Hyperspectral image multi-class change detection method based on global-local joint modeling

Abstract

The application discloses a hyperspectral image multi-class change detection method based on global-local joint modeling, belongs to the technical field of remote sensing image processing, and solves the problems of calculation redundancy and receptive field limitation of a local learning method and insufficient modeling capability of a global learning method in the prior art.The method comprises the following steps: constructing a hyperspectral image change detection model; inputting a joint hyperspectral image to be detected into the hyperspectral image change detection model; obtaining change class prediction results of each pixel; and generating a hyperspectral image multi-class change detection map.The embodiment of the application constructs a hyperspectral image change detection model, utilizes the linear calculation complexity advantage of a Mamba model, efficiently establishes long-range dependence of spatial information on the whole image, realizes fast reasoning without dividing image blocks, enhances the capturing capability of the model for subtle change regions and edges, and significantly improves the precision and robustness of hyperspectral image multi-class change detection.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically relating to a method for detecting multiple types of changes in hyperspectral images based on global-local joint modeling. Background Technology

[0002] Hyperspectral image change detection, a crucial research area in remote sensing, plays a vital role in applications such as agricultural monitoring, ecological assessment, and disaster analysis. Traditional change detection methods, such as Change Vector Analysis (CVA), primarily rely on manually set thresholds, making them highly sensitive to image noise and limiting their generalization ability when dealing with complex changes in ground features. In recent years, with the development of deep learning technology, methods based on Convolutional Neural Networks (CNNs) and Transformers have demonstrated significant advantages in hyperspectral image feature extraction, automatically learning multi-level feature representations and improving the accuracy of change detection to some extent.

[0003] However, existing deep learning-based hyperspectral change detection methods still have the following limitations:

[0004] Local learning methods suffer from computational redundancy and receptive field limitations: Existing methods employ a patch-based input approach, dividing the entire image into multiple local regions for processing. This approach results in a large number of overlapping pixels between adjacent image patches, leading to redundant computation and increased time and storage overhead. Furthermore, due to the size limitations of the image patches, the model's receptive field is restricted, making it difficult to capture long-range dependencies across the entire image and affecting the recognition performance in large, continuously changing regions.

[0005] Global learning methods have insufficient modeling capabilities: To avoid block-based operations, some methods attempt to build global learning models based on the entire image. Among them, CNN-based global methods (patch-free) can improve inference speed, but due to the limitation of convolutional kernel locality, they need to stack multiple layers of networks to establish long-range dependencies, resulting in complex model structures, a large number of parameters, and deep networks are prone to gradient vanishing or information smoothing problems.

[0006] Limitations of Mamba Models in Preserving Local Details: In recent years, the state-space model (Mamba) has been introduced into visual tasks, attracting attention for its ability to model long-range dependencies of sequences with linear computational complexity. However, Mamba's one-dimensional scanning mechanism is prone to destroying local spatial structure when processing two-dimensional images, leading to the loss of details such as edges and textures, and its ability to detect subtle changes in hyperspectral images is relatively weak.

[0007] The feature fusion mechanism is not effective enough: Existing methods often lack an adaptive trade-off mechanism when fusing global semantic information and local detailed features, which can easily lead to information redundancy or feature conflict, and make it difficult to effectively suppress background interference while highlighting the changing areas.

[0008] In summary, to address the issues of computational redundancy and receptive field limitations in existing local learning methods, insufficient modeling capabilities in global learning methods, deficiencies in local detail preservation, and ineffective feature fusion mechanisms, we propose a hyperspectral image multi-class change detection method based on global-local joint modeling. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for detecting multiple categories of changes in hyperspectral images based on global-local joint modeling. This method solves the problems of computational redundancy and receptive field limitations in local learning methods, insufficient modeling capabilities in global learning methods, defects in local detail preservation, and ineffective feature fusion mechanisms in existing technologies.

[0010] This invention is implemented as follows: a multi-category change detection method for hyperspectral images based on global-local joint modeling, wherein the multi-category change detection method for hyperspectral images based on global-local joint modeling includes:

[0011] Step S1: Acquire dual-temporal hyperspectral image data of the same area, and perform radiometric correction, band selection and normalization preprocessing on the original dual-temporal hyperspectral image data to obtain standardized dual-temporal hyperspectral images.

[0012] Step S2: Align the standardized dual-temporal hyperspectral images in the spatial dimension, combine pixels at the same spatial location along the spectral dimension, and construct a joint hyperspectral image as input data for change detection.

[0013] Step S3: Based on the idea of ​​non-blocking global learning, a hyperspectral image change detection model is constructed for multi-category change detection in hyperspectral images. The hyperspectral image change detection model is based on a global-local patch-free MambaGL network architecture and adopts an encoder-decoder structure. The hyperspectral image change detection model is used to realize pixel-level feature representation and multi-scale feature modeling.

[0014] Step S4: Divide the joint hyperspectral image into training samples and test samples according to a preset ratio, and input the training samples into the constructed hyperspectral image change detection model for iterative training until the parameters of the hyperspectral image change detection model converge.

[0015] Step S5: Input the joint hyperspectral image to be detected into the trained hyperspectral image change detection model, obtain the change category prediction results of each pixel, and generate a multi-category change detection map of the hyperspectral image.

[0016] Preferably, in step S1, when performing radiometric correction, band selection and normalization preprocessing on the original dual-temporal hyperspectral image data, the spectral mean and standard deviation are calculated one by one with the band changes, so that the original dual-temporal hyperspectral image data to be processed conforms to the standard normal distribution.

[0017] Preferably, step S2 further includes: constructing a random hierarchical sample sequence using joint hyperspectral imagery;

[0018] The method for constructing a random hierarchical sample sequence using joint hyperspectral imagery includes: randomly selecting training samples from the entire joint hyperspectral imagery proportionally to form a training set, and then selecting labeled samples proportionally. Divided into different sample lists, forming A list of samples is a random stratified sample sequence.

[0019] Preferably, the method for constructing the hyperspectral image change detection model includes:

[0020] Step S3.1: In the hyperspectral image change detection model, a global feature modeling branch is introduced to perform global sequence modeling on the entire joint hyperspectral image in order to establish long-range dependencies between arbitrary spatial locations and realize the learning of global spatial context information.

[0021] Step S3.2: In the hyperspectral image change detection model, a local context feature extraction branch is introduced to dynamically aggregate pixel features within the local spatial range, so as to enhance the model's ability to express local change areas and detailed texture information.

[0022] Step S3.3: In the hyperspectral image change detection model, global and local features are enhanced and adaptively fused to highlight the features of the changed area, suppress the interference of the invariant background, and dynamically adjust the feature weights according to the image content to generate a fused feature representation that combines global context and local details.

[0023] Preferably, in step S3.1, the global feature modeling branch performs pixel-by-pixel segmentation on the input feature map and performs two-dimensional selective scanning along four specific directions: top left to bottom right, bottom right to top left, top right to bottom left, and bottom left to top right, to obtain a one-dimensional sequence. It is then input into a Mamba block for global information modeling. The calculation process is as follows:

[0024]

[0025]

[0026]

[0027]

[0028] Where Φ_chunk(·) represents the chunking operation; A, B, C, and D are the parameters of the discrete state-space model (SSM). Let be the hidden state at time t; and y represents the content branch and gating branch in the global Mamba block, respectively; z is the intermediate variable that serves as the intermediate output of the Mamba block, and y is the final output of the Mamba block. By dividing the input into the content stream and the gating stream, the Mamba block implements input-related selective state updates, enabling tokens with large amounts of information to be effectively propagated, while suppressing irrelevant or noisy information. This improves the flexibility and stability of modeling while maintaining linear computational complexity.

[0029] Preferably, in step S3.2, the local context feature extraction branch divides the joint hyperspectral image into non-overlapping local windows, generates cluster centers within each local window, calculates the cosine similarity matrix S between the data points and the cluster centers, and performs feature aggregation and updating based on the similarity. The formula for calculating the aggregated feature g is:

[0030]

[0031]

[0032] Where v represents the numerical representation of the cluster center, α and β are learnable scaling and translation parameters, σ(.) is the Sigmoid function, used to modulate the contribution of each point to the aggregation result according to similarity; the normalization factor T is used to suppress the influence of changes in cluster size, thereby stabilizing the aggregation features. Feature aggregation integrates semantically similar points into a consistent context representation, constructs robust cluster-level features, effectively suppresses noise, enhances local structural consistency, and provides an efficient context sharing mechanism for subsequent feature back-distribution.

[0033] The pixels are then updated using the following method:

[0034]

[0035] in To represent similarity, the same process as described above is used, employing a fully connected (FC) layer to transfer the feature dimension from the value space dimension. Project onto the original dimension c.

[0036] Preferably, in step S3.3, the enhancement and adaptive fusion of global and local features is achieved through a feature enhancement and adaptive fusion module. This is implemented using a feature adaptive enhancement fusion strategy, which first calculates feature differences to enhance texture details, and then uses an adaptive weighting mechanism to fuse global features. and local features The enhanced features are calculated as follows:

[0037]

[0038]

[0039] in, and These are the global and local original features, respectively. For coarse-grained fusion of global and local original features, GAP represents global average pooling, and δ is the Sigmoid function; the final fused features The calculation is as follows:

[0040]

[0041] in, and These are the global and local feature weights learned through the multilayer perceptron, respectively.

[0042] Preferably, in step S4, when training the hyperspectral image change detection model, the network training optimizer is stochastic gradient descent, the loss function is cross-entropy loss function, the batch size is set to 1, the training epochs are set to 600, the initial learning rate is set to 0.001, and the learning rate decays by a factor of 5 every 100 epochs.

[0043] Preferably, in step S5, obtaining the change category prediction results of each pixel includes: inputting the entire joint hyperspectral image into a trained hyperspectral image change detection model, wherein the hyperspectral image change detection model outputs a change detection result map with the same spatial size as the input image.

[0044] Compared with the prior art, the embodiments of this application have the following main advantages:

[0045] This invention overcomes the limitations of traditional patch-based methods in terms of local receptive field and computational redundancy in overlapping areas. It constructs a hyperspectral image change detection model based on a global-local patch-free MambaGL network architecture. Leveraging the linear computational complexity of the Mamba model, it efficiently establishes long-range spatial dependencies across the entire image, avoiding the high computational cost of traditional global methods and enabling rapid inference without the need for image patch division. Simultaneously, a local contextual feature extraction branch is introduced. By dynamically aggregating features within a local window, it effectively compensates for the neglect of local spatial structure and texture details by the global linear scanning mechanism, enhancing the model's ability to capture subtle changes and edges. Furthermore, by combining a feature adaptive enhancement fusion strategy, it fully mines and utilizes the global semantics and local details of the image, dynamically adjusting the fusion weights based on the image content. This promotes the complementary advantages of long-range dependencies and short-range details, significantly improving the accuracy and robustness of multi-class change detection in hyperspectral images.

[0046] The hyperspectral image change detection model based on the selective state-space model Mamba establishes long-range dependencies between pixels in the entire image through a two-dimensional cross-scanning mechanism, achieving global context modeling with linear computational complexity. The local context feature extraction branch performs dynamic feature clustering within a local window, enhancing the representation of neighborhood details. Furthermore, the enhancement and adaptive fusion of global and local features are achieved through a feature adaptive enhancement fusion strategy. By employing difference enhancement and weight adaptive mechanisms, global semantic and local texture features are fused to form complementary feature representations, making the prediction results of the hyperspectral image change detection model closer to the true results and exhibiting superior prediction performance. The overall framework of the hyperspectral image change detection model adopts an encoder-decoder structure, performing global scanning and local clustering in parallel at various scales. After multi-level fusion and up / downsampling, pixel-level change categories are output through a convolutional classifier head. The final method outperforms existing methods in overall accuracy (OA), average accuracy (AA), and Kappa coefficient on multiple hyperspectral change detection datasets, while also offering lower computational cost and faster inference speed. Furthermore, this invention is applicable to remote sensing applications such as agricultural monitoring, ecological assessment, and disaster analysis, providing an efficient and accurate change detection solution. Attached Figure Description

[0047] Figure 1 A schematic diagram comparing the feature interaction mechanism of the method of the present invention with that of the prior art is shown.

[0048] Figure 2 A diagram comparing the principles of Patch-free and Patch-Based learning frameworks is shown.

[0049] Figure 3 The diagram illustrates the implementation process of the hyperspectral image multi-category change detection method based on global-local joint modeling.

[0050] Figure 4 A schematic diagram of the architecture of the global feature modeling branch in an embodiment of the present invention is shown.

[0051] Figure 5 A schematic diagram of the architecture of the local context feature extraction branch in an embodiment of the present invention is shown.

[0052] Figure 6 A schematic diagram of the feature enhancement and adaptive fusion module in an embodiment of the present invention is shown.

[0053] Figure 7 A false-color composite image of dual-temporal hyperspectral images of the target region selected in an embodiment of the present invention is shown.

[0054] Figure 8 The diagram shows the hyperspectral image change detection results obtained by the method of this embodiment of the invention.

[0055] Figure 9 A ground reference map showing the changes in manually marked pixels in the target area according to an embodiment of the present invention is shown. Detailed Implementation

[0056] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0057] Regarding existing technologies, Figure 1 This diagram illustrates a comparison of the architecture of the feature interaction mechanism between the method of this invention (i.e., the Global-Local method) and existing technical methods. The existing technical methods include CNN models, Transformer models, and Global ViM models, such as... Figure 1As shown, existing image feature processing methods mainly employ three different approaches: CNN, Transformer, and Global ViM. The first, Convolutional Neural Network (CNN), conceptualizes the image as a structured feature grid, using convolutional layers to slide across local space with a fixed stride for feature extraction. However, its receptive field is limited by the kernel size, making it difficult to directly establish global dependencies. The second, Transformer, uses a self-attention mechanism to treat the image as discrete tokens, allowing each token to interact with all others. While achieving a global receptive field, its computational complexity increases quadratically, resulting in high computational costs when processing high-resolution images. The third, Global ViM, uses a cross-scanning module based on VMamba to integrate pixel information from different directions. It achieves a global receptive field with linear complexity through a selective state-space model, but its one-dimensional scanning mechanism has limited ability to preserve two-dimensional spatial structure. The method proposed in this invention combines the cross-scanning module with a contextual clustering layer, treating the image as a set of data points and dynamically grouping and clustering them within a local window. This preserves the ability to model global long-range dependencies while enhancing the extraction of local contextual information. Figure 2 The diagram illustrates a comparison of the principles of two learning frameworks: patch-free and patch-based. It visually demonstrates the core improvements of this invention compared to traditional methods. Figure 2 (a) represents a patch-based local learning framework, which divides an image into multiple patches (such as an S×S region) and then encodes each patch to learn local features. Figure 2 (b) represents the Patch-free global learning framework. This framework does not rely on image patch division, but directly encodes and decodes the entire image to learn global features. Figure 2 (c) represents Limitation 1: Overlapping pixels cause large computation, specifically, Limitation 1 based on patch-based processing: overlapping pixels lead to extensive computation. When processing images, overlapping pixels between image patches cause redundant computation, increasing the computational load; while Figure 2(d) represents Limitation 2: Limited patch size focuses on local information, i.e., Limitation 2 based on patch-based methods: limited patch size focuses on local information. Due to the limited size of the image patch, the model can only capture local information and is unable to obtain global information. In summary, existing local learning methods suffer from computational redundancy and receptive field limitations, while global learning methods have insufficient modeling capabilities, deficiencies in preserving local details, and ineffective feature fusion mechanisms. Specifically, to address these issues, we propose a hyperspectral image multi-class change detection method based on global-local joint modeling. In short, the method first performs radiometric correction, band selection, and normalization preprocessing on the original dual-temporal hyperspectral image data to construct a joint hyperspectral image as input data for change detection. A hyperspectral image change detection model for multi-class change detection is then constructed. The joint hyperspectral image to be detected is input into the trained hyperspectral image change detection model to obtain the change category prediction results for each pixel and generate a hyperspectral image multi-class change detection map. In this embodiment of the invention, the hyperspectral image change detection model based on the selective state-space model Mamba establishes long-range dependencies between pixels in the entire image through a two-dimensional cross-scanning mechanism, achieving global context modeling with linear computational complexity. The local context feature extraction branch performs dynamic feature clustering within a local window, enhancing the representation of neighborhood details. Furthermore, the enhancement and adaptive fusion of global and local features are achieved through a feature adaptive enhancement fusion strategy. Through difference enhancement and weight adaptive mechanisms, global semantic and local texture features are fused to ultimately form complementary feature representations, making the prediction results of the hyperspectral image change detection model closer to the true results and exhibiting superior prediction performance. The overall framework of the hyperspectral image change detection model adopts an encoder-decoder structure, performing global scanning and local clustering in parallel at various scales. After multi-level fusion and up / downsampling, pixel-level change categories are output through a convolutional classifier head. The final method outperforms existing methods in overall accuracy (OA), average accuracy (AA), and Kappa coefficient on multiple hyperspectral change detection datasets, while also having lower computational cost and faster inference speed. Furthermore, this invention is applicable to remote sensing applications such as agricultural monitoring, ecological assessment, and disaster analysis, providing an efficient and accurate change detection solution.

[0058] This invention provides a method for detecting multiple categories of changes in hyperspectral images based on global-local joint modeling. Figure 3 This diagram illustrates the implementation flow of the hyperspectral image multi-category change detection method based on global-local joint modeling. The method specifically includes:

[0059] Step S1: Acquire dual-temporal hyperspectral image data of the same area, and perform radiometric correction, band selection and normalization preprocessing on the original dual-temporal hyperspectral image data to obtain standardized dual-temporal hyperspectral images.

[0060] In step S1, when performing radiometric correction, band selection and normalization preprocessing on the original dual-temporal hyperspectral image data, the spectral mean and standard deviation are calculated one by one with the band changes to ensure that the original dual-temporal hyperspectral image data to be processed conforms to a standard normal distribution.

[0061] Step S2: Align the standardized dual-temporal hyperspectral images in the spatial dimension, combine pixels at the same spatial location along the spectral dimension, and construct a joint hyperspectral image as input data for change detection. This invention facilitates subsequent training and learning by calculating the spectral mean and the standard deviation of the sample data.

[0062] It should be noted that step S2 further includes: constructing a random stratified sample sequence using joint hyperspectral images;

[0063] The method for constructing a random hierarchical sample sequence using joint hyperspectral imagery includes: randomly selecting training samples from the entire joint hyperspectral imagery proportionally to form a training set, and then selecting labeled samples proportionally. Divided into different sample lists, forming The sample list is a random hierarchical sample sequence. This invention classifies the training set samples to form a random hierarchical sample sequence, which facilitates subsequent random training and learning, ensures that the subsequent training model can converge effectively, and increases the robustness of the global learning model training.

[0064] Step S3: Based on the idea of ​​non-blocking global learning, a hyperspectral image change detection model for multi-category change detection in hyperspectral images is constructed. The hyperspectral image change detection model is based on a global-local patch-free MambaGL network architecture and adopts an encoder-decoder structure. The hyperspectral image change detection model is used to realize pixel-level feature representation and multi-scale feature modeling. Furthermore, the hyperspectral image change detection model constructed in this invention based on a global-local patch-free MambaGL network architecture achieves collaborative learning of global spatial context information and local detail texture information by combining global feature modeling branch, local context feature extraction branch, and feature enhancement and adaptive fusion strategy.

[0065] Step S4: Divide the joint hyperspectral image into training samples and test samples according to a preset ratio, and input the training samples into the constructed hyperspectral image change detection model for iterative training until the parameters of the hyperspectral image change detection model converge.

[0066] In step S4, during the training of the hyperspectral image change detection model, the optimizer used is stochastic gradient descent, the loss function is cross-entropy loss, the batch size is set to 1, the training epochs are set to 600, and the initial learning rate is set to 0.001, decreasing by a factor of 5 every 100 epochs. This invention optimizes the training process using the cross-entropy loss function, making the results of the hyperspectral image change detection model closer to the true results, improving its predictive performance, and increasing the predictive power of the hyperspectral image change detection model.

[0067] Step S5: Input the joint hyperspectral image to be detected into the trained hyperspectral image change detection model, obtain the change category prediction results of each pixel, and generate a multi-category change detection map of the hyperspectral image. This invention completes hyperspectral image detection by inputting the entire joint hyperspectral image into the trained hyperspectral image change detection model, performing change detection, and outputting the change detection result map. The detection results are output and displayed to facilitate detailed comparison by staff.

[0068] In this embodiment of the invention, the hyperspectral image change detection model based on the selective state-space model Mamba establishes long-range dependencies between pixels in the entire image through a two-dimensional cross-scanning mechanism, achieving global context modeling with linear computational complexity. The local context feature extraction branch performs dynamic feature clustering within a local window, enhancing the representation of neighborhood details. Furthermore, the enhancement and adaptive fusion of global and local features are achieved through a feature adaptive enhancement fusion strategy. Through difference enhancement and weight adaptive mechanisms, global semantic and local texture features are fused to ultimately form complementary feature representations, making the prediction results of the hyperspectral image change detection model closer to the true results and exhibiting superior prediction performance. The overall framework of the hyperspectral image change detection model adopts an encoder-decoder structure, performing global scanning and local clustering in parallel at various scales. After multi-level fusion and up / downsampling, pixel-level change categories are output through a convolutional classifier head. The final method outperforms existing methods in overall accuracy (OA), average accuracy (AA), and Kappa coefficient on multiple hyperspectral change detection datasets, while also having lower computational cost and faster inference speed. Furthermore, this invention is applicable to remote sensing applications such as agricultural monitoring, ecological assessment, and disaster analysis, providing an efficient and accurate change detection solution.

[0069] In this embodiment, in step S5, obtaining the change category prediction results of each pixel includes: inputting the entire joint hyperspectral image into a trained hyperspectral image change detection model, wherein the hyperspectral image change detection model outputs a change detection result map with the same spatial size as the input image.

[0070] It should be noted that in existing technologies, the input to local learning architectures is image patches of limited size. The hyperspectral image change detection model of this invention breaks through the limitation of image patch size. By constructing a hyperspectral image change detection model based on the global information of the entire image, redundant calculation of overlapping pixels is avoided, and long-range spatial information dependencies of the entire image are established. By making full use of global receptive field information, rapid change detection of images without dividing the image into patches is achieved. In addition, this invention can also fully explore the information in the spectral, spatial, and temporal domains of the image. Combined with the global semantic information of the entire image, the influence of unchanged background areas on changed foreground areas is weakened, and the positions of changed pixels are strengthened, further improving the detection accuracy. According to the pixel-level detection characteristics of dual-temporal images, the global training samples of the entire image are transformed into random hierarchical sample sequences to obtain diverse stochastic gradients, ensuring effective convergence of the model and enhancing the robustness of the hyperspectral image change detection model training.

[0071] In a further preferred embodiment of the present invention, the method for constructing the hyperspectral image change detection model includes:

[0072] Step S3.1: A global feature modeling branch is introduced into the hyperspectral image change detection model to perform global sequence modeling on the entire joint hyperspectral image, thereby establishing long-range dependencies between arbitrary spatial locations and learning global spatial context information. The hyperspectral image change detection model adopts an encoder-decoder structure, where both the encoder network and the decoder network are unique, and together they constitute the fully convolutional network of the hyperspectral image change detection model. Figure 4 The diagram illustrates the architecture of the global feature modeling branch in an embodiment of the present invention. The present invention utilizes the global feature modeling branch, through four-directional two-dimensional selective scanning and Mamba block modeling, and employs an input-related selective state update mechanism to establish global long-range dependencies while maintaining linear computational complexity, thereby improving the model's flexibility and stability in modeling global information.

[0073] In step S3.1, the global feature modeling branch performs pixel-by-pixel segmentation on the input feature map and performs two-dimensional selective scanning along four specific directions: top left to bottom right, bottom right to top left, top right to bottom left, and bottom left to top right, to obtain a one-dimensional sequence. It is then input into a Mamba block for global information modeling. The calculation process is as follows:

[0074]

[0075]

[0076]

[0077]

[0078] Where Φ_chunk(·) represents the chunking operation; A, B, C, and D are the parameters of the discrete state-space model (SSM). Let be the hidden state at time t; and y represents the content branch and gating branch in the global Mamba block, respectively; z is the intermediate variable that serves as the intermediate output of the Mamba block, and y is the final output of the Mamba block. By dividing the input into the content stream and the gating stream, the Mamba block implements input-related selective state updates, enabling tokens with large amounts of information to be effectively propagated, while suppressing irrelevant or noisy information. This improves the flexibility and stability of modeling while maintaining linear computational complexity.

[0079] Step S3.2, introduce a local context feature extraction branch into the hyperspectral image change detection model, wherein, Figure 5 The diagram illustrates the architecture of the local context feature extraction branch in this embodiment of the invention. Pixel features are dynamically aggregated within a local spatial range to enhance the model's ability to express local variation regions and detailed texture information. This invention utilizes the local context feature extraction branch to generate cluster centers within a local window and perform feature aggregation and updates based on similarity, thereby constructing robust cluster-level features to suppress noise and enhance local structural consistency.

[0080] In step S3.2, the local context feature extraction branch divides the joint hyperspectral image into non-overlapping local windows, generates cluster centers within each local window, calculates the cosine similarity matrix S between the data points and the cluster centers, and performs feature aggregation and updating based on the similarity. The formula for calculating the aggregated feature g is as follows:

[0081]

[0082]

[0083] Where v represents the numerical representation of the cluster center, α and β are learnable scaling and translation parameters, σ(.) is the Sigmoid function, used to modulate the contribution of each point to the aggregation result according to similarity; the normalization factor T is used to suppress the influence of changes in cluster size, thereby stabilizing the aggregation features. Feature aggregation integrates semantically similar points into a consistent context representation, constructs robust cluster-level features, effectively suppresses noise, enhances local structural consistency, and provides an efficient context sharing mechanism for subsequent feature back-distribution.

[0084] The pixels are then updated using the following method:

[0085]

[0086] in To represent similarity, the same process as described above is used, employing a fully connected (FC) layer to transfer the feature dimension from the value space dimension. Project onto the original dimension c.

[0087] Step S3.3 involves enhancing and adaptively fusing global and local features in the hyperspectral image change detection model. This highlights features of changed areas, suppresses interference from invariant backgrounds, and dynamically adjusts feature weights based on image content to generate a fused feature representation that combines global context and local details. Specifically, the feature enhancement and adaptive fusion module employs an adaptive feature enhancement fusion strategy to calculate feature differences, enhance texture details, and dynamically adjusts the fusion of global and local features using an adaptive weight mechanism to generate complementary fused feature representations. This effectively highlights changed areas and suppresses background interference. Figure 6 A schematic diagram of the feature enhancement and adaptive fusion module in an embodiment of the present invention is shown.

[0088] In this embodiment, in step S3.3, the enhancement and adaptive fusion of global and local features is achieved through a feature enhancement and adaptive fusion module. The feature adaptive enhancement fusion strategy is introduced, which first calculates feature differences to enhance texture details, and then uses an adaptive weighting mechanism to fuse global features. and local features The enhanced features are calculated as follows:

[0089]

[0090]

[0091] in, and These are the global and local original features, respectively. For coarse-grained fusion of global and local original features, GAP represents global average pooling, and δ is the Sigmoid function; the final fused features The calculation is as follows:

[0092]

[0093] in, and These are the global and local feature weights learned through the multilayer perceptron, respectively.

[0094] Compared with existing technologies, the embodiments of this invention overcome the limitations of local receptive field and computational redundancy in overlapping areas of traditional patch-based methods. A hyperspectral image change detection model is constructed, based on a global-local patch-free MambaGL network architecture. Leveraging the linear computational complexity advantage of the Mamba model, long-range spatial dependencies are efficiently established across the entire image, avoiding the high computational cost of traditional global methods and enabling rapid inference without the need for image patch division. Simultaneously, a local contextual feature extraction branch is introduced, dynamically aggregating features within local windows to effectively compensate for the neglect of local spatial structure and texture details by the global linear scanning mechanism, enhancing the model's ability to capture subtle changes and edges. Furthermore, an adaptive feature enhancement fusion strategy is combined to fully mine and utilize the global semantics and local details of the image, dynamically adjusting the fusion weights based on the image content. This promotes the complementary advantages of long-range dependencies and short-range details, significantly improving the accuracy and robustness of multi-class change detection in hyperspectral images.

[0095] Performance Testing: To verify the actual performance of the method proposed in this invention, hyperspectral image data of the target region was selected for testing and verification. The testing process strictly followed the technical procedures proposed in this invention. First, dual-temporal hyperspectral image data of the same region was acquired. After preprocessing operations such as radiometric correction, band selection, and normalization, a standardized joint hyperspectral image was constructed as the model input. Figure 7 This shows a false-color composite image of dual-temporal hyperspectral images of the selected target area in an embodiment of the present invention. Figure 7 The overall distribution of ground features and spectral characteristics of the test area are presented intuitively. The preprocessed joint hyperspectral imagery is input into a pre-trained hyperspectral image change detection model. This model, based on a global-local joint modeling mechanism, establishes long-range dependencies between pixels through global scanning extraction units and enhances the ability to express detailed features by combining local context clustering units, ultimately outputting pixel-level change detection results. Figure 8 The image shows the hyperspectral image change detection results obtained by the method of this embodiment of the invention. Figure 8It is evident that the hyperspectral image change detection model can accurately identify the specific pixel locations where changes have occurred, and demonstrates good distinguishing ability for different types of change areas.

[0096] To further verify the detection accuracy, the output results of the hyperspectral image change detection model were compared with the manually labeled ground reference values. Figure 9 A comparative analysis was conducted, in which... Figure 9 A ground reference map of artificially marked changed pixels in the target area according to an embodiment of the present invention is shown. Quantitative evaluation reveals that the method of the present invention achieves excellent levels in key indicators such as overall accuracy (OA), average accuracy (AA), and Kappa coefficient, with significantly lower false negative and false positive rates for changed pixels compared to traditional methods. In particular, it demonstrates outstanding detection performance in edge regions and targets with subtle changes, proving the effectiveness of the global-local feature fusion strategy.

[0097] Experimental results show that the proposed hyperspectral image multi-category change detection method based on global-local joint modeling can not only achieve accurate identification of change areas, but also maintain high computational efficiency. The processing time of the entire image is significantly shorter than that of image patch-based methods, providing reliable technical support for the practical application of large-scale hyperspectral image change detection.

[0098] In summary, this invention provides a method for detecting multi-class changes in hyperspectral images based on global-local joint modeling. In this embodiment, the hyperspectral image change detection model based on the selective state-space model Mamba establishes long-range dependencies between pixels in the entire image through a two-dimensional cross-scanning mechanism, achieving global context modeling with linear computational complexity. Meanwhile, the local context feature extraction branch performs dynamic feature clustering within a local window, enhancing the representation of neighborhood details. Furthermore, the enhancement and adaptive fusion of global and local features are achieved through a feature adaptive enhancement fusion strategy. Through difference enhancement and weight adaptive mechanisms, global semantic and local texture features are fused, ultimately forming complementary feature representations. This makes the prediction results of the hyperspectral image change detection model closer to the true results, and its prediction performance is superior. The overall framework of the hyperspectral image change detection model adopts an encoder-decoder structure, performing global scanning and local clustering in parallel at various scales. After multi-level fusion and up / downsampling, pixel-level change categories are output through a convolutional classifier head. The final method outperforms existing methods in overall accuracy (OA), average accuracy (AA), and Kappa coefficient on multiple hyperspectral change detection datasets, while also offering lower computational costs and faster inference speed. Furthermore, this invention is applicable to remote sensing applications such as agricultural monitoring, ecological assessment, and disaster analysis, providing an efficient and accurate change detection solution.

[0099] This invention overcomes the limitations of traditional patch-based methods in terms of local receptive field and computational redundancy in overlapping areas. It constructs a hyperspectral image change detection model based on a global-local patch-free MambaGL network architecture. Leveraging the linear computational complexity of the Mamba model, it efficiently establishes long-range spatial dependencies across the entire image, avoiding the high computational cost of traditional global methods and enabling rapid inference without the need for image patch division. Simultaneously, a local contextual feature extraction branch is introduced. By dynamically aggregating features within a local window, it effectively compensates for the neglect of local spatial structure and texture details by the global linear scanning mechanism, enhancing the model's ability to capture subtle changes and edges. Furthermore, by combining a feature adaptive enhancement fusion strategy, it fully mines and utilizes the global semantics and local details of the image, dynamically adjusting the fusion weights based on the image content. This promotes the complementary advantages of long-range dependencies and short-range details, significantly improving the accuracy and robustness of multi-class change detection in hyperspectral images.

[0100] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0101] It should be understood that the disclosed apparatus can be implemented in other ways, given the several embodiments provided in this application. For example, the apparatus embodiments described above are merely illustrative; the division of units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or communication connections shown or discussed may be through some interfaces; the indirect coupling or communication connections between devices or units may be telecommunications or other forms.

[0102] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not depart from the concept of the present invention. These technical solutions are also within the scope of protection of the present invention.

Claims

1. A method for detecting multiple categories of changes in hyperspectral images based on global-local joint modeling, characterized in that, The method includes: Step S1: Acquire dual-temporal hyperspectral image data of the same area, and perform radiometric correction, band selection and normalization preprocessing on the original dual-temporal hyperspectral image data to obtain standardized dual-temporal hyperspectral images. Step S2: Align the standardized dual-temporal hyperspectral images in the spatial dimension, combine pixels at the same spatial location along the spectral dimension, and construct a joint hyperspectral image as input data for change detection. Step S3: Based on the idea of ​​non-blocking global learning, a hyperspectral image change detection model is constructed for multi-category change detection in hyperspectral images. The hyperspectral image change detection model is based on a global-local patch-free MambaGL network architecture and adopts an encoder-decoder structure. The hyperspectral image change detection model is used to realize pixel-level feature representation and multi-scale feature modeling. Step S4: Divide the joint hyperspectral image into training samples and test samples according to a preset ratio, and input the training samples into the constructed hyperspectral image change detection model for iterative training until the parameters of the hyperspectral image change detection model converge. Step S5: Input the joint hyperspectral image to be detected into the trained hyperspectral image change detection model, obtain the change category prediction results of each pixel, and generate a multi-category change detection map of the hyperspectral image.

2. The hyperspectral image multi-category change detection method based on global-local joint modeling as described in claim 1, characterized in that: In step S1, when performing radiometric correction, band selection and normalization preprocessing on the original dual-temporal hyperspectral image data, the spectral mean and standard deviation are calculated one by one with the band changes, so that the original dual-temporal hyperspectral image data to be processed conforms to the standard normal distribution.

3. The hyperspectral image multi-category change detection method based on global-local joint modeling as described in claim 1, characterized in that: Step S2 further includes: constructing a random hierarchical sample sequence using joint hyperspectral images; The method for constructing a random hierarchical sample sequence using joint hyperspectral imagery includes: randomly selecting training samples from the entire joint hyperspectral imagery proportionally to form a training set, and then selecting labeled samples proportionally. Divided into different sample lists, forming A list of samples is a random stratified sample sequence.

4. The hyperspectral image multi-category change detection method based on global-local joint modeling as described in claim 1, characterized in that: The method for constructing the hyperspectral image change detection model includes: Step S3.1: In the hyperspectral image change detection model, a global feature modeling branch is introduced to perform global sequence modeling on the entire joint hyperspectral image in order to establish long-range dependencies between arbitrary spatial locations and realize the learning of global spatial context information. Step S3.2: Introduce a local context feature extraction branch into the hyperspectral image change detection model to dynamically aggregate pixel features within the local spatial range; Step S3.3: In the hyperspectral image change detection model, global and local features are enhanced and adaptively fused to highlight the features of the changed area, suppress the interference of the invariant background, and dynamically adjust the feature weights according to the image content to generate a fused feature representation that combines global context and local details.

5. The hyperspectral image multi-category change detection method based on global-local joint modeling as described in claim 4, characterized in that: In step S3.1, the global feature modeling branch performs pixel-by-pixel segmentation on the input feature map and performs two-dimensional selective scanning along four specific directions: top left to bottom right, bottom right to top left, top right to bottom left, and bottom left to top right, to obtain a one-dimensional sequence. It is then input into a Mamba block for global information modeling. The calculation process is as follows: Where Φ_chunk(·) represents the block operation; A, B, C, and D are the parameters of the discrete state-space model (SSM). Let be the hidden state at time t; and y represents the content branch and gating branch in the global Mamba block, respectively; z is an intermediate variable used as the intermediate output of the Mamba block, and y is the final output of the Mamba block.

6. The hyperspectral image multi-category change detection method based on global-local joint modeling as described in claim 5, characterized in that: In step S3.2, the local context feature extraction branch divides the joint hyperspectral image into non-overlapping local windows, generates cluster centers within each local window, calculates the cosine similarity matrix S between the data points and the cluster centers, and performs feature aggregation and updating based on the similarity. The formula for calculating the aggregated feature g is as follows: Where v represents the numerical representation of the cluster center, α and β are learnable scaling and translation parameters, σ(.) is the Sigmoid function, and the normalization factor T is used to suppress the influence of changes in cluster size.

7. The hyperspectral image multi-category change detection method based on global-local joint modeling as described in claim 6, characterized in that: In step S3.3, the enhancement and adaptive fusion of global and local features is achieved through a feature enhancement and adaptive fusion module. This is implemented using a feature adaptive enhancement fusion strategy, which first calculates feature differences to enhance texture details, and then uses an adaptive weighting mechanism to fuse global features. and local features The enhanced features are calculated as follows: in, and These are the global and local original features, respectively. For coarse-grained fusion of global and local original features, GAP represents global average pooling, and δ is the Sigmoid function; the final fused features The calculation is as follows: in, and These are the global and local feature weights learned through the multilayer perceptron, respectively.

8. The hyperspectral image multi-category change detection method based on global-local joint modeling as described in claim 1, characterized in that: In step S4, when training the hyperspectral image change detection model, the network training optimizer is stochastic gradient descent, the loss function is cross-entropy loss function, the batch size is set to 1, the training epochs are set to 600, the initial learning rate is set to 0.001, and the learning rate decays by a factor of 5 every 100 epochs.

9. The hyperspectral image multi-category change detection method based on global-local joint modeling as described in claim 8, characterized in that: In step S5, obtaining the change category prediction results for each pixel includes: inputting the entire joint hyperspectral image into a trained hyperspectral image change detection model, wherein the hyperspectral image change detection model outputs a change detection result map with the same spatial size as the input image.

Images