A method and system for detecting changes in remote sensing images based on asymmetric state-space modeling

Abstract

The present application relates to the technical field of image analysis, in particular to a remote sensing image change detection method and system based on asymmetric state space modeling, the method of the present application firstly performs RGB channel standardization and channel dimension splicing on the double-phase remote sensing images of the same area which have been geometrically registered, adopts a four-level down-sampling encoder and matches a symmetric double residual detail forward propagation path to extract features, which can obtain multi-scale features in layers while avoiding the loss of spatial details caused by down-sampling; then, the deep layer coding features are split, processed by a time sequence interaction module and an asymmetric double-phase Mamba module, and then fused, so as to capture the differences between double-phase features and strengthen the time sequence correlation; finally, the resolution is restored through a four-level up-sampling decoder combined with a lightweight adaptive fusion and a symmetric semantic feedback path, so that the remote sensing image change detection result is accurately obtained, and the accuracy of remote sensing image change analysis and detection is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, specifically to a method and system for detecting changes in remote sensing images based on asymmetric state space modeling. Background Technology

[0002] Remote sensing image change detection aims to automatically identify differences in the surface condition of the same geographical area at different points in time. It is a core technology for applications such as land resource monitoring, urban expansion analysis, and disaster assessment. With the explosive growth of high-resolution remote sensing data, higher demands are placed on the accuracy, efficiency, and generalization ability of change detection algorithms.

[0003] Deep learning-based change detection methods have become mainstream. Visual Transformers and their variants (such as ChangeFormer and SwinSUNet) have been used to model long-range dependencies and have made significant progress. However, these methods have new bottlenecks: most attention mechanisms are still based on symmetrical feature extraction backbones and have failed to fundamentally change the symmetrical nature of bi-temporal affinity computation; the self-attention mechanism of the standard Transformer has quadratic computational complexity, and the memory and computational overhead is huge when processing high-resolution remote sensing images, making it difficult to adapt to large-scale applications; the complex model structure leads to a large number of parameters, high training data requirements, and limited generalization.

[0004] Therefore, there is an urgent need for a new architecture that can: (1) achieve bi-temporal difference modeling at the computational core level; (2) maintain efficient long-range dependency capture capability; (3) achieve adaptive feature fusion and enhancement with low overhead; and (4) build a lightweight, high-performance and end-to-end trainable complete system. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for detecting changes in remote sensing images based on asymmetric state-space modeling.

[0006] The technical solution of this invention is as follows: A remote sensing image change detection method based on asymmetric state-space modeling includes the following operations: S1. Acquire geometrically registered remote sensing images of the same area at different times, and after standardization processing of the three RGB channels, stitch them together according to the channel dimension to obtain a dual-temporal stitched image. S2. A four-level downsampling encoder is used to extract layered features from the two-phase stitched images and output four-level encoded features. Each encoding layer includes a convolutional layer, a batch normalization layer and an activation function. A max pooling layer is set between the upper and lower encoding layers. The encoder has a built-in symmetrical double residual detail forward propagation path, which includes injecting the first-level encoded features into the second-level encoding layer and the fourth-level encoding layer respectively. S3. The fourth-level coding features are split into features to obtain dual-temporal features. After being processed by the shared attention-based temporal interaction module and the asymmetric dual-temporal Mamba module, the features are merged to obtain the bottleneck layer fused features. S4. A four-level upsampling decoder is used to restore the bottleneck layer fusion features step by step with increasing resolution, outputting the first-level decoded features. After classification processing, the remote sensing image change detection results are obtained. Each decoding layer consists of a max unpooling layer, a deconvolution block, and a lightweight adaptive fusion module. Each decoding layer fuses the corresponding layer's encoded features through the lightweight adaptive fusion module. The decoder has a built-in symmetric semantic feedback cross-layer path, which includes injecting the bottleneck layer fusion features into the third-level decoding layer and the first-level decoding layer, respectively.

[0007] In S2, the operation of injecting the first-level coding features into the second-level coding layer is achieved by performing convolution on the first-level coding features, adjusting and learning them with a learnable parameter α, and concatenating them with the first-level coding pooling features to obtain the first-level coding detail injection features, which are then used for processing in the second-level coding layer to obtain the second-level coding features; the first-level coding pooling features are obtained by max pooling the first-level coding features.

[0008] The processing steps in the shared attention-based temporal interaction module in S3 are as follows: the dual-temporal features are concatenated along the channel dimension, processed by convolutional dimensionality reduction and a feedforward network to extract common features; the common features are processed by 1×1 convolution and Sigmaod activation function to generate an attention map; based on the attention map, the dual-temporal features are fused to obtain dual-temporal interaction features.

[0009] The operations of the asymmetric dual-temporal Mamba module in S3 are as follows: the dual-temporal interaction features are processed by serialization, linear projection, convolution, and activation functions, and then processed by asymmetric selective state space to obtain dual-temporal state space features; simultaneously, the dual-temporal interaction features are processed by multi-path parallel local perception to obtain dual-temporal perception features; the dual-temporal perception features and the dual-temporal state space features are fused according to their corresponding temporal order to obtain dual-temporal state space enhancement features; the dual-temporal interaction features are dual-temporal features obtained by processing the temporal interaction module based on shared attention.

[0010] In the asymmetric differential state update function of the asymmetric selective state space, the L1 norm is used for the first time-corresponding feature and the L2 norm is used for the second time-corresponding feature to perform differential state update.

[0011] Before the asymmetric dual-temporal Mamba module processing operation, the dual-temporal interaction features are further subjected to external asymmetric feature enhancement processing to obtain dual-temporal interaction enhancement features, which are used to perform the asymmetric dual-temporal Mamba module processing operation. In the external asymmetric feature enhancement processing operation, the first temporal interaction feature in the dual-temporal interaction features is processed by a multilayer perceptron, and the second temporal interaction feature is enhanced by channel attention recalibration.

[0012] In S4, the operation of injecting the bottleneck layer fusion features into the third-level decoding layer is as follows: the bottleneck layer fusion features are convolved, then adjusted and learned with a learnable parameter β, and concatenated with the fourth decoding depooling features to obtain the third-level decoding detail injection features. These features are then processed by the lightweight adaptive fusion module to obtain the third fusion features. The third fusion features are processed by the deconvolution block to obtain the third-level decoding features. The fourth decoding depooling features are obtained by processing the fourth-level decoding features through the max depooling layer.

[0013] A remote sensing image change detection system based on asymmetric state-space modeling, used to implement the aforementioned remote sensing image change detection method based on asymmetric state-space modeling, includes: The dual-temporal mosaic image generation unit is used to acquire geometrically registered remote sensing images of the same area at different times. After being standardized by the three RGB channels, the images are mosaicked along the channel dimension to obtain the dual-temporal mosaic image. The encoding unit is used to extract hierarchical features from the two-phase stitched images using a four-level downsampling encoder and output four-level encoded features. Each encoding layer includes a convolutional layer, a batch normalization layer and an activation function, and a max pooling layer is provided between the upper and lower encoding layers. The encoder has a built-in symmetrical double residual detail forward propagation path, which includes injecting the first-level encoded features into the second-level encoding layer and the fourth-level encoding layer respectively. The bottleneck unit is used to split the fourth-level coding features to obtain dual-temporal features. After being processed by the shared attention-based temporal interaction module and the asymmetric dual-temporal Mamba module, the features are then merged to obtain the bottleneck layer fused features. The decoding unit employs a four-level upsampling decoder to progressively restore the bottleneck layer fusion features at each level, outputting the first-level decoded features. After classification processing, the remote sensing image change detection results are obtained. Each decoding layer consists of a max-unpooling layer, a deconvolution block, and a lightweight adaptive fusion module. Each decoding layer fuses the corresponding layer's encoded features through the lightweight adaptive fusion module. The decoder has a built-in symmetric semantic feedback cross-layer path, which includes injecting the bottleneck layer fusion features into the third-level decoding layer and the first-level decoding layer, respectively.

[0014] A remote sensing image change detection device based on asymmetric state space modeling includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the aforementioned remote sensing image change detection method based on asymmetric state space modeling.

[0015] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned remote sensing image change detection method based on asymmetric state-space modeling.

[0016] The beneficial effects of this invention are as follows: This invention provides a remote sensing image change detection method based on asymmetric state-space modeling. First, RGB channel standardization and channel dimension stitching are performed on geometrically registered dual-temporal remote sensing images of the same region. This unifies the image pixel scale and completely preserves the original information of both temporal phases, providing a standardized and complete input for change detection. Then, a four-level downsampling encoder combined with a symmetric dual-residual detail forward propagation path is used to extract features. This allows for the acquisition of multi-scale features layer by layer while avoiding the loss of spatial details caused by downsampling, balancing deep semantics and shallow detail information. Next, the deep encoded features are split, processed by a temporal interaction module and an asymmetric dual-temporal Mamba module, and then fused. This accurately captures the differences in features between the two temporal phases, strengthens temporal correlation, and achieves refined perception of change information. Finally, a resolution is restored by a four-level upsampling decoder combined with lightweight adaptive fusion and a symmetric semantic feedback path. This dynamically adapts the feature fusion ratio and ensures the consistency of the output semantics, ultimately obtaining accurate remote sensing image change detection results and effectively improving the accuracy and completeness of the detection. This invention provides a remote sensing image change detection method based on asymmetric state-space modeling. The method includes acquiring and preprocessing dual-temporal remote sensing image data; performing state-space recursive calculations on dual-temporal features using an asymmetric dual-temporal Mamba module; applying the L1 norm to the features corresponding to the first temporal phase and the L2 norm to the features corresponding to the second temporal phase for differentiated calculation and state updates; and relying on a globally symmetric dual residual information loop architecture, employing four cross-layer paths of detail forward propagation and semantic back-feedback to achieve efficient long-range dependency capture capabilities and realize bidirectional flow of detail and semantic information during encoding and decoding. Simultaneously, a lightweight shared attention-based temporal interaction module, a static multi-path parallel local perception module, and an adaptive fusion module are combined to perform temporal interaction, local information extraction, and adaptive fusion of features, obtaining remote sensing image change detection results based on the fused features. This invention provides a remote sensing image change detection method based on asymmetric state space modeling. By deeply integrating asymmetric differential modeling with the state space model, it achieves refined perception of change information. At the same time, by optimizing feature expression and fusion efficiency through multi-module collaborative optimization, it improves the accuracy and practicality of remote sensing image change detection. This invention provides a remote sensing image change detection method based on asymmetric state space modeling. The core asymmetric dual-temporal Mamba module has linear complexity, while all other auxiliary modules adopt a minimalist design with a very small number of parameters. The entire model is very lightweight while maintaining high performance, making it easy to train and deploy. It is suitable for end-to-end training and applicable to the rapid processing of large-scale remote sensing images. Attached Figure Description

[0017] The solutions and advantages of this application will become clear to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] In the attached diagram: Figure 1 This is a flowchart illustrating the method of this embodiment. Figure 2 As shown in the embodiments, the remote sensing image change detection effect of the method in this embodiment under different scenarios is illustrated. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the exemplary embodiments of this application clearer, the technical solutions in the exemplary embodiments of this application are described clearly and completely below. Obviously, the described exemplary embodiments are only some embodiments of this application, and not all embodiments.

[0020] This embodiment provides a remote sensing image change detection method based on asymmetric state-space modeling. (See also...) Figure 1 This includes the following operations: S1. Acquire geometrically registered remote sensing images of the same area at different times, and after standardization processing of the three RGB channels, stitch them together according to the channel dimension to obtain a dual-temporal stitched image. S2. A four-level downsampling encoder is used to extract layered features from the two-phase stitched images and output four-level encoded features. Each encoding layer includes a convolutional layer, a batch normalization layer and an activation function. A max pooling layer is set between the upper and lower encoding layers. The encoder has a built-in symmetrical double residual detail forward propagation path, which includes injecting the first-level encoded features into the second-level encoding layer and the fourth-level encoding layer respectively. S3. The fourth-level coding features are split into features to obtain dual-temporal features. After being processed by the shared attention-based temporal interaction module and the asymmetric dual-temporal Mamba module, the features are merged to obtain the bottleneck layer fused features. S4. A four-level upsampling decoder is used to restore the bottleneck layer fusion features step by step with increasing resolution, outputting the first-level decoded features. After classification processing, the remote sensing image change detection results are obtained. Each decoding layer consists of a max unpooling layer, a deconvolution block, and a lightweight adaptive fusion module. Each decoding layer fuses the corresponding layer's encoded features through the lightweight adaptive fusion module. The decoder has a built-in symmetric semantic feedback cross-layer path, which includes injecting the bottleneck layer fusion features into the third-level decoding layer and the first-level decoding layer, respectively. The specific steps are detailed below.

[0021] S1. Acquire geometrically registered remote sensing images of the same area at different times, and after standardization processing of the three RGB channels, stitch them together according to the channel dimension to obtain a dual-temporal stitched image.

[0022] The normalization process for the three RGB channels can be achieved using the following formula: , in, For the first c Standardized values ​​for each channel, For the original remote sensing image c The pixel values ​​of each channel. For the first c The average pixel value of each channel. For the first c The pixel variance values ​​of each channel, the pixel mean of the three RGB channels μ=[0.485,0.456,0.406], and the pixel variance values ​​of the three RGB channels σ=[0.229,0.224,0.225] are the pixel mean and standard deviation of the ImageNet dataset.

[0023] S2. A four-level downsampling encoder is used to extract layered features from the two-phase stitched images and output four-level coded features. Each coding layer includes a convolutional layer, a batch normalization layer and an activation function. A max pooling layer is set between the upper and lower coding layers. The encoder has a built-in symmetrical double residual detail forward propagation path, which includes injecting the first-level coded features into the second-level coding layer and the fourth-level coding layer respectively.

[0024] The four-level downsampling encoder consists of four coding layers from top to bottom: a first-level coding layer, a second-level coding layer, a third-level coding layer, and a fourth-level coding layer. Max-pooling layers are used between coding layers to progressively abstract features and reduce spatial size. Each coding layer includes two convolutional layers, a batch normalization layer, and an activation function for feature extraction. Additionally, the encoder incorporates a symmetric dual-residual detail forward propagation path. This structure consists of two parallel detail forward propagation paths, which inject the first-level encoded features into the second and fourth-level coding layers respectively. This efficiently transfers the highest-resolution detail information from the shallowest layer (i.e., the first-level encoded features) to the middle and deeper layers, mitigating spatial information loss caused by downsampling. The specific operation of the four-level downsampling encoder is as follows.

[0025] First encoding stage: The two-phase stitched images are processed by the first-level encoding layer (two convolutional layers, a batch normalization layer, and an activation function) to obtain the first-level encoded features. The calculation formula is as follows: , Where X represents a two-phase mosaic image. These are intermediate features of the first-level encoding. For the first-level encoding features, Conv is convolution, BN is layer normalization, and ReLU is the ReLU activation function.

[0026] The second encoding stage: The first-level encoded features are processed by max pooling to obtain the first-level encoded pooled features. In the second encoding stage, the first detail forward propagation residual connection is introduced, that is, the first-level encoded features are injected into the second-level encoding layer. This can be achieved by processing the first-level encoded features through convolution, adjusting the learning parameter α, and concatenating it with the first-level encoded pooled features to obtain the first-level encoded detail-injected features. These features are then processed by the second-level encoding layer (two convolutional layers, a batch normalization layer, and an activation function) to obtain the second-level encoded features. The calculation formula is as follows: , in, This is the first-level encoded pooling feature. For the first pooled index, For the 1×1 convolutional adapter layer, the number of channels was adjusted from 16 to 32. Let α be a learnable parameter, and let be the injection strength parameter that can be learned in the second-level encoding. Inject features into the first-level coding details. E1 is a second-level coding intermediate feature, and E2 is a second-level coding feature.

[0027] The third encoding stage: The second-level encoded features are processed by max pooling and the third-level encoding layer (two convolutional layers, batch normalization layer and activation function) to obtain the third-level encoded pooled features.

[0028] Fourth encoding stage: The third-level encoded features are processed by max pooling to obtain the third-level encoded pooled features. In the fourth encoding stage, a second detail forward propagation residual connection is introduced to inject the first-level encoded features into the fourth-level encoding layer. This achieves efficient transfer of high-resolution detail information from the shallowest layer to the deeper features, mitigating the spatial information loss caused by multiple downsampling. This can be achieved by processing the first-level encoded features through convolution, adjusting the learning parameter β, and concatenating it with the third-level encoded pooled features to obtain the third-level encoded detail-injected features. These features are then processed by the fourth-level encoding layer (two convolutional layers, a batch normalization layer, and an activation function) to obtain the fourth-level encoded features.

[0029] S3. The fourth-level coding features are split into features to obtain dual-temporal features. After being processed by the time-series interaction module based on shared attention and the asymmetric dual-temporal Mamba module, the features are merged to obtain the bottleneck layer fused features.

[0030] First, the fourth-level coding features are split into two phase features, including the first phase feature and the second phase feature. .

[0031] Then, the dual-temporal features are processed by a shared attention-based temporal interaction module to guide the network to focus on potential regions of common interest, thereby enhancing the accuracy of subsequent change detection, resulting in dual-temporal interaction features, including a first temporal interaction feature and a second temporal interaction feature.

[0032] The processing steps in the time-series interaction module based on shared attention are as follows: the dual-temporal features are concatenated along the channel dimension, and then processed by convolutional dimensionality reduction and a feedforward network to extract common features; the common features are processed by 1×1 convolution and Sigmaod activation function to generate an attention map; based on the attention map, the dual-temporal features are fused to achieve feature enhancement and obtain dual-temporal interaction features.

[0033] The feature fusion operation based on attention maps is implemented using the following formula: , U k This represents the k-th temporal feature in a dual-temporal feature set, where DWConv is a depthwise separable convolution. The k-th temporal separable convolutional feature, A is the attention map, γ is the learnable enhancement coefficient, and V... k Let be the k-th temporal interaction feature in the dual temporal interaction features, and ⊙ represent element-wise multiplication.

[0034] Next, the dual-temporal interaction features, including the first temporal interaction features and the second temporal interaction features, are processed by the asymmetric dual-temporal Mamba module to accurately capture the changing signals and obtain dual-temporal state space enhancement features.

[0035] The asymmetric dual-temporal Mamba module operates as follows: The dual-temporal interaction features are processed through serialization, linear projection, convolution, and activation functions (preferably SiLU activation function). Then, they undergo asymmetric selective state space processing to obtain dual-temporal state space features, including first and second temporal state space features. Simultaneously, to address Mamba's potential shortcomings in local spatial relationship modeling, the dual-temporal interaction features are processed through multi-path parallel local perception to obtain dual-temporal perception features. Finally, the dual-temporal perception features and the dual-temporal state space features are fused according to their corresponding temporal order to obtain dual-temporal state space enhanced features. The feature fusion process is achieved by performing layer normalization and flattening on the dual-temporal perception features, followed by a weighted summation of the dual-temporal state space features according to their corresponding temporal order.

[0036] The above serialization and linear projection are achieved through the following formula: , S k V represents the flattened feature of the k-th time series. k For the k-th time series normalized feature, U k The projection feature of the k-th time series is used for state-space computation, R. k Let be the residual path feature of the k-th time series, Flatten is the flattening operation, LayerNorm is the layer normalization operation, Linear is the linear processing operation, and Split is the splitting operation.

[0037] The operation of asymmetric selective state-space processing is implemented through the following formula: , , These are the state vectors at times t and t-1, respectively. , Let be the first phase input vectors at times t and t-1, respectively. These are the projection vectors of the first and second time series sequences, respectively, obtained by processing the projection features of the first and second time series sequences through convolution and activation functions, respectively. Let be the second phase input vector at time t. , Let be the first state-space characteristics and the second state-space characteristics at time t. for, These are the output projection matrices for the first time phase and the output projection matrices for the second time phase, respectively. This is the discretized state transition matrix. , These are the first asymmetric differential state update function and the second asymmetric differential state update function defined in this embodiment. In the asymmetric differential state update function in asymmetric selective state space processing, the L1 norm is used for the first time-corresponding feature and the L2 norm is used for the second time-corresponding feature to perform differential state updates. The L1 norm tends to produce a strong response to sparse, drastic pixel-level changes, while the L2 norm is more sensitive to gentle, regional overall changes. The combination of the two allows the model to comprehensively perceive changes from complementary mathematical properties.

[0038] The formula for calculating the asymmetric differential state update function is as follows: , and These are the L1 norm and L2 norm, respectively, and η = 14.0 is the amplification factor. .

[0039] The operation of multi-path parallel local sensing processing is implemented through the following formula: , V k The dual-temporal interaction features include the k-th temporal interaction feature. BN is used for normalization, and GELU is the GELU activation function. , Let w1 and w2 be the first and second nonlinear features of the k-th time series, respectively, and let L be the first and second weights. k This is the k-th time-series sensing feature in the dual-time-series sensing features.

[0040] Furthermore, prior to the asymmetric dual-temporal Mamba module processing, an external asymmetric feature enhancement process is performed on the dual-temporal interaction features to obtain dual-temporal interaction enhancement features, which are then used to perform the asymmetric dual-temporal Mamba module processing. In the external asymmetric feature enhancement process, the first temporal interaction feature in the dual-temporal interaction features undergoes multilayer perceptron processing to preserve details, while the second temporal interaction feature is enhanced using channel attention recalibration.

[0041] The channel attention recalibration enhancement operation is implemented using the following formula: , V2 represents the second temporal interaction feature, GAP represents global max pooling, g represents the second temporal interaction max pooling feature, MLP represents multilayer perceptron processing, σ represents the sigmoid function, Expand represents channel expansion, and w represents the attention value. This enhances the features of the second temporal interaction.

[0042] Finally, the first and second phase state space enhancement features from the dual-phase state space enhancement features are merged. This can be achieved through linear processing, reshaping (using the Reshape function), and concatenation to obtain the bottleneck layer fusion features. The calculation formula is as follows: , O k To enhance the linear characteristics of the state space in the k-th phase, Y k This refers to the k-th phase state space enhancement feature in the dual-phase state space enhancement feature set. For O k transpose, This represents the state space reshaping feature of the k-th phase. As a bottleneck layer fusion feature, These are the state space reshaping features of the first time phase and the state space reshaping features of the second time phase, respectively.

[0043] S4. A four-level upsampling decoder is used to restore the bottleneck layer fusion features step by step with increasing resolution, outputting the first-level decoded features. After classification processing, the remote sensing image change detection results are obtained. Each decoding layer consists of a max unpooling layer, a deconvolution block, and a lightweight adaptive fusion module. Each decoding layer fuses the corresponding layer's encoded features through the lightweight adaptive fusion module. The decoder has a built-in symmetric semantic feedback cross-layer path, which includes injecting the bottleneck layer fusion features into the third-level decoding layer and the first-level decoding layer, respectively.

[0044] The four-level upsampling decoder consists of the first-level decoding layer, the second-level decoding layer, the third-level decoding layer, and the fourth-level decoding layer from top to bottom. The operation of the four-level downsampling encoder is as follows.

[0045] Fourth-level decoding layer: The bottleneck layer fusion features are processed by the max unpooling layer to obtain the fourth unpooling features; the fourth unpooling features and the fourth-level encoding features are processed by the lightweight adaptive fusion module to obtain the fourth fused features; the fourth fused features are processed by the deconvolution block to obtain the fourth-level decoding features.

[0046] The lightweight adaptive fusion module processes operations using the following formula: , F4 is the fourth fusion feature, w A w Bw3 and w4 represent the encoding transformation weight, depooling transformation weight, and non-linear weight, respectively. a4 * indicates channel adjustment operation, E4 and H4 are the fourth-level encoding feature and the fourth unpooling feature, respectively, ReLU is the ReLU activation function, and Contact is concatenation.

[0047] The third-level decoding layer: The fourth-level decoding features are processed by the max unpooling layer to obtain the fourth decoding unpooling features; the bottleneck layer fusion features are injected into the third-level decoding layer, that is, the bottleneck layer fusion features are convolved and then adjusted with a learnable parameter β, and then concatenated with the fourth decoding unpooling features to obtain the third-level decoding detail injection features; the third-level decoding detail injection features and the third-level encoding features are processed by a lightweight adaptive fusion module to obtain the third fusion features; the third fusion features are processed by a deconvolution block to obtain the third-level decoding features.

[0048] Second-level decoding layer: The third-level decoding features are processed by the max unpooling layer to obtain the third decoding unpooling features, which are then processed by the second-level encoding features through a lightweight adaptive fusion module to obtain the second fused features; the second fused features are then processed by a deconvolution block to obtain the second-level decoding features.

[0049] First-level decoding layer: The second-level decoding features are processed by a max unpooling layer to obtain the second decoding unpooling features; the bottleneck layer fusion features are injected into the first-level decoding layer, that is, the bottleneck layer fusion features are convolved and then adjusted with a learnable parameter β, and then concatenated with the second decoding unpooling features to obtain the first-level decoding detail injection features; the first-level decoding detail injection features and the first-level encoding features are processed by a lightweight adaptive fusion module to obtain the first fusion features; the first fusion features are processed by a deconvolution block to obtain the first-level decoding features.

[0050] Finally, the first-level decoded features, after classification processing, can be obtained through convolution and softmax functions to produce binarized remote sensing image change detection results. See the results below. Figure 2 The white areas in the image represent the detected areas of change.

[0051] In practice, the model formed by the aforementioned four-level downsampling encoder, bottleneck layer, four-level upsampling decoder, and classification layer can be used as a remote sensing image change detection model. This model is trained using a training set consisting of multiple bi-temporal stitched images and corresponding real-world binary change labels. The remote sensing image to be processed is then processed by this model to obtain the change detection result. The training loss function is constructed based on cross-entropy loss and Lovász-Softmax loss.

[0052] To verify the detection effect of this embodiment, the following experiment was conducted.

[0053] The AdamW optimizer was used in the experiment, with an initial learning rate of Weight decay The learning rate was scheduled using cosine annealing, and the results were evaluated on the LEVIR-CD+ test set. The method of this embodiment was compared with the current state-of-the-art baseline methods. The experimental results are shown in Table 1.

[0054] Table 1. Summary of experimental performance comparison of different methods

[0055] As shown in Table 1, the method in this embodiment achieves the best results in F1 score, Precision, Recall and IoU, demonstrating its comprehensive advantages in accuracy and efficiency.

[0056] This embodiment provides a remote sensing image change detection system based on asymmetric state-space modeling, used to implement the aforementioned remote sensing image change detection method based on asymmetric state-space modeling, including: The dual-temporal mosaic image generation unit is used to acquire geometrically registered remote sensing images of the same area at different times. After being standardized by the three RGB channels, the images are mosaicked along the channel dimension to obtain the dual-temporal mosaic image. The encoding unit is used to extract hierarchical features from the two-phase stitched images using a four-level downsampling encoder and output four-level encoded features. Each encoding layer includes a convolutional layer, a batch normalization layer and an activation function, and a max pooling layer is provided between the upper and lower encoding layers. The encoder has a built-in symmetrical double residual detail forward propagation path, which includes injecting the first-level encoded features into the second-level encoding layer and the fourth-level encoding layer respectively. The bottleneck unit is used to split the fourth-level coding features to obtain dual-temporal features. After being processed by the shared attention-based temporal interaction module and the asymmetric dual-temporal Mamba module, the features are then merged to obtain the bottleneck layer fused features. The decoding unit employs a four-level upsampling decoder to progressively restore the bottleneck layer fusion features at each level, outputting the first-level decoded features. After classification processing, the remote sensing image change detection results are obtained. Each decoding layer consists of a max-unpooling layer, a deconvolution block, and a lightweight adaptive fusion module. Each decoding layer fuses the corresponding layer's encoded features through the lightweight adaptive fusion module. The decoder has a built-in symmetric semantic feedback cross-layer path, which includes injecting the bottleneck layer fusion features into the third-level decoding layer and the first-level decoding layer, respectively.

[0057] This embodiment also provides a remote sensing image change detection device based on asymmetric state space modeling, including a processor and a memory, wherein the processor executes the computer program stored in the memory to implement the above-mentioned remote sensing image change detection method based on asymmetric state space modeling.

[0058] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described remote sensing image change detection method based on asymmetric state-space modeling.

[0059] This embodiment provides a remote sensing image change detection method based on asymmetric state-space modeling. First, RGB channels of geometrically registered dual-temporal remote sensing images of the same area are standardized and channel dimensions are stitched together. This unifies the image pixel scale and completely preserves the original information of both temporal phases, providing a standardized and complete input for change detection. Then, a four-level downsampling encoder is used in conjunction with a symmetric dual-residual detail forward propagation path to extract features. This allows for the acquisition of multi-scale features at different levels while avoiding the loss of spatial details caused by downsampling, taking into account both deep semantics and shallow detail information. Next, the deep encoded features are split, processed by a temporal interaction module and an asymmetric dual-temporal Mamba module, and then fused. This accurately captures the differences in features between the two temporal phases, strengthens temporal correlation, and achieves refined perception of change information. Finally, a resolution is restored by a four-level upsampling decoder combined with lightweight adaptive fusion and a symmetric semantic feedback path. This dynamically adapts the feature fusion ratio and ensures the consistency of the output semantics, ultimately obtaining accurate remote sensing image change detection results and effectively improving the accuracy and completeness of detection.

[0060] This embodiment provides a remote sensing image change detection method based on asymmetric state-space modeling. The method includes acquiring and preprocessing dual-temporal remote sensing image data; performing state-space recursive calculations on dual-temporal features using an asymmetric dual-temporal Mamba module; applying the L1 norm to the features corresponding to the first temporal phase and the L2 norm to the features corresponding to the second temporal phase for differentiated calculation and state updates; and relying on a globally symmetric dual residual information loop architecture, employing four cross-layer paths of detail forward propagation and semantic back-feedback to achieve efficient long-range dependency capture capabilities and realize bidirectional flow of detail and semantic information during encoding and decoding. Simultaneously, a lightweight shared attention-based temporal interaction module, a static multi-path parallel local perception module, and an adaptive fusion module are combined to perform temporal interaction, local information extraction, and adaptive fusion of features, obtaining the remote sensing image change detection result based on the fused features.

[0061] This embodiment provides a remote sensing image change detection method based on asymmetric state space modeling. By deeply integrating asymmetric differential modeling with the state space model, it achieves refined perception of change information. At the same time, by optimizing feature expression and fusion efficiency through multi-module collaborative optimization, it improves the accuracy and practicality of remote sensing image change detection.

[0062] This embodiment provides a remote sensing image change detection method based on asymmetric state space modeling. The core asymmetric dual-temporal Mamba module has linear complexity, while all other auxiliary modules adopt a minimalist design with a very small number of parameters. The entire model is very lightweight while maintaining high performance, making it easy to train and deploy. It is suitable for end-to-end training and applicable to the rapid processing of large-scale remote sensing images.

[0063] While exemplary embodiments of the invention have been described herein, many other variations or modifications conforming to the principles of the invention can be directly determined or derived from the disclosure of this invention without departing from its spirit and scope. Therefore, the scope of the invention should be understood and recognized to cover all such other variations or modifications.

Claims

1. A method for detecting changes in remote sensing images based on asymmetric state-space modeling, characterized in that, This includes the following operations: S1. Acquire geometrically registered remote sensing images of the same area at different times, and after standardization processing of the three RGB channels, stitch them together according to the channel dimension to obtain a dual-temporal stitched image. S2. A four-level downsampling encoder is used to extract layered features from the two-phase stitched images and output four-level encoded features. Each encoding layer includes a convolutional layer, a batch normalization layer and an activation function. A max pooling layer is set between the upper and lower encoding layers. The encoder has a built-in symmetrical double residual detail forward propagation path, which includes injecting the first-level encoded features into the second-level encoding layer and the fourth-level encoding layer respectively. S3. The fourth-level coding features are split into features to obtain dual-temporal features. After being processed by the shared attention-based temporal interaction module and the asymmetric dual-temporal Mamba module, the features are merged to obtain the bottleneck layer fused features. S4. A four-level upsampling decoder is used to restore the bottleneck layer fusion features step by step with increasing resolution, outputting the first-level decoded features. After classification processing, the remote sensing image change detection results are obtained. Each decoding layer consists of a max unpooling layer, a deconvolution block, and a lightweight adaptive fusion module. Each decoding layer fuses the corresponding layer's encoded features through the lightweight adaptive fusion module. The decoder has a built-in symmetric semantic feedback cross-layer path, which includes injecting the bottleneck layer fusion features into the third-level decoding layer and the first-level decoding layer, respectively.

2. The remote sensing image change detection method based on asymmetric state-space modeling according to claim 1, characterized in that, In S2, the operation of injecting the first-level coding features into the second-level coding layer is achieved by performing convolution on the first-level coding features, adjusting and learning them with a learnable parameter α, and concatenating them with the first-level coding pooling features to obtain the first-level coding detail injection features, which are then used for processing in the second-level coding layer to obtain the second-level coding features; the first-level coding pooling features are obtained by max pooling the first-level coding features.

3. The remote sensing image change detection method based on asymmetric state-space modeling according to claim 1, characterized in that, The processing steps in the shared attention-based temporal interaction module in S3 are as follows: the dual-temporal features are concatenated along the channel dimension, processed by convolutional dimensionality reduction and a feedforward network to extract common features; the common features are processed by 1×1 convolution and Sigmaod activation function to generate an attention map; based on the attention map, the dual-temporal features are fused to obtain dual-temporal interaction features.

4. The remote sensing image change detection method based on asymmetric state-space modeling according to claim 1, characterized in that, The operation of the asymmetric dual-temporal Mamba module in S3 is as follows: the dual-temporal interaction features are processed by serialization, linear projection, convolution and activation function, and then processed by asymmetric selective state space to obtain dual-temporal state space features; at the same time, the dual-temporal interaction features are processed by multi-path parallel local perception to obtain dual-temporal perception features. By performing corresponding temporal feature fusion processing on dual-temporal sensing features and dual-temporal state space features, dual-temporal state space enhancement features are obtained. The dual-temporal interaction features are dual-temporal features, obtained by processing the temporal interaction module based on shared attention.

5. The remote sensing image change detection method based on asymmetric state-space modeling according to claim 4, characterized in that, In the asymmetric differential state update function of the asymmetric selective state space, the L1 norm is used for the first time-corresponding feature and the L2 norm is used for the second time-corresponding feature to perform differential state update.

6. The remote sensing image change detection method based on asymmetric state-space modeling according to claim 1, characterized in that, Before the operations of the asymmetric dual-temporal Mamba module, there is also an enhancement process for the external asymmetric features of the dual-temporal interaction features to obtain dual-temporal interaction enhancement features, which are used to perform the operations of the asymmetric dual-temporal Mamba module. In the external asymmetric feature enhancement process, the first temporal interaction feature in the dual temporal interaction features is processed by a multilayer perceptron, and the second temporal interaction feature is enhanced by channel attention recalibration.

7. The remote sensing image change detection method based on asymmetric state-space modeling according to claim 1, characterized in that, In S4, the operation of injecting the bottleneck layer fusion features into the third-level decoding layer is as follows: after convolution processing of the bottleneck layer fusion features, they are adjusted and learned with the learnable parameter β, and then concatenated with the fourth decoding depooling features to obtain the third-level decoding detail injection features, which are used with the third-level encoding features and processed by the lightweight adaptive fusion module to obtain the third fusion features. The third fusion feature is processed by a deconvolution block to obtain the third-level decoding feature; The fourth decoding depooling feature is obtained by processing the fourth-level decoding feature through the maximum depooling layer.

8. A remote sensing image change detection system based on asymmetric state-space modeling, used to implement the remote sensing image change detection method based on asymmetric state-space modeling as described in claim 1, characterized in that, include: The dual-temporal mosaic image generation unit is used to acquire geometrically registered remote sensing images of the same area at different times. After being standardized by the three RGB channels, the images are mosaicked along the channel dimension to obtain the dual-temporal mosaic image. The encoding unit is used to extract hierarchical features from the two-phase stitched images using a four-level downsampling encoder and output four-level encoded features. Each encoding layer includes a convolutional layer, a batch normalization layer and an activation function, and a max pooling layer is provided between the upper and lower encoding layers. The encoder has a built-in symmetrical double residual detail forward propagation path, which includes injecting the first-level encoded features into the second-level encoding layer and the fourth-level encoding layer respectively. The bottleneck unit is used to split the fourth-level coding features to obtain dual-temporal features. After being processed by the shared attention-based temporal interaction module and the asymmetric dual-temporal Mamba module, the features are then merged to obtain the bottleneck layer fused features. The decoding unit is used to perform progressive resolution recovery of the bottleneck layer fusion features using a four-level upsampling decoder, output the first-level decoded features, and obtain the remote sensing image change detection results after classification processing. Each decoding layer consists of a max unpooling layer, a deconvolution block, and a lightweight adaptive fusion module. Each decoding layer fuses the corresponding layer's encoded features through the lightweight adaptive fusion module. The decoder has a built-in symmetric semantic feedback cross-layer path, which includes injecting the bottleneck layer fusion features into the third-level decoding layer and the first-level decoding layer, respectively.

9. A remote sensing image change detection device based on asymmetric state-space modeling, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the remote sensing image change detection method based on asymmetric state space modeling as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the remote sensing image change detection method based on asymmetric state space modeling as described in any one of claims 1-7.

Images