Remote sensing image change detection method based on CNN-Mama hybrid network

Abstract

The invention provides a remote sensing image change detection method based on a CNN-Mama hybrid network, and relates to the technical field of image detection, and the method comprises the following steps: carrying out the preprocessing of a disclosed remote sensing image change detection data set, and constructing a dual-temporal high-resolution remote sensing image training set and a dual-temporal high-resolution remote sensing image test set; the preprocessing comprises image registration, cutting, normalization and data enhancement; constructing a change detection network model based on a CNN-Mama trunk, and performing training on the training set to obtain an optimal model; the network model comprises a twin encoder, a dynamic grouping attention module, a space adaptive fusion module, a multi-scale edge enhancement module and a decoder; and inputting a dual-temporal remote sensing image of a to-be-detected area into the optimal model, and outputting a binary change detection image with the same size as the input image through feature extraction, feature interaction, cross-layer fusion and step-by-step decoding. According to the method, the convolutional neural network and the state space modeling network are combined, and the local feature capture capability of the CNN and the long-range dependence modeling characteristic complementation of Mama are combined.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image detection, and in particular, relates to a remote sensing image change detection method based on a CNN-Mamba hybrid network. BACKGROUND

[0002] With the continuous growth of population and the continuous expansion of city size, the influence of human activities on the earth's environment has reached a historical high, and remote sensing image change detection, as an important technical means to identify surface changes by comparing different time images, plays a key role in the fields of city planning, building expansion monitoring, road detection and disaster assessment. Early traditional change detection methods (such as difference method, change vector analysis and threshold segmentation) rely on spectral information and artificial rules, although they are easy to implement, but are easily disturbed by external factors such as light, season and atmospheric conditions, and the stability and reliability of the detection results are low; while shallow machine learning methods (such as support vector machine, K nearest neighbor and random forest) have improved the feature representation ability to a certain extent, but their feature extraction depends on artificial design, and the generalization and automation level are still limited.

[0003] With the development of deep learning technology, convolutional neural network (CNN) is widely used in remote sensing image change detection tasks. CNN can automatically extract multi-level features of images through end-to-end learning mechanism, which significantly improves the detection accuracy and robustness, and the local receptive field characteristics of convolution operation can effectively capture the detailed features such as building edges and textures, which performs outstandingly in building boundary segmentation and small-scale target recognition; however, due to the limitation of local convolution kernel, CNN has insufficient modeling ability for long-range dependence and global context, which leads to missed detection or background interference in large-scale or complex scenes, and the traditional convolution structure is also prone to loss of detailed information in the multiple downsampling process.

[0004] To address the shortcomings of CNNs in global modeling, the Transformer architecture has been introduced into remote sensing image change detection tasks in recent years. Its self-attention mechanism possesses powerful global modeling capabilities, enabling it to capture long-range dependencies in large-scale images and demonstrating significant advantages in cross-temporal feature alignment and global change perception. Compared to CNNs, Transformers excel in global context modeling and feature representation flexibility. However, a consequence is that its computational complexity increases quadratically with input resolution, becoming excessively costly for high-resolution remote sensing tasks. Against this backdrop, state-space models (such as VMamba) have emerged as a new research direction. They maintain linear complexity while possessing strong long-range dependency modeling capabilities, efficiently extracting global spatiotemporal features, making them particularly suitable for large-scale, high-resolution remote sensing image scenarios. Although VMamba has significant advantages in global feature modeling, its ability to characterize local spatial features and boundaries remains insufficient when used alone. Therefore, further research and improvements are urgently needed to enhance its performance. Summary of the Invention

[0005] To address the technical problem mentioned in the background that the ability to characterize local spatial features and boundaries is still insufficient when used alone, this invention provides a remote sensing image change detection method based on a CNN-Mamba hybrid network. This invention uses a combination of ResNet and Mamba to form the backbone network, organically combining local representation capabilities with global modeling capabilities. An adaptive multi-scale fusion module enables adaptive fusion of multi-scale features. An edge feature enhancement module improves the discriminative power and continuity of boundary regions. A dynamic grouping attention module, through a dynamic channel allocation mechanism, achieves flexibility in modeling features of different granularities, improving feature representation capabilities and adaptability to complex environments.

[0006] The technical means employed in this invention are as follows: A method for detecting changes in remote sensing images based on a CNN-Mamba hybrid network includes the following steps: Step 1: Preprocess the publicly available remote sensing image change detection dataset to construct a dual-temporal high-resolution remote sensing image training set and a test set; the preprocessing includes: image registration, cropping, normalization and data augmentation; Step 2: Construct a change detection network model based on the CNN-Mamba backbone and train it on the training set to obtain the optimal model; the network model includes: a Siamese encoder, a dynamic grouping attention module, a spatial adaptive fusion module, a multi-scale edge enhancement module, and a decoder; Step 3: Input the dual-temporal remote sensing image of the area to be detected into the optimal model. After feature extraction, feature interaction, cross-layer fusion and step-by-step decoding, output a binary change detection map of the same size as the input image.

[0007] Furthermore, the multi-scale edge enhancement module, based on the edge enhancement module, expands the receptive field through dilated convolution to extract multi-scale information before handing it over to the edge enhancement module for processing; The multi-scale edge enhancement module processes the input features through standard convolution and then into dilated convolution, generating three branches with different receptive fields: ; in, These represent the multi-scale dilated convolution outputs corresponding to different dilation rates; This represents the feature map input to the current decoding layer; Different branches of the receptive field are input into independent edge enhancement modules for edge enhancement: ; in, Represents the boundary features after processing by the edge enhancement module at the i-th scale; This represents the output of the convolution at the i-th scale; Indicates multi-scale paths; The output is concatenated through channels and then fused using a 1×1 convolution: ; in, This represents the multi-scale boundary enhancement features after fusion; This represents edge enhancement features corresponding to three different scales; The global residual output is formed by combining the original input features: ; in, This represents the output of the multi-scale edge enhancement module.

[0008] Furthermore, the dynamic grouping attention module enhances its ability to discriminate complex and changing regions through adaptive channel grouping and multi-type attention mechanisms; Let the input features be The dynamic grouping attention module first performs global statistics to generate channel partitioning weights. : ; in, This indicates a global average pooling operation; Represents input features; Channel dimension based on weight The data is dynamically divided into four subsets, which are then fed into four attention branches to form an adaptive multi-scale feature stream. The outputs {Y1, Y2, Y3, Y4} of the four attention branches are concatenated along the channel dimension and then fused through a 1×1 convolution and a nonlinear transformation.

[0009] The final output features form a residual connection with the input features: ; in, This represents the fused attention feature map; This represents the intermediate features after fusion.

[0010] Furthermore, the four parallel branches include: a dot product attention branch, a window attention branch, a region attention branch, and a global attention branch; Furthermore, the spatial adaptive fusion module achieves fine-grained adaptive fusion between shallow detail features and deep semantic features; Let the low-level features of the input be... High-level characteristics are The two sets of features are concatenated along the channel dimension to obtain a joint representation: ; Spatial weight graph generated using a lightweight fusion network : ; This weighted graph is used to perform weighted fusion of high- and low-level features in the spatial dimension: ; When α approaches 1, the fusion calculation at this position... The coefficient is relatively large, and The coefficient is relatively small, and the system at this location is mainly composed of... The final weighted result is dominated by α, thus tending to retain shallow details; conversely, when α is close to 0, the fusion calculation at that location... The coefficient is relatively large, and The coefficient is relatively small, and the system at this location is mainly composed of... It dominates the final weighted result, and therefore pays more attention to the semantic consistency of the higher level; The input features are divided into multiple subgroups by channel, and each subgroup is processed by an independent SAC unit and CBAM attention unit before being concatenated and fused. ; Finally, feature compression and spatial consistency restoration are achieved through 1×1 convolution and residual thinning layers, as shown in the equation: ; in, This represents the result of splicing multiple sets of fused features; This represents the final output feature of the spatial adaptive fusion module.

[0011] Compared with the prior art, the present invention has the following advantages: This invention combines convolutional neural networks with state-space modeling networks, complementing the local feature capture capabilities of CNNs with the long-range dependency modeling characteristics of Mamba.

[0012] This invention introduces a dynamic grouping attention module, which achieves the gradual fusion of local texture information and global semantic information through hybrid attention collaborative modeling, thereby maintaining accurate recognition of subtle changes in regions even in complex backgrounds.

[0013] This invention enhances boundary features by designing an edge enhancement module and utilizing a channel-space collaborative attention mechanism, enabling the model to accurately perceive changes in building outlines and subtle structures. This design effectively suppresses background noise interference, improves the clarity and continuity of boundaries in changing areas, and reduces false detections and false alarms caused by blurred boundaries.

[0014] This invention introduces a spatial adaptive fusion module that dynamically generates fusion weights for each spatial location, achieving fine-grained alignment and adaptive fusion of deep semantics and shallow details. This mechanism enables the model to adaptively select to retain semantic or detailed information based on different regional features, thereby significantly improving small target detection and boundary repair capabilities. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the network structure of the present invention.

[0017] Figure 2 This is a schematic diagram of the structure of the multi-scale edge enhancement module of the present invention.

[0018] Figure 3 This is a schematic diagram of the dynamic lightweight gating module of the present invention.

[0019] Figure 4 This is a schematic diagram of the spatial adaptive fusion module of the present invention.

[0020] Figure 5 This is a comparison chart of change detection results for different models on the LEVIR dataset.

[0021] Figure 6 This is a comparison chart of change detection results for different models using the WHU-CD dataset. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] like Figures 1 to 4As shown, this application proposes a remote sensing image change detection method based on a CNN-Mamba hybrid network. Its overall framework adopts a Siamese symmetric network structure, consisting of an encoder, skip connections, and a decoder, aiming to simultaneously capture global context and fine-grained boundary features, thereby achieving accurate change recognition in complex scenes. In the encoder stage, the first three stages of ResNet18 are used as the shallow feature extraction backbone network to extract local texture and edge details from the image. To further improve the shortcomings of CNN in long-range dependency modeling, VMamba's VSS_Block is subsequently introduced as a global feature extraction unit in the deep semantic modeling part. VSS_Block is based on the state-space modeling principle and adopts a four-directional selective scanning structure to achieve effective capture of long-range dependencies with linear complexity. This module achieves consistent modeling of global features and efficient aggregation of multi-scale context through cross-channel and cross-spatial information interaction. This application introduces a Dynamic Group Attention (DGAM) module between ResNet18 and VSS_Block. This module enhances the expressive power of local details and edge features while maintaining computational efficiency by combining multiple attention mechanisms in parallel (including dot product attention, window attention, region attention, and global attention), thus achieving a gradual fusion of features from local details to global semantics. In the VSS_Block output stage, a Dynamic Group Attention (DGAM) module is embedded again for semantic alignment and fine-grained enhancement of deep features. Through its dynamic channel partitioning and branch adaptive selection mechanism, this module achieves a synergistic enhancement of global semantic consistency and refined expression of local features, providing more discriminative and structured fused features for subsequent decoding stages. In the shallow paths, an Edge Enhancement (DEUM) module is introduced, specifically for capturing edge information in shallow features. This module uses lightweight convolution and feature differencing operations to saliencyize edge features, effectively improving the network's ability to perceive the boundaries of changing regions. In the decoding stage, a progressive upsampling method is used to restore spatial resolution, and a Spatial Adaptive Fusion (SACM) module is embedded in each layer's fusion process. This module generates independent fusion weights for each spatial location, enabling adaptive weighting of shallow detail features and deep semantic features, thus achieving fine-grained feature alignment and fusion at the spatial level. When the fusion weights approach 1, the model tends to retain shallow detail information; when they approach 0, it relies more on high-level semantic structures, thus maintaining edge integrity while ensuring global consistency. Finally, the network restores the spatial resolution to the same level as the input image through multi-layer SACM fusion and a 3×3 convolutional prediction head, outputting binarized change detection results.This design fully integrates the local feature extraction capabilities of CNNs with the long-range dependency modeling advantages of Mamba structures. By combining the multi-attention dynamic aggregation of DGAMs with the spatial adaptive fusion mechanism of SACM, it achieves multi-level feature collaborative modeling from local edges to global semantics, significantly improving the accuracy of change detection and boundary clarity in complex scenes.

[0025] Remote sensing image change detection, such as Figure 1 As shown, the encoder module uses the first three stages of ResNet18 as the shallow feature extraction backbone network, and VSS_Block is the Visual State Space Block (VSSB), which implements spatiotemporal state modeling based on the VMamba architecture. Unlike traditional convolutional neural networks with local convolutional kernels or Transformer fixed self-attention, VSSB efficiently captures long-range dependencies while maintaining linear computational complexity through state space parameterization. Specifically, VSS_Block employs four-directional selective scan 2D on the two-dimensional feature map, including four directions: top left to bottom right, bottom right to top left, top right to bottom left, and bottom left to top right, thereby establishing bidirectional dependencies across regions. This mechanism can effectively characterize the continuity and global consistency of long-distance structures such as building clusters in large-scale remote sensing images.

[0026] This application also includes a method for detecting changes in remote sensing images based on a CNN-Mamba hybrid network, comprising the following steps: Step 1: Preprocess the publicly available remote sensing image change detection dataset to construct a dual-temporal high-resolution remote sensing image training set and a test set; the preprocessing includes: image registration, cropping, normalization and data augmentation; Step 2: Construct a change detection network model based on the CNN-Mamba backbone and train it on the training set to obtain the optimal model; the network model includes: a Siamese encoder, a dynamic grouping attention module, a spatial adaptive fusion module, a multi-scale edge enhancement module, and a decoder; Step 3: Input the dual-temporal remote sensing image of the area to be detected into the optimal model. After feature extraction, feature interaction, cross-layer fusion and step-by-step decoding, output a binary change detection map of the same size as the input image.

[0027] Specifically, the DEMU module is low-level, such as... Figure 2As shown, a multi-scale edge enhancement module (DEUM) is proposed to further improve the expressive ability of boundary regions in change detection. The input features are first jointly recalibrated in the channel space by CBAM to suppress low-frequency noise and background interference and enhance the significant response of the boundary region. Subsequently, the features processed by CBAM and the input global mean pooling result are differentially processed to extract high-frequency edge components and obtain its expression, as shown in Equation (1).

[0028] , , (1) Next, the features are nonlinearly refined by 3×3 convolution and GELU activation function to obtain their expression, as shown in equation (2).

[0029] (2) And form a residual connection with the original input to achieve feature fidelity.

[0030] Finally, edge features are generated through 1×1 convolution, while skip connections are preserved for subsequent fusion.

[0031] To further adapt to target boundaries of different scales, this application designs a multi-scale edge enhancement structure.

[0032] First, the input features are fed into dilated convolution (DConv) after standard convolution, generating three branches with different receptive fields to capture edge context features at different scales. The expression is shown in Equation (3).

[0033] (3) Subsequently, each scale branch is input into an independent EUM unit for edge enhancement, as shown in equation (4).

[0034] (4) The three EUM outputs are concatenated by channel splicing and then fused by 1×1 convolution, as shown in equation (5).

[0035] (5) The global residual output is formed by combining the original input features, as shown in equation (6).

[0036] (6) This module enables the model to maintain global consistency while possessing multi-scale boundary perception capabilities and progressive features at detailed levels.

[0037] To fully integrate feature dependencies across different scales and ranges, this application introduces a Dynamic Group Attention (DGAM) module between CNN and Mamba to achieve collaborative modeling of local fine features and global context. This module, through adaptive channel grouping and multi-type attention mechanisms, balances edge detail enhancement with global semantic consistency, thereby significantly improving the model's ability to discriminate complex and changing regions.

[0038] Let the input features be First, perform global statistics on the channel to generate channel-specific weights: .

[0039] GAP represents global average pooling. The channel dimension is based on the weights. The feature is dynamically divided into four subsets, which are then fed into four attention branches to form an adaptive multi-scale feature stream.

[0040] The four parallel branches correspond to different types of attention mechanisms: The Dot Attention branch captures fine local features by calculating short-range dependencies through standard self-attention, and its formula is shown in Equation (7).

[0041] (7) This branch is suitable for modeling fine-grained information such as building edges and textures.

[0042] The Window Attention branch divides the features into local windows and performs self-attention calculations within each window. Its formula is shown in Equation (8).

[0043] ; (8) The Area Attention branch divides the input features into several areas in space and performs attention calculations only within each area to achieve spatially adaptive feature fusion. An attention score is calculated within each area, as shown in equation (9).

[0044] (9) The final output region weighted features are obtained by formula (10).

[0045] (10) This branch achieves information aggregation within the region while maintaining global efficiency, thereby improving structural stability and semantic consistency.

[0046] The Global Attention branch uses a global feature compression and channel weighting mechanism to construct global dependencies, and its formula is shown in equation (11). (11) This branch ensures that features maintain global consistency over long distances.

[0047] The outputs {Y1,Y2,Y3,Y4} of the four attention branches are concatenated along the channel dimension and then fused through a 1×1 convolution and a nonlinear transformation. .

[0048] The final output features form a residual connection with the input features: .

[0049] Through the above design, DGAM achieves progressive dependency modeling from local textures to regions and then to the global while keeping computational complexity under control. This module effectively compensates for the shortcomings of CNNs in long-range relationship modeling and provides a bridge and contextual dependencies for the transition from CNNs to Mamba.

[0050] To avoid the low-level features being overwhelmed by the high-level semantics when fusing high-level and low-level features, and to enhance the model's performance in complex scenarios, we propose a spatial adaptive fusion module to achieve fine-grained adaptive fusion between shallow detail features and deep semantic features.

[0051] Let the low-level features of the input be... High-level characteristics are Low-level features contain rich texture and edge details, while high-level features have a stronger global semantic expression capability.

[0052] The module first concatenates the two sets of features along the channel dimension to obtain a joint representation, the expression of which is shown in equation (12).

[0053] (12) A spatial weight graph is then generated using a lightweight fusion network. For specific calculations, see equation (13).

[0054] (13) Then, the weighted graph is used to perform weighted fusion of high and low layer features in the spatial dimension: the specific calculation is shown in equation (14).

[0055] (14) when When the value is close to 1, the module tends to retain shallow details; when... When the value approaches 0, greater emphasis is placed on semantic consistency at higher levels. This mechanism enables the model to adaptively adjust information sources based on spatial location, achieving accurate modeling of building details and the edges of changing areas.

[0056] To further enhance the discriminative power of the fused features, the SACM module is combined with an attention enhancement mechanism (such as CBAM) within a block-based structure. Specifically, the input features are divided into multiple subgroups according to channels. Each group is processed by an independent SAC unit and CBAM attention unit before being spliced ​​and fused to obtain the fusion formula, as shown in equation (15).

[0057] (15) Finally, feature compression and spatial consistency restoration are achieved through 1×1 convolution and residual refinement layer, as shown in Equation (16).

[0058] (16) The SACM module is used at the feature fusion of various scales during the decoding stage. By adaptively adjusting the fusion ratio of low-level and high-level features, it not only preserves detailed texture information but also enhances deep semantic structure, thereby significantly improving the boundary accuracy of changing region detection and the performance of adapting to complex environments.

[0059] This application employs binary cross-entropy loss. With Dice loss The combined strategy leverages the complementary properties of the two loss functions, demonstrating synergistic effects in multi-class imbalanced sample processing scenarios. BCE accurately measures pixel-by-pixel classification error, which is beneficial for overall convergence; Dice directly measures the overlap between the predicted and ground truth regions, effectively suppressing bias caused by class imbalance and strengthening the learning of boundaries and small targets. The two are naturally complementary, thus improving the detection accuracy of changing edges. The total loss is: .

[0060] Binary cross-entropy loss (BCE) is used for pixel-level binary classification. It measures the difference between the model's predicted probability and the true label, improving the model's accurate discrimination between changed regions (positive class) and unchanged regions (negative class). Let the i-th... Pixel real label is The original logits output by the model are , The predicted probability is obtained after applying the Sigmoid function. The total number of pixels in the sample is Thus, we obtain equation (17).

[0061] (17) For dual-temporal remote sensing imagery, unchanged pixels typically far outnumber changed pixels, resulting in a significant class imbalance. If only BCE is used, the model tends to favor the unchanged class, leading to missed detections of changed regions (especially narrow boundaries and small targets). Therefore, this application introduces Soft Dice loss to directly optimize region overlap, alleviating the imbalance problem and improving boundary fit. Let... If the constant is a minimum smoothing constant (to prevent the denominator from being zero), then we get to equation (18).

[0062] , (18) In the above formula Represents pixels The true label, among which Indicates "changes have occurred". =0 means "no change"; For the model to pixels The predicted probability of belonging to "change"; The total number of pixels involved in the loss calculation.

[0063] The joint loss simultaneously constrains pixel-level classification consistency and region overlap consistency during the optimization process, effectively improving the detection capability for changing boundaries and small-area changes under imbalanced class conditions, and reducing overdetection and underdetection. Furthermore, if stronger imbalance exists on a specific dataset, robustness can be further enhanced by setting weights or introducing positive and negative sample weights into the BCE component.

[0064] Example 1 In order to objectively and comprehensively compare the performance of the network proposed in this application, this embodiment selects mainstream and advanced change detection methods in the field of change detection for comparison.

[0065] Table 1. Quantitative evaluation results of different methods on the LEVIR-CD and WHU-CD datasets:

[0066] On the two mainstream high-resolution remote sensing change detection datasets, LEVIR-CD and WHU-CD, the method of this invention performs excellently in the two core comprehensive indicators, F1-score and IOU.

[0067] On the LEVIR-CD dataset, our method achieved an F1 score of 91.97% and an IoU of 85.13%, both higher than all comparable models. On the WHU-CD dataset, our method performed exceptionally well in F1 (94.31%) and IoU (89.23%), again leading in IoU. Furthermore, our model maintained a leading position in recall (90.82%), only slightly behind CGNet's 90.96%. However, while CGNet strives for the highest recall, its F1 (91.89%) and IoU (85.00%) are both lower than our model. This clearly demonstrates that our method has a superior overall detection quality and localization accuracy, effectively balancing recall and overall precision.

[0068] This method significantly improves recall while maintaining high precision. This balance is achieved through the synergistic effect of the network structure: high recall is due to the multi-group attention mechanism introduced by the dynamic grouping attention module between CNN and Mamba, which achieves progressive fusion from local texture to global semantics, ensuring accurate identification of subtle changes in complex backgrounds and effectively reducing false negatives. Higher boundary clarity is achieved by the multi-scale edge enhancement module, which strengthens boundary features, effectively suppresses background noise interference, and improves the clarity and continuity of boundary changes. Spatial adaptive fusion dynamically generates fusion weights for each spatial location during the decoding stage, achieving fine-grained alignment and adaptive fusion of deep semantic features and shallow detail features. This mechanism enables the model to adaptively balance semantic and detail information according to different region features, significantly improving small object detection and boundary repair capabilities. In summary, the proposed model achieves significant improvements in change detection accuracy, boundary integrity, and robustness in complex scenes, fully validating the effectiveness and innovation of the proposed hybrid network structure.

[0069] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the above embodiments of the present invention, the descriptions of each embodiment have their own emphasis; parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. It should be understood that the disclosed technical content in the several embodiments provided in this application can be implemented in other ways.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A remote sensing image change detection method based on a CNN-Mamba hybrid network, characterized in that, The method comprises the following steps: Step 1, preprocessing the public remote sensing image change detection dataset, and constructing a dual-time high-resolution remote sensing image training set and a test set; The preprocessing comprises image registration, cropping, normalization and data enhancement; Step 2, constructing a change detection network model based on a CNN-Mamba backbone and training the model on the training set to obtain an optimal model; the network model comprises a twin encoder, a dynamic grouping attention module, a spatial adaptive fusion module, a multi-scale edge enhancement module and a decoder; Step 3, inputting dual-time remote sensing images of a region to be detected into the optimal model, and outputting a binary change detection map of the same size as the input images through feature extraction, feature interaction, cross-layer fusion and step-by-step decoding.

2. The remote sensing image change detection method based on the CNN-Mamba hybrid network according to claim 1, characterized in that, The multi-scale edge enhancement module, on the basis of the edge enhancement module, first extracts multi-scale information through a dilated convolution to expand the receptive field and then processes the information in the edge enhancement module; The multi-scale edge enhancement module inputs the features into a dilated convolution after standard convolution to generate three branches of different receptive fields: ； wherein, denote the multi-scale dilated convolution outputs corresponding to different dilation rates, respectively; denotes the feature map of the current decoding layer input; The branches of different receptive fields are input into independent edge enhancement modules for edge enhancement: ； wherein, represents a boundary feature processed by an edge enhancement module at the first scale; represents a boundary feature processed by an edge enhancement module at the second scale; represents a multi-scale path; The output is channel spliced and fused through a 1x1 convolution: ； wherein, denotes the fused multi-scale boundary enhanced feature; denotes the edge enhanced feature corresponding to the three scales respectively; The original input features form a global residual output: ； wherein, denotes the output of the multi-scale edge enhancement module.

3. The CNN-Mamba hybrid network-based remote sensing image change detection method according to claim 1, characterized in that, The dynamic grouping attention module improves the discrimination ability of complex change regions through adaptive channel grouping and multi-type attention mechanism; Let the input feature be The dynamic grouping attention module first performs global statistics to generate channel division weights : ； wherein, represents a global average pooling operation; represents an input feature; channel dimension according to weight The dynamic is divided into four subsets, respectively, into four attention branches, forming an adaptive multi-scale feature flow, the output of the four attention branches {Y1, Y2, Y3, Y4} is spliced in the channel dimension, and then fused by 1×1 convolution and nonlinear transformation: The final output features and the input features form a residual connection: ； wherein, represents a fused attention feature map; represents a fused intermediate feature.

4. The CNN-Mamba hybrid network-based remote sensing image change detection method according to claim 3, characterized in that, The four parallel branches comprise a dot product attention branch, a window attention branch, a region attention branch and a global attention branch.

5. The remote sensing image change detection method based on CNN-Mamba hybrid network according to claim 1, characterized in that, The spatial adaptive fusion module realizes fine-grained adaptive fusion between shallow detail features and deep semantic features; Let the input low-level features be , and the high-level features be ; the joint representation is obtained by concatenating the two groups of features in the channel dimension: ； Generating spatial weight maps by light-weight fusion network : ； The weight map is used to weight and fuse high-level and low-level features in the spatial dimension: ； When α approaches 1, the fusion calculation at this position... The coefficient is relatively large, and The coefficient is relatively small, and the system at this location is mainly composed of... The final weighted result is dominated by α, thus tending to retain shallow details; conversely, when α is close to 0, the fusion calculation at that location... The coefficient is relatively large, and The coefficient is relatively small, and the system at this location is mainly composed of... It dominates the final weighted result, and therefore pays more attention to the semantic consistency of the higher level; The input features are divided into multiple subgroups by channel, and each subgroup is processed by an independent SAC unit and a CBAM attention unit and then spliced and fused: ； Finally, the feature compression and spatial consistency recovery are completed through a 1x1 convolution and a residual refinement layer as shown in the formula: ； wherein, denotes the concatenation result of multiple groups of fusion features; denotes the final output feature of the spatial adaptive fusion module.

Images