Remote sensing change detection method based on teacher-student framework and multi-space distillation

Abstract

The application discloses a remote sensing change detection method based on a teacher-student framework and multi-space distillation, relates to the technical field of remote sensing image processing, and comprises the following steps: constructing a text prototype library and a visual prototype library; performing zero sample prediction on an image to be detected to generate text guided features; extracting multi-scale feature maps, global semantic features and spatial level features by a student network; calculating absolute differences, similarity differences and text guided differences to splice initial change features; respectively calculating channel attention and spatial attention of the initial change features to fuse multi-scale fusion features to splice spliced features; and inputting the decoder to obtain a binary change detection map. The method effectively improves the precision of remote sensing image change detection and provides important technical support for the application in the fields of city planning and environment monitoring.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a remote sensing image change detection method based on multimodal large model prior knowledge and incorporating visual-linguistic knowledge distillation. It aims to detect changes in ground features at the same location over different times using dual-temporal high-resolution images. Background Technology

[0002] Remote Sensing Change Detection (RSCD) is a key technology that uses remote sensing images of the same area acquired at different times to automatically identify and analyze information on changes in land surface features. It is widely used in urban planning, land resource management, disaster assessment, and ecological environment monitoring. With the rapid development of remote sensing technology, acquiring high-resolution remote sensing images has become increasingly convenient, providing a rich data foundation for the refined monitoring of land surface changes.

[0003] Traditional remote sensing change detection methods mainly rely on algebraic operations (such as image interpolation and ratio) or shallow machine learning classifiers (such as support vector machines and random forests). These methods are usually based on manually designed low-level features (such as spectrum and texture). When faced with complex surface environments and high-resolution imagery, they often struggle to handle interference from illumination changes, registration errors, and heterogeneity within land cover classes, resulting in limited detection accuracy and robustness.

[0004] In recent years, deep learning techniques, represented by Convolutional Neural Networks (CNNs) and Transformers, have made significant progress in change detection tasks. Current mainstream deep learning methods typically employ Siamese network structures to extract visual features from bi-temporal images and generate change maps through feature fusion and difference discrimination. However, most of these methods are limited to a single-modal learning paradigm of "vision-to-vision," exhibiting the following significant limitations: Insufficient semantic understanding: Existing methods mainly rely on pixel-level visual appearance differences for judgment, lacking an understanding of the higher-level semantics of ground features. Therefore, they are prone to misjudging non-structural appearance changes (i.e., "pseudo-changes") caused by seasonal changes (such as changes in vegetation color), differences in lighting, or shadows as real changes.

[0005] Limited generalization ability: Most models rely excessively on densely labeled data from specific datasets for supervised training, lacking prior guidance from external knowledge bases. When applied to cross-regional or cross-scene data, performance often drops significantly due to a lack of semantic understanding of the essential attributes of ground features.

[0006] Weak ability to perceive subtle changes: In complex scenes, real changes in ground features are often submerged in complex background noise, and it is difficult to capture fine-grained semantic change boundaries by simply subtracting or splicing visual features.

[0007] While recent research has attempted to incorporate visual-language models (such as contrastive language-image pre-trained CLIPs) to assist change detection, most existing methods employ simple feature concatenation or implicit alignment strategies, failing to explicitly construct a stable semantic space structure. This results in low efficiency in cross-modal knowledge transfer and susceptibility to noise in the training data. Therefore, effectively transferring the general semantic prior knowledge inherent in large-scale visual-language models to remote sensing change detection tasks, while suppressing spurious changes and maintaining a keen perception of real changes, remains a pressing technical challenge in this field. Summary of the Invention

[0008] This application provides a remote sensing change detection method based on a teacher-student framework and multi-spatial distillation to address the problems in the prior art.

[0009] This application provides a remote sensing change detection method based on a teacher-student framework and multi-spatial distillation, including: In the teacher network, a text prototype library is constructed using preset land cover categories, and a visual prototype library is initialized. During the training process, the momentum of the visual prototype library is updated based on the similarity between the training samples and the text prototype library. In the student network, the image encoder in the teacher network performs zero-shot prediction on the image to be detected, and generates text-guided features based on the prediction results. The image to be detected is input into the student network, and multi-scale feature maps, global semantic features, and spatial features are extracted. The difference perception module calculates absolute difference and similarity difference based on the multi-scale feature maps, and calculates text-guided difference based on spatial features and text-guided features. The absolute difference, similarity difference, and text-guided difference are concatenated with the original features of the image to be detected to form initial variation features. The channel-space attention module calculates channel attention and spatial attention on the initial variation features, and multi-scale fusion features are obtained by fusing channel attention and spatial attention. The multi-scale fusion features are concatenated with global semantic features to form concatenated features. The concatenated features are input into the decoder to obtain a binary change detection map; During the training of the student network, a triple semantic alignment loss function is established based on basic task loss, visual spatial distillation loss, textual spatial distillation loss, and prototype semantic structure alignment loss. The student network is then trained using the triple semantic alignment loss function.

[0010] The remote sensing change detection method based on the teacher-student framework and multi-spatial distillation in this application has the following advantages: 1. Significantly suppresses spurious change interference: This application utilizes a large visual-language model as the teacher network and introduces environment-invariant text semantic priors. By calculating text-guided differences and similarity differences, the model can effectively distinguish between unstructured appearance changes caused by seasonal changes (such as vegetation color changes) and light changes and true semantic changes, significantly reducing the false detection rate.

[0011] 2. Enhanced semantic consistency and robustness: This application constructs a text prototype library (TPB) and a visual prototype library (VPB). By explicitly constructing these semantic prototypes as "anchors" and using triple semantic alignment loss for constraints, the student network is forced to learn structured semantic distributions, rather than just pixel-level differences, thereby significantly improving the model's generalization ability in complex scenarios.

[0012] 3. Achieved fine perception of multi-granular features: The difference perception (DP) module proposed in this application considers the differences in three dimensions: amplitude, direction and semantics. Combined with the channel-spatial attention (CSA) mechanism, the model can capture macroscopic semantic changes while preserving minute details of ground object boundaries. It has achieved detection accuracy superior to existing technologies on the SYSU-CD (Sun Yat-sen University aerial image change detection dataset) and LEVIR-CD (Earth Vision and Intelligent Remote Sensing Laboratory change detection dataset) datasets (mF1 and mIoU indices are significantly improved).

[0013] This application, through the above-mentioned technical solution, effectively improves the accuracy of remote sensing image change detection, and provides important technical support for applications in urban planning, environmental monitoring and other fields. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is an architecture diagram of the remote sensing change detection method based on a teacher-student framework and multi-spatial distillation provided in the embodiments of this application.

[0016] Figure 2 This is a diagram illustrating the architecture of multi-space distillation provided in an embodiment of this application.

[0017] Figure 3 This is an architecture diagram of the teacher network and feature distillation provided for embodiments of this application.

[0018] Figure 4 This is an architecture diagram of the DP module provided in an embodiment of this application.

[0019] Figure 5 This is an architecture diagram of the CSA module provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Figure 1 This is an architecture diagram of the remote sensing change detection method based on a teacher-student framework and multi-spatial distillation provided in this application embodiment. This application embodiment provides a remote sensing change detection method based on a teacher-student framework and multi-spatial distillation, including: In S100, within the teacher network, a text prototype library is constructed using preset land cover categories, and a visual prototype library is initialized. During training, the momentum of the visual prototype library is updated based on the similarity between the training samples and the text prototype library.

[0022] For example, the teacher network employs a CLIP model pre-trained on a large-scale image-text dataset, such as the ViT-B / 16 (16-pixel block basic visual transformer) architecture. During the training of the student network, the parameters of the teacher network remain frozen, do not participate in gradient updates, and are only used to build text and visual prototype libraries.

[0023] Furthermore, such as Figure 3 As shown, the method for building a text prototype library is as follows: Fill each preset land feature category into the preset prompt template to generate the corresponding prompt text description; The text encoder extracts features from the prompt text description, performs L2 normalization to obtain the corresponding text prototype vectors, and all text prototype vectors form a text prototype library.

[0024] Specifically, select common remote sensing scenarios K Each land cover category, in the embodiments of this application K =56, such as forests, grasslands, buildings, and water bodies. Label each category. c k , k =1,2,…, KEnter the preset prompt template, such as "A remote sensing image of c k This will generate the corresponding prompt text description.

[0025] The CLIP text encoder in the frozen teacher model is used to extract features from the above prompt text description, and L2 normalization is performed to obtain... K A static text prototype library TPB is composed of several text prototype vectors. }

[0026] Furthermore, the method for updating the momentum of the visual prototype library is as follows: Image features are obtained by using the frozen CLIP visual encoder in the teacher model to extract features from the input image. ; Calculate the cosine similarity between the training sample and all text prototype vectors in the text prototype library, and find the text category index with the highest cosine similarity. Update the visual prototype in the visual prototype library corresponding to the text category index with the highest cosine similarity using the features of the training samples:

[0027] in, and These are the visual prototypes before and after the update. The momentum coefficient, These are the image features output by the visual encoder in the teacher network.

[0028] Specifically, initialization K Visual prototype vectors constitute the visual prototype library VPB={ m 1, m 2,…, m K The VPB is maintained using an Online Momentum Update (VPB) strategy: During training, for each input training sample v, the cosine similarity between v and all text prototype vectors in the TPB is calculated to find the best-matching text category index. j The visual prototype of the corresponding category is updated using the current training sample v.

[0029] This step enables VPB to capture robust visual centers for each type of land cover under environmental changes.

[0030] In S110, within the student network, the image encoder in the teacher network performs zero-shot prediction on the image to be detected, generating text-guided features based on the prediction results. The image to be detected is input into the student network, where multi-scale feature maps, global semantic features, and spatial features are extracted. The difference perception module calculates absolute difference and similarity difference based on the multi-scale feature maps, and calculates text-guided difference based on spatial features and text-guided features. The absolute difference, similarity difference, and text-guided difference are concatenated with the original features of the image to be detected to form initial variation features. The channel-space attention module calculates channel attention and spatial attention on the initial variation features, and multi-scale fusion features are obtained by fusing channel attention and spatial attention. The multi-scale fusion features are then concatenated with global semantic features to form concatenated features.

[0031] For example, firstly, the dual-temporal remote sensing image pair to be detected is acquired. I (A) and I (B) These correspond to phases A and B, respectively, and the corresponding pixel-level binary change truth labels. The input images are uniformly adjusted to a fixed size of 256×256 pixels and then standardized.

[0032] In this embodiment, the student network uses ResNet-50 as the backbone network and is initialized using CLIP image encoder weights to maintain initial alignment with the teacher network feature space.

[0033] Furthermore, the method for generating text guidance features is as follows: The image encoder in the teacher network is used to perform zero-shot prediction on the image to be detected, so as to obtain the probability that the image to be detected belongs to each land cover category; Multiple land cover categories with the highest probabilities are selected (nine categories are selected in this embodiment), sorted from high to low probabilities, and used to form natural language prompts describing the content of the image to be detected, such as "An aerial image of [class1], [class2]..."; The natural language prompts are input into the text encoder to extract text guidance features, providing high-level semantic priors for subsequent modules.

[0034] Simultaneously, the dual-temporal images to be detected are input into the ResNet-50 backbone of the student network to extract multi-scale features across four stages (Stages 1-4). Then, a Feature Pyramid Network (FPN) is used to fuse the multi-scale features generated from multiple stages, resulting in a multi-scale feature map F with a unified channel dimension. s The number of stages s∈{1,2,3,4}.

[0035] Furthermore, such as Figure 4 As shown, while the student network extracts multi-scale feature maps, attention pooling is performed on the features extracted in stage four of the student network to extract global semantic features G from the A-phase image and the B-phase image, respectively. A and G B Meanwhile, to preserve semantic information in spatial structure, the features extracted by ResNet50 in stage four (s=4) are used to generate spatial-level features S for time-phase images A and B through convolutional layers. A and S B Through convolution mapping, Stage 4 outputs feature maps for phases A and B. and It is converted into spatial-level features.

[0036] The absolute difference is expressed as:

[0037] in, For students in the network stage s The absolute difference and These are the multi-scale features of phases A and B, respectively. This indicates taking the absolute value.

[0038] This difference mainly reflects the changes in local feature amplitude caused by the addition, disappearance or significant structural changes of ground features, but it is more sensitive to changes in illumination and noise.

[0039] The similarity difference is expressed as:

[0040] in, Represents cosine similarity, [ h,w [ represents spatial location coordinates] and These are two phases in the stage. s The feature vector at that position, Represents the dot product of vectors; for The sigmoid function is used to map the difference in similarity. .

[0041] This difference, to some extent, eliminates the influence of feature scale variations and focuses more on consistent changes in feature direction, thus providing stronger robustness against spurious changes caused by variations in imaging conditions.

[0042] The differences in text guidance are represented as follows:

[0043] in, For the text-guided differences in the student network at stage s and and are the spatial-level features of phase A and phase B respectively, and are the text-guided features of phase A and phase B respectively. In Figure 4 , F A and F B are respectively and abbreviations at any stage s , D abs is abbreviation at any stage s , D sim is abbreviation at any stage s , D text is abbreviation at any stage s . [[ID=X]] [[ID=Y]]

[0044] This difference introduces high-level semantic priors, explicitly models the matching change degree of dual-phase images in the same semantic space, and is used to distinguish pseudo-changes.

[0045] Concatenate the above three difference features with the multi-scale features extracted by the student network along the channel dimension to generate initial change features , then apply the channel attention mechanism to adaptively re-weight and fuse the concatenated features, so as to generate the final output features at each stage. The DP module performs the above operations s at each stage, thereby generating initial change features Y s [[ID=6X]].

[0046] Furthermore, as Figure 5 shown, in order to adaptively screen key change information from the initial change features, the embodiment of the present application also introduces a channel-spatial attention module, and this model sequentially performs the following operations: (1) Channel attention: Perform global average pooling (GAP) on the input initial change features Y s to obtain channel attention features , c is the number of channels, and obtain channel weights through a multi-layer perceptron (MLP), emphasizing the channels containing important semantic changes.

[0047] (2) Spatial attention: GAP processing is performed on each channel, and then the data is input into the MLP. The sigmoid function is selected to complete the weighted processing of the features and obtain the channel refined features. Based on channel-weighted features Spatial weights are generated through convolution operations. Locate areas of significant change.

[0048] (3) Multi-scale aggregation: Through adaptive fusion of different s The CSA module obtains the final multi-scale fused features from the features extracted by the dual attention mechanism. Y CSA .

[0049] (4) Fusing features across multiple scales Y CSA With global semantic features G A and G B Broadcast splicing is performed to obtain splicing features.

[0050] S120: Input the spliced ​​features into the decoder to obtain a binary change detection map.

[0051] For example, before inputting to the decoder, the concatenated features obtained in S110 are first fused through a convolutional layer to obtain fused concatenated features. Then, the resolution of the features is restored through the decoder, and finally a binary change detection map is generated. This detection head can reflect the changes in the remote sensing image.

[0052] Specifically, the decoder employs a progressive upsampling structure, restoring spatial resolution through multiple convolutional and interpolation operations. Simultaneously, skip connections are introduced to fuse low-level detail information. Finally, change detection results are generated using 1×1 convolutions and the sigmoid function.

[0053] S130: During the training of the student network, a triple semantic alignment loss function is established based on the basic task loss, visual spatial distillation loss, textual spatial distillation loss, and prototype semantic structure alignment loss. The student network is then trained using the triple semantic alignment loss function.

[0054] For example, such as Figure 2 As shown, the triple semantic alignment loss function is expressed as:

[0055] in, This is a triple semantic alignment loss function. 、 、 and These are the basic task loss, visual spatial distillation loss, textual spatial distillation loss, and prototype semantic structure alignment loss, respectively. and They are 0.5 and 0.4 respectively. Figure 2 as well as Figure 3 In this context, t is short for Text Prototype Vector in TPB, v is short for Image Features extracted by the Image Encoder, m is short for Visual Prototype Vector in VPB, and G is a collective term for Global Semantic Features. Figure 3 middle, v 1, v 2,…, v n Image features extracted by the image encoder n The dimension of image features. t 3, t 2, t k , t 9,…, t 7 represents the text category with the highest cosine similarity obtained during the momentum update process of the visual prototype library. t 1, t 2,…, t k This is the text prototype vector in TPB.

[0056] Specifically, the basic task loss uses weighted cross-entropy loss to supervise the consistency between pixel-level prediction results and ground truth labels.

[0057] The visual spatial distillation loss uses KL divergence to constrain the visual feature distribution of the student network to approximate the visual feature distribution of the teacher network, thus conveying generalized knowledge, as shown below:

[0058] in, Denotes KL divergence, s oftmax function, and These represent teacher characteristics and student visual characteristics, respectively. and These represent the temperature parameters used to soften the probability distributions of the teacher and student models, respectively, in the embodiments of this application. and =4.

[0059] Text spatial distillation loss calculation of text guidance features generated by student networks Teacher text features extracted by the text encoder in the teacher network The L2 distance between them ensures the semantic accuracy of the student network in understanding the image content, as shown below:

[0060] in, This represents the L2 distance.

[0061] The prototype semantic structure alignment loss is a key aspect of this application. This loss is calculated by plotting the similarity distribution between the teacher's text features and all prototypes in the VPB. P T This distribution is used as a soft label, as shown below:

[0062] in, Image features With visual prototype m j The similarity distribution Represents cross-entropy loss, Represents the similarity function. Image features representing teacher characteristics, i.e. v k , This represents the visual prototype in VPB. Indicates temperature parameter, =2.

[0063] Calculate the similarity distribution between student text features obtained from the visual encoder and text encoder in the student network and all prototypes in the VPB. As shown below:

[0064] in, for T T Visual prototypes in VPB m j The visual similarity distribution.

[0065] Forced using cross-entropy loss Approaching As shown below:

[0066] This loss forces the student network to maintain the same topological structure, i.e., clustering relationship, as the teacher network in the semantic space, thereby enabling it to distinguish between real semantic changes and pseudo-changes.

[0067] Simulation Experiment 1. Experiment setup.

[0068] (1) Data set.

[0069] SYSU-CD: This dataset contains 4448 pairs of rigorously registered multi-temporal image pairs. Each pair is 512×512 pixels in size and includes corresponding change mask annotations to represent pixel-level variation regions.

[0070] LEVIR-CD: This dataset consists of remote sensing images and corresponding pixel-level change annotations, totaling 637 image pairs, each image being 1024×1024 pixels in size. The dataset contains change information on different buildings in urban areas, featuring high spatial resolution and complex backgrounds.

[0071] (2) Evaluation indicators.

[0072] OA: Overall classification accuracy; mIoU: Average value of the overlap between the predicted result and the actual changed area; mF1: Measures the overall performance of the model in detecting changes; Rec: The proportion of actual changed areas that are correctly detected; Pre: The proportion of predicted changes that are actually correct in the changed area.

[0073] (3) Hyperparameters.

[0074] Learning rate: 1×10 4 .

[0075] Batch Size: 16 (NVIDIA A100 GPU).

[0076] Training epochs: 200 epochs, policy (patience=20).

[0077] 2. Comparative experiment.

[0078] Seven models were selected and compared on the SYSU-CD and LEVIR-CD datasets: four were pure vision models (ISDANet, CFNet, RhighNet, and MPNet), and three were vision-language models (ChangeCLIP, MGCR, and MMChange). The experimental data obtained on the SYSU-CD dataset are shown in Table 1.

[0079] Table 1 Results of each comparison method on SYSU-CD

[0080] As shown in Table 1, the SYSU-CD dataset presents significant challenges due to complex seasonal variations and lighting interference. The method presented in this application achieves the best overall performance, with an mF1 score of 89.26% and an mIoU of 81.10%.

[0081] Compared to purely visual methods, the proposed method significantly outperforms RhighNet and ISDANet, improving mF1 by 1.5% and 1.02%, respectively. This improvement primarily stems from the introduction of textual semantic priors. Purely visual methods are susceptible to being misled by seasonal vegetation changes (e.g., leaves turning yellow), while this proposed method effectively suppresses such spurious changes by leveraging semantic consistency in the textual description, which acts as a semantic gating mechanism. This advantage is particularly evident in the improved recall score.

[0082] Compared to ChangeCLIP, the method in this application improves the mF1 score by 1.5%. Although both methods utilize CLIP, ChangeCLIP lacks explicit structural constraints in the feature space. In contrast, the method in this application constructs an environment-invariant visual prototype library, stabilizing feature representations under complex conditions, thereby achieving more accurate change detection.

[0083] The experimental data obtained on the LEVIR-CD dataset are shown in Table 2.

[0084] Table 2 Results of each comparison method on LEVIR-CD

[0085] As shown in Table 2, the LEVIR-CD dataset focuses on high-resolution building variations, imposing stringent requirements on edge integrity and geometric accuracy. Our proposed method again achieves the best overall performance, with an mF1 score of 95.77% and an mIoU of 92.14%.

[0086] Compared to purely visual methods, although methods such as RhighNet and MPNet have already demonstrated strong performance on this dataset (mF1 scores exceeding 95.5%, indicating performance saturation), our proposed method still achieves further improvements, with a 0.46% increase in mIoU compared to RhighNet. This result demonstrates that our proposed method is not only effective in addressing semantic ambiguity but also remains competitive in refined object detection against relatively simple backgrounds.

[0087] Purely visual methods often produce jagged edges or overly smooth boundaries when segmenting buildings. In contrast, the method in this application, through the synergy of a difference perception module and a channel-spatial attention module, effectively preserves pixel-level geometric details while incorporating high-level semantic information. This design allows the method in this application to achieve an optimal balance between accuracy and recall.

[0088] Compared to ChangeCLIP, the method in this application outperforms ChangeCLIP in both mF1 and mIoU. Although ChangeCLIP benefits from CLIP's generalization ability, its direct transfer strategy often reduces the spatial resolution of feature maps, making it difficult to accurately depict fine-grained building outlines. In contrast, the method in this application employs a multi-spatial distillation strategy, which not only transfers semantic knowledge from the teacher network but also enhances attention to local details through multi-scale feature fusion, ensuring accurate boundary depiction in high-resolution scenes.

[0089] 3. Ablation experiment.

[0090] To verify the effectiveness of the teacher-student framework and the proposed multi-space distillation strategy in the proposed method, stepwise ablation experiments were conducted on the SYSU-CD and LEVIR-CD datasets. The baseline model is defined as using only the basic task loss L... seg The student network was trained. Following the multi-spatial distillation strategy described in step S130, visual spatial distillation loss, textual spatial distillation loss, and prototype semantic structure alignment loss were progressively introduced. Experimental results are summarized in Table 3: Table 3 Ablation Experiment Results

[0091] The following conclusions can be drawn from Table 3: (1) The effectiveness of visual-spatial distillation loss. Introducing L... VIS Subsequently, the mF1 score on SYSU-CD increased from 87.76% to 89.26% (+1.50%). This improvement demonstrates that transferring general visual representations from CLIP to student networks can effectively enhance the generalization ability of low-level features.

[0092] However, relying solely on instance-level visual alignment is insufficient to fully distinguish between appearance changes caused by lighting and true semantic changes. Therefore, performance improvements gradually plateau, revealing the limitations of visual distillation alone.

[0093] (2) The Importance of Text Spatial Distillation Loss. Building upon visual spatial distillation, L is further introduced... TEXThis resulted in a significant performance improvement. On SYSU-CD, the mF1 score further increased to 89.11% (+0.74%), and the recall rate also improved significantly, from 81.07% to 82.43%.

[0094] When an image undergoes non-semantic appearance changes (such as seasonal color changes), the semantic consistency of the text description helps the model recover regions misclassified by purely visual features, thereby effectively suppressing spurious change signals.

[0095] (3) Contribution of prototype semantic structure alignment loss. The complete model introduces L based on TPB / VPB. SEM On SYSU-CD, the mF1 score reached 89.26%, and the mIoU was 81.10%. Although the improvement in mIoU was relatively small on the relatively simple LEVIR-CD dataset (an improvement of +0.59% compared to the baseline), the module's performance was more significant on the more challenging SYSU-CD dataset.

[0096] This behavior can be attributed to the characteristics of the dataset. LEVIR-CD primarily focuses on building variations, with relatively limited semantic categories and clear object boundaries, allowing the baseline model to learn reasonable distinguishable features. In contrast, SYSU-CD involves complex multi-class semantic ambiguity, where the advantages of our method are more pronounced. This also explains the larger performance improvement observed on SYSU-CD.

[0097] Meanwhile, these results validate the necessity of constructing an environment-invariant VPB. By constraining student features to focus on a stable semantic center, the model exhibits stronger robustness under extreme imaging conditions (such as cloud shadows and viewpoint changes), successfully achieving the crucial transition from pixel-level alignment to semantic-level structural alignment.

[0098] Experiments show that the proposed method achieves optimal performance in semantic change detection tasks of remote sensing images across multiple scenarios, validating the effectiveness of the triple encoding / decoding framework. Ablation experiments further quantify the contribution of each module, providing direction for subsequent optimization.

[0099] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0100] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A remote sensing change detection method based on a teacher-student framework and multi-spatial distillation, characterized in that, include: In the teacher network, a text prototype library is constructed using preset land cover categories, and a visual prototype library is initialized. During the training process, the visual prototype library is updated with momentum based on the similarity between the training samples and the text prototype library. In the student network, the image encoder in the teacher network is used to perform zero-shot prediction on the image to be detected, and text guidance features are generated based on the prediction results; the image to be detected is input into the student network to extract multi-scale feature maps, global semantic features and spatial features. The difference perception module calculates absolute difference and similarity difference based on the multi-scale feature map, and calculates text guidance difference based on the spatial features and the text guidance features. The absolute difference, similarity difference, and text guidance difference are then concatenated with the original features of the image to be detected to form initial change features. The channel-spatial attention module calculates channel attention and spatial attention on the initial change features, and the channel attention and spatial attention are fused to obtain multi-scale fusion features. The multi-scale fusion features are then concatenated with the global semantic features to form concatenated features. The spliced ​​features are input into the decoder to obtain a binary change detection map; During the training of the student network, a triple semantic alignment loss function is established based on the basic task loss, visual spatial distillation loss, textual spatial distillation loss, and prototype semantic structure alignment loss. The student network is then trained using the triple semantic alignment loss function.

2. The remote sensing change detection method based on the teacher-student framework and multi-spatial distillation according to claim 1, characterized in that, During the training of the student network, the parameters of the teacher network remain frozen and are used only to construct the text prototype library and the visual prototype library.

3. The remote sensing change detection method based on the teacher-student framework and multi-spatial distillation according to claim 1, characterized in that, The method for constructing the text prototype library is as follows: Fill each of the preset land feature categories into the preset prompt template to generate the corresponding prompt text description; The features of the prompt text description are extracted using a text encoder, and L2 normalization is performed to obtain the corresponding text prototype vectors. All the text prototype vectors constitute the text prototype library.

4. The remote sensing change detection method based on the teacher-student framework and multi-spatial distillation according to claim 1, characterized in that, The method for updating the momentum of the aforementioned visual prototype library is as follows: Calculate the cosine similarity between the training sample and all text prototype vectors in the text prototype library, and find the text category index with the highest cosine similarity. Update the visual prototype in the visual prototype library corresponding to the text category index with the highest cosine similarity using the features of the training samples: in, and These are the visual prototypes before and after the update, respectively. The momentum coefficient, The image features are output by the visual encoder in the teacher network.

5. The remote sensing change detection method based on the teacher-student framework and multi-spatial distillation according to claim 1, characterized in that, The method for generating the text guidance features is as follows: The image encoder in the teacher network is used to perform zero-shot prediction on the image to be detected, so as to obtain the probability that the image to be detected belongs to each of the land cover categories; Select the multiple land cover categories with the highest probabilities, sort them from high to low probabilities, and form a natural language prompt describing the content of the image to be detected; The natural language prompts are input into a text encoder to extract the text guidance features.

6. The remote sensing change detection method based on teacher-student framework and multi-spatial distillation according to claim 1, characterized in that, While the student network extracts the multi-scale feature map, attention pooling is performed on the features extracted in stage four of the student network to extract the global semantic features; at the same time, the features extracted in stage four are used to generate the spatial features through convolutional layers.

7. The remote sensing change detection method based on teacher-student framework and multi-spatial distillation according to claim 1, characterized in that, The absolute difference is expressed as: in, For the student network in the stage s The absolute difference mentioned above, and The multi-scale features of phases A and B are respectively. Indicates taking the absolute value; The similarity difference is represented as: in, Represents cosine similarity, [ h,w [ ] represents spatial location coordinates. and These are two phases in the stage. s The feature vector at that position, Represents the dot product of vectors; for The similarity difference is mapped to the sigmoid function. The text-guided differences are represented as follows: in, For the student network in the stage s The differences in the text guidance mentioned above. and The spatial-level features for phases A and B are respectively. and The text guidance features are for phases A and B, respectively.

8. The remote sensing change detection method based on teacher-student framework and multi-spatial distillation according to claim 1, characterized in that, The triple semantic alignment loss function is expressed as follows: in, The triple semantic alignment loss function is... , , and These are the basic task loss, the visual spatial distillation loss, the textual spatial distillation loss, and the prototype semantic structure alignment loss, respectively. and They are 0.5 and 0.4 respectively.

Images