Remote sensing image change detection method based on space-time multi-scale perception

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a time-space multi-scale perception remote sensing image change detection method. The method mainly comprises the following steps: acquiring a multi-temporal remote sensing image pair from a known data set, and preprocessing to construct a training set; constructing a remote sensing image change detection model; training the constructed remote sensing image change detection model by using the training data; and performing change detection on an input multi-temporal image pair by using the trained remote sensing image change detection model. The invention designs a space-time interleaving enhancement module and a multi-scale attention reweighting mechanism, which are used for realizing efficient feature interaction and characterization optimization in remote sensing change detection. The space-time interleaving enhancement module realizes time sequence fusion through bidirectional cross-time domain feature interaction; according to the multi-scale attention reweighting mechanism, single-scale local difference reservation and multi-scale semantic consistency are integrated, and a self-adaptive weighting strategy is adopted, so that the feature discrimination ability is enhanced while the structural integrity of a change graph is ensured.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method for detecting changes in remote sensing images with spatiotemporal multi-scale perception. Background Technology

[0002] The combined effects of human activities and natural forces, particularly the acceleration of global climate change and urbanization, have led to significant changes in land cover and ecosystem structure. These changes highlight the urgent need for monitoring technologies to promote sustainable disaster reduction, resource management, and urban planning. Against this backdrop, remote sensing image change detection (CD) has become a key technology for identifying surface changes by comparing multi-temporal satellite imagery.

[0003] Traditional classification methods (CD) are generally divided into three categories: algebra-based pixel comparison methods, transformation-based methods that project features into a change-sensitive subspace, and supervised learning-based classification methods. However, these traditional methods often rely on hand-designed features and empirical thresholds, which limits their generalization ability, especially in complex geographic environments.

[0004] With the development of deep learning, convolutional neural network (CNN)-based architectures have become mainstream. The CD task is popular due to its powerful feature representation capabilities and end-to-end learning framework. Subsequent architectural innovations such as multi-scale fusion, attention mechanisms, deep supervision, and dense connections have further improved model performance. However, the inherent locality of convolutional operations limits these models' ability to capture long-range dependencies, leading to serious limitations in large-scale landscape analysis. This limitation has spurred the development of Transformer-based methods, which effectively capture global contextual relationships using self-attention mechanisms. However, their practical deployment faces significant challenges due to the quadratic computational complexity of attention operations, especially when processing high-resolution remote sensing images. Recent breakthroughs in state-space models (SSMs), particularly the Mamba architecture, have achieved linearization of computational complexity while maintaining excellent sequence modeling performance, providing innovative solutions for related technological fields. Furthermore, the VMamba architecture, by extending the Mamba framework to the domain of two-dimensional image data processing, has demonstrated significant technical advantages in computer vision tasks. Inspired by these studies, recent work has extended Mamba-based architectures to the CD task.

[0005] Despite the progress made by existing Mamba-based methods in enhancing the global receptive field, they still suffer from the following problems:

[0006] 1. Existing methods have significant shortcomings in cross-temporal feature interaction, making it difficult to effectively model the spatiotemporal correlation between two temporal images. This results in the generated change detection map being susceptible to background interference factors such as seasonal changes and illumination changes, leading to significant pseudo-change noise and causing the problem of missing real ground feature changes.

[0007] 2. Existing methods have shortcomings in the fusion mechanism of local detailed features and global contextual information. This insufficient ability to integrate multi-scale features leads to two typical problems in change detection results in dense prediction tasks: first, the boundary localization of the changed region is unclear; second, a large number of false detections are generated. These problems ultimately severely restrict the overall accuracy of change detection. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a method for detecting changes in remote sensing images based on spatiotemporal multi-scale perception.

[0009] The technical solution adopted in this invention is as follows:

[0010] A method for detecting changes in remotely sensed images at multiple spatiotemporal scales includes the following steps:

[0011] S1. Obtain multi-temporal image pairs from the known dataset to construct the training dataset;

[0012] S2. Construct a remote sensing image change detection model, which includes a multi-level feature pair extraction module, a spatiotemporal interaction enhancement module, a single-scale feature fusion module, a multi-scale feature aggregation module, and a classifier.

[0013] The multi-level feature pair extraction module employs a shared-weight dual-branch feature extractor based on a pre-trained VMamba model to extract multi-level features from the input dual-temporal remote sensing image pairs, thus obtaining multi-level feature pairs.

[0014] The spatiotemporal interaction enhancement module processes multi-level feature pairs into preliminary change features. The processing procedure is as follows: first, multi-level feature pairs are stacked alternately along the width dimension at the same level to construct a new sequence. After processing the new sequence using visual state space blocks, it is then divided into two temporal features along the width dimension. Finally, the two temporal features are placed alternately along the channel dimension to obtain preliminary change features.

[0015] The single-scale feature fusion module fuses the two temporal features and the preliminary change features obtained from the spatiotemporal interaction enhancement module into a single-level change feature. The processing procedure is as follows: first, the two temporal features are fused with the preliminary change features respectively to obtain two single-temporal change features, and then the two single-temporal change features are fused to obtain a single-level change feature.

[0016] The multi-scale feature aggregation module processes single-level variation features into aggregated features. The processing steps are as follows: first, bilinear interpolation is used to upsample the high-level features of the single-level variation features to obtain multiple resolution-aligned features with consistent channels. Then, the weights of the multiple aligned features are calculated and adjusted through a channel attention mechanism. Finally, the obtained weights are assigned to the corresponding aligned features. Then, the weighted aligned features are summed element by element to generate the final aggregated features.

[0017] The classifier performs a linear projection of the aggregated features along the channel dimension to generate a category prediction result for each spatial location, thereby generating the final change detection map.

[0018] S3. The remote sensing image change detection model is trained using the training data to obtain a trained remote sensing image change detection model. The loss function used in the training process is a composite loss function of cross-entropy loss and Lovász-Softmax loss.

[0019] S4. Use the trained remote sensing image change detection model to perform change detection on the input multi-temporal image pairs.

[0020] Furthermore, in S1, when constructing the training dataset, the acquired image pairs are cropped into 256×256 pixel image patches, and the acquired image pairs are augmented using data augmentation methods.

[0021] Furthermore, in S2, the multi-level feature pair extraction module includes a four-stage extraction process, resulting in four levels of feature maps with spatial resolutions of 64 × 64, 32 × 32, 16 × 16, and 8 × 8, respectively. The first stage preprocesses the input data using an initial backbone block and models the global context information using a Visual State Space (VSS) block. The latter three stages each consist of downsampling and VSS blocks; that is, the latter three stages first downsample the data and then extract features using VSS blocks. The input multi-temporal image pair is defined as a pair of bi-temporal images input at given times T1 and T2. and The multi-level features extracted by the multi-level feature extraction module are as follows: and , This represents a pre-trained VMamba feature extractor.

[0022] Furthermore, in S2, the spatiotemporal interaction enhancement module includes a spatiotemporal interleaving enhancement unit and a channel interaction unit, and the input of the spatiotemporal interleaving enhancement unit is... and Where C, H, and W represent the channel, height, and width dimensions, respectively. Representing the real number field, spatiotemporal interleaving enhancement units are alternately stacked along the width dimension at the same level i to form a new sequence, represented as:

[0023] ,

[0024] in, , and These represent the visual state space block function, the splicing and splitting operations along the width dimension, respectively; It is a new sequence merged along the width dimension. Yes Features obtained through processing visual state space blocks These are two temporal features split along the width dimension;

[0025] The input to the channel interaction unit is Features are placed alternately along the channel dimension, represented as:

[0026] ,

[0027] in, This indicates the features after processing by the channel interaction unit, i.e., the initial change features.

[0028] Furthermore, in S2, the input to the single-scale feature fusion module is... and First, through the dynamic fusion module, respectively... and , and The two feature maps are fused to obtain two single-phase variation features. Then, a dynamic fusion module fuses these two single-phase variation features to obtain a single-level variation feature. This dynamic fusion module employs a multi-scale attention mechanism (EMA) and combines it with an attention-guided adaptive feature fusion strategy to complete the adaptive fusion of the two feature maps. The two feature maps are defined as follows: and The fusion process of the dynamic fusion module is represented as follows:

[0029]

[0030]

[0031]

[0032] in, This indicates that features are cascaded along the channel dimension. and These represent the EMA function and separation along the channel dimension, respectively. , as well as It is a weighted matrix. These are the characteristics after fusion.

[0033] Furthermore, in S2, the processing procedure of the multi-scale feature aggregation module is expressed mathematically as follows:

[0034]

[0035]

[0036]

[0037]

[0038] in, This indicates bilinear interpolation upsampling. This represents a 1 × 1 convolution operation. Represents the channel attention function. It is a weighted matrix. Corresponding to single-level features, This represents the multi-level features after aggregation; the implementation of CA is expressed mathematically as follows:

[0039]

[0040] in, This represents the input feature map, where C is the number of channels, and H and W are the height and width of the feature map, respectively. and These represent performing global average pooling and max pooling operations on the input feature map, respectively, to generate two channel descriptors. and All of their dimensions are ;in and These consist of two shared fully connected layers. W1 reduces the channel dimension from C to C / r, while W2 increases it back to C, with ReLU activation function applied in between. A non-linear mapping is applied to the pooled channel descriptors to extract inter-channel attention features. The two mapped channel attention vectors are then added and normalized using the Sigmoid function to obtain the final channel attention weight map. .

[0041] Furthermore, in S2, the classifier is a 1×1 convolutional classifier used to map the input 128-channel feature map into a 2-channel output.

[0042] Furthermore, the composite loss function described in S3 is expressed as:

[0043]

[0044]

[0045]

[0046] in, It is a composite loss function. and These are the cross-entropy loss function and the Lovász-Softmax loss function, respectively. It is the true label of category c. This is the class probability predicted by the model, where class C=2, corresponding to change / no change. is the subgradient of the Jaccard error for class c, and m(c) is the sorting error vector associated with class c.

[0047] The beneficial effects of this invention are:

[0048] (1) Enhanced robustness of the model to noise and artifacts: By leveraging VMamba’s long-range dependency modeling and the full cross-temporal feature interaction of the spatiotemporal interleaving enhancement module, the model effectively suppresses background interference and reduces the false negative rate.

[0049] (2) Enhanced fine-grained change representation of the model: This invention designs a multi-scale attention reweighting mechanism. This mechanism not only maintains the structural integrity of the change graph by synergistically fusing the local difference preservation ability of the single-scale feature fusion module with the cross-scale semantic consistency of the multi-scale feature aggregation module, but also achieves a more refined representation of the change features, while significantly reducing the false detection rate. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the network structure of the present invention.

[0051] Figure 2 This is a visual comparison diagram of the network and the comparison method of the present invention. Detailed Implementation

[0052] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0053] like Figure 1 As shown, the spatiotemporal multi-scale sensing remote sensing image change detection method of the present invention includes the following steps:

[0054] S1. Obtain multi-temporal image pairs from the known dataset, preprocess them, and then construct the training dataset;

[0055] S2. Construct a remote sensing image change detection model;

[0056] S3. Train the constructed remote sensing image change detection model using the training data;

[0057] S4. Use the trained remote sensing image change detection model to perform change detection on the input multi-temporal image pairs.

[0058] Furthermore, step S1 uses three public datasets, as described below:

[0059] 1. SYSU-CD dataset: A large-scale remote sensing change detection dataset containing 20,000 pairs of 256×256 images, covering a variety of change types and rich spatiotemporal change features, used to evaluate the robustness of algorithms in complex scenarios.

[0060] 2. LEVIR-CD dataset: A high-resolution urban building change detection dataset consisting of 637 pairs of Google Earth images, cropped to generate 256×256 tiles, suitable for long-term building change analysis.

[0061] 3. WHU-CD dataset: Focusing on building change detection, it is constructed from a set of ultra-high resolution aerial images, and 256×256 tiles are generated by cropping. It is suitable for fine recognition tasks of urban structure updates.

[0062] The data preprocessing is as follows:

[0063] For the LEVIR-CD and WHU-CD datasets (excluding the SYSU-CD dataset), multi-temporal image pairs and their corresponding labels are first cropped into 256×256 pixel image patches to meet the size requirements of the network training input. During the training phase, to improve the model's generalization ability, data augmentation methods such as random rotation, horizontal flipping, and vertical flipping are used to perform diverse transformations on the image patches. During the testing phase, the trained network is directly used to perform inference on the original-sized test data, avoiding block boundary effects and ensuring the continuity and accuracy of prediction results, thus achieving an effective balance between training efficiency and detection performance.

[0064] Furthermore, the change detection model proposed in this invention in step S2 includes, in sequence, an input layer, a spatiotemporal interaction enhancement module, a single-scale feature fusion module, a multi-scale feature aggregation module, an output layer, and a loss function, which are specifically implemented as follows;

[0065] The input layer employs a shared-weight dual-branch feature extractor based on a pre-trained VMamba model to extract four levels of feature maps from the input bi-temporal remote sensing image pairs, with spatial resolutions of 64 × 64, 32 × 32, 16 × 16, and 8 × 8, respectively. The feature extractor implements a four-stage progressive learning process. The first stage preprocesses the input data using an initial backbone block and models global contextual information using a VSS block. In the following three stages, the input data is first downsampled and then feature extraction is performed using the VSS block. The bi-temporal image pairs input at given times T1 and T2 are denoted as... and The extracted four-level features are represented as follows: and . This represents a pre-trained VMamba feature extractor. and This represents the extracted multi-level features.

[0066] The spatiotemporal interaction enhancement module is used to promote interaction between bi-temporal features, including a spatiotemporal interleaving enhancement unit and a channel interaction unit. The spatiotemporal interleaving enhancement unit is used to fully capture the spatiotemporal relationship between bi-temporal feature pairs and enhance feature alignment capabilities to reduce spurious change interference. The specific process of the spatiotemporal interleaving enhancement unit is as follows: First, bi-temporal features at the same level are alternately stacked along the width dimension to construct a new sequence. Then, the sequence is processed using a visual state space block (VSS) to capture long-distance spatiotemporal relationships through global modeling. Finally, the resulting sequence is split into bi-temporal branches along the width dimension, effectively promoting interaction between bi-temporal features. The input bi-temporal feature pair is denoted as... and At the same level, units i are alternately stacked along the width dimension to form a new sequence, where C, H, and W represent the channel, height, and width dimensions, respectively. The processing of the spatiotemporal interleaving enhancement unit can be expressed as the following formula:

[0067]

[0068] , and These represent the stitching and splitting operations along the width dimension of the Visual State Space (VSS) blocks. The channel interaction unit operates on the output of the spatiotemporal interleaving enhancement unit to extract preliminary change cues for subsequent refinement of differential features. The specific process of the channel interaction unit is as follows: by alternating the placement of features along the channel dimension, adjacent channels originate from different temporal features, and the channel interaction mechanism promotes local interaction between features of dual-temporal images. The processing of the channel interaction unit can be expressed by the following formula:

[0069]

[0070] This represents the features obtained after processing by the channel interaction unit. Compared with traditional methods such as direct concatenation or difference, this interleaving method can achieve local mixing of temporal features while effectively avoiding redundant information. Furthermore, unlike global fusion strategies, this method does not require additional parameters, and the computational cost is negligible, providing a lightweight yet effective extraction method for discriminative feature representation.

[0071] The single-scale feature fusion module adaptively fuses the bi-temporal features enhanced by the spatiotemporal interleaving enhancement unit with the coarse variation features obtained by the channel interaction unit, generating accurate single-level variation feature maps while preserving local difference details. Specifically, the single-scale feature fusion module uses a dynamic fusion module (DFM) to adaptively fuse the enhanced bi-temporal features with the coarse features. The dynamic fusion module utilizes the difference between the coarse features and the single-temporal image features to dynamically guide the single-temporal features to focus more on the discriminative regions, thereby generating more accurate single-temporal variation features. Then, the obtained fused features are further refined using an additional dynamic fusion module to generate accurate single-level variation features. The dynamic fusion module employs an efficient multi-scale attention mechanism (EMA) and combines it with an attention-guided adaptive feature fusion strategy to complete the adaptive fusion of the two feature maps. The information regions in the two feature maps are dynamically emphasized through the attention weights learned by the EMA. Through its channel reorganization strategy and parallel multi-scale path design, the EMA can enhance the model's ability to capture global and local information while maintaining computational efficiency. Given two feature maps... and The processing of the dynamic fusion module can be expressed by the following formula:

[0072]

[0073]

[0074]

[0075] Features are cascaded along the channel dimension. and These represent the EMA function and the separation along the channel dimension, respectively. , as well as It is a weighted matrix. These are the characteristics after fusion;

[0076] The multi-scale feature aggregation module adaptively aggregates multi-level features to ensure semantic consistency across scales. In this module, an effective channel attention (CA) mechanism and an attention-guided adaptive feature aggregation strategy are designed to integrate multi-level features, effectively addressing information differences between different levels and mitigating semantic degradation typically introduced by top-down progressive integration. The implementation of the multi-scale feature aggregation module consists of three stages: feature alignment, attention-based reweighting, and adaptive feature aggregation. In the feature alignment stage, bilinear interpolation is used to upsample high-level features to match the resolution of low-level features. After 1 × 1 convolution, batch normalization (BN) and ReLU activation are performed, followed by channel dimension compression to generate resolution-aligned features with consistent channels. In the reweighting stage, a CA mechanism is designed to calculate weights for multi-level feature maps, dynamically adjusting the contribution of each layer of features. Finally, in the adaptive feature aggregation stage, the calculated weights are assigned to the corresponding feature maps, and the weighted feature maps are summed element-wise to generate the final aggregated feature map. The implementation of the multi-scale feature aggregation module can be expressed mathematically as follows:

[0077]

[0078]

[0079]

[0080]

[0081] in, This indicates bilinear interpolation upsampling. This represents a 1 × 1 convolution operation. Represents the channel attention function. It is a weighted matrix. Corresponding to single-level features, This represents the multi-level features after aggregation; the implementation of CA is expressed mathematically as follows:

[0082]

[0083] in, This represents the input feature map, where C is the number of channels, and H and W are the height and width of the feature map, respectively. and These represent performing global average pooling and max pooling operations on the input feature map, respectively, to generate two channel descriptors. and All of their dimensions are .in and These consist of two shared fully connected layers (equivalent to 1×1 convolutions). W1 reduces the channel dimension from C to C / r, while W2 increases it back to C, with ReLU activation applied in between. This part performs a non-linear mapping on the pooled channel descriptors to extract inter-channel attention features. The two mapped channel attention vectors are then summed and normalized using a sigmoid function to obtain the final channel attention weight map. .

[0084] The output layer inputs the aggregated feature representation into a 1×1 convolutional classifier, which maps the input 128-channel feature map into a 2-channel output, performs linear projection on the features along the channel dimension, thereby generating the category prediction result for each spatial location, and finally generating the change detection map.

[0085] The loss function described in S3 employs a composite architecture of cross-entropy loss and Lovász-Softmax loss, achieving precise optimization for change detection tasks through weighted fusion. Specifically, firstly, the standard binary cross-entropy loss is calculated based on the model's original output, automatically ignoring invalid pixel regions with a label value of 255 during the calculation. Then, the network output is sigmoid-normalized, and the Lovász-Softmax loss function is introduced to directly optimize the Jaccard index for changed regions. This function enhances sensitivity to broken boundaries and small-scale changed regions by calculating the subgradient of the prediction error vector. Finally, the two losses are linearly combined with a weight ratio of 0.75:1, where the cross-entropy loss ensures pixel-level classification accuracy, while the Lovász loss focuses on improving boundary segmentation quality. The specific formula is as follows:

[0086]

[0087]

[0088]

[0089] and These are the cross-entropy loss function and the Lovász-Softmax loss function, respectively. It is the true label of category c (one-hot encoded). This is the class probability predicted by the model, where class C = 2 (changed / unchanged). is the subgradient of the Jaccard error for class c, and m(c) is the sorting error vector associated with class c.

[0090] The Adam optimizer was used during training, with an initial learning rate of 1×10^(-4), a weight decay coefficient of 5×10^(-4), a batch size of 8, and 240,000 training iterations. During training, validation was performed on a validation set every 500 iterations to select the optimal model parameters.

[0091] The steps in S4 to train the network using the training dataset and obtain the trained detection network performance are as follows:

[0092] First, a VMamba-based feature extractor is used to extract four levels of feature maps from the input bi-temporal remote sensing image pairs, with spatial resolutions of 64 × 64, 32 × 32, 16 × 16, and 8 × 8, respectively.

[0093] Then, at each feature layer, the spatiotemporal interaction enhancement module uses the spatiotemporal interleaving enhancement unit to fully model the spatiotemporal relationship between the two temporal features, thereby achieving effective cross-temporal feature interaction, and uses the channel interaction unit to generate preliminary change cues for subsequent differential feature refinement.

[0094] The single-scale feature fusion module maintains local differences by dynamically reweighting enhanced bi-temporal features from the spatiotemporally interleaved enhancement unit and coarse variation features from the channel interaction unit. Subsequently, the multi-scale feature aggregation module adaptively aggregates multi-level features to generate a high-precision variation representation with cross-level semantic consistency.

[0095] Finally, the aggregated feature representations are passed through a linear projection layer to generate the final change detection map.

[0096] A visual comparison experiment was conducted between this invention and six benchmark remote sensing image change detection methods to verify the superiority of this invention. Two methods based on convolutional neural networks (CNN): FC-EF and SELFNet, two methods based on Transformer: BIT and ChangeFormer, and two methods based on Mamba: ChangeMamba and MF-VMamba were selected as comparison methods. Figure 2 The image presents a visual comparison of six benchmark methods and our proposed method based on three public datasets. White patches represent accurately detected areas of change, black patches represent correctly identified areas of no change, red patches represent false positives (misjudged as changed areas when they are actually unchanged), and green patches represent missed detections (actual changes that were not detected). Figure 2As shown, traditional CNN and Transformer-based methods exhibit significant limitations, including a high number of false positives (red patches) and false negatives (green patches), primarily due to their sensitivity to seasonal variations, lighting conditions, and limited training samples. Recently, Mamba-based methods have demonstrated strong long-range modeling capabilities, helping to reduce false positive and false negative rates; however, they still struggle to capture fine-grained variation information, especially when dealing with small-scale changes and dense prediction tasks.

Claims

1. A method for detecting changes in remotely sensed images at multiple spatiotemporal scales, characterized in that, Includes the following steps: S1. Obtain multi-temporal image pairs from the known dataset to construct the training dataset; S2. Construct a remote sensing image change detection model, which includes a multi-level feature pair extraction module, a spatiotemporal interaction enhancement module, a single-scale feature fusion module, a multi-scale feature aggregation module, and a classifier. The multi-level feature pair extraction module employs a shared-weight dual-branch feature extractor based on a pre-trained VMamba model to extract multi-level features from the input dual-temporal remote sensing image pairs, thus obtaining multi-level feature pairs. The spatiotemporal interaction enhancement module processes multi-level feature pairs into preliminary change features. The processing procedure is as follows: first, multi-level feature pairs are stacked alternately along the width dimension at the same level to construct a new sequence. After processing the new sequence using visual state space blocks, it is then divided into two temporal features along the width dimension. Finally, the two temporal features are placed alternately along the channel dimension to obtain preliminary change features. The single-scale feature fusion module fuses the two temporal features and the preliminary change features obtained from the spatiotemporal interaction enhancement module into a single-level change feature. The processing procedure is as follows: first, the two temporal features are fused with the preliminary change features respectively to obtain two single-temporal change features, and then the two single-temporal change features are fused to obtain a single-level change feature. The multi-scale feature aggregation module processes single-level variation features into aggregated features. The processing steps are as follows: first, bilinear interpolation is used to upsample the high-level features of the single-level variation features to obtain multiple resolution-aligned features with consistent channels. Then, the weights of the multiple aligned features are calculated and adjusted through a channel attention mechanism. Finally, the obtained weights are assigned to the corresponding aligned features. Then, the weighted aligned features are summed element by element to generate the final aggregated features. The classifier performs a linear projection of the aggregated features along the channel dimension to generate a category prediction result for each spatial location, thereby generating the final change detection map. S3. The remote sensing image change detection model is trained using the training data to obtain a trained remote sensing image change detection model. The loss function used in the training process is a composite loss function of cross-entropy loss and Lovász-Softmax loss. S4. Use the trained remote sensing image change detection model to perform change detection on the input multi-temporal image pairs.

2. The method for detecting changes in remote sensing images based on spatiotemporal multi-scale perception according to claim 1, characterized in that, In S1, when constructing the training dataset, the acquired image pairs are cropped into 256×256 pixel image patches, and the acquired image pairs are augmented using data augmentation methods.

3. The method for detecting changes in remote sensing images based on spatiotemporal multi-scale perception according to claim 1, characterized in that, In S2, the multi-level feature extraction module includes a four-stage extraction process, thereby obtaining four levels of feature maps with spatial resolutions of 64 × 64, 32 × 32, 16 × 16, and 8 × 8, respectively. The first stage preprocesses the input data using an initial backbone block and models the global context information using a Visual State Space (VSS) block. The subsequent three stages all consist of downsampling and VSS blocks; that is, the data is first downsampled in each of the last three stages, and then features are extracted using a VSS block. The input multi-temporal image pair is defined as a pair of two-temporal images input at a given time T1 and T2. and The multi-level features extracted by the multi-level feature extraction module are as follows: and , This represents a pre-trained VMamba feature extractor.

4. The method for detecting changes in remote sensing images with spatiotemporal multi-scale perception according to claim 3, characterized in that, In S2, the spatiotemporal interaction enhancement module includes a spatiotemporal interleaving enhancement unit and a channel interaction unit. The input of the spatiotemporal interleaving enhancement unit is... and Where C, H, and W represent the channel, height, and width dimensions, respectively. Representing the real number field, spatiotemporal interleaving enhancement units are alternately stacked along the width dimension at the same level i to form a new sequence, represented as: ， in, , and These represent the visual state space block function, the splicing and splitting operations along the width dimension, respectively; It is a new sequence merged along the width dimension. Yes Features obtained through processing visual state space blocks These are two temporal features split along the width dimension; The input to the channel interaction unit is Features are placed alternately along the channel dimension, represented as: ， in, This indicates the features after processing by the channel interaction unit, i.e., the initial change features.

5. The method for detecting changes in remote sensing images with spatiotemporal multi-scale perception according to claim 4, characterized in that, In S2, the input of the single-scale feature fusion module is and First, through the dynamic fusion module, respectively... and , and The two feature maps are fused to obtain two single-phase variation features. Then, a dynamic fusion module fuses these two single-phase variation features to obtain a single-level variation feature. This dynamic fusion module employs a multi-scale attention mechanism (EMA) and combines it with an attention-guided adaptive feature fusion strategy to complete the adaptive fusion of the two feature maps. The two feature maps are defined as follows: and The fusion process of the dynamic fusion module is represented as follows: ， ， ， in, This indicates that features are cascaded along the channel dimension. and These represent the EMA function and separation along the channel dimension, respectively. , as well as It is a weighted matrix. These are the characteristics after fusion.

6. The method for detecting changes in remote sensing images based on spatiotemporal multi-scale perception according to claim 5, characterized in that, In S2, the processing procedure of the multi-scale feature aggregation module is expressed mathematically as follows: ， ， ， ， in, This indicates bilinear interpolation upsampling. This represents a 1 × 1 convolution operation. Represents the channel attention function. It is a weighted matrix. Corresponding to single-level features, This represents the multi-level features after aggregation; the implementation of CA is expressed mathematically as follows: ， in, This represents the input feature map, where C is the number of channels, and H and W are the height and width of the feature map, respectively. and These represent performing global average pooling and max pooling operations on the input feature map, respectively, to generate two channel descriptors. and All of their dimensions are ;in and These consist of two shared fully connected layers. W1 reduces the channel dimension from C to C / r, while W2 increases it back to C, passing through a ReLU activation function in between. A non-linear mapping is applied to the pooled channel descriptors to extract inter-channel attention features. The two mapped channel attention vectors are then summed and normalized using a sigmoid function to obtain the final channel attention weight map. .

7. The method for detecting changes in remote sensing images with spatiotemporal multi-scale perception according to claim 6, characterized in that, In S2, the classifier is a 1×1 convolutional classifier used to map the input 128-channel feature map into a 2-channel output.

8. The method for detecting changes in remote sensing images based on spatiotemporal multi-scale perception according to claim 7, characterized in that, The composite loss function described in S3 is expressed as follows: ， ， ， in, It is a composite loss function. and These are the cross-entropy loss function and the Lovász-Softmax loss function, respectively. It is the true label of category c. This is the class probability predicted by the model, where class C=2, corresponding to change / no change. is the subgradient of the Jaccard error for class c, and m(c) is the sorting error vector associated with class c.

Images