A remote sensing image change detection method based on change-aware Mamba

Abstract

This invention proposes a change detection method for remote sensing images based on change-aware Mamba. Addressing the problem that existing uniform scanning strategies cannot adapt to the sparsity of changed regions, this method includes the following steps: 1) Extracting features from dual-temporal remote sensing images using a dual-branch feature extraction network, and using a change-aware adaptive scanning module combined with spatial change priors to adjust scanning behavior, focusing on change-related regions and suppressing background noise; 2) Using a selective scanning fusion module to interactively fuse the extracted dual-temporal features, expanding the features into multi-directional sequences and performing path-dependent sequential processing through Mamba blocks to accumulate change evidence; 3) Upsampling and reconstructing the fused features using a channel-averaged visual state spatial decoder, and outputting the final change detection map. This invention significantly improves the accuracy, structural consistency, and robustness of change detection in remote sensing images under complex backgrounds by changing the uniform scanning strategy.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer vision technology, and mainly to a method for detecting changes in remote sensing images based on change-aware Mamba. Background Technology

[0002] Change detection (CD) aims to identify differences in the same geographic area at different times and is a fundamental task in remote sensing fields such as urban monitoring, disaster assessment, and ecosystem analysis. In recent years, deep learning methods have transformed change detection from pixel-based comparison to feature-level modeling.

[0003] Early mainstream methods were primarily based on Convolutional Neural Networks (CNNs). While these could extract local detail features, they were limited by their finite receptive field, making it difficult to establish long-range dependencies in the global space. This resulted in spurious changes and fragmented predictions in complex scenes. To overcome the limitations of CNNs, researchers introduced the Transformer architecture, which uses a self-attention mechanism to establish global correlations, achieving better results. However, the self-attention mechanism suffers from limitations that are quadratic inversely proportional to image resolution. The increasing computational complexity places a significant computational burden on the processing of high-resolution remote sensing images.

[0004] Recently, the Mamba architecture, based on State Space Models (SSMs), has gained attention for its combination of global modeling capabilities and linear time complexity. The computational efficiency of Mamba has garnered significant attention in visual tasks. However, existing Mamba-based change detection methods directly inherit the "uniform scanning strategy" from general visual backbone networks, treating all spatial locations and scanning paths as equally important. This uniform scanning approach contradicts the sparsity and geometric variability of real-world changes in remote sensing images, leading to diffusion and dispersion in the model's response in changed regions and introducing substantial background noise. Furthermore, existing methods largely rely on static feature fusion, failing to fully leverage Mamba's inherent advantages in processing dynamic, path-dependent sequence information.

[0005] To address the aforementioned issues, there is an urgent need for a remote sensing image change detection method that can adaptively perceive changing regions and dynamically adjust the state space sequence scanning process, so as to accurately capture the structural features of real changes while reducing background interference. Summary of the Invention

[0006] To overcome the shortcomings of existing Mamba-based change detection technologies, such as inaccurate change response, susceptibility to complex background interference, and blurred edges caused by uniform scanning strategies, this invention provides a remote sensing image change detection method and system based on change-aware adaptive Mamba. This invention reconstructs change detection as a change-aware sequential state space modeling process, achieving high-precision and robust detection of changed areas through dynamic adjustment of the scanning path and feature fusion mechanism. To achieve the above objectives, the technical solution of this invention is as follows: A remote sensing image change detection method based on change-aware Mamba, mainly including the following steps: Step 1: Extracting change space prior weights, acquiring dual-temporal remote sensing images of the same region, inputting them into the change prior module (CPM), jointly extracting the intensity difference and texture consistency between the dual-temporal images, generating a multi-channel change weight map as a soft space guide for the scanning process. Step 2: Change-aware adaptive scanning and feature extraction, inputting the dual-temporal remote sensing images into a dual-branch feature extraction network, which consists of hierarchically stacked change-aware adaptive scanning visual state space blocks (CAAS VSS Block). Using the change weight map obtained in Step 1, a learnable injection mechanism is used to dynamically modulate and weight the four-way scan stream, enabling the state transition process to adaptively focus on change-related regions and suppress background regions, thereby extracting dual-temporal features. Step 3: Path-dependent selective scan fusion. The extracted dual-temporal features are concatenated and input into the selective scan fusion module (S2FM). This module expands the dual-temporal features into a one-dimensional sequence and inputs weight-shared Mamba blocks, modeling the feature interaction as a path-dependent sequence modeling process. It accumulates real change evidence on each scan path by relying on hidden states. Subsequently, a gated residual fusion mechanism is used to balance the sequence features and local spatial cues, outputting the fused change feature map. Step 4: Feature Reconstruction and Change Detection Map Generation. The fused change feature map is input into a decoder composed of Channel Average Visual State Space (CAVSS) blocks. Deep semantics and shallow spatial details are combined through progressive upsampling and skip connections. The result is optimized using a multi-scale depth supervision strategy, and the final remote sensing image binary change detection map is output.

[0007] Compared with the prior art, the present invention has the following beneficial effects: I. Higher detection accuracy and stronger noise resistance. By adopting a change-aware adaptive scanning mechanism, the traditional uniform scanning method of Mamba is changed. The model is guided to adaptively focus on the changing area based on spatial priors and suppress background noise, effectively solving the problems of change response diffusion and false change interference in complex scenes.

[0008] II. More accurate spatiotemporal feature interaction and accumulation. By utilizing the selective scan fusion module to capture the nonlinear evolutionary relationship between features in two time phases, and by deeply mining and accumulating change evidence through the sequential memory characteristics of the state-space model, the feature modeling capability for complex changing targets is significantly improved.

[0009] III. By combining the above modules, this invention improves the accuracy of change detection and enhances the continuity of the spatial structure of the change detection results while maintaining a low computational load and parameter count, resulting in a detection map with clearer boundaries and better robustness. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the various embodiments disclosed in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only a few embodiments disclosed in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 The diagram shows the overall network structure of a remote sensing image change detection method based on change-aware adaptive Mamba, as provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the Change Prior Module (CPM) provided in an embodiment of the present invention. Figure 3 A structural comparison diagram of the Change-Aware Visual State Space Block (CAAS VSS Block) provided in this embodiment of the invention and a traditional module. Figure 4 This is a schematic diagram illustrating the principle of change-aware two-dimensional selective scanning (CA-SS2D) provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the internal structure of the selective scan fusion module (S2FM) provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0013] This invention proposes a change detection method for remote sensing images based on change-aware Mamba. Addressing the irregularity and sparsity of the true change distribution in remote sensing images, it innovatively introduces a change-aware adaptive scanning (CAAS) mechanism and a selective scan fusion (S2FM) mechanism. For example... Figure 1 As shown, this method adopts an end-to-end encoder-decoder architecture, and the specific implementation steps are as follows: Step 1: Extract the prior weights of spatial variation. Input a pair of registered dual-temporal remote sensing images of the same region. ,in and These represent the image height and width, respectively, with 3 representing the number of channels. To avoid the blind and uniform scanning method in the subsequent Mamba model, this step first uses a lightweight variation prior module (CPM) to acquire prior knowledge. CPM calculates the pixel-wise intensity difference and texture consistency difference based on local windows between the two images: The absolute difference map, the element-wise product map, and the original dual-temporal remote sensing image are concatenated through channels to form a joint feature tensor: The joint feature tensor is processed through a convolutional neural network to output the corresponding multi-channel change prior weight map: The weight map shows high values ​​in potentially changing areas and low values ​​in unchanged background areas.

[0014] Step 2: Change-aware adaptive scanning and feature extraction. A parameter-sharing Siamese network is used to perform... and Feature extraction is performed. The network consists of multiple layers, each containing several Change-Aware Adaptive Scanning Visual State Space Blocks (CAAS VSS Blocks). For example... Figure 3 As shown, traditional VSS blocks use standard two-dimensional selective scanning (SS2D), while the CAAS VSS block in this embodiment replaces it with change-aware two-dimensional selective scanning (CA-SS2D). In the CAAS VSSBlock, let the input feature tensor be... The CAAS VSS block flattens it out to a length of sequences, such as Figure 4 As shown, in CA-SS2D, the sequence is scanned in four directions (from left to right, from right to left, from top to bottom, and from bottom to top). This embodiment introduces the aforementioned change weight map based on this. Specifically: (1) Downsampling to the spatial resolution of the current feature tensor yields ; (2) In Mamba's discretized state-space equations, the state transition matrix of the traditional method and input matrix Depends only on input This invention will The value is injected as a spatial modulation signal: in, It is the time step parameter. and These represent the preset weight matrices in the state-space model. and It is a linear mapping function. Through the above modulation, when the scan path passes through... When the background region is low, the feature accumulation process is suppressed; when the path passes through... In regions of high variation, the model adaptively allocates more attention to accurately capture bi-temporal features.

[0015] To adaptively and dynamically adjust the injection strength based on prior changes in different scenarios, a learnable injection coefficient is introduced. : pass Convolution transforms the multi-channel variation prior weight map into spatial weight maps corresponding to the four scanning directions. The feature flow of each scanning direction, its corresponding spatial weight map, and the injection coefficients are weighted and modulated, and feature aggregation is performed in the following manner: in This represents the scan flow in the k-th direction. This is the corresponding spatial weight map.

[0016] Step 3: Path-dependent selective scan fusion. For example... Figure 5 As shown, dual-temporal features are obtained. Furthermore, traditional splicing or subtraction methods neglect the spatiotemporal evolution relationships between features. This step employs a selective scan fusion module (S2FM) for deep interaction.

[0017] (1) First, By splicing along the channel dimension, we obtain ; (2) Channel dimensionality reduction and feature representation optimization are performed sequentially through the Ghost Block and SE Block to obtain a refined feature map. This refined feature map is then flattened along the spatial dimension into a one-dimensional sequence and input into a fusion memory unit composed of multi-directional Mamba blocks. During this process, the hidden state... Continuously updated along the scan path: in Let represent the input feature at time t in a one-dimensional sequence. This indicates the hidden state at the current moment. This represents the hidden state at the previous time step. This mechanism allows the model to capture and store the temporal differences in features evolving along the scan path by updating the hidden state; (3) In order to preserve local spatial details, a gated residual connection mechanism was set up: Finally, the gated output result is obtained: The gated output sequences from each direction are then summed and reshaped into two-dimensional spatial features. Finally, residual connections from the refined feature map are introduced to output a fused and transformed feature map. .

[0018] Step 4: Feature Reconstruction and Change Detection Map Generation. For example... Figure 1 As shown in the Decoder section, Input decoder. The decoder employs a hierarchical upsampling architecture. To address the redundancy introduced by high-dimensional channels, this embodiment utilizes a Channel Average Visual State Space Block (CAVSS Block) in the decoder. Within the CAVSS block, before performing state space sequence modeling, the average value of the feature map across the channel dimension is calculated, or adaptive channel pooling is employed to guide the model to focus on the continuity of the spatial structure and smooth the edges of changing regions. Skip connections to the feature extraction branch are added between decoder layers. Finally, the network passes through a... The convolutional kernel outputs a prediction map with 1 prediction channel. After Sigmoid activation, it is binarized by setting a threshold to output the final detection result. During the training phase, this example combines Focal Loss and Dice Loss for multi-scale deep supervised training.

Claims

1. A method for detecting changes in remote sensing images based on change-aware Mamba, characterized in that, The process includes the following steps: Step 1: Acquire dual-temporal remote sensing images and input them into a weight-sharing dual-branch feature extraction network for multi-scale feature extraction; the feature extraction network includes a change-aware adaptive scanning module, which uses a generated change prior weight map to dynamically modulate the scanning path of the visual state space block to enhance the features of change-related areas and suppress background interference; Step 2: Use a selective scanning fusion module to interactively fuse the dual-temporal features output in Step 1. This module unfolds the spatial features into a multi-directional one-dimensional sequence, gathers historical change information through a state space model, and outputs fused features using a gated residual mechanism; Step 3: Input the fused features obtained in Step 2 into the channel average visual state space decoder, and output the predicted change detection map through progressive upsampling and feature reconstruction.

2. The method according to claim 1, characterized in that, The specific process of generating the change prior weight map in step 1 is as follows: constructing a change prior module and calculating the absolute difference map of the intensity of the two-temporal remote sensing images. Element-wise product map with texture consistency : in , The images are two-phase images before and after the change. For element-wise multiplication calculations, the absolute difference map, the element-wise product map, and the original dual-temporal remote sensing image are channel-concatenated to form a joint feature tensor. : in,[ The channel concatenation operation processes the joint feature tensor through a convolutional neural network to output the corresponding multi-channel change prior weight map. : in, This is a convolution mapping operation.

3. The method according to claim 2, characterized in that, The specific operation of the change-aware adaptive scanning module is as follows: In the change-aware two-dimensional selective scanning, learnable injection coefficients are introduced. Adjusting the injection intensity based on prior changes: in For learnable parameter values, This represents the Sigmoid activation function, through... Convolution incorporates the multi-channel variation prior weight map Converted into spatial weight maps corresponding to the four scanning directions The feature streams of each scanning direction are weighted and modulated with their corresponding spatial weight maps and the injection coefficients, and feature aggregation is performed in the following manner: in This represents the scan flow in the k-th direction. For the corresponding spatial weight map, {ud, du, lr, rl} represent the scan flow in the four directions: top to bottom, bottom to top, left to right, and right to left, respectively. This represents a full-1 tensor with the same dimension as the spatial weight graph. This is the final feature map output after feature aggregation.

4. The method according to claim 1, characterized in that, The specific processing flow of the selective scan fusion module in step 2 includes: concatenating the extracted dual-temporal features in the channel dimension, and sequentially performing channel dimensionality reduction and feature representation optimization through Ghost Block and SE Block to obtain a refined feature map; expanding the refined feature map into four one-dimensional sequences along four preset scanning directions; inputting the above four one-dimensional sequences in parallel into a weight-shared Mamba block, using the internal hidden state as a cyclic accumulator, and aggregating feature difference information along each scanning path.

5. The method according to claim 4, characterized in that, The gated residual mechanism in step 2 is implemented as follows: Let the one-dimensional sequence of the input Mamba block be... The corresponding sequential output sequence is Learnable gating weights are generated by calculating the linear transformation after concatenating the two. : in ( () represents the linear mapping layer, and after calculation, the gated output result is obtained. : The gated output sequences in each direction are then summed and reshaped into two-dimensional spatial features, and finally, residual connections of the refined feature map are introduced.

6. The method according to claim 1, characterized in that, The channel-average visual state space decoder used in step 3 combines a channel attention mechanism, using global average pooling and max pooling to generate adaptive weights for channel recalibration.

7. The method according to claim 1, characterized in that, The proposed method employs a multi-scale deep supervision strategy during the model training phase. Its hybrid loss function is composed of a weighted sum of Focal loss and Dice loss, and it jointly optimizes the change detection prediction maps of multiple scales output by the decoder at different stages.

Images