for geographic information systems and sustainable development decision-making.

[0003] In recent years, with the rapid development of deep learning and the continuous improvement in the performance and number of remote sensing satellites, the use of change detection technology to extract valuable information from massive amounts of remote sensing data has become increasingly important. Currently, many studies focus on how to effectively extract changed areas in dual-temporal remote sensing imagery. Based on the use of labels in the change detection training set, remote sensing image change detection can be mainly divided into the following four categories: supervised change detection methods that use pixel-level labels; unsupervised change detection methods that do not require pixel-level labels; weakly supervised change detection methods based on image-level labels; and semi-supervised change detection methods that combine a small amount of pixel-level labels with a large amount of unlabeled data. Semi-supervised change detection methods have gained widespread attention in recent years due to their ability to effectively combine a small amount of labeled data with a large amount of unlabeled data, demonstrating unique advantages in the absence of labeled data.

[0004] Currently, researchers have proposed numerous excellent methods in the field of semi-supervised change detection, demonstrating promising results in utilizing unlabeled data and improving the performance of change detection models. These methods can be roughly categorized into two types: GAN-based semi-supervised remote sensing image change detection methods and consistency regularization-based semi-supervised remote sensing image change detection methods. The former utilizes generative adversarial networks to capture the underlying distribution of unlabeled data, thereby generating features of the changed regions. The latter significantly improves the robustness of the model and the efficiency of utilizing unlabeled data by maintaining consistent predictions under different augmentation conditions.

[0005] With the continuous development of remote sensing technology and semi-supervised learning technology, semi-supervised change detection methods based on consistency regularization have been widely used. However, existing methods usually rely solely on thresholds to filter pseudo-labels during the consistency regularization process. For example, the FPA method uses the alignment between features and prediction results to utilize a large amount of unlabeled data, and the ST-RCL method utilizes unlabeled data by rotating and unrotating. However, these methods often rely solely on fixed thresholds to filter pseudo-labels in the subsequent self-training process, ignoring the spatial relationship between pixels, resulting in incomplete pseudo-labels. As the training progresses, erroneous pseudo-labels will continue to accumulate, making it impossible for the model to fully extract effective feature information from unlabeled data, further affecting the performance of the change detection model. Summary of the Invention

[0006] The present invention provides a semi-supervised change detection method for high-resolution remote sensing images based on label expansion. This method addresses the problem in existing techniques where consistency regularization often relies on threshold filtering of pseudo-labels, ignoring the spatial relationships between pixels and resulting in incomplete pseudo-labels. The accumulation of erroneous pseudo-labels prevents the model from fully extracting valid feature information from unlabeled data, further impacting the performance of the change detection model.

[0007] On the one hand, an embodiment of the present invention provides a semi-supervised high-resolution remote sensing image change detection method based on label expansion, comprising:

[0008] Setting an input dataset, wherein the input dataset includes a labeled dataset and an unlabeled dataset;

[0009] Build an encoder-decoder change detection model;

[0010] Constraining the consistency of prediction results between weak enhancement and strong enhancement in constructing the encoder-decoder change detection model;

[0011] Constraining the consistency of prediction results between the two strong enhancements in constructing the encoder-decoder change detection model;

[0012] Extending pseudo labels through position interaction graphs in constructing the encoder-decoder change detection model;

[0013] Performing data optimization via supervised and unsupervised losses in constructing the encoder-decoder change detection model;

[0014] Extracting the input data set through the encoder of the encoder-decoder change detection model to perform feature extraction to obtain differential features;

[0015] The differential features are decoded by a decoder of the encoder-decoder change detection model to obtain a change detection probability map.

[0016] In one possible implementation, constraining the consistency of prediction results between weak enhancement and strong enhancement in constructing the encoder-decoder change detection model and constraining the consistency of prediction results between two strong enhancements in constructing the encoder-decoder change detection model include:

[0017] Performing a weak enhancement operation on the unlabeled dual-time remote sensing image of the unlabeled dataset to obtain a weakly enhanced image;

[0018] Copying the weakly enhanced image three times;

[0019] Predicting the first weakly enhanced image using a change detection network model to obtain a weakly enhanced image change probability map;

[0020] Creating the pseudo label for the weakly enhanced image change probability map;

[0021] performing a strong enhancement operation on the second and third weakly enhanced images to obtain a strongly enhanced image;

[0022] Predicting the strongly enhanced image using the change detection network model to obtain a strongly enhanced image change probability map;

[0023] The pseudo labels constrain the consistency of the prediction results of the change detection network model for the weakly enhanced image and the strongly enhanced image.

[0024] In a possible implementation, the weak enhancement operation is to simultaneously perform data enhancement on the unlabeled dual-time remote sensing image;

[0025] The data augmentation is random flipping.

[0026] In a possible implementation, the strong enhancement operation is to perform random enhancement on each of the unlabeled dual-time remote sensing images respectively;

[0027] The random enhancement includes color jittering and Gaussian blur.

[0028] In one possible implementation, performing data optimization using supervised loss and unsupervised loss in constructing the encoder-decoder change detection model includes:

[0029] Calculating the supervised loss of the encoder-decoder change detection model by calculating a cross-entropy loss and updating network weights through backpropagation;

[0030] Calculating the unsupervised loss of the encoder-decoder change detection model by calculating a cross entropy loss and a position interaction loss;

[0031] Obtaining a total loss function based on the calculation of the supervised loss and the unsupervised loss;

[0032] The encoder-decoder change detection model is data optimized according to the total loss function.

[0033] In a possible implementation, extracting the input data set by an encoder of the encoder-decoder change detection model to perform feature extraction to obtain differential features includes:

[0034] A ResNet-50-based twin encoder is used to extract feature information of the dual-temporal high-resolution remote sensing image of the input dataset;

[0035] The feature information extracted by the twin encoder through a four-layer structure;

[0036] The differential features are composed of the feature information at different levels.

[0037] In one possible implementation, decoding the differential features by a decoder of the encoder-decoder change detection model to obtain a change detection probability map includes:

[0038] The decoder is composed of a convolutional layer and a dilated spatial pyramid pooling module;

[0039] The decoder decodes the feature information of different levels of the differential feature to obtain the change detection probability map.

[0040] The semi-supervised high-resolution remote sensing image change detection method based on label expansion in the present invention has the following advantages:

[0041] (1) A one-weak-two-strong consistency regularization framework is proposed to constrain the consistency of the prediction results between weak enhancement and strong enhancement, as well as between two strong enhancements.

[0042] (2) We introduce a position interaction graph and use the global-local relationship between pixels to mine the inherent consistency of pseudo labels, thereby improving model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 A flowchart of a semi-supervised high-resolution remote sensing image change detection method based on label expansion provided by an embodiment of the present application;

[0045] Figure 2 A semi-supervised high-resolution remote sensing image change detection framework diagram of a semi-supervised high-resolution remote sensing image change detection method based on label expansion provided by an embodiment of the present application;

[0046] Figure 3 A network structure diagram of a change detection model of a semi-supervised high-resolution remote sensing image change detection method based on label expansion provided by an embodiment of the present application;

[0047] Figure 4 A comparison diagram of visual results on a LEVIR-CD data set of a semi-supervised high-resolution remote sensing image change detection method based on label expansion provided by an embodiment of the present application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0049] Figure 1 A flowchart of a semi-supervised high-resolution remote sensing image change detection method based on label expansion provided by an embodiment of the present application; the embodiment of the present application provides a semi-supervised high-resolution remote sensing image change detection method based on label expansion, which comprises:

[0050] An input data set is set, which comprises a labeled data set and an unlabeled data set;

[0051] An encoder-decoder change detection model is constructed;

[0052] The consistency of the prediction results between weak enhancement and strong enhancement is constrained in the construction of the encoder-decoder change detection model;

[0053] The consistency of the prediction results between two strong enhancements is constrained in the construction of the encoder-decoder change detection model;

[0054] The pseudo-labels are expanded by a position interaction graph in the construction of the encoder-decoder change detection model;

[0055] The data is optimized by supervised loss and unsupervised loss in the construction of the encoder-decoder change detection model;

[0056] Extracting the input data set through the encoder of the encoder-decoder change detection model to perform feature extraction to obtain differential features;

[0057] The differential features are decoded by a decoder of the encoder-decoder change detection model to obtain a change detection probability map.

[0058] Constraining the consistency of prediction results between a weak enhancement and a strong enhancement in constructing the encoder-decoder change detection model and constraining the consistency of prediction results between two strong enhancements in constructing the encoder-decoder change detection model include:

[0059] Performing a weak enhancement operation on the unlabeled dual-time remote sensing image of the unlabeled dataset to obtain a weakly enhanced image;

[0060] Copying the weakly enhanced image three times;

[0061] Predicting the first weakly enhanced image using a change detection network model to obtain a weakly enhanced image change probability map;

[0062] Creating the pseudo label for the weakly enhanced image change probability map;

[0063] performing a strong enhancement operation on the second and third weakly enhanced images to obtain a strongly enhanced image;

[0064] Predicting the strongly enhanced image using the change detection network model to obtain a strongly enhanced image change probability map;

[0065] The pseudo labels constrain the consistency of the prediction results of the change detection network model for the weakly enhanced image and the strongly enhanced image.

[0066] The weak enhancement operation is to simultaneously perform data enhancement on the unlabeled dual-time remote sensing image;

[0067] The data augmentation is random flipping.

[0068] The strong enhancement operation is to randomly enhance each image of the unlabeled dual-time remote sensing image;

[0069] The random enhancement includes color jittering and Gaussian blur.

[0070] Data optimization through supervised and unsupervised losses in building the encoder-decoder change detection model includes:

[0071] Calculating the supervised loss of the encoder-decoder change detection model by calculating a cross-entropy loss and updating network weights through backpropagation;

[0072] calculating the encoder-decoder change detection model through the cross-entropy loss calculation and the position interaction loss calculation;

[0073] calculating a total loss function according to the supervised loss and the unsupervised loss;

[0074] optimizing the encoder-decoder change detection model according to the total loss function.

[0075] the feature extraction of the input data set by the encoder of the encoder-decoder change detection model includes:

[0076] adopting a ResNet-50-based twin encoder to extract feature information of the dual-time high-resolution remote sensing image of the input data set;

[0077] the feature information extracted by the twin encoder through a four-layer structure;

[0078] the difference feature is composed of different levels of the feature information.

[0079] decoding the difference feature by the decoder of the encoder-decoder change detection model to obtain a change detection probability map includes:

[0080] the decoder is composed of a convolutional layer and a dilated spatial pyramid pooling module;

[0081] decoding different levels of the feature information of the difference feature by the decoder to obtain the change detection probability map.

[0082] As shown in Figure 1 , 2 the input data set is first split and cropped into 256x256 size pictures without overlapping between each picture, and a total of 10192 pairs of images are obtained. At the same time, the data set is divided into a training set, a validation set and a test set according to a 7:1:2 ratio, and then 80% of the training set is divided into an unlabeled data set and a 20% labeled data set according to the ratio. The labeled data set is specifically represented as , wherein represents the i-th pair of dual-time remote sensing images, represents the image at T1, represents the image at T2, represents the corresponding binary pixel-level label, and n represents the size of the data set. The unlabeled data set is specifically represented as , wherein represents the i-th pair of bitemporal remote sensing images, and N represents the size of the dataset.

[0083] Then, we build an encoder-decoder change detection model, such as Figure 3 As shown in the figure, the network uses a twin encoder based on ResNet-50 to extract the feature information of dual-time high-resolution remote sensing images, and uses a decoder based on DeepLabv3+ to decode the differential feature information extracted by the encoder to generate the final change detection probability map.

[0084] Encoder: The encoder is based on the ResNet-50 network structure, where Layer 1, Layer 2, Layer 3, and Layer 4 correspond to the four main feature extraction stages in ResNet-50. Each layer corresponds to a series of 1×1, 3×3, and 1×1 convolutions. Layer 1 extracts low-level feature information, which mainly includes image boundaries, textures, and contours. Layer 4 extracts high-level feature information and obtains the global information and higher-level semantic information of the image. Specifically, the dual-time remote sensing images at time T1 and T2 are After the four layers of ResNet-50, four layers of feature maps are obtained respectively { , , , }and{ , , , }.in, Represents the feature map of the image at time T1 after being processed by the i-th layer, The feature map representing the image at time T2 after being processed by layer i. Next, the feature differences of Layer 1 and Layer 4 are calculated to obtain the low-level features. and high-level features .

[0085] (1)

[0086] (2)

[0087] Decoder: The decoder is composed of convolutional layers and Atrous Spatial Pyramid Pooling (ASPP) modules. It processes the feature information of different levels extracted by the decoder to obtain the final change detection probability map. The ASPP module contains a 1×1 convolution, three 3×3 dilated convolutions with ratios of 6, 12, and 18, and a global pooling operation. First, the low-level differential feature map After 1x1 convolution operation, the number of channels of features is reduced, which is more conducive to the fusion with subsequent high-level features. Meanwhile, the high-level difference feature maps After the processing of the ASPP module, multi-scale semantic information is captured. Subsequently, the high-level difference feature maps processed by the ASPP module are restored to a higher resolution through upsampling operation. Then, the low-level difference feature maps and the high-level difference feature maps are spliced. In order to further refine the change detection results after fusion, two 3x3 convolution operations are used to process the fused features. After two convolutions, the final output feature map is mapped through a 1x1 convolution kernel to generate a change detection probability map . Here, 2 represents two categories of change and no change. Next, the p is scaled to [0, 1] along the category dimension at the pixel level by using the softmax function.

[0088] (3)

[0089] In the training stage of the encoder-decoder change detection model, according to whether the data set has a label, it is divided into supervised training part and unsupervised training part. The unlabeled dual-time remote sensing images are subjected to weak enhancement (such as random flipping) operation to obtain images, which are then copied into three parts. One part directly passes through the change detection network model (CD network) to obtain the predicted change probability map , After that, pseudo labels are created. The second and third parts are subjected to strong enhancement operations as in formulas 4 and 5, where is a predefined strong data enhancement operation, including random operations of color jitter (ColorJitter) and Gaussian blur (Gaussian Blur). Unlike weak enhancement, strong enhancement is a random enhancement operation on each image to obtain and , that is, The enhancement effect of is not equivalent to , and the enhancement effect of is not equivalent to , which improves the diversity of input images. Subsequently, they are input into the change detection network model (CD network) to obtain the corresponding predicted probability maps and , and then the pseudo labels are used to constrain the prediction results of the two parts.

[0090] (4)

[0091] (5)

[0092] To further improve the effectiveness of pseudo labels, the present application proposes a pseudo label expansion method based on a locational interaction map (LIM). As shown in Figure 3 , first, the high-level features containing semantic information are reduced in dimension, and two 1x1 convolution kernels are used to decompose the feature map to obtain two feature vectors and . Here and are not equal and both contain information about each pixel point in the image. Among them, C 1 represents the feature channel dimension, H 1 W 1 refers to the number of feature vectors. To capture the pixel relationship between different spatial positions, the pixel relationship matrix is calculated using formula 6, and then the locational interaction map is further constructed using formula 7. Among them, T represents the transpose of Z1, D is the scaling quantity of P, and M contains the mutual relationship between each pair of pixel points.

[0093] (6)

[0094] (7)

[0095] To apply the locational interaction map M between pixels to pseudo label expansion, first, the shape of the change probability map output by the model needs to be matched using a bilinear difference operation, and then the predicted result after label expansion is obtained using formula 8 , where represents a bilinear difference operation.

[0096] (8)

[0097] Therefore, according to the model prediction probability result after label expansion and , the locational interaction loss can be calculated, as shown in formula 9.

[0098] (9)

[0099] Among them, represents a fixed threshold value for screening pseudo labels, represents the predicted probability greater than is considered as a high-quality pseudo label and recorded as 1; otherwise, it is set to 0. CE is the cross entropy.

[0100] During the experimental training phase, the optimization of the model relies on the supervision loss and unsupervised loss Among them, the supervision loss Mainly used for learning with labeled data without supervised loss The combination of the two can effectively improve the change detection performance of the model.

[0101] The specific contents are as follows:

[0102] (1) Supervision losses Computation: In the supervised branch, the model uses a labeled dataset By comparing with the true labels, the model can learn the difference between the changed and non-changed areas. Specifically, the output prediction result of the model is With label The cross-entropy loss (CE) is calculated accordingly, and the network weights are updated through backpropagation, so that the model can gradually improve its ability to detect changed areas. The loss used in the supervision part can be expressed as:

[0103] (10)

[0104] Among them, H and W represent the height and width of the input image respectively. express and The corresponding spatial pixel position.

[0105] (2) Unsupervised loss Calculation: The basic loss function of the unsupervised part still uses cross-entropy loss (CE), which can be expressed as:

[0106] (11)

[0107] Here, here Represents a fixed threshold used to filter pseudo labels. Indicates that the predicted probability is greater than , which is considered as a high-quality pseudo label and recorded as 1; otherwise, it is set to 0.

[0108] Unsupervised loss By cross entropy loss and position interaction loss Composition, which is calculated as follows:

[0109] (12)

[0110] Among them, the cross entropy loss According to formula 11, this loss directly measures the prediction results of the model output and pseudo labels In semi-supervised learning, the quality of pseudo labels determines the effectiveness of unsupervised training. In order to improve the reliability of pseudo labels, this paper adopts a strong enhancement strategy and calculates the cross entropy loss under two enhancement methods. and , and finally take the average value:

[0111] (13)

[0112] (3) Position interaction loss Calculation: Position Interaction Loss The purpose of the calculation based on Formula 9 is to optimize the quality of pseudo labels by utilizing the spatial relationship between pixels. Since the changes in remote sensing images are usually regionally consistent, relying solely on pixel-level confidence screening may result in low-quality pseudo labels. To overcome this problem, this paper introduces the Locational Interaction Map (LIM) to further expand the scope of pseudo labels by calculating the spatial relationship between pixels. When calculating the location interaction loss, the LIM between pixels is first constructed and applied to the model prediction probability map. , get the updated prediction probability map Then, according to the three branches of the unsupervised part, the corresponding three position interaction losses can be obtained: 、 and The final loss calculation formula is as follows:

[0113] (14)

[0114] Therefore, the final total loss function L is composed of the supervision loss and unsupervised loss The calculation formula is as follows:

[0115] (15)

[0116] In one possible embodiment, in order to accurately evaluate the performance of the proposed semi-supervised change detection method, five indicators including IoU, F1, Kappa, TPR and TNR are selected for evaluation.

[0117] The detailed explanation is as follows:

[0118]

[0119]

[0120]

[0121]

[0122]

[0123]

[0124]

[0125]

[0126]

[0127] In the above formula, the basic calculation units TP, TN, FP and FN represent the correct prediction of changed pixels, the correct prediction of unchanged pixels, the unchanged pixels being wrongly predicted as changed pixels, and the changed pixels being wrongly predicted as unchanged pixels, respectively. OA represents the overall classification accuracy of the model. IoU can reflect the degree of overlap between the predicted region and the real region. In the experiment, the IoU of the changed pixels is used instead of the global IoU. c F1 evaluates the overall performance of the model. Kappa evaluates the consistency between the model and the actual observation. TPR evaluates the ability of the model to predict changed pixels, i.e. sensitivity. TNR evaluates the ability of the model to predict unchanged pixels, i.e. specificity. The higher the above five average indicators, the better the performance of the model.

[0128] In one possible embodiment, simulation experiments are carried out.

[0129] (1) Simulation conditions:

[0130] A single NVIDIA GeForce GTX 3080Ti GPU with 12G memory is used for training and testing. In the experiment, the batchsize is set to 4, and the LEVIR-CD dataset is used for the dataset.

[0131] (2) Simulation content:

[0132] In order to prove the effectiveness of the present application in improving the performance of semi-supervised high-resolution remote sensing image change detection, it is compared with six excellent semi-supervised change detection methods in recent years. The selected comparison methods include S4GAN, SemiCDNet, SemiCD, FPA, Unimatch and C 2 F 2In addition, the effects of supervised training using only a small amount of labels (Only-sup) and supervised training using all label data (Fully-sup) are compared to comprehensively evaluate the advantages of the method in change detection. To further illustrate the utilization of unlabeled data sets, the experimental results under 1%, 5%, 10% and 20% labeled data set proportions are compared. As shown in Tables 1, 2, 3 and 4.

[0133] Table 1 Results of different methods on LEVIR-CD test set under 1% labeled data set

[0134]

[0135] Table 2 Results of different methods on LEVIR-CD test set under 5% labeled data set

[0136]

[0137] Table 3 Results of different methods on LEVIR-CD test set under 10% labeled data set

[0138]

[0139] Table 4 Results of different methods on LEVIR-CD test set under 20% labeled data set

[0140]

[0141] From Tables 1-4, the effectiveness of the present application under different proportions of labeled training sets can be seen, which is verified on the LEVIR-CD data set. In particular, from Table 4, it can be observed that under the semi-supervised training setting using only 20% labeled data and 80% unlabeled data, the proposed method can still achieve a performance close to that obtained using 100% labeled data in full supervision (Fully-sup), further proving its practicality and robustness under low labeling resource conditions. From the qualitative analysis point of view, as shown in the detection results of different methods on the LEVIR-CD data set: (a) T1 time image, (b) T2 time image, (c) ground truth, (d) S4GAN, (e) SemiCDNet, (f) SemiCD, (g) FPA, (h) Unimatch, (i) C Figure 4 2 F 2 , (j) the method of the present application. Among them, the technical solution of the present application shows strong smoothness and accuracy in processing the boundary of the change area, and can more accurately identify the real boundary of the change area.

[0142] ​While the preferred embodiments of the application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they have the benefit of the present disclosure without departing from the spirit and scope of the application. Accordingly, it is intended that the appended claims include all such modifications and variations as fall within the scope of the present application.

[0143] It is apparent that those skilled in the art can make various changes and modifications to the application without departing from the spirit and scope of the application. It is therefore intended that the present application cover all such changes and modifications that are within its scope.

Claims

1. A semi-supervised high-resolution remote sensing image change detection method based on label expansion, characterized in that: include: Setting an input dataset, wherein the input dataset includes a labeled dataset and an unlabeled dataset; Build an encoder-decoder change detection model; Constraining the consistency of prediction results between weak enhancement and strong enhancement in constructing the encoder-decoder change detection model; Constraining the consistency of prediction results between the two strong enhancements in constructing the encoder-decoder change detection model; wherein, performing a weak enhancement operation on the unlabeled dual-time remote sensing image of the unlabeled dataset to obtain a weakly enhanced image; Copying the weakly enhanced image three times; Predicting the first weakly enhanced image using a change detection network model to obtain a weakly enhanced image change probability map; Creating a pseudo label for the weakly enhanced image change probability map; performing a strong enhancement operation on the second and third weakly enhanced images to obtain a strongly enhanced image; Predicting the strongly enhanced image using the change detection network model to obtain a strongly enhanced image change probability map; Constraining the consistency of prediction results of the change detection network model for the weakly enhanced image and the strongly enhanced image through the pseudo labels; Extending pseudo labels through position interaction graphs in constructing the encoder-decoder change detection model; The step of extending the pseudo labels by using the position interaction graph in constructing the encoder-decoder change detection model includes: High-level features containing semantic information Perform dimensionality reduction; Use two 1×1 convolution kernels to decompose the feature map to obtain two feature vectors and , and Not equal, and Contains the information of each pixel in the feature map, where C 1 represents the feature channel dimension, H 1 W 1 refers to the number of eigenvectors; According to the formula Calculate pixel relationship matrix ; By formula The position interaction graph is constructed as , where T represents the transposition of Z1, D is the scaling of P, and M contains the relationship between the two pixels; performing shape matching on the change probability map output by the encoder-decoder change detection model through a bilinear interpolation operation; By formula Get the prediction results after label expansion ,in Represents a bilinear difference operation; According to the formula and the predicted probability results of the encoder-decoder change detection model after the label expansion and Calculate position interaction loss ,in, Represents a fixed threshold used to filter pseudo labels. Indicates that the predicted probability is greater than is considered as a high-quality pseudo label and recorded as 1; otherwise, it is set to 0. CE is the cross entropy; Performing data optimization via supervised and unsupervised losses in constructing the encoder-decoder change detection model; Extracting the input data set through the encoder of the encoder-decoder change detection model to perform feature extraction to obtain differential features; The differential features are decoded by a decoder of the encoder-decoder change detection model to obtain a change detection probability map.

2. The semi-supervised high-resolution remote sensing image change detection method based on label expansion according to claim 1 is characterized in that: The weak enhancement operation is to simultaneously perform data enhancement on the unlabeled dual-time remote sensing image; The data augmentation is random flipping.

3. The semi-supervised high-resolution remote sensing image change detection method based on label expansion according to claim 1 is characterized in that: The strong enhancement operation is to randomly enhance each image of the unlabeled dual-time remote sensing image; The random enhancement includes color jittering and Gaussian blur.

4. The semi-supervised high-resolution remote sensing image change detection method based on label expansion according to claim 1 is characterized in that: Data optimization through supervised and unsupervised losses in building the encoder-decoder change detection model includes: Calculating the supervised loss of the encoder-decoder change detection model by calculating a cross-entropy loss and updating network weights through backpropagation; Calculating the unsupervised loss of the encoder-decoder change detection model by calculating a cross entropy loss and a position interaction loss; Obtaining a total loss function based on the calculation of the supervised loss and the unsupervised loss; The encoder-decoder change detection model is data optimized according to the total loss function.

5. The semi-supervised high-resolution remote sensing image change detection method based on label expansion according to claim 1 is characterized in that: The extracting the input data set by the encoder of the encoder-decoder change detection model to perform feature extraction to obtain differential features includes: A ResNet-50-based twin encoder is used to extract feature information of the dual-temporal high-resolution remote sensing image of the input dataset; The feature information extracted by the twin encoder through a four-layer structure; The differential features are composed of the feature information at different levels.

6. The semi-supervised high-resolution remote sensing image change detection method based on label expansion according to claim 5 is characterized in that: Decoding the differential features by the decoder of the encoder-decoder change detection model to obtain a change detection probability map includes: The decoder is composed of a convolutional layer and a dilated spatial pyramid pooling module; The decoder decodes the feature information of different levels of the differential feature to obtain the change detection probability map.

Images