Remote sensing image semantic change detection method based on semantic adaptation and time domain alignment
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-26
- Publication Date
- 2026-08-11
AI Technical Summary
[0003]现有的遥感变化检测算法采用二值变化检测任务，即系统输出的结果通常仅指出图像中哪些区域发生了变化，而未提供关于这些变化的具体语义信息，如变化的类型、性质或具体类别；这种信息的缺失在某些应用场景中可能无法满足实际需求，特别是在遥感图像的解译、土地利用变化监测和灾害响应等领域，需要更具辨识度和解释性的检测结果；为此，语义变化检测作为一种更具信息含量的图像分析任务被提出，该任务不仅识别出图像中发生变化的区域，还能进一步将这些变化归类为若干预定义的语义类别，从而生成包含“从-到”语义标签的变化图
[0039]本发明相对于现有技术具备的有益效果为：本发明提供一种基于语义自适应和时域对齐的遥感图像语义变化检测方法，能够缓解遥感图像中的语义信息不足，通过设计一个原型时域对齐模块来减少类间差异和域差，该模块有效地最小化了双时相图像对之间的域差，从而减轻了伪变化引起的误检；本发明通过在模型中引入频域感知特征融合解码器，减少了上采样过程中的类内不一致性并且恢复了下采样过程中丢失的高频边界细节，从而提高了模型检测变化物体的边界精度，整体模型易于实施，准确率高。
Smart Images
Figure CN122551206A_ABST
Abstract
Description
Technical Field
[0001] This invention provides a method for detecting semantic changes in remote sensing images based on semantic adaptation and temporal alignment, belonging to the field of remote sensing image semantic change detection technology. Background Technology
[0002] Change detection is the process of identifying differences in the state of an object or phenomenon by observing it at different times. Change detection based on remote sensing images is an important method for detecting changes on the Earth's surface and has wide applications in urban planning, environmental monitoring, agricultural surveys, disaster assessment, and map revision.
[0003] Existing remote sensing change detection algorithms employ binary change detection tasks, meaning the system output typically only indicates which regions in the image have changed, without providing specific semantic information about these changes, such as the type, nature, or specific category of the change. This lack of information may not meet practical needs in certain application scenarios, particularly in areas such as remote sensing image interpretation, land use change monitoring, and disaster response, where more discriminative and interpretable detection results are required. To address this, semantic change detection has been proposed as a more information-rich image analysis task. This task not only identifies regions in the image that have changed but also further categorizes these changes into several predefined semantic categories, thereby generating a change map containing "from-to" semantic labels.
[0004] Despite the rapid development of high spatial resolution and multi-temporal remote sensing technologies, existing semantic change detection methods still face challenges such as insufficient semantic information and domain gaps between two-temporal images. They also suffer from a lack of fine-grained identification and classification of change types, making it urgent to develop a semantic change detection scheme based on semantic adaptation and temporal alignment. Summary of the Invention
[0005] To address the technical problems existing in the background art, the present invention adopts the following technical solution: providing a method for detecting semantic changes in remote sensing images based on semantic adaptation and temporal alignment, comprising the following detection steps:
[0006] Step 1: Obtain the pre-change temporal remote sensing image T1 and the post-change temporal remote sensing image T2, and input the pre-change and post-change images into the trained semantic adaptation and temporal alignment network SATANet;
[0007] The semantic adaptation and temporal alignment network SATANet uses the Siamese feature extraction network SAM2 as the encoder to extract multi-scale features from the pre-change temporal remote sensing image T1 and the post-change temporal remote sensing image T2, respectively, to obtain the corresponding multi-level feature maps. and ;
[0008] Step 2: Convert multi-level feature maps and The input is fed into the prototype temporal alignment module PTA in the semantic adaptation and temporal alignment network SATANet. By constructing a dual-temporal feature prototype and performing cross-temporal semantic alignment, it aggregates similar land cover features in the images before and after the change, distinguishes different types of land cover features, reduces intra-class differences and enhances inter-class separability, and obtains multi-level feature maps before and after the change after temporal alignment.
[0009] Step 3: Input the time-domain aligned, before-and-after multi-level feature maps into the frequency-domain perceptual feature fusion decoder FFDecoder to obtain the upsampled feature map;
[0010] Step 4: Input the upsampled feature map into the change detection head and the semantic segmentation head to obtain the final predicted change detection map and semantic segmentation map. During the training phase, the semantic segmentation loss function and the binary change detection loss are used to optimize the semantic adaptation function, the semantic adaptation and temporal alignment network SATANet. During the inference phase, the predicted semantic change detection results are obtained directly.
[0011] Step 5: Use the trained and optimized feature extraction network SAM2 to obtain multi-level features of the pre-change time-phase remote sensing image T1 and the post-change time-phase remote sensing image T2. Input the multi-level features into the trained and optimized frequency domain sensing feature fusion decoder FFDecoder to obtain upsampled features.
[0012] The change detection head and semantic segmentation head are used to perform change detection and semantic segmentation on the upsampled dual-temporal remote sensing image to be detected, and the corresponding change detection map and semantic segmentation map are obtained.
[0013] The specific method for step 1 is as follows:
[0014] To adapt the pre-trained feature extraction network SAM2 to scene characteristics such as varying land cover scales and bi-temporal distribution shifts caused by differences in illumination and season in remote sensing images, low-rank adaptive LoRA weights are integrated into each self-attention layer of the pre-trained ViT. For specific remote sensing image semantic change detection tasks, low-rank decomposition matrices A and B are integrated into the pre-trained query and value projection matrices of each self-attention layer. and In the self-attention module, low-rank adaptive LoRA modifies the pre-trained query and value projection matrices by merging low-rank updates. and These updates are parameterized using low-rank decomposition to ensure computational efficiency during the fine-tuning process, which is expressed as follows:
[0015] ;
[0016] in, and Representing the input and output features respectively; matrix and Define a low-rank decomposition matrix, which is applied to the pre-trained weights. Weight update Modeling is performed.
[0017] The specific method for step 2 is as follows:
[0018] In the time domain of the T1 image, the feature centroid class prototype of each class The expression is:
[0019] ;
[0020] Where N represents the number of T1 images; H and W are feature maps. The shape; This represents the pixel-level features of the T1 image extracted by the feature extractor. This represents the corresponding class label; c is the index of the number of classes C. Represents an index function, when When the condition is met, the index function equals 1; otherwise, it equals 0. Given a T2 image, the feature map is obtained after passing through the feature extractor. , With each prototype Similarity between Calculate as follows:
[0021] ;
[0022] In the time-domain T2 image, annotated pixel features are attracted to prototypes of their respective categories while being repelled by prototypes of other categories. This process is guided by a contrastive loss function, expressed as:
[0023] ;
[0024] To improve the domain invariance of the prototype, a dynamic update mechanism is adopted. This mechanism integrates feature information from two time domains and refines the prototype iteratively to ensure robust alignment. The expression is:
[0025] ;
[0026] in, It is a hyperparameter used to balance the contributions from the time domain of the T1 and T2 images; This represents the estimated time-domain prototype T1. This represents the estimated time-domain prototype T2.
[0027] The frequency domain-aware feature fusion decoder FFDecoder used in step 3 specifically includes: an adaptive low-pass filter generator, an offset generator, and an adaptive high-pass filter generator.
[0028] The specific method for step 4 is as follows:
[0029] To achieve semantic change detection, the expression for the overall loss L is defined as follows:
[0030] ;
[0031] in, It is a loss due to a binary change. It is a comparison of losses. It is a semantic segmentation loss, used to perform multi-class semantic classification on each pixel of the image before and after the change. By minimizing this loss function, the model learns the ability to accurately predict semantic categories.
[0032] The cross-entropy loss function is used to calculate the error of pixel-level semantic classification, and the expression is as follows:
[0033] ;
[0034] Where M represents the total number of semantic categories, This represents the true probability of the i-th category. This represents the predicted probability of assigning a pixel to the i-th category; when That is, the ground truth map indicates which pixel belongs to which category. At that time, the loss will drive the model to increase. This ensures that pixels are classified into the correct categories;
[0035] Binary change loss The binary cross-entropy loss function is used to achieve binary classification based on pixel-level changes / no changes, and the expression is:
[0036] ;
[0037] Where N is the total number of pixels in the image; when That is, when the ground truth map indicates changes, the loss will penalize the model by assigning a lower value. This value encourages the model to detect changes more accurately.
[0038] The above Ensure that each pixel is correctly classified into the correct semantic category. This helps the model distinguish between changed and unchanged regions at the pixel level. It improves feature alignment between two temporal images, which can minimize domain differences and reduce false detections in change detection.
[0039] The advantages of this invention compared to existing technologies are as follows: This invention provides a semantic change detection method for remote sensing images based on semantic adaptation and temporal alignment, which can alleviate the lack of semantic information in remote sensing images. By designing a prototype temporal alignment module to reduce inter-class differences and domain differences, this module effectively minimizes the domain difference between two temporal image pairs, thereby reducing false detections caused by spurious changes. This invention introduces a frequency domain-aware feature fusion decoder into the model, reducing intra-class inconsistencies during upsampling and restoring high-frequency boundary details lost during downsampling, thereby improving the boundary accuracy of the model in detecting changing objects. The overall model is easy to implement and has high accuracy. Attached Figure Description
[0040] The present invention will be further described below with reference to the accompanying drawings:
[0041] Figure 1 This is an architecture diagram of the semantic adaptation and temporal alignment network of this invention;
[0042] Figure 2 This is a diagram of the low-rank adaptive architecture proposed in the SAM2 feature extraction network of this invention;
[0043] Figure 3 This is a schematic diagram of the frequency domain sensing feature fusion decoder of the present invention;
[0044] Figure 4 This is a schematic diagram of the generator structure in the frequency domain sensing feature fusion decoder of the present invention. Detailed Implementation
[0045] like Figures 1 to 4 As shown, this invention addresses the lack of precise identification and classification of change types in existing change detection methods by proposing a semantic change detection method for remote sensing images based on the semantic adaptation and temporal alignment network SATANet. This method alleviates the insufficient semantic information in remote sensing images. The prototype temporal alignment module PTA is designed to reduce inter-class differences and domain discrepancies. This module effectively minimizes the domain discrepancy between pairs of two temporal images, thereby mitigating false detections caused by spurious changes. By introducing the frequency domain-aware feature fusion decoder FFDecoder into the model, intra-class inconsistencies during upsampling are reduced, and high-frequency boundary details lost during downsampling are recovered, thus improving the model's accuracy in detecting the boundaries of changing objects.
[0046] This invention presents a self-designed semantic adaptation and temporal alignment network, SATANet, for detecting semantic changes in remote sensing images. Before detection, a SATANet network is first constructed, comprising a feature extraction network SAM2, a prototype temporal alignment module PTA, and a frequency-domain perceptual feature fusion decoder FFDecoder. The constructed SATANet network needs to be trained and optimized using remote sensing change detection training samples before it can be used for change detection tasks. After training, the optimized SATANet is then used to perform change detection and semantic segmentation on the dual-temporal remote sensing images to be detected.
[0047] Furthermore, a preferred embodiment of the present invention includes the following steps:
[0048] Step 1: Feature Map Extraction
[0049] The Semantic Adaptation and Temporal-domain Alignment Network (SATANet) used in this invention is Figure 1 Both the before-and-after images are optical remote sensing images, each with a resolution of 512×512 pixels, RGB three channels, and a spatial resolution ranging from 0.5 to 3 meters per pixel; the two images are input into the twin feature extraction network SAM2 ( Figure 1 The SAM2 feature extraction network, pre-trained on a large number of natural images, performs excellently in extracting semantic features. However, due to the inherent domain gap between natural images and remotely sensed images, Low Rank Adaptation (LoRA) is used to fine-tune the pre-trained SAM2 feature extraction network.
[0050] Semantic adaptation of the SAM2 feature extraction network:
[0051] To effectively transfer the Visual Foundation Model (VFM) trained on natural images to change detection tasks on remote sensing images, the SATANet framework employs a popular parameter-efficient fine-tuning technique called LoRA (LoRa). Figure 2 LoRA (Local Language Automation) is designed to achieve semantic adaptation between remotely sensed images and natural images. Originally designed for fine-tuning large language models, LoRA operates on the assumption that the weight updates of pre-trained VFMs are sparse and exhibit low intrinsic rank during fine-tuning. The core concept of LoRA is to freeze the pre-trained model weights and incorporate trainable low-rank matrices into each layer. This approach significantly reduces the number of trainable parameters required for fine-tuning, thereby reducing computational resource demands while maintaining model performance.
[0052] To address the specific challenges of semantic change detection in remote sensing images, LoRA weights are integrated into each self-attention layer of a pre-trained ViT, such as... Figure 2 As shown. For a specific remote sensing image semantic change detection task, the low-rank decomposition matrices (denoted as A and B) are integrated into the pre-trained query and value projection matrices (denoted as A and B) of each self-attention layer. and In the self-attention module, LoRA modifies the pre-trained query and value projection matrices by merging low-rank updates. and These updates are parameterized using low-rank decomposition, thus ensuring computational efficiency during the fine-tuning process. This process is represented as:
[0053] ;
[0054] in, and Representing the input and output features respectively; matrix and Define a low-rank decomposition matrix, which is applied to the pre-trained weights. Weight update Modeling is performed.
[0055] The rank r of these matrices is set to be significantly smaller than the feature dimension d, thus ensuring that the number of trainable parameters remains minimal. Furthermore, the scaling factor... Controlling the magnitude of LoRA updates prevents excessive deviation from pre-trained weights while providing sufficient flexibility for domain adaptation.
[0056] Step 2: Prototype temporal alignment:
[0057] This invention addresses the change detection task in dual-temporal images. Semantic change detection in remote sensing images is severely challenged by domain gaps caused by differences in time, environment, or instrumentation between dual-temporal images. These differences often lead to inconsistencies in feature representations, resulting in poor semantic alignment and an increased likelihood of spurious changes. To address these challenges, this invention designs a Prototype Domain Adaptation (PTA) module.
[0058] In semantic change detection tasks, features within the same semantic class often exhibit considerable variation due to temporal differences. For example... Figure 3As shown, the PTA module addresses this issue by reducing intra-class variability, bringing features of the same class closer to their corresponding class prototypes. However, remote sensing images often contain similar features from different classes (e.g., trees and low vegetation), which can lead to ambiguous classification. To address this, the PTA module introduces a repulsion mechanism that pushes features from different classes apart based on their prototypes. By increasing the separation between inter-class features, this mechanism enhances the model's ability to distinguish classes and reduces misclassifications caused by feature overlap.
[0059] At the heart of the PTA module is its contrastive learning framework, where prototypes act as anchors to attract features from the same class (positive samples) and repel features from different classes (negative samples). This process not only improves the robustness of feature representations but also enhances the model's ability to capture meaningful semantic differences, further improving the accuracy of semantic change detection outputs. Real-world remote sensing images often contain multiple intra-class variations, manifested as differences in texture, shape, or spectral properties within the same land cover category. The PTA module effectively preserves this diversity by dynamically updating prototypes to encapsulate the comprehensive features of each class while ensuring feature consistency across the temporal domain. This adaptive approach allows the module to remain sensitive to subtle intra-class variations while maintaining semantic alignment across time.
[0060] The PTA module is a domain-adaptive technique designed to enhance inter-class separability and intra-class consistency while improving segmentation performance in remote sensing image semantic change detection tasks. The PTA module achieves this by leveraging class prototype modeling, which mitigates domain differences between bitemporal images and enhances feature alignment. In the temporal domain of the T1 image, the feature centroid of each class (called the class prototype)... ) Use the following formula to calculate:
[0061] (1);
[0062] Where N represents the number of T1 images; H and W are feature maps. The shape; This represents the pixel-level features of the T1 image extracted by the feature extractor. This represents the corresponding class label; c is the index of the number of classes C. Represents an index function, when When the condition is met, the index function equals 1; otherwise, it equals 0. Given a T2 image, the feature map is obtained after passing through the feature extractor. , With each prototype Similarity between It can be calculated as follows:
[0063] (2);
[0064] In the time-domain T2 image, annotated pixel features are attracted to prototypes of their respective classes while being repelled by prototypes of other classes. This process is guided by a contrastive loss function, defined as:
[0065] (3);
[0066] To improve the domain invariance of the prototype, a dynamic update mechanism is adopted; this mechanism integrates feature information from two time domains and refines the prototype iteratively to ensure robust alignment.
[0067] (4);
[0068] in, It is a hyperparameter used to balance the contributions from the time domain of the T1 and T2 images; This represents the estimated time-domain prototype T1. The estimated time-domain T2 prototypes are calculated using formula (1). By dynamically adjusting the prototype representation, this mechanism ensures that the prototypes remain representative across domains, thereby mitigating spurious changes caused by domain differences. Furthermore, the PTA module is applied to the feature encoder to facilitate semantic alignment. By simultaneously minimizing intra-class differences and maximizing inter-class separability, the PTA module effectively enhances the model's robustness to domain-induced differences, reduces the ambiguity of semantic boundaries, and improves the accuracy of the generated semantic change map. This feature is particularly beneficial for addressing challenges in remote sensing images, such as the presence of class diversity and small irregular features.
[0069] Step 3: Use a decoder to perform frequency domain feature fusion:
[0070] This invention addresses the semantic change detection task of remote sensing images, a dense prediction task that combines pixel-level binary change detection with semantic segmentation. These tasks require accurate preservation of spatial boundaries in high-resolution images while effectively extracting information from specific classes. However, during feature fusion in high-resolution remote sensing image tasks, boundary information often becomes blurred or misaligned, leading to inaccurate boundary depiction. To address these challenges, this invention proposes a novel Frequency-aware Feature Fusion Decoder (FFDecoder). By introducing the FFDecoder, the loss of boundary information during upsampling is mitigated. By preserving high-frequency details, the decoder ensures improved edge integrity in the generated change map, thereby more accurately depicting the changed regions. Figure 3As shown, it utilizes frequency domain analysis and feature manipulation to mitigate boundary misalignment issues and improve the performance of dense prediction tasks. The pixel shuffle operation restores spatial resolution by rearranging the feature map from 4×C×H×W to C×2H×2W. Furthermore, in the initial and final fusion stages, the adaptive low-pass filter (ALPF) generator and the adaptive high-pass filter (AHPF) share the same parameter settings, ensuring consistency in feature refinement.
[0071] FFDecoder consists of three key components (such as...) Figure 4 As shown), each component is designed to address key challenges related to semantic change detection. The pixel unshuffle operation halves the spatial dimension of the feature map while expanding the channel dimension by a factor of four, dividing it into four groups, for example, changing from C×2H×2W to 4×C×H×W; where:
[0072] (1) Adaptive Low-Pass Filter (ALPF) generator (e.g.) Figure 4 (as shown in part a)
[0073] This module smooths high-frequency noise to enhance intra-class consistency while preserving fine-grained details. It generates dynamic low-pass filters using convolutional layers and a softmax mechanism, spatially adapting to local features in different regions. The filtered high-level features are then upsampled to ensure cohesion and consistency of intra-class features across different spatial resolutions.
[0074] (2) Offset generator (e.g.) Figure 4 (as shown in part b)
[0075] To refine boundary details and correct feature misalignment, this module calculates local offsets based on feature similarity. High-level features are then dynamically resampled using offsets with adjusted orientation and magnitude within a defined space. By replacing low-similarity regions with more consistent neighboring features, the offset generator effectively sharpens boundary definitions and aligns semantic features across spatial domains.
[0076] (3) Adaptive High-Pass Filter (AHPF) generator (e.g.) Figure 4 (as shown in part c)
[0077] This module recovers high-frequency details lost during upsampling, thereby improving boundary sharpness and edge accuracy. A dynamic high-pass filter is generated through a combination of convolution operations and filter inversion. This process is applied to low-level features, emphasizing high-frequency edge information to recover fine boundary details, which is crucial for improving the accuracy of semantic change detection.
[0078] By combining these three components, FFDecoder addresses common limitations in semantic change detection tasks, including blurred boundaries, inconsistent features, and loss of fine-grained spatial details. This decoder ensures that dense prediction models preserve accurate spatial boundaries, robust feature alignment, and class-specific discriminability, ultimately improving the accuracy and reliability of semantic change detection in high-resolution remote sensing images.
[0079] Step 4: Joint training based on multiple loss functions:
[0080] To optimize the semantic change detection model SATANet proposed in this invention, three loss functions are employed: semantic segmentation loss, binary change detection loss, and prototype contrast loss. These losses aim to jointly address the unique challenges of semantic change detection, such as maintaining semantic accuracy, recognizing binary changes, and ensuring feature alignment across time domains. The overall loss is defined as follows:
[0081] (5);
[0082] in, It is a loss due to a binary change. It is the contrast loss shown in Formula 3. This is a semantic segmentation loss that ensures the accuracy of multi-class predictions of semantic categories across bitemporal images. It is implemented using cross-entropy loss, expressed as:
[0083] (6);
[0084] Where M represents the total number of semantic categories, This represents the true probability of the i-th category. This represents the predicted probability of assigning a pixel to the i-th category; when (That is, the ground truth map indicates that a pixel belongs to a category) When this happens, the loss will drive the model to increase. This ensures that pixels are classified into the correct categories;
[0085] Binary change loss Focusing on identifying pixel-level changes between two temporal images, it achieves binary classification of pixel-level changes / no changes by employing a binary cross-entropy loss function. It evaluates the model's ability to classify pixels as "changed" or "unchanged," expressed as:
[0086] (7);
[0087] Where N is the total number of pixels in the image; when When the ground truth map indicates a change, the loss will penalize the model with a lower assignment. This value encourages the model to detect changes more accurately.
[0088] The above Ensure that each pixel is correctly classified into the correct semantic category. This helps the model distinguish between changed and unchanged regions at the pixel level. Improved feature alignment between bi-temporal images, minimized domain differences and reduced false detections in change detection; these loss functions work together to ensure the model can achieve accurate and robust semantic change detection.
[0089] Step 5: Perform change detection using the trained SATANet model:
[0090] After the training phase, the multi-level features of images T1 and T2 are obtained using the trained network. First, the upsampled multi-level features are detected and segmented using a change detection head and a semantic segmentation head to obtain the corresponding change detection map and semantic segmentation map.
[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for detecting semantic changes in remote sensing images based on semantic adaptation and temporal alignment, characterized in that: The testing steps include the following: Step 1: Obtain the pre-change temporal remote sensing image T1 and the post-change temporal remote sensing image T2, and input the pre-change and post-change images into the trained semantic adaptation and temporal alignment network SATANet; The semantic adaptation and temporal alignment network SATANet uses the Siamese feature extraction network SAM2 as the encoder to extract multi-scale features from the pre-change temporal remote sensing image T1 and the post-change temporal remote sensing image T2, respectively, to obtain the corresponding multi-level feature maps. and ; Step 2: Convert multi-level feature maps and The input is fed into the prototype temporal alignment module PTA in the semantic adaptation and temporal alignment network SATANet. By constructing a dual-temporal feature prototype and performing cross-temporal semantic alignment, it aggregates similar land cover features in the images before and after the change, distinguishes different types of land cover features, reduces intra-class differences and enhances inter-class separability, and obtains multi-level feature maps before and after the change after temporal alignment. Step 3: Input the time-domain aligned, before-and-after multi-level feature maps into the frequency-domain perceptual feature fusion decoder FFDecoder to obtain the upsampled feature map; Step 4: Input the upsampled feature map into the change detection head and the semantic segmentation head to obtain the final predicted change detection map and semantic segmentation map. During the training phase, the semantic segmentation loss function and the binary change detection loss are used to optimize the semantic adaptation function, the semantic adaptation and temporal alignment network SATANet. During the inference phase, the predicted semantic change detection results are obtained directly. Step 5: Use the trained and optimized feature extraction network SAM2 to obtain multi-level features of the pre-change time-phase remote sensing image T1 and the post-change time-phase remote sensing image T2. Input the multi-level features into the trained and optimized frequency domain sensing feature fusion decoder FFDecoder to obtain upsampled features. The change detection head and semantic segmentation head are used to perform change detection and semantic segmentation on the upsampled dual-temporal remote sensing image to be detected, and the corresponding change detection map and semantic segmentation map are obtained.
2. The remote sensing image semantic change detection method based on semantic adaptation and temporal alignment according to claim 1, characterized in that: The specific method for step 1 is as follows: To adapt the pre-trained feature extraction network SAM2 to scene characteristics such as varying land cover scales and bi-temporal distribution shifts caused by differences in illumination and season in remote sensing images, low-rank adaptive LoRA weights are integrated into each self-attention layer of the pre-trained ViT. For specific remote sensing image semantic change detection tasks, low-rank decomposition matrices A and B are integrated into the pre-trained query and value projection matrices of each self-attention layer. and In the self-attention module, low-rank adaptive LoRA modifies the pre-trained query and value projection matrices by merging low-rank updates. and These updates are parameterized using low-rank decomposition to ensure computational efficiency during the fine-tuning process, which is expressed as follows: ； in, and Representing the input and output features respectively; matrix and Define a low-rank decomposition matrix, which is applied to the pre-trained weights. Weight update Modeling is performed.
3. The remote sensing image semantic change detection method based on semantic adaptation and temporal alignment according to claim 2, characterized in that: The specific method for step 2 is as follows: In the time domain of the T1 image, the feature centroid class prototype of each class The expression is: ； Where N represents the number of T1 images; H and W are feature maps. The shape; This represents the pixel-level features of the T1 image extracted by the feature extractor. This represents the corresponding class label; c is the index of the number of categories C. Represents an index function, when When the condition is met, the index function equals 1; otherwise, it equals 0. Given a T2 image, the feature map is obtained after passing through the feature extractor. , With each prototype similarity between Calculate as follows: ； In the time-domain T2 image, annotated pixel features are attracted to prototypes of their respective categories while being repelled by prototypes of other categories. This process is guided by a contrastive loss function, expressed as: ； To improve the domain invariance of the prototype, a dynamic update mechanism is adopted. This mechanism integrates feature information from two time domains and refines the prototype iteratively to ensure robust alignment. The expression is: ； in, It is a hyperparameter used to balance the contributions from the time domain of the T1 and T2 images; This represents the estimated time-domain prototype T1. This represents the estimated time-domain prototype T2.
4. The remote sensing image semantic change detection method based on semantic adaptation and temporal alignment according to claim 3, characterized in that: The frequency domain-aware feature fusion decoder FFDecoder used in step 3 specifically includes: an adaptive low-pass filter generator, an offset generator, and an adaptive high-pass filter generator.
5. The remote sensing image semantic change detection method based on semantic adaptation and temporal alignment according to claim 4, characterized in that: The specific method for step 4 is as follows: To achieve semantic change detection, the expression for the overall loss L is defined as follows: ； in, It is a loss due to a binary change. It is a comparison of losses. It is a semantic segmentation loss, used to perform multi-class semantic classification on each pixel of the image before and after the change. By minimizing this loss function, the model learns the ability to accurately predict semantic categories. The cross-entropy loss function is used to calculate the error of pixel-level semantic classification, and the expression is as follows: ； Where M represents the total number of semantic categories, This represents the true probability of the i-th category. This represents the predicted probability of assigning a pixel to the i-th category; when That is, the ground truth map indicates which pixel belongs to which category. At that time, the loss will drive the model to increase. This ensures that pixels are classified into the correct categories; Binary change loss The binary cross-entropy loss function is used to achieve binary classification based on pixel-level changes / no changes, and the expression is: ； Where N is the total number of pixels in the image; when That is, when the ground truth map indicates changes, the loss will penalize the model by assigning a lower value. This value encourages the model to detect changes more accurately. The above Ensure that each pixel is correctly classified into the correct semantic category. This helps the model distinguish between changed and unchanged regions at the pixel level. It improves feature alignment between two temporal images, which can minimize domain differences and reduce false detections in change detection.