Multi-level visual state space remote sensing semantic change detection system and method based on dynamic expansion mechanism

Abstract

The invention relates to the technical field of computer vision and artificial intelligence, in particular to a multi-level visual state space remote sensing semantic change detection system and method based on a dynamic expansion mechanism, and the method comprises the steps: extracting a multi-stage feature map based on an SSM coding backbone network according to front and back time phase remote sensing images; according to the multi-stage feature map, an AWVSS module is adopted for decoding to generate a binary transformation feature map and a semantic transformation feature map; multiplying the binary transformation feature map serving as a mask by the semantic transformation feature map to generate a mask semantic transformation feature map; and calculating loss according to the mask semantic transformation feature map and the real label semantic transformation segmentation map, and obtaining remote sensing semantic change information. Compared with the existing CNN and Transform methods, the method provided by the invention has significant advantages in the aspects of detection precision, calculation efficiency, multi-scale adaptability and the like.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and artificial intelligence technology, and in particular to a multi-level visual state space remote sensing semantic change detection system and method based on a dynamic expansion mechanism. Background Art

[0002] Remote sensing change detection (RSCD) is the process of identifying changes in the Earth's surface over time using satellite or aerial imagery. This process is crucial for monitoring environmental changes, urban development, farmland changes, and other fields [1]. The advancement of remote sensing technology has enabled the acquisition of a large number of remote sensing images with different resolutions and modes, increasing the demand for automatic land change detection in dual-temporal imagery.

[0003] Based on the semantic label information required in the output change map, change detection tasks can be divided into binary change detection (BCD) and semantic change detection (SCD). BCD aims to distinguish between changed and unchanged pixels in bitemporal imagery by applying binary labels. In contrast, the SCD task is inherently more complex than traditional BCD because it not only identifies the areas of change but also classifies the types of change, providing detailed "from-to" information at the pixel level. Therefore, SCD provides a detailed and information-rich approach for monitoring land cover change in the context of remote sensing.

[0004] To effectively capture complex patterns in high-resolution optical remote sensing imagery, convolutional neural networks (CNNs) and visual transformers (ViTs) are two major architectures commonly used for back-convolutional coherence and sequential coherence (BCD) approaches. Compared to CNNs, ViTs generally exhibit superior modeling capabilities, primarily due to the integration of a self-attention mechanism that enables dynamic prediction of global receptive fields and weights. Their computational complexity increases with the square of the input size, resulting in significant overhead in downstream tasks involving large spatial resolutions. To address this challenge, significant efforts have been made to improve the efficiency of attention mechanisms, primarily by limiting the size of the computational window, or stride. While effective, these approaches inherently balance the trade-off between receptive field and computational efficiency, limiting their ability to capture long-range dependencies in visual data.

[0005] Recently, structured state-space sequence models have emerged as a promising class of architectures for sequence modeling. They introduce a new parameterization for state-space models from a continuous-time, recurrent, and convolutional perspective, enabling efficient and principled modeling of long-range dependencies. They propose to address the S4 challenge by introducing weight modulation during the propagation process, significantly expanding the effective receptive field. State-space models have proven to be very effective in modeling long sequences with linear complexity, achieving promising results in computer vision tasks and medical image segmentation. However, despite their advantages, these scanning strategies introduce spatial discontinuities between adjacent pixels, which hinders the efficient flow of information within the sequence. Furthermore, these models have limitations when applied to remote sensing imagery, highlighting the need to explore their potential in various change detection subtasks.

[0006] Convolutional neural networks and visual transformers are two major architectures commonly used for semantic change detection in remote sensing. Compared to convolutional neural networks, ViTs generally exhibit superior modeling capabilities, primarily due to the integration of a self-attention mechanism, which enables dynamic prediction of global receptive fields and weights. However, self-attention introduces a significant side effect: its computational complexity increases quadratically with the input size, resulting in significant overhead in downstream tasks involving large spatial resolutions. To address this challenge, limiting the size of the computational window, or stride, is often employed to improve the efficiency of the attention mechanism. While effective, these methods inherently balance the trade-off between receptive field and computational efficiency, limiting their ability to capture long-range dependencies in visual data. Summary of the Invention

[0007] The technical problem solved by the present invention is that the local correlation constraints of the convolution operation in the prior art limit the CNN's ability to capture long-range dependencies between multi-level features, especially when dealing with targets with significant scale changes and spatial distribution diversity (such as agricultural land diversity), which will cause serious missed detection and false detection problems. Although the Transformer shows theoretical advantages in global context capture (such as cross-regional correlation modeling), the huge computational overhead brought by the multi-head attention mechanism makes it face the Pareto frontier dilemma of accuracy-speed in time-sensitive change detection tasks. The present invention proposes a multi-level visual state space remote sensing semantic change detection system and method based on a dynamic expansion mechanism. The method can effectively obtain spatiotemporal relationships, local and global contextual information from multi-temporal features, integrate multi-scale contextual information, effectively detect changes of different categories, and has linear complexity. It solves the inherent contradiction between existing CNN and Transformer in long-range dependency capture and computational efficiency, and achieves coordinated optimization of accuracy and speed.

[0008] The first object of the present invention is to provide a multi-level visual state space remote sensing semantic change detection method based on a dynamic expansion mechanism, comprising: Acquire remote sensing images of before and after phases; Extract multi-stage feature maps based on the SSM-based encoding backbone network of the remote sensing images of the previous and next phases; the resolution of the feature maps of each stage is different; According to the multi-stage feature map, the AWVSS module is used to decode and generate a binary transformation feature map and a semantic transformation feature map; Use the binary transformation feature map as a mask to multiply it with the semantic transformation feature map to generate a masked semantic transformation feature map; The loss is calculated based on the mask semantic transformation feature map and the true label semantic transformation segmentation map to obtain remote sensing semantic change information.

[0009] In one embodiment, the multi-stage feature map is obtained according to the following steps: Based on the twin network architecture, the input image is divided into blocks, and then multi-level features are extracted sequentially. Each network stage consists of downsampling and VSS blocks to generate feature maps with different resolutions of H / 4×W / 4, H / 8×W / 8, H / 16×W / 16, and H / 32×W / 32.

[0010] In one embodiment, the binary transformation feature map is obtained by performing binary change detection based on the AWVSS module according to the multi-stage feature map.

[0011] In one embodiment, the binary transformation feature map is obtained according to the following steps: Arrange multi-stage feature maps from low resolution to high resolution; After the low-resolution feature map is independently scanned by the AWVSS module, it is upsampled and passed to the next stage; the next stage includes fusing the output features of the feature map of the current stage independently scanned by the AWVSS module with the feature map of the previous stage, upsampling, and then passing it to the next stage until the feature maps of all stages are fused and upsampling is completed to obtain the binary transformation feature map.

[0012] In one embodiment, the semantic transformation feature map is obtained by performing semantic change detection based on the AWVSS module according to the multi-stage feature map.

[0013] In one embodiment, the AWVSS module is based on the mechanism of the Atrous window visual state space, which uses an expanded window to perform progressive scanning to extract features. It is specifically composed of two branches. One branch is used for feature extraction, which uses a depth-wise separable convolution followed by an AWSS2D layer; the other branch is used to calculate the multiplication gated signal, which is obtained through a linear mapping followed by an activation layer.

[0014] In one embodiment, the AWSS2D layer processing includes an expansion window scanning module, a feature extraction module based on the SSM (S6) Block model, and an expansion window scanning merging module. The specific processing process includes: The feature map output by the encoding stage is input and the image is divided into four groups of windows of different sizes through the expansion window scanning module; The image blocks in each group of windows are expanded and arranged into sequences according to the horizontal traversal path. The feature extraction module based on the SSM (S6) Block model processes them in parallel. The processed sequences are reconstructed and merged through the expansion window scanning and merging module to generate the output feature map.

[0015] In one embodiment, the dilated window scan merging module is used to recalibrate the weights of the feature map and selectively activate the most relevant variation features from the multi-window scan features.

[0016] A second object of the present invention is to provide a computer program product, comprising a computer program, which, when executed by a processor, implements a multi-level visual state space remote sensing semantic change detection method based on a dynamic expansion mechanism.

[0017] A third object of the present invention is to provide a multi-level visual state space remote sensing semantic change detection system based on a dynamic expansion mechanism, comprising: The data acquisition module is used to obtain remote sensing images of the preceding and following time phases; a multi-stage feature map is extracted based on the encoding backbone network of the SSM based on the remote sensing images of the preceding and following time phases; wherein the resolution of the feature map of each stage is different; A feature extraction module is used to generate a binary transformation feature map and a semantic transformation feature map by decoding the multi-stage feature map using the AWVSS module; the binary transformation feature map is used as a mask to multiply the semantic transformation feature map to generate a masked semantic transformation feature map; The change detection module is used to calculate the loss based on the mask semantic transformation feature map and the true label semantic transformation segmentation map to obtain remote sensing semantic change information.

[0018] The multi-level visual state space remote sensing semantic change detection method and system based on dynamic expansion mechanism proposed in this paper have significant advantages over existing CNN and Transformer methods in terms of detection accuracy, computational efficiency, and multi-scale adaptability. The specific beneficial effects are as follows: (1) Improve change detection accuracy, reduce missed detections and false detections, and enhance long-range dependency modeling capabilities: ① Adaptively adjust the receptive field of the convolution kernel through a dynamic expansion mechanism to overcome the local limitations of traditional CNNs and effectively capture large-scale changes in land features (such as farmland fragmentation and urban expansion). ② A multi-level visual state space model (VSSM) explicitly models spatiotemporal feature dependencies, reducing missed detections caused by changes in target scale (such as small buildings vs. large farmland). Multi-scale feature fusion optimization: A hierarchical feature interaction mechanism is used to fuse low-level details (such as road edges) with high-level semantics (such as land class conversion), reducing false detection rates in complex scenarios (such as cloud occlusion and shadow interference).

[0019] (2) Computational efficiency is significantly improved, suitable for real-time applications, with linear computational complexity (O(N)): The state-space model (SSM) replaces the self-attention mechanism of the Transformer, avoiding the O(N²) computational overhead. Memory usage is reduced: Through hierarchical feature compression and shared state space parameters, the model memory usage is reduced (compared to the Swin Transformer) (3) Enhanced robustness in multiple scenarios and adaptive scale changes: A dynamic expansion mechanism automatically adjusts the convolution kernel expansion rate (adjustable from 1 to 9) to adapt to targets of different resolutions (e.g., small roads vs. large water bodies). Cross-sensor compatibility: The model performs stably on multiple sources of data, such as Landsat and GF2, and can be migrated without retraining, making it suitable for large-scale remote sensing image processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The semantic change detection network based on AtrousMamba provided by the present invention; Figure 2 Encode the backbone network for the SSM change detection framework; Figure 3 Diagram of the Atrous window scanning mechanism; (a) Atrous window visual state space, (b) Atrous window SS2D, (c) Atrous window scanning merging module; Figure 4 It is a binary change detection network based on VSS and AWVSS. DETAILED DESCRIPTION

[0021] In order to illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description with reference to the embodiments.

[0022] The purpose of the present invention is to provide a multi-level visual state space remote sensing semantic change detection system and method based on a dynamic expansion mechanism. The present invention extracts multi-stage image features from the before-after phase remote sensing images using an SSM-based encoding backbone network. The feature images at different stages are decoded using AWVSS to capture the local and global spatial context of the input data, generating binary transformation feature maps and semantic transformation feature maps. The feature maps are upsampled and fused with lower-level feature map information with higher resolution. The binary transformation feature map is used as a mask to multiply the semantic transformation feature map with the semantic transformation feature map to generate a masked semantic transformation feature map. The fused feature map is smoothed using a residual layer. Finally, the semantic transformation feature map and the true label semantic transformation segmentation map are used to calculate the loss value.

[0023] See also Figure 1 As shown, the present invention provides a multi-level visual state space remote sensing semantic change detection method based on a dynamic expansion mechanism, which extracts multi-stage feature maps based on an SSM-based encoding backbone network of the previous and next phase remote sensing images; wherein the resolution of the feature map of each stage is different; the AWVSS module is used to decode the multi-stage feature map to generate a binary transformation feature map and a semantic transformation feature map; the binary transformation feature map is used as a mask to multiply the semantic transformation feature map to generate a masked semantic transformation feature map; the loss is calculated based on the masked semantic transformation feature map and the true label semantic transformation segmentation map to obtain remote sensing semantic change information.

[0024] The semantic change detection network in this paper adopts a weight-sharing Siamese architecture based on VMamba, and the decoder is designed using the AtrousMamba framework. VMamba is a general-purpose visual backbone network built on a state-space model (SSM), with a global receptive field and dynamic weights. The AtrousMamba architecture enables the model to effectively capture global contextual information while retaining fine-grained local features, thereby enhancing its change detection capabilities.

[0025] To achieve the above objectives, the present invention provides a multi-level visual state space remote sensing semantic change detection method based on a dynamic expansion mechanism, comprising: S1, obtaining remote sensing images of the preceding and following phases; Pre- and post-phase remote sensing images refer to remote sensing images of the same area captured at different points in time. In remote sensing images, the term "phase" refers to the time or time period in which the image was captured. Pre- and post-phase remote sensing images correspond to two specific time points or time periods before and after a change.

[0026] S2, extracting multi-stage feature maps based on the SSM-based encoding backbone network of the remote sensing images of the previous and next phases; wherein the resolution of the feature maps of each stage is different; The multi-stage feature map is obtained according to the following steps: Based on the twin network architecture, the input image is divided into blocks, and then multi-level features are extracted sequentially. Each network stage consists of downsampling and VSS blocks to generate feature maps with different resolutions of H / 4×W / 4, H / 8×W / 8, H / 16×W / 16, and H / 32×W / 32.

[0027] In this embodiment, see Figure 2 As shown in the figure, the input image is first divided into multiple patches, generating a 2D feature map with spatial dimensions of H / 4 × W / 4. The Siamese encoder extracts multi-level features from the input image. Each network stage consists of a VSS block, generating feature maps with different resolutions of H / 8 × W / 8, H / 16 × W / 16, and H / 32 × W / 32. These features are then passed to a customized variational decoder. By gradually expanding the receptive field, the variational decoder effectively learns the spatiotemporal relationships in the multi-level features, ultimately achieving accurate BCD results.

[0028] S3, decoding the multi-stage feature map using the AWVSS module to generate a binary transformation feature map and a semantic transformation feature map; The AWVSS module is based on the Atrous window visual state space mechanism and uses an expanded window for progressive scanning to extract features. It consists of two branches: one for feature extraction, which uses a depthwise separable convolution followed by an AWSS2D layer; the other for calculating the multiplication gated signal, which is obtained through a linear mapping followed by an activation layer.

[0029] The AWSS2D layer processing includes an expansion window scanning module, a feature extraction module based on the SSM (S6) Block model, and an expansion window scanning merging module. The specific processing process includes: The feature map output by the encoding stage is input and the image is divided into four groups of windows of different sizes through the expansion window scanning module; The image blocks in each group of windows are expanded and arranged into sequences according to the horizontal traversal path. The feature extraction module based on the SSM (S6) Block model processes them in parallel. The processed sequences are reconstructed and merged through the expansion window scanning and merging module to generate the output feature map.

[0030] The expansion window scan merging module is used to recalibrate the weights of the feature map and selectively activate the most relevant variation features from the multi-window scan features.

[0031] See also Figure 3 As shown, Figure 3(a) shows the Atrous window visual state space, progressively scanned using an expanding window. The residual network consists of two branches: one for feature extraction, using depthwise separable convolution followed by an AWSS2D layer; the other for computing a multiplicative gating signal, obtained via a linear mapping followed by an activation layer. Our approach incorporates a selective scanning mechanism using an expanding window with an adjustable expansion rate, enabling the scanning range to be gradually expanded. This design enables the model to simultaneously capture fine-grained local features and global context, improving its performance in change detection tasks.

[0032] Figure 3 Figure (b) shows Atrous Window SS2D. The data processing pipeline through the dilated window SS2D module consists of three steps: the dilated window scanning module (AWScan), the selective scanning module based on the S6 module, and the dilated window scanning merging module (AWSMerg). The combination of dilated window scanning and merging is collectively referred to as the dilated window scanning module (AWSM). Given the feature map output from the encoding stage, AWSS2D first segments the image into four sets of windows of different sizes using four different ratios (e.g., 2, 5, 7, and 9). The image blocks within each set of windows are expanded horizontally and arranged into sequences. These four sequences are then processed in parallel by independent S6 modules, with each scan window independently capturing relevant information. This progressive dilated window scanning method ensures comprehensive analysis within each selective scan block, balancing global context and local details. Finally, the processed sequences are reconstructed and merged using AWSMerge to generate the output feature map.

[0033] To enhance the interaction between channels, an AWSMerge method is proposed to recalibrate the weights of feature maps and selectively activate the most relevant variation features from multi-window scanning features. Figure 3 As shown in (c), it has spatial dimensions The feature map is compressed into a vector through the global average pooling operation Therefore, each element in the vector contains the global information of its corresponding channel. i The calculation formula for each element is as follows:

[0034] Where H and W are the length and width of the feature map X, respectively.

[0035] The weight vector is generated by two consecutive fully connected layers. The formula for this process is as follows:

[0036] in, f 1 and f2 represents two fully connected layers. The weight vector V is element-wise multiplied with the input feature Y to generate the final output tensor. This operation can be expressed as: =V ⊙ Y AWSMerge calculates importance weights for each feature channel, enabling the network to automatically enhance responses to important features while suppressing the influence of irrelevant or minor features.

[0037] Among them, the binary transformation feature map is obtained by performing binary change detection based on the multi-stage feature map based on the AWVSS module.

[0038] The binary transformation feature map is obtained by following the steps below: Arrange multi-stage feature maps from low resolution to high resolution; After the low-resolution feature map is independently scanned by the AWVSS module, it is upsampled and passed to the next stage; the next stage includes fusing the output features of the feature map of the current stage independently scanned by the AWVSS module with the feature map of the previous stage, upsampling, and then passing it to the next stage until the feature maps of all stages are fused and upsampling is completed to obtain the binary transformation feature map.

[0039] See also Figure 4 As shown in Figure 1, the SSM-based encoding backbone network extracts multi-stage image features from the preceding and following temporal remote sensing images. The decoding stage has three branches: the middle branch is often referred to as the change detection (CD) head, and the top and bottom branches correspond to the semantic segmentation (SS) head. AWVSS is used to decode the feature images at different stages to generate binary transformed feature maps. The specific structure of the change decoder is based on the proposed AWVSS module. We first rearrange the preceding and following temporal feature maps obtained during the encoding stage. The computational process is described as follows:

[0040]

[0041]

[0042] Where b is the batch size; c is the channel; h and w are the length and width of the feature map respectively; W is the total width after the feature map is arranged; m is a natural number.

[0043] It should be noted that after the entire encoding stage and before the decoding stage, that is, before decoding, the feature maps of each stage are rearranged and then pass through each sub-module of the decoder.

[0044] Then, use the AWVSS module to scan independently Z 1、Z 2 and Z 3 Each module effectively captures fine-grained local features and global context. The three scans are concatenated and passed through a convolutional layer for channel alignment and feature extraction. Finally, the output features of the current stage are fused with the feature map from the previous stage, upsampled, and passed to the next stage. By fully leveraging the spatiotemporal correlation information in the feature maps from the four encoder stages, high-precision binary change detection results are ultimately produced.

[0045] In this embodiment, the semantic transformation feature map is obtained by performing semantic change detection based on the AWVSS module according to the multi-stage feature map.

[0046] According to the remote sensing images of the previous and next phases and the real semantic segmentation labels, the SSM-based encoding backbone network extracts multi-stage image features; the feature images of different stages are decoded using AWVSS to obtain the semantic transformation feature maps.

[0047] S4, using the binary transformation feature map as a mask to multiply the semantic transformation feature map to generate a masked semantic transformation feature map; See also Figure 1 As shown in the figure, based on the remote sensing images of the previous and next time phases, the real binary transformation labels and the semantic segmentation labels, the SSM-based encoding backbone network extracts multi-stage image features; the feature images of different stages are decoded using AWVSS to capture the local and global spatial context of the input data, generating binary transformation feature maps and semantic transformation feature maps; the feature maps are upsampled and fused with the low-level feature map information with higher resolution; the binary transformation feature map is used as a mask to multiply the semantic transformation feature map to generate a masked semantic transformation feature map; the fused feature map is smoothed through the residual layer. The calculation steps are as follows:

[0048]

[0049]

[0050]

[0051] Where F4, F3, F2, and F1 are the calculated semantic transformation segmentation maps; ConvBlock is the convolution block, and ResBlock is the residual connection block.

[0052] S5. Calculate the loss based on the mask semantic transformation feature map and the true label semantic transformation segmentation map to obtain remote sensing semantic change information.

[0053] In this embodiment, the loss value is calculated using the semantic transformation feature map and the true label semantic transformation segmentation map; Loss calculation formula: (1) Loss function for BCD: Since binary change detection (BCD) is essentially a binary classification task, we use a classification loss function to guide network training. Specifically, the binary cross entropy (BCE) loss provides supervision for each pixel by comparing the predicted result with the true label (GT), which is defined as follows:

[0054] in, P c represents the predicted probability, Y c represents the true label.

[0055] (2) Loss function for SCD: The semantic loss is optimized using the cross entropy loss function, which is defined as follows:

[0056] in, N is the number of semantic categories, Y i and Pi Respectively represent i The present invention introduces a semantic change loss based on the contrastive learning principle to strengthen the semantic consistency between the two predicted semantic segmentation maps. Its calculation formula is:

[0057] in, Y c represents the true label, X 1 and X 2 represents the pixel vectors of the two semantic segmentation maps. The total loss calculation formula is:

[0058] According to the calculated total loss, the loss value is made as small as possible, that is, the target value is optimized.

[0059] This paper dynamically adjusts the convolution kernel sampling position through a learnable offset, replacing the fixed dilation rate dilated convolution. It can adapt to irregularly shaped targets (such as curved roads).

[0060] The present invention is based on the collaborative design of a dynamic expansion mechanism and a multi-level visual state space model (VSSM), specifically including: Based on a dynamic dilated convolution mechanism, this invention employs a convolution kernel with an adaptive dilation rate (adjustable from 1 to 8), addressing the scale sensitivity issue inherent in traditional fixed dilation rates. This method also incorporates an attention mechanism to dynamically select dilation regions (for example, focusing on feature edges or large, uniform areas). This overcomes the local receptive field limitations of CNNs and enables multi-scale feature capture (for example, simultaneously detecting small roads and large farmlands).

[0061] This invention, based on multi-level visual state space modeling (VSSM), explicitly models spatiotemporal dependencies through a hierarchical state space design (pixel level → object level → scene level). It also replaces the quadratic computations of the Transformer with linear state transition equations (O(N)). While maintaining global context, it improves computational efficiency by more than four times (compared to ViT).

[0062] The present invention provides a computer program product, comprising a computer program. When the computer program is executed by a processor, a multi-level visual state space remote sensing semantic change detection method based on a dynamic expansion mechanism is implemented.

[0063] The present invention provides a multi-level visual state space remote sensing semantic change detection system based on a dynamic expansion mechanism, comprising: The data acquisition module is used to obtain remote sensing images of the preceding and following time phases; a multi-stage feature map is extracted based on the encoding backbone network of the SSM based on the remote sensing images of the preceding and following time phases; wherein the resolution of the feature map of each stage is different; A feature extraction module is used to generate a binary transformation feature map and a semantic transformation feature map by decoding the multi-stage feature map using the AWVSS module; the binary transformation feature map is used as a mask to multiply the semantic transformation feature map to generate a masked semantic transformation feature map; The change detection module is used to calculate the loss based on the mask semantic transformation feature map and the true label semantic transformation segmentation map to obtain remote sensing semantic change information.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-level visual state space remote sensing semantic change detection method based on dynamic expansion mechanism, characterized by: include: Acquire remote sensing images of before and after phases; Extract multi-stage feature maps based on the SSM-based encoding backbone network of the remote sensing images of the previous and next phases; the resolution of the feature maps of each stage is different; According to the multi-stage feature map, the AWVSS module is used to decode and generate a binary transformation feature map and a semantic transformation feature map; Use the binary transformation feature map as a mask to multiply it with the semantic transformation feature map to generate a masked semantic transformation feature map; The loss is calculated based on the mask semantic transformation feature map and the true label semantic transformation segmentation map to obtain remote sensing semantic change information.

2. The multi-level visual state space remote sensing semantic change detection method based on dynamic expansion mechanism according to claim 1 is characterized in that: The multi-stage feature map is obtained according to the following steps: Based on the twin network architecture, the input image is divided into blocks, and then multi-level features are extracted sequentially. Each network stage consists of downsampling and VSS blocks to generate feature maps with different resolutions of H / 4×W / 4, H / 8×W / 8, H / 16×W / 16, and H / 32×W / 32.

3. The multi-level visual state space remote sensing semantic change detection method based on dynamic expansion mechanism according to claim 1 is characterized in that: The binary transformation feature map is obtained by performing binary change detection based on the multi-stage feature map based on the AWVSS module.

4. The multi-level visual state space remote sensing semantic change detection method based on dynamic expansion mechanism according to claim 3 is characterized in that: The binary transformation feature map is obtained by following the steps below: Arrange multi-stage feature maps from low resolution to high resolution; After the low-resolution feature map is independently scanned by the AWVSS module, it is upsampled and passed to the next stage; The next stage includes independently scanning the feature map of the current stage through the AWVSS module, outputting features that are fused with the feature map of the previous stage, and then upsampling, and then passing it to the next stage until the feature maps of all stages are fused and upsampling is completed to obtain the binary transformation feature map.

5. The multi-level visual state space remote sensing semantic change detection method based on dynamic expansion mechanism according to claim 1 is characterized in that: The semantic transformation feature map is obtained by performing semantic change detection based on the multi-stage feature map based on the AWVSS module.

6. The multi-level visual state space remote sensing semantic change detection method based on dynamic expansion mechanism according to claim 1 is characterized in that: The AWVSS module is based on the Atrous window visual state space mechanism and uses an expanded window for progressive scanning to extract features. It consists of two branches: one for feature extraction, which uses a depthwise separable convolution followed by an AWSS2D layer; the other for calculating the multiplication gated signal, which is obtained through a linear mapping followed by an activation layer.

7. The multi-level visual state space remote sensing semantic change detection method based on dynamic expansion mechanism according to claim 6 is characterized in that: The AWSS2D layer processing includes an expansion window scanning module, a feature extraction module based on the SSM (S6) Block model, and an expansion window scanning merging module. The specific processing process includes: The feature map output by the encoding stage is input and the image is divided into four groups of windows of different sizes through the expansion window scanning module; The image blocks in each group of windows are expanded and arranged into sequences according to the horizontal traversal path. The feature extraction module based on the SSM (S6) Block model processes them in parallel. The processed sequences are reconstructed and merged through the expansion window scanning and merging module to generate the output feature map.

8. The multi-level visual state space remote sensing semantic change detection method based on dynamic expansion mechanism according to claim 7 is characterized in that: The expansion window scan merging module is used to recalibrate the weights of the feature map and selectively activate the most relevant variation features from the multi-window scan features.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the multi-level visual state space remote sensing semantic change detection method based on a dynamic expansion mechanism according to any one of claims 1 to 8 is implemented.

10. A multi-level visual state space remote sensing semantic change detection system based on dynamic expansion mechanism, characterized in that: include: Data acquisition module, used to obtain remote sensing images of the previous and next phases; Extract multi-stage feature maps based on the SSM-based encoding backbone network of the remote sensing images of the previous and next phases; the resolution of the feature maps of each stage is different; A feature extraction module is used to generate a binary transformation feature map and a semantic transformation feature map by decoding the multi-stage feature map using the AWVSS module; the binary transformation feature map is used as a mask to multiply the semantic transformation feature map to generate a masked semantic transformation feature map; The change detection module is used to calculate the loss based on the mask semantic transformation feature map and the true label semantic transformation segmentation map to obtain remote sensing semantic change information.

Images