Semi-supervised remote sensing image change detection method and system based on neighborhood relation propagation

Abstract

This invention discloses a semi-supervised remote sensing image change detection method and system based on neighborhood relationship propagation, belonging to the field of computer vision and remote sensing image processing technology. It includes: acquiring dual-temporal remote sensing image pairs and dividing them into labeled and unlabeled datasets; constructing a teacher-student framework semi-supervised model, calculating the squared Euclidean distance between the center pixel and its neighboring pixels, converting it into a relevance weight matrix using Softmax, and performing neighborhood-weighted aggregation on the initial prediction map to obtain a corrected prediction map; calculating the supervised loss using labeled samples, and calculating the unsupervised loss using unlabeled samples through weak perturbation to generate pseudo-labels and strong perturbation for consistency learning; jointly optimizing the student model parameters and updating the teacher model using exponential moving average; inputting the dual-temporal images into the trained student model, outputting a probability map, and generating a binary change detection mask. This invention enhances spatial consistency through neighborhood relationship propagation, improving detection performance when labels are limited.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and remote sensing image processing technology, specifically to a semi-supervised remote sensing image change detection method and system based on neighborhood relationship propagation. Background Technology

[0002] Change detection aims to identify pixel-level changes between two images of the same geographical location taken at different times. This task has significant application value in fields such as natural disaster management, urban sprawl monitoring, and environmental monitoring. In recent years, with the rapid development of deep learning technology, supervised learning change detection methods based on convolutional neural networks (CNNs) and Transformers have emerged in large numbers and demonstrated excellent performance in various benchmark tests. However, the success of these supervised learning methods largely depends on large-scale, high-quality pixel-level labeled datasets. In the field of remote sensing image processing, obtaining accurate pixel-level change detection labels usually requires a significant investment of manpower and time, while unlabeled remote sensing image data is relatively abundant and readily available. To alleviate the dependence on large amounts of labeled data, semi-supervised learning methods that can simultaneously utilize a small amount of labeled data and a large amount of unlabeled data have become a current research hotspot.

[0003] Existing semi-supervised change detection methods mainly include network architectures specifically designed for change detection and methods improved based on general semi-supervised frameworks (such as consistency regularization). Among them, methods based on consistency regularization show outstanding performance by using pseudo-labels generated from weakly perturbed images to guide the learning of strongly perturbed images. Nevertheless, current semi-supervised change detection techniques based on consistency regularization still have significant limitations. Specifically, most existing methods ignore the spatial dependencies between pixels, i.e., the fact that adjacent pixels often have similar change attributes. Most existing schemes treat pixels in the image as independent individuals, directly generating prediction results from the difference feature map. This isolated processing severs the structural consistency and spatial context of the local image. In addition, change detection is essentially a comparative task involving accurate alignment and difference analysis of two-phase images, while existing methods often directly adopt general semi-supervised classification strategies without optimizing for this characteristic of change detection. This makes it difficult for the model to effectively capture the correlation between pixels and their neighboring pixels in terms of change patterns. This not only reduces the accuracy of the model when detecting small changes, but also easily leads to problems such as spatial discontinuity, blurred boundaries, and high false alarm rates in the prediction results. Summary of the Invention

[0004] The purpose of this invention is to provide a semi-supervised remote sensing image change detection method and system based on neighborhood relationship propagation, so as to solve the problems of poor spatial consistency and limited detection accuracy caused by ignoring pixel neighborhood correlation in the existing technology.

[0005] To achieve the above objectives, the technical solution provided by this invention is: a semi-supervised remote sensing image change detection method based on neighborhood relationship propagation, comprising the following steps: S1: Obtain bi-temporal remote sensing image pairs of the land cover change area, construct a training dataset and divide it into a labeled dataset and an unlabeled dataset; the labeled dataset includes bi-temporal remote sensing image pairs and their corresponding pixel-level real change labels, while the unlabeled dataset only includes bi-temporal remote sensing image pairs. S2: Construct a semi-supervised change detection model based on a teacher-student framework, in which both the student model and the teacher model include a shared encoder, feature decoder, and prediction head. The student model also includes a neighborhood relationship propagation module. S3: In the labeled dataset, the dual-temporal remote sensing image pairs are input into the student model. The initial prediction map is output through the shared encoder, feature decoder, and prediction head. Then, the neighborhood relationship propagation module outputs the corrected prediction map. The neighborhood relationship propagation module calculates the squared Euclidean distance between the center pixel and its neighboring pixels based on the differential features of the dual-temporal remote sensing image pairs. The distance is converted into a correlation weight matrix through the Softmax function. The initial prediction map is then subjected to neighborhood weighted aggregation to generate the corrected prediction map. The supervision loss between the initial prediction map, the corrected prediction map, and the pixel-level true change label is calculated. In the unlabeled dataset, weak perturbations are applied to pairs of bi-temporal remote sensing images and then input into the teacher model to obtain a predicted probability map. After filtering by a preset confidence threshold, pseudo-labels are generated. Different strong perturbations are applied to the same pair of bi-temporal remote sensing images and then input into the student model to obtain the corresponding initial prediction map and corrected prediction map. The unsupervised consistency loss between each initial prediction map, corrected prediction map and pseudo-label is calculated. The total loss function is set as the sum of the supervised loss and the unsupervised consistency loss. The semi-supervised change detection model is trained to obtain the trained student model. S4: Input the dual-temporal remote sensing image to be detected into the trained student model, output a pixel-level change probability map, and then process it to obtain the spatial location information of the change area of ​​ground features.

[0006] To optimize the above technical solution, the specific measures also include: In step S1, the construction of the training dataset and its division into labeled and unlabeled datasets specifically involves: Building a training dataset The training dataset is divided into labeled datasets. and unlabeled datasets ;in, Indicates the first Remote sensing images of each sample acquired in the first time phase; Indicates the first Remote sensing images of the samples acquired in the second time phase; Indicates the first Each labeled sample has a pixel-level ground real change label, where the changed area is labeled as the foreground and the unchanged area is labeled as the background. This represents the total number of labeled training samples; Indicates the first element in the unlabeled dataset. Remote sensing images of each sample acquired in the first time phase; Indicates the first element in the unlabeled dataset. Remote sensing images of the samples acquired in the second time phase; This represents the total number of unlabeled training samples.

[0007] In step S2, the shared encoder is a pre-trained visual feature extraction network that extracts multi-scale semantic features of dual-temporal remote sensing image pairs; the feature decoder is a decoder based on multi-level feature fusion, which obtains multi-level differential features by performing differential operations on the multi-scale semantic features output by the shared encoder, and then fuses them through upsampling and skip connections to generate a high-resolution differential map.

[0008] Further, in step S3, the specific process of calculating the squared Euclidean distance between the center pixel and its neighboring pixels based on the differential features of the dual-temporal remote sensing image pair is as follows: The shared encoder of the student model extracts bi-temporal images respectively. and Multiscale semantic features and Then, the absolute differences of the corresponding features are calculated to obtain the difference features. , to differentiate features The two parallel convolutional branches are mapped to the central features respectively. and neighborhood features Set the sliding window size to The feature map is unfolded in the spatial dimension to obtain the feature vector of the center pixel. and neighborhood pixel feature vectors Square Euclidean distance The calculation formula is: ; in, Indicates the first In the nth sample, the nth The central pixel and its neighborhood of the first... The squared Euclidean distance between pixels in the feature space; For batch indexing; This represents the number of feature channels.

[0009] In step S3, the squared Euclidean distance is converted into a relevance weight matrix using the Softmax function. This matrix is ​​then used to perform neighborhood weighted aggregation on the initial prediction map, generating a revised prediction map. The specific process is as follows: Using negative The function converts the squared Euclidean distance into a normalized correlation weight matrix. The formula is: ; Furthermore, the initial prediction map is spatially expanded to the same size as the neighborhood window to obtain the neighborhood prediction representation. Using the correlation weight matrix The weighted aggregation expression is: ; in, For the first In the nth sample, the nth The central pixel and its neighborhood of the first... The squared Euclidean distance between pixels in the feature space; To adjust the sliding window size; This indicates that after neighborhood-weighted aggregation, the first... In the nth sample Category in Updated predicted values ​​for each pixel position; In the initial prediction graph output by the student model, the first... In the nth sample Category in the first Within the neighborhood of the nth center pixel as reference The predicted value for each pixel position.

[0010] In step S3, the weak perturbations applied to the unlabeled dataset include random flipping and random cropping, while the strong perturbations applied to the unlabeled dataset include random flipping, random cropping, color perturbation, and random Gaussian noise.

[0011] In step S3, the total loss function is set as the sum of the supervised loss and the unsupervised consistency loss, and the specific expression is as follows: Total loss function : ; Monitoring losses : ; Furthermore, unsupervised consistency loss : ; in, and These are the batch sizes with and without labels, respectively. Cross-entropy loss; and These are the initial prediction map and the corrected prediction map for labeled samples, respectively. This is a real label; and The initial prediction plots are shown under two strongly disturbed views. and This is the corresponding corrected prediction graph; A pixel-level predicted probability map with weak perturbations output by the teacher model; These are pseudo-tags; The confidence threshold; This is an indicator function.

[0012] Further, in step S3, the process of generating pseudo-labels after filtering based on a preset confidence threshold is as follows: The teacher model makes predictions on pairs of bi-temporal remote sensing images with weak perturbations, resulting in pixel-level prediction probability maps. Each pixel location contains the predicted probability distribution for each category; the predicted probability map... By selecting the maximum value along the category dimension, the maximum predicted probability value for each pixel location is obtained. and the corresponding category index; compare the maximum predicted probability value with the preset confidence threshold. When comparing, The corresponding category index is used as the pseudo-label for that pixel. ;when If the pixel is invalid, it will be marked as an invalid pixel and will not be included in the subsequent unsupervised consistency loss calculation.

[0013] As another important technical solution, the present invention also provides a semi-supervised remote sensing image change detection system based on neighborhood relationship propagation, used to implement the disclosed semi-supervised remote sensing image change detection method based on neighborhood relationship propagation, including: The dual-temporal dataset construction module is used to acquire dual-temporal remote sensing image pairs of areas of land cover change, construct training datasets and divide them into labeled datasets and unlabeled datasets; the labeled datasets include dual-temporal remote sensing image pairs and their corresponding pixel-level real change labels, while the unlabeled datasets only include dual-temporal remote sensing image pairs. The teacher-student semi-supervised change detection model construction module is used to build a semi-supervised change detection model based on the teacher-student framework. Both the student model and the teacher model include a shared encoder, feature decoder, and prediction head. The student model also includes a neighborhood relationship propagation module. A semi-supervised joint training module is used to input bi-temporal remote sensing image pairs into the student model on labeled datasets. The model outputs an initial prediction map via a shared encoder, feature decoder, and prediction head. A neighborhood relationship propagation module then outputs a corrected prediction map. This module calculates the squared Euclidean distance between the center pixel and its neighboring pixels based on the differential features of the bi-temporal remote sensing image pairs. This distance is then converted into a relevance weight matrix using a Softmax function. The initial prediction map is then weighted and aggregated to generate the corrected prediction map. Finally, the supervised loss between the initial and corrected prediction maps and the pixel-level true change labels is calculated. In the unlabeled dataset, weak perturbations are applied to pairs of bi-temporal remote sensing images and then input into the teacher model to obtain a predicted probability map. After filtering by a preset confidence threshold, pseudo-labels are generated. Different strong perturbations are applied to the same pair of bi-temporal remote sensing images and then input into the student model to obtain the corresponding initial prediction map and corrected prediction map. The unsupervised consistency loss between each initial prediction map, corrected prediction map and pseudo-label is calculated. The total loss function is set as the sum of the supervised loss and the unsupervised consistency loss. The semi-supervised change detection model is trained to obtain the trained student model. Probability of change Figure 2 The value-based output module is used to input the dual-temporal remote sensing image to be detected into the trained student model, output a pixel-level change probability map, and then process it to obtain the spatial location information of the change area of ​​ground features.

[0014] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements a semi-supervised remote sensing image change detection method based on neighborhood relationship propagation as described above.

[0015] The present invention also proposes a computer-readable storage medium storing a computer program that enables a computer to execute a semi-supervised remote sensing image change detection method based on neighborhood relation propagation as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention incorporates a neighborhood relationship propagation module in the student model. It calculates the squared Euclidean distance between the center pixel and its neighboring pixels using the difference features of dual-temporal images, and transforms the distance into a correlation weight using the Softmax function. This performs neighborhood-weighted aggregation on the initial prediction map, enabling the prediction results of each pixel to incorporate its local neighborhood information. This enhances the spatial consistency of the changed areas and helps to obtain more continuous and clearer change detection results.

[0017] This invention employs a teacher-student semi-supervised framework. The teacher model generates pseudo-labels for weakly perturbated images, while the student model performs consistency learning on multiple strongly perturbated versions of the same image. The neighborhood relationship propagation module applies spatial constraints to the supervision signals of the pseudo-labels, effectively utilizing unlabeled data to assist model training. Even with limited labeled samples, it maintains good detection performance, thereby reducing the workload of manual labeling.

[0018] Comparative experiments on multiple publicly available change detection datasets show that, compared with existing semi-supervised change detection methods, this invention achieves a higher Intersection over Union (IoU) ratio under different annotation ratios during training settings. In particular, the performance advantage is more obvious when the proportion of labeled samples is low. This invention has stronger generalization ability and robustness in scenarios where annotations are scarce.

[0019] The method proposed in this invention can be implemented based on commonly used encoders (such as pre-trained visual feature extraction networks) and decoder structures, without relying on a specific network backbone, and has good versatility and scalability. Attached Figure Description

[0020] Figure 1 : A schematic diagram of the structure of the present invention.

[0021] Figure 2 : A schematic diagram of the process in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below through specific embodiments, but it should not be construed as limiting the scope of the subject matter of the present invention to the following embodiments. All technologies implemented based on the above content of the present invention fall within the scope of the present invention.

[0023] In some implementations, such as Figure 2 As shown, this invention provides a semi-supervised remote sensing image change detection method based on neighborhood relationship propagation, comprising the following steps: S1: Obtain two-temporal remote sensing image pairs of the land cover change area, construct a training dataset, and divide the original dataset into labeled dataset and unlabeled dataset according to a preset ratio; wherein, the labeled dataset includes two-temporal remote sensing image pairs and their corresponding pixel-level real change labels, and the unlabeled dataset only includes two-temporal remote sensing image pairs. The dual-temporal remote sensing images were uniformly cropped and normalized to construct a training dataset. The training dataset is divided into labeled datasets. and unlabeled datasets ;in, Indicates the first Remote sensing images of each sample acquired in the first time phase; Indicates the first Remote sensing images of the samples acquired in the second time phase; Indicates the first Each labeled sample has a pixel-level ground real change label, and the changed areas are labeled as foreground and the unchanged areas are labeled as background. This represents the total number of labeled training samples; Indicates the first element in the unlabeled dataset. Remote sensing images of each sample acquired in the first time phase; Indicates the first element in the unlabeled dataset. Remote sensing images of the samples acquired in the second time phase; This represents the total number of unlabeled training samples.

[0024] S2: Construct a semi-supervised change detection model based on a teacher-student framework, wherein both the student model and the teacher model structure include a shared encoder, a feature decoder, and a prediction head; the student model also includes a neighborhood relationship propagation module to assist supervised learning and consistency constraints. In some implementations, the shared encoder is a pre-trained visual feature extraction network, DinoV2, that extracts bi-temporal images. and Multiscale semantic features and , , Represents the encoder's first Layered output; the feature decoder adopts a DPT-style layered dense prediction decoding structure, and combines multi-scale feature projection, resolution alignment, and progressive refinement fusion mechanisms to achieve dense prediction. Layer-by-layer differencing operations are performed on the multi-scale semantic features output by the shared encoder to obtain multi-level differencing features. Subsequently, the feature decoder constructs a bottom-up, progressively upsampling path and combines it with cross-layer skip connections to perform layer-by-layer fusion and reconstruction of the differential features at each level, ultimately generating a high-resolution differential map.

[0025] The multi-level features output by the shared encoder are first channel-mapped using 1×1 convolutions to achieve cross-layer feature space alignment. Explicit spatial resampling of features at different scales is then performed through upsampling or downsampling operations, enabling feature fusion across layers within a unified resolution space. In the feature integration stage, a top-down, hierarchical decoding path is introduced. High-level semantic features are progressively upsampled and then fused with encoder features at corresponding scales via cross-layer skip connections. At each fusion stage, a RefineNet-style fusion module is used to interactively model the upsampled and skip features, thereby refining semantic information and spatial details layer by layer. This process propagates from deep semantic layers to shallow high-resolution layers, ultimately yielding feature representations with rich contextual semantics and fine spatial structure. A high-resolution difference map is then generated using a convolutional prediction head. Among them, difference graph The final feature representation is formed by decoding and fusing multi-level differential features layer by layer.

[0026] The neighborhood relationship propagation module calculates the squared Euclidean distance between the center pixel and its neighboring pixels based on the differential features of the two-temporal remote sensing image pairs. It then converts this squared Euclidean distance into a correlation weight matrix using the Softmax function. This matrix is ​​used to perform neighborhood-weighted aggregation on the initial prediction map output by the prediction head, generating a corrected prediction map. Figure 1 As shown.

[0027] In some implementations, multi-level differential features are used. The central features are mapped through two parallel convolutional branches, Conv1 and Conv2. and neighborhood features Set the sliding window size to For central features The spatial dimensions are flattened using the Flatten operation and then processed by Unsqueeze to obtain the feature vector of the center pixel. For neighborhood features The feature vector of neighboring pixels is obtained by extracting local blocks through the Unfold operation and then reshaping them. This invention uses squared Euclidean distance as a correlation metric between the center pixel and its neighboring pixels. The calculation formula is: ; in, Indicates the first In the nth sample, the nth The central pixel and its neighborhood of the first... The squared Euclidean distance between pixels in the feature space; This represents the number of feature channels.

[0028] Using negative The function converts the squared Euclidean distance into a normalized correlation weight matrix. Let this be denoted as the neighborhood correlation map A, which is used to characterize the correlation distribution between the center pixel and its neighboring pixels. Its calculation formula is: ; Among them, the neighborhood correlation graph The dimension is , Indicates batch size, Indicates the number of pixels within the neighborhood window. This represents the total number of spatial pixels in the feature map.

[0029] In some implementations, a neighborhood correlation map is used to spatially guide the prediction results, enhancing the consistency between the center pixel prediction and its neighborhood structure, and enabling prediction correction based on local correlation. The initial prediction map output by the prediction head is spatially expanded to a size consistent with the neighborhood window to obtain a neighborhood prediction representation. Using the correlation weight matrix The weighted aggregation expression is: ; in, It represents the degree of correlation between pixels within a local window, where the matrix elements represent the correlation weights between corresponding pixels; For the first In the nth sample, the nth The central pixel and its neighborhood of the first... The squared Euclidean distance between pixels in the feature space; This indicates that after neighborhood-weighted aggregation, the first... In the nth sample Category in Updated predicted values ​​for each pixel position; In the initial prediction graph output by the student model, the first... In the nth sample Category in the first Within the neighborhood of the nth center pixel as reference The predicted value for each pixel position is obtained. The initial prediction map is updated using the weighting effect of the neighborhood correlation graph, resulting in a corrected prediction result, thus achieving prediction result optimization based on neighborhood relationship propagation.

[0030] S3: Using samples from the labeled dataset, input the bi-temporal remote sensing image pairs into the student model. A shared encoder extracts multi-scale feature maps of the bi-temporal remote sensing image pairs. A feature decoder performs difference operations and feature fusion on the multi-scale feature maps to obtain a difference map. This difference map is then input into the prediction head. The prediction head sequentially passes through a 3×3 convolutional layer (Conv3x3), a ReLU activation layer, and a 1×1 convolutional layer (Conv1x1) for non-linear mapping and dimensionality compression, outputting a pixel-level initial prediction map with the number of channels equal to the number of classes. The initial prediction graph is then used as input to the neighborhood relationship propagation module for weighted correction.

[0031] Using samples from an unlabeled dataset, weak perturbations are applied to pairs of bi-temporal remote sensing images and then input into the teacher model to obtain a weakly perturbated prediction probability map. After filtering by a preset confidence threshold, pseudo-labels are generated. Different strong perturbations are applied to the same pair of bi-temporal remote sensing images and then input into the student model to obtain corresponding initial prediction maps and corrected prediction maps. The unsupervised consistency loss between each initial prediction map, corrected prediction map, and pseudo-label is calculated. In some implementations, weak perturbations are lightweight data augmentations that keep the semantic information of the main subject of the bi-temporal remote sensing images largely unchanged, including random flipping and random cropping, to generate stable teacher model predictions; strong perturbations are data augmentations that further superimpose appearance changes and noise interference on the basis of weak perturbations, including random flipping, random cropping, color perturbation and random Gaussian noise, to enhance the robustness of the model to complex data distribution changes.

[0032] The teacher model makes predictions on pairs of bi-temporal remote sensing images with weak perturbations, resulting in pixel-level prediction probability maps. Each pixel location contains the predicted probability distribution for each category; the predicted probability map... By selecting the maximum value along the category dimension, the maximum predicted probability value for each pixel location is obtained. and the corresponding category index; compare the maximum predicted probability value with the preset confidence threshold. When comparing, The corresponding category index is used as the pseudo-label for that pixel. ;when If the pixel is invalid, it will be marked as an invalid pixel and will not be included in the subsequent unsupervised consistency loss calculation.

[0033] The total loss function is constructed by summing the supervised loss and the unsupervised consistency loss. The student model parameters are updated by backpropagation algorithm with the goal of minimizing the total loss function. The teacher model parameters are updated by exponential moving average of the student model parameters. The process continues until the preset number of training rounds is reached or the total loss function converges, resulting in a trained student model. Total loss function : ; Monitoring losses Using cross-entropy loss, the expression is: ; Unsupervised consistency loss For two strong disturbance views and The consistency between the initial prediction map and the corrected prediction map and the pseudo-label is calculated separately, using the following expression: ; in, and These are the batch sizes with and without labels, respectively. Cross-entropy loss; and These are the initial prediction map and the corrected prediction map for the labeled dataset samples, respectively. Labels representing pixel-level real-world changes; and The initial prediction plots are shown under two strongly disturbed views. and This is the corresponding corrected prediction graph; This is an indicator function.

[0034] S4: The student model trained on the input of dual-temporal remote sensing images is used only with a shared encoder, feature decoder, and prediction head to output a pixel-level change probability map. This probability map uses a dual-channel output to map the predicted probability values ​​of the change class (representing areas where land cover type or land cover status changes significantly under multi-temporal observations) and the non-change class (representing areas where land cover or land cover status maintains consistency and stability under multi-temporal observations). Subsequently, a pixel-by-pixel Argmax operation is performed on the class dimension of the dual-channel probability result. By comparing the confidence weights of each spatial unit belonging to the above two attributes, the class with the higher probability is selected as the final predicted label of the pixel, thereby generating a binarized change detection mask, and finally realizing the accurate extraction of the spatial distribution results of land cover change areas.

[0035] Example 1 In the implementation phase of this invention, two NVIDIA RTX 3080 Ti graphics cards are used as the computing platform, and PyTorch is used as the development framework for deep learning models.

[0036] We obtained the publicly available change detection datasets LEVIR-CD and WHU-CD through an academic authorization platform. Both datasets contain remote sensing image pairs from different time periods and their corresponding change annotations.

[0037] The dual-temporal remote sensing images were uniformly cropped and normalized to meet the network input requirements. The dataset was divided into training, validation, and test sets according to a preset ratio. The training set was further divided into labeled and unlabeled datasets for use in a semi-supervised training framework. The input images were randomly flipped, randomly cropped, and had their colors perturbed to improve the model's generalization ability.

[0038] A semi-supervised change detection model based on a teacher-student framework is constructed, in which both the student model and the teacher model structure include a shared encoder, a feature decoder, and a prediction head; the student model also includes a neighborhood relationship propagation module, and the teacher model parameters are updated from the student model parameters through an exponential moving average (EMA) method.

[0039] The shared encoder employs a pre-trained DINOv2 model to extract multi-scale semantic features, thereby enhancing the expressive power of regions of change.

[0040] Regarding model training configuration, this embodiment sets the total number of training epochs to 60 to ensure sufficient convergence of the model even with a small number of labeled samples. The batch size is set to 8 to balance memory usage and gradient update stability.

[0041] The model training employed stochastic gradient descent with momentum (SGD) for optimization, with the momentum coefficient set to 0.995 to enhance the stability of gradient updates and reduce oscillations. The initial learning rate was set to 5 × 10⁻⁶. 6 Furthermore, a cosine annealing strategy is employed to decay the learning rate, causing it to gradually approach zero at the end of training, thereby improving the model's convergence stability.

[0042] Introducing a neighborhood window mechanism, the window size is set to... =9×9, used to construct local neighborhood relationships and perform feature propagation and weighted aggregation.

[0043] The model performance was quantitatively evaluated using a general change detection evaluation index, with the intersection-over-union ratio (IoU) used as the evaluation index to measure the model's performance in detecting changed regions.

[0044] The model performance is evaluated using the validation set that was not updated during training.

[0045] The testing process is consistent with the forward propagation in the training phase. The dual-temporal remote sensing images are input into the model. After encoder feature extraction and decoding prediction, the change detection results are directly output without backpropagation and parameter updates.

[0046] The final output is a pixel-level change probability map, which is then used to generate a binary change map through thresholding.

[0047] To verify the impact of key parameters of the neighborhood relationship propagation module on change detection performance in this invention, experiments were conducted on different neighborhood window sizes and similarity measurement methods while keeping other network structures and training strategies unchanged. The specific experimental settings are as follows: neighborhood window sizes were set to 5×5, 9×9, 13×13, 17×17, and 21×21, and similarity measurement methods included dot product, cosine similarity, and squared Euclidean distance. The experimental results are shown in Table 1.

[0048] Table 1. Impact of different neighborhood window sizes and similarity measurement methods on model performance;

[0049] To ensure fairness and reproducibility in the experimental comparisons, the default configurations from their published papers or official implementations were used for each comparison method. Specifically, the encoder (feature extraction backbone network) settings for each method are as follows: SemiCDNet uses UNet++ as the encoder; SemiCD uses ResNet50; Unimatch uses ResNet50; SemiCD-VL uses ResNet50; and UnimatchV2 uses DINOv2. The method of this invention uses DINOv2 as a shared encoder.

[0050] Table 2. Performance comparison of different methods on the LEVIR-CD dataset under different annotation ratios;

[0051] Table 3. Performance comparison of different methods on the WHU-CD dataset under different annotation ratios;

[0052] In summary, the ablation experiments in Table 1 show that the neighborhood window size and correlation measurement method have a significant impact on change detection performance. The combination of a 9×9 window and squared Euclidean distance achieves optimal performance while maintaining computational efficiency. Tables 2 and 3 compare the performance of the proposed method with existing representative methods under different annotation ratios. The results show that the proposed method outperforms existing methods on both the LEVIR-CD and WHU-CD datasets, especially with significant performance improvement under scarce annotation conditions. Therefore, Tables 1, 2, and 3 fully demonstrate that the proposed neighborhood relationship propagation module and overall method can effectively improve the accuracy and robustness of remote sensing image change detection, verifying the practicality and superiority of the invention.

[0053] In another embodiment, this invention proposes a semi-supervised remote sensing image change detection system based on neighborhood relation propagation, used to implement the disclosed semi-supervised remote sensing image change detection method based on neighborhood relation propagation, comprising: The dual-temporal dataset construction module is used to acquire dual-temporal remote sensing image pairs of areas of land cover change, construct training datasets and divide them into labeled datasets and unlabeled datasets; the labeled datasets include dual-temporal remote sensing image pairs and their corresponding pixel-level real change labels, while the unlabeled datasets only include dual-temporal remote sensing image pairs. The teacher-student semi-supervised change detection model construction module is used to build a semi-supervised change detection model based on the teacher-student framework. Both the student model and the teacher model include a shared encoder, feature decoder, and prediction head. The student model also includes a neighborhood relationship propagation module. A semi-supervised joint training module is used to input bi-temporal remote sensing image pairs into the student model on labeled datasets. The model outputs an initial prediction map via a shared encoder, feature decoder, and prediction head. A neighborhood relationship propagation module then outputs a corrected prediction map. This module calculates the squared Euclidean distance between the center pixel and its neighboring pixels based on the differential features of the bi-temporal remote sensing image pairs. This distance is then converted into a relevance weight matrix using a Softmax function. The initial prediction map is then weighted and aggregated to generate the corrected prediction map. Finally, the supervised loss between the initial and corrected prediction maps and the pixel-level true change labels is calculated. In the unlabeled dataset, weak perturbations are applied to pairs of bi-temporal remote sensing images and then input into the teacher model to obtain a predicted probability map. After filtering by a preset confidence threshold, pseudo-labels are generated. Different strong perturbations are applied to the same pair of bi-temporal remote sensing images and then input into the student model to obtain the corresponding initial prediction map and corrected prediction map. The unsupervised consistency loss between each initial prediction map, corrected prediction map and pseudo-label is calculated. The total loss function is set as the sum of the supervised loss and the unsupervised consistency loss. The semi-supervised change detection model is trained to obtain the trained student model. Probability of change Figure 2 The value-based output module is used to input the dual-temporal remote sensing image to be detected into the trained student model, output a pixel-level change probability map, and then process it to obtain the spatial location information of the change area of ​​ground features.

[0054] In another embodiment of the present invention, an electronic device is proposed, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a semi-supervised remote sensing image change detection method based on neighborhood relationship propagation as described above.

[0055] In another embodiment of the present invention, a computer-readable storage medium is provided storing a computer program that causes a computer to execute a semi-supervised remote sensing image change detection method based on neighborhood relation propagation as described above.

[0056] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent substitutions, and improvements made by those skilled in the art to the above embodiments without departing from the scope of the technical solution of the present invention, based on the technical essence of the present invention, shall still fall within the protection scope of the technical solution of the present invention.

Claims

1. A semi-supervised remote sensing image change detection method based on neighborhood relation propagation, characterized in that, Includes the following steps: S1: Obtain bi-temporal remote sensing image pairs of the land cover change area, construct a training dataset and divide it into a labeled dataset and an unlabeled dataset; the labeled dataset includes bi-temporal remote sensing image pairs and their corresponding pixel-level real change labels, while the unlabeled dataset only includes bi-temporal remote sensing image pairs. S2: Construct a semi-supervised change detection model based on a teacher-student framework, in which both the student model and the teacher model include a shared encoder, feature decoder, and prediction head. The student model also includes a neighborhood relationship propagation module. S3: In the labeled dataset, the dual-temporal remote sensing image pairs are input into the student model. The initial prediction map is output through the shared encoder, feature decoder, and prediction head. Then, the neighborhood relationship propagation module outputs the corrected prediction map. The neighborhood relationship propagation module calculates the squared Euclidean distance between the center pixel and its neighboring pixels based on the differential features of the dual-temporal remote sensing image pairs. The distance is converted into a correlation weight matrix through the Softmax function. The initial prediction map is then subjected to neighborhood weighted aggregation to generate the corrected prediction map. The supervision loss between the initial prediction map, the corrected prediction map, and the pixel-level true change label is calculated. In the unlabeled dataset, weak perturbations are applied to pairs of bi-temporal remote sensing images and then input into the teacher model to obtain a predicted probability map. After filtering by a preset confidence threshold, pseudo-labels are generated. Different strong perturbations are applied to the same pair of bi-temporal remote sensing images and then input into the student model to obtain the corresponding initial prediction map and corrected prediction map. The unsupervised consistency loss between each initial prediction map, corrected prediction map and pseudo-label is calculated. The total loss function is set as the sum of the supervised loss and the unsupervised consistency loss. The semi-supervised change detection model is trained to obtain the trained student model. S4: Input the dual-temporal remote sensing image to be detected into the trained student model, output a pixel-level change probability map, and then process it to obtain the spatial location information of the change area of ​​ground features. 2.The semi-supervised remote sensing image change detection method based on neighborhood relation propagation according to claim 1, wherein: In step S2, the shared encoder is a pre-trained visual feature extraction network that extracts multi-scale semantic features of dual-temporal remote sensing image pairs; the feature decoder is a decoder based on multi-level feature fusion, which obtains multi-level differential features by performing differential operations on the multi-scale semantic features output by the shared encoder, and then fuses them through upsampling and skip connections to generate a differential map.

3. The semi-supervised remote sensing image change detection method based on neighborhood relation propagation according to claim 1, characterized in that: In step S3, the calculation of the squared Euclidean distance between the center pixel and its neighboring pixels based on the differential features of the dual-temporal remote sensing image pair is specifically as follows: The shared encoder of the student model extracts multi-scale semantic features of the double-phase images respectively, calculates absolute difference values of the corresponding features to obtain difference features, and maps the difference features through two parallel convolution branches to center features and neighborhood features respectively and neighborhood features , sets a sliding window size, expands the feature maps in the spatial dimension to obtain center pixel feature vectors and neighborhood pixel feature vectors , and square Euclidean distances The calculation formula is: ； wherein, represents the squared Euclidean distance in the feature space between the i-th center pixel and the j-th pixel within its neighborhood in the i-th sample; is a batch index; is the number of feature channels.​​​ 4. The semi-supervised remote sensing image change detection method based on neighborhood relation propagation according to claim 1, characterized in that: In step S3, the squared Euclidean distance is converted into a relevance weight matrix using the Softmax function. This matrix is ​​then used to perform neighborhood weighted aggregation on the initial prediction map, generating a revised prediction map. The specific process is as follows: Using negative functions to convert squared Euclidean distances into normalized relevance weight matrices , where: ； spatially expanding the initial prediction map by a size consistent with the neighborhood window to obtain a neighborhood prediction representation using a correlation weight matrix for weighted aggregation, expressed as: ； wherein, is the square Euclidean distance in the feature space between the th center pixel in the th sample and the th pixel in its neighborhood; is the size of the sliding window; represents the updated prediction value of the th class at the th pixel position in the th sample after neighborhood weighted aggregation; is the prediction value of the th class at the th pixel position in the neighborhood with reference to the th center pixel in the initial prediction map output by the student model in the th sample.

5. The semi-supervised remote sensing image change detection method based on neighborhood relation propagation according to claim 1, characterized in that: In step S3, the weak perturbations applied to the unlabeled dataset include random flipping and random cropping, while the strong perturbations applied to the unlabeled dataset include random flipping, random cropping, color perturbation, and random Gaussian noise.

6. The semi-supervised remote sensing image change detection method based on neighborhood relation propagation according to claim 1, characterized in that: In step S3, the total loss function is set as the sum of the supervised loss and the unsupervised consistency loss, and the specific expression is as follows: Total loss function : ； Supervised loss : ； Unsupervised consistency loss : ； wherein, and are annotated and unannotated batch size, respectively, is cross-entropy loss; and are initial prediction map and corrected prediction map of annotated samples, respectively; is true label; and are initial prediction maps under two strong perturbation views; and are corresponding corrected prediction maps; is pixel-level prediction probability map of weak perturbation output by the teacher model; is pseudo label; is confidence threshold; is indicator function.

7. The semi-supervised remote sensing image change detection method based on neighborhood relationship propagation according to claim 1, characterized in that: In step S3, the pseudo-labels are generated after filtering based on a preset confidence threshold. The specific process is as follows: The teacher model makes predictions on pairs of bi-temporal remote sensing images with weak perturbations, resulting in pixel-level prediction probability maps. Each pixel location contains the predicted probability distribution for each category; the predicted probability map... By selecting the maximum value along the category dimension, the maximum predicted probability value for each pixel location is obtained. and the corresponding category index; compare the maximum predicted probability value with the preset confidence threshold. When comparing, The corresponding category index is used as the pseudo-label for that pixel. ;when If the pixel is invalid, it will be marked as an invalid pixel and will not be included in the subsequent unsupervised consistency loss calculation.

8. A semi-supervised remote sensing image change detection system based on neighborhood relation propagation, used to implement the semi-supervised remote sensing image change detection method based on neighborhood relation propagation as described in any one of claims 1 to 7, characterized in that, include: The dual-temporal dataset construction module is used to acquire dual-temporal remote sensing image pairs of areas of land cover change, construct training datasets and divide them into labeled datasets and unlabeled datasets; the labeled datasets include dual-temporal remote sensing image pairs and their corresponding pixel-level real change labels, while the unlabeled datasets only include dual-temporal remote sensing image pairs. The teacher-student semi-supervised change detection model construction module is used to build a semi-supervised change detection model based on the teacher-student framework. Both the student model and the teacher model include a shared encoder, feature decoder, and prediction head. The student model also includes a neighborhood relationship propagation module. A semi-supervised joint training module is used to input bi-temporal remote sensing image pairs into the student model on labeled datasets. The model outputs an initial prediction map via a shared encoder, feature decoder, and prediction head. A neighborhood relationship propagation module then outputs a corrected prediction map. This module calculates the squared Euclidean distance between the center pixel and its neighboring pixels based on the differential features of the bi-temporal remote sensing image pairs. This distance is then converted into a relevance weight matrix using a Softmax function. The initial prediction map is then weighted and aggregated to generate the corrected prediction map. Finally, the supervised loss between the initial and corrected prediction maps and the pixel-level true change labels is calculated. In the unlabeled dataset, weak perturbations are applied to pairs of bi-temporal remote sensing images and then input into the teacher model to obtain a predicted probability map. After filtering by a preset confidence threshold, pseudo-labels are generated. Different strong perturbations are applied to the same pair of bi-temporal remote sensing images and then input into the student model to obtain the corresponding initial prediction map and corrected prediction map. The unsupervised consistency loss between each initial prediction map, corrected prediction map and pseudo-label is calculated. The total loss function is set as the sum of the supervised loss and the unsupervised consistency loss. The semi-supervised change detection model is trained to obtain the trained student model. The change probability map binarization output module is used to input the dual-temporal remote sensing image to be detected into the trained student model, output a pixel-level change probability map, and then process it to obtain the spatial location information of the change area of ​​ground features.

9. An electronic device, characterized in that, include: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements a semi-supervised remote sensing image change detection method based on neighborhood relationship propagation as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program causes the computer to execute a semi-supervised remote sensing image change detection method based on neighborhood relationship propagation as described in any one of claims 1 to 7.

Images