Hyperspectral image change detection method based on frequency-space interaction and time domain difference correction

Abstract

The application discloses a hyperspectral image change detection method based on frequency-space interaction and time domain difference correction, and solves the technical problems of insufficient frequency-space feature cooperation and single multi-temporal feature fusion in the prior art. The method comprises the following steps: obtaining two-time-phase preliminary extraction features of preliminary feature extraction; performing spatial domain coding and frequency domain coding on the two-time-phase preliminary extraction features respectively, and obtaining frequency-space cooperative features of the two-time phases according to a frequency-space interaction attention mechanism; performing channel-by-channel addition and channel-by-channel difference on the two-time-phase preliminary extraction features, and obtaining difference features based on channel fusion attention; obtaining correction features based on single-phase attention according to one of the two-time-phase preliminary extraction features; adding the difference features and the correction features to obtain corrected time domain features; obtaining a difference feature map based on the two-time-phase frequency-space cooperative features and the corrected time domain features; and inputting the difference feature map into a multilayer perception machine for classification to obtain a detection result.

Description

Technical Field

[0001] This invention belongs to the field of image data processing technology, specifically, it relates to hyperspectral image change detection technology, and more specifically, it relates to a hyperspectral image change detection method based on frequency-space interaction and temporal differential correction. Background Technology

[0002] Remote sensing image change detection technology plays a crucial role in key areas such as military surveillance, marine resource utilization, land planning, and monitoring of changes in land features. To obtain more detailed information about land features, hyperspectral images (HSI) have become a key research focus in the field of change detection due to their unique technological advantages. Hyperspectral images typically contain hundreds of consecutive spectral bands, providing richer and more detailed spectral information than traditional multispectral images. This allows them to capture subtle differences in spectral features caused by changes in the material or state of land features, thus demonstrating a significant advantage in identifying and detecting minute changes.

[0003] To effectively process the high-dimensional and complex data of hyperspectral images, deep learning-based methods have become the mainstream research direction in this field. Although deep learning-based change detection methods have made some progress in certain areas, there are still two pressing technical challenges in terms of performance and robustness for practical applications: 1) Insufficient synergistic utilization of frequency and spatial domain features: When extracting frequency domain features, related methods often suffer from high computational complexity, easy loss of effective information, and susceptibility to noise interference. More importantly, while frequency and spatial domain features are highly complementary, there is currently a lack of technical mechanisms to efficiently enhance them synergistically, resulting in insufficient comprehensive perception capabilities of the model for complex scenes and difficulty in accurately identifying all changing regions; 2) Simple and limited effectiveness of multi-temporal feature fusion strategies: The essence of change detection lies in the effective comparison and fusion of features from different temporal phases. Commonly used methods often employ simple differencing of feature maps from two temporal phases. This approach is rather crude, making it difficult to explicitly model and enhance the characteristic differences caused by changes in real land features. At the same time, it cannot effectively distinguish and suppress spurious change signals caused by non-land feature factors such as changes in illumination and atmospheric conditions.

[0004] Therefore, how to provide a hyperspectral image change detection method based on more accurate feature fusion to maximize the highlighting of real change areas and suppress background noise interference is the current key technical bottleneck for improving change detection performance. Summary of the Invention

[0005] One of the objectives of this invention is to provide a hyperspectral image change detection method based on frequency-space interaction and time-domain differential correction, which aims to achieve deep complementary fusion of frequency domain and spatial domain information and dynamic and accurate calibration of time-domain differences, thereby effectively improving the accuracy and robustness of hyperspectral image change detection.

[0006] To achieve the above-mentioned objectives, the present invention employs the following technical solution:

[0007] A hyperspectral image change detection method based on frequency-space interaction includes:

[0008] Two temporal hyperspectral image data are acquired, and data preprocessing is performed on each to obtain the first temporal input data and the second temporal input data.

[0009] The first and second time-phase input data are input into the trained detection network model for the following detection process to obtain the detection results:

[0010] Preliminary feature extraction is performed on the first phase input data and the second phase input data to obtain preliminary extracted features for the first phase and preliminary extracted features for the second phase.

[0011] Spatial domain encoding and frequency domain encoding are performed on the first and second phase preliminary extracted features, respectively, to obtain first phase spatial domain features, second phase spatial domain features, first phase frequency domain features, and second phase frequency domain features. Based on the first phase spatial domain features, a value vector for the first phase frequency-space interaction attention is obtained; based on the first phase frequency domain features, a query vector and a key vector for the first phase frequency-space interaction attention are obtained; based on the value vector, query vector, and key vector of the first phase frequency-space interaction attention, a first phase frequency-space collaborative feature is obtained. Based on the second phase spatial domain features, a value vector for the second phase frequency-space interaction attention is obtained; based on the second phase frequency domain features, a query vector and a key vector for the second phase frequency-space interaction attention are obtained; based on the value vector, query vector, and key vector of the second phase frequency-space interaction attention, a second phase frequency-space collaborative feature is obtained.

[0012] The first and second phase preliminary extracted features are added channel by channel, and the query vector and key vector of the channel fusion attention are obtained based on the addition result. The first and second phase preliminary extracted features are then differencing channel by channel, and the value vector of the channel fusion attention is obtained based on the difference result. Differential features are obtained based on the query vector, key vector, and value vector of the channel fusion attention. The query vector, key vector, and value vector of the single-phase attention are obtained based on the first or second phase preliminary extracted features, and the correction features are obtained based on the query vector, key vector, and value vector of the single-phase attention. The differential features and the correction features are then added to obtain the corrected temporal domain features.

[0013] The first time-phase frequency-space coordination feature and the second time-phase frequency-space coordination feature are added to the corrected time-domain feature to obtain the first time-phase calibrated feature and the second time-phase calibrated feature; the first time-phase calibrated feature and the second time-phase calibrated feature are differentially processed to obtain a differential feature map;

[0014] The differential feature map is input into a multilayer perceptron for classification to obtain the detection result.

[0015] In some embodiments of this application, spatial domain encoding is performed on the first temporal preliminary extracted features and the second temporal preliminary extracted features to obtain the first temporal spatial domain features and the second temporal spatial domain features, including:

[0016] The features initially extracted in the first time phase are processed sequentially using 1×1 convolution and multi-scale convolution. The results of the multi-scale convolution are then concatenated and processed again using 1×1 convolution to obtain the first path processing result for the first time phase. Layer normalization is then performed on the features initially extracted in the first time phase. The processed results are then processed using a self-attention mechanism to obtain the query vector, key vector, and value vector of the self-attention for the first time phase. Based on the query vector, key vector, and value vector of the self-attention for the first time phase, the second path processing result for the first time phase is obtained. The first path processing result and the second path processing result for the first time phase are then concatenated to obtain the spatial domain features of the first time phase.

[0017] The features initially extracted in the second time phase are processed sequentially using 1×1 convolution and multi-scale convolution. The results of the multi-scale convolution are then concatenated and processed again using 1×1 convolution to obtain the first path processing result of the second time phase. Layer normalization is then performed on the features initially extracted in the second time phase. The processed results are then processed using a self-attention mechanism to obtain the query vector, key vector, and value vector of the self-attention in the second time phase. Based on the query vector, key vector, and value vector of the self-attention in the second time phase, the second path processing result of the second time phase is obtained. The first path processing result of the second time phase and the second path processing result of the second time phase are then concatenated to obtain the spatial domain features of the second time phase.

[0018] In some embodiments of this application, frequency domain encoding is performed on the first and second phase preliminary extracted features to obtain the first and second phase frequency domain features, including:

[0019] The features of the first time phase are initially extracted and then subjected to 3×3 convolution and two-dimensional fast Fourier transform to obtain the frequency domain results of the first time phase. The amplitude and phase of the first time phase are separated from the frequency domain results of the first time phase. The amplitude and phase of the first time phase are then subjected to convolution and two-dimensional inverse fast Fourier transform to obtain the frequency domain features of the first time phase.

[0020] The features of the second time phase are initially extracted and then subjected to 3×3 convolution and two-dimensional fast Fourier transform to obtain the frequency domain results of the second time phase. The amplitude and phase of the second time phase are separated from the frequency domain results of the second time phase. The amplitude and phase of the second time phase are then subjected to convolution and two-dimensional inverse fast Fourier transform to obtain the frequency domain features of the second time phase.

[0021] In some embodiments of this application, the differential features are obtained based on the query vector, key vector, and value vector of the channel fusion attention, including:

[0022] Based on the query vector and key vector of the channel fusion attention, the channel fusion attention matrix is ​​obtained;

[0023] The differential feature is obtained by multiplying the channel fusion attention matrix with the value vector of the channel fusion attention.

[0024] In some embodiments of this application, the correction features are obtained based on the query vector, key vector, and value vector of the uniphase attention, including:

[0025] Based on the query vector and key vector of the uniphase attention, obtain the uniphase attention matrix;

[0026] The standard attention matrix is ​​obtained by fusing the channel fusion attention matrix and the single-phase attention matrix through differential operations;

[0027] The correction feature is obtained by multiplying the standard attention matrix with the value vector of the uniphase attention.

[0028] In some embodiments of this application, preliminary feature extraction is performed on the first time-phase input data and the second time-phase input data to obtain preliminary extracted features of the first time-phase and preliminary extracted features of the second time-phase, including:

[0029] The first phase input data is sequentially processed by 3×3 convolution feature extraction, batch normalization, and ReLU activation function transformation to obtain the preliminary extracted features of the first phase.

[0030] The second phase input data is sequentially processed by 3×3 convolution feature extraction, batch normalization, and ReLU activation function transformation to obtain the preliminary extracted features of the second phase.

[0031] In some embodiments of this application, two temporal hyperspectral image data are acquired, and data preprocessing is performed on each to obtain first temporal input data and second temporal input data, including:

[0032] The two temporal hyperspectral image data are sequentially subjected to dimensionality reduction and standardization processing to obtain standardized data;

[0033] By sliding a sliding window over the standardized data, image patch samples are constructed to obtain the first phase input data and the second phase input data, respectively.

[0034] In some embodiments of this application, the method further includes:

[0035] When training the detection network model using the training set, the cross-entropy loss function is used as the optimization objective, and the Adam optimizer is used to iteratively optimize the model parameters.

[0036] In some embodiments of this application, the method further includes:

[0037] The test set data is input into the trained detection network model to obtain the test detection results, and the test detection results are then visualized.

[0038] The trained detection network model is validated using the set performance metrics and the test detection results.

[0039] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the above-described hyperspectral image change detection method is implemented.

[0040] Compared with the prior art, the advantages and positive effects of the present invention are:

[0041] This invention proposes a hyperspectral image change detection method based on frequency-space interaction and temporal differential correction. For multi-temporal hyperspectral image data, it utilizes spatial domain encoding to extract multi-scale local and global features from the spatial domain, and frequency domain encoding to extract structural and periodic features from the frequency domain. Through a frequency-space interaction attention mechanism, it obtains frequency-space collaborative features, achieving cross-domain feature enhancement. This fully leverages the complementarity of frequency-space features, improving the detection network model's ability to perceive subtle changes in complex scenes, while reducing the computational complexity of frequency domain feature extraction and overcoming the shortcomings of traditional methods in frequency-space collaboration. Furthermore, by setting up a multi-channel temporal differential correction structure with dual branches, it extracts differential features through a channel fusion attention mechanism. A single-phase attention mechanism extracts correction features, which are then dynamically fused to achieve temporal feature calibration, obtaining corrected temporal features. This avoids information loss caused by simple difference and can accurately capture real change features, suppressing spurious change interference introduced by illumination, noise, etc., significantly improving the accuracy and robustness of feature fusion across different time phases. Finally, the frequency-space co-factor features are fused with the corrected temporal features to form a differential feature map. A multilayer perceptron is used to classify the differential feature map, achieving accurate detection of hyperspectral image changes. This method is suitable for multi-source remote sensing image change detection scenarios such as military surveillance, agricultural monitoring, urban planning, and disaster assessment. It can meet the remote sensing data processing needs of complex ground feature changes and has significant advantages in improving detection accuracy and anti-interference capabilities.

[0042] Other features and advantages of the present invention will become clearer after reading the detailed embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0044] Figure 1 Flowcharts of some embodiments of the hyperspectral image change detection method provided by the present invention;

[0045] Figure 2 A schematic diagram of the overall architecture of the detection network model for some embodiments of the hyperspectral image change detection method provided by the present invention;

[0046] Figure 3 A schematic diagram of the architecture of the frequency-space interaction attention module in the detection network model of some embodiments of the hyperspectral image change detection method provided by the present invention;

[0047] Figure 4A schematic diagram of the architecture of a multi-channel temporal domain difference correction module in the detection network model of some embodiments of the hyperspectral image change detection method provided by the present invention;

[0048] Figure 5 This is a schematic diagram of the visualization results of changing the first image dataset using multiple methods, including the hyperspectral image change detection method provided in this invention.

[0049] Figure 6 This is a schematic diagram showing the visualization results of changing a second image dataset using multiple methods, including the hyperspectral image change detection method provided in this invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0051] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they can be implemented by those skilled in the art. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0052] To address the common technical problems in existing deep learning-based hyperspectral image change detection techniques, such as insufficient frequency-spatial feature coordination and a single multi-temporal feature fusion strategy, this invention proposes a hyperspectral image change detection method based on frequency-spatial interaction and temporal differential correction. By innovatively designing a frequency-spatial interaction attention module and a multi-channel temporal differential correction module, a highly efficient and collaborative network architecture is constructed. This fundamentally solves the key pain points of low frequency domain feature extraction efficiency and insufficient temporal feature fusion in existing methods, significantly improving the model's ability to perceive subtle changes in complex scenes and its anti-interference performance. The technical solution of this invention is applicable to various practical scenarios of hyperspectral image change detection, such as farmland monitoring, land use planning, ecological environment assessment, and disaster emergency response. It can handle both subtle change detection tasks between short temporal phases and complex change recognition needs over long temporal intervals.

[0053] Figure 1 The diagram shows flowcharts of some embodiments of the hyperspectral image change detection method provided by the present invention. Specifically, it shows flowcharts of some embodiments of the hyperspectral image change detection method based on frequency-space interaction and temporal difference correction.

[0054] like Figure 1 As shown, this embodiment uses the following process to achieve hyperspectral image change detection.

[0055] Step S11: Acquire two temporal hyperspectral image data, perform data preprocessing respectively, and obtain the first temporal input data and the second temporal input data.

[0056] Data preprocessing includes, but is not limited to, data dimensionality reduction and data standardization. The specific implementation of this step can adopt existing technologies, and this embodiment does not limit it.

[0057] The first and second phase input data are input into the trained detection network model to perform the detection process in steps S12-S15, and finally obtain the change detection result.

[0058] Step S12: Perform preliminary feature extraction on the first phase input data and the second phase input data to obtain the preliminary extracted features of the first phase and the preliminary extracted features of the second phase.

[0059] By performing preliminary feature extraction on the two-phase input data, the fundamental spectral-spatial features in the data are captured. This step can be implemented using existing techniques, and this embodiment does not limit its implementation.

[0060] Step S13: Perform frequency-space interaction processing based on the preliminary extracted features of the first time phase and the preliminary extracted features of the second time phase to obtain the frequency-space collaborative features of the two time phases; also perform time-domain differential correction processing based on the preliminary extracted features of the first time phase and the preliminary extracted features of the second time phase to obtain the corrected time-domain features.

[0061] This process is the core processing procedure of this embodiment.

[0062] Frequency-space interaction processing is performed to obtain two-phase frequency-space coordinated features, thereby extracting complementary features from the frequency and spatial domains and co-enhancing them. The specific method is as follows:

[0063] First, spatial domain encoding is performed on the preliminary extracted features of the first time phase and the preliminary extracted features of the second time phase respectively to obtain the spatial domain features of the first time phase and the spatial domain features of the second time phase. Then, frequency domain encoding is performed on the preliminary extracted features of the first time phase and the preliminary extracted features of the second time phase respectively to obtain the frequency domain features of the first time phase and the frequency domain features of the second time phase.

[0064] Then, the frequency-space interaction attention module is used to obtain frequency-space cooperative features. Based on the first time-phase spatial domain features, the value vector of the first time-phase frequency-space interaction attention is obtained; based on the first time-phase frequency domain features, the query vector and key vector of the first time-phase frequency-space interaction attention are obtained; then, based on the value vector, query vector, and key vector of the first time-phase frequency-space interaction attention, the frequency-space cooperative features of the first time-phase are obtained. Based on the second time-phase spatial domain features, the value vector of the second time-phase frequency-space interaction attention is obtained; based on the second time-phase frequency domain features, the query vector and key vector of the second time-phase frequency-space interaction attention are obtained; then, based on the value vector, query vector, and key vector of the second time-phase frequency-space interaction attention, the frequency-space cooperative features of the second time-phase are obtained.

[0065] Furthermore, temporal differential correction processing is performed to obtain corrected temporal features, and the expression of features in the changing region is enhanced through channel dimension difference modeling and dynamic calibration. In this embodiment, for the problem of dual-temporal feature fusion, differential features and correction features are extracted separately through a dual-branch structure to achieve dynamic calibration. The specific method is as follows:

[0066] Extracting Differential Features: In the differential feature extraction branch, the preliminary extracted features of the first and second time phases are added channel by channel to capture the common contextual information of the dual-time phase data. The preliminary extracted features of the first and second time phases are also differen channel by channel to quantify the absolute change magnitude between the two time phases. Then, the query vector and key vector of the channel fusion attention are obtained based on the addition result, and the value vector of the channel fusion attention is obtained based on the difference result. Finally, based on the attention mechanism, the differential features are obtained from the query vector, key vector, and value vector of the channel fusion attention.

[0067] Extracting correction features: In the correction feature extraction branch, the query vector, key vector, and value vector of single-phase attention are obtained based on the preliminary feature extraction in the first phase or the preliminary feature extraction in the second phase; then, the correction features are obtained based on the query vector, key vector, and value vector of single-phase attention.

[0068] Finally, the differential features extracted by the differential feature extraction branch and the correction features extracted by the correction feature extraction branch are added together to obtain the corrected time-domain features.

[0069] Step S14: Obtain the differential feature map based on the two-phase frequency-space coordination characteristics and the corrected time-domain characteristics.

[0070] Specifically, the first temporal-spatial coordinated feature is added to the corrected time-domain feature to obtain the first temporal-calibrated feature; the second temporal-spatial coordinated feature is added to the corrected time-domain feature to obtain the second temporal-calibrated feature. Then, the first and second temporal-calibrated features are differentially processed to obtain a differential feature map. The obtained differential feature map integrates the spatial-frequency coordinated feature extracted by the spatial-frequency interaction attention module and the corrected time-domain feature extracted by the multi-channel time-domain differential correction module. Utilizing the synergistic effect of the two modules, efficient feature extraction and accurate fusion of dual-temporal hyperspectral image data are achieved.

[0071] Step S15: Input the differential feature map into the multilayer perceptron for classification to obtain the detection result.

[0072] This step can be implemented using existing technologies, and this embodiment does not limit it.

[0073] The method described in the above embodiments for hyperspectral image change detection achieves accurate detection of hyperspectral image changes through frequency-space collaborative feature extraction and temporal differential correction of dual-temporal hyperspectral images. It is applicable to multi-source remote sensing image change detection scenarios such as military surveillance, agricultural monitoring, urban planning, and disaster assessment. It can handle the remote sensing data processing needs of complex ground feature changes and has significant advantages in improving detection accuracy and anti-interference capabilities. Furthermore, during frequency-space collaborative feature extraction, the query vector and key vector of frequency-space interactive attention are calculated using the global characteristics of frequency domain features. Attention allocation is guided based on the overall structural features and texture similarity of the scene. Meanwhile, the value vector of frequency-space interactive attention is obtained from the spatial domain features carrying the specific content of the image, and high-precision detail information is extracted and aggregated in a targeted manner. This mechanism overcomes the limitations of spatial distance, establishing connections between structurally similar regions globally while fully preserving high-fidelity information of spatial domain features, thereby generating fusion features with both strong robustness and high discriminative power. It can not only significantly enhance the feature response of changed areas in the image, but also effectively suppress the influence of irrelevant background noise and various interference information, ultimately greatly improving the model's sensitivity to small changes and overall detection performance in complex scenes.

[0074] In other embodiments, the temporal hyperspectral image data is preprocessed to obtain corresponding temporal input data, including:

[0075] The acquired temporal hyperspectral image data undergoes dimensionality reduction and standardization sequentially to obtain standardized data. In some embodiments, Principal Component Analysis (PCA) is used to reduce the dimensionality of the hyperspectral image data by calculating the covariance matrix and eigenvalue decomposition to select core principal components, thus preserving key information while reducing computational complexity. In some embodiments, the dimensionality-reduced data is standardized to convert it into a distribution with a mean of 0 and a variance of 1, eliminating dimensional differences and improving model stability.

[0076] Image patch samples are constructed by sliding a sliding window across the standardized data to obtain the corresponding temporal input data. In some embodiments, a sliding window of a specified size (typically 11 and an odd number) is used to slide across the standardized data to generate image patch samples corresponding to each location, ensuring that the samples contain complete spatial context information.

[0077] Figure 2 The diagram shows the overall architecture of the detection network model of some embodiments of the hyperspectral image change detection method provided by the present invention.

[0078] like Figure 2 As shown, the detection network model in this embodiment includes a preliminary feature extraction layer, a frequency-space interaction attention module, a multi-channel temporal difference correction module, and a multilayer perceptron (MLP) classifier.

[0079] After preprocessing the two acquired hyperspectral image data from different time phases, the first time phase input data is obtained. Second phase input data The two phase input data are respectively input into the preliminary feature extraction layer to perform preliminary feature extraction, obtaining the preliminary extracted features of the first phase and the preliminary extracted features of the second phase, which are labeled as follows. and In this embodiment, taking the first time phase as an example, the preliminary feature extraction quickly captures the basic spectral-spatial features of the highlight image through lightweight convolution operations. The specific extraction process includes: First, processing the input data of the first time phase... 3×3 convolutional feature extraction is performed, sliding over each spatial location of the input data to preserve local neighborhood information correlation while compressing redundant dimensions, effectively extracting spectral differences and spatial texture features between pixels. Then, batch normalization is applied to the convolutionally extracted features to adjust the distribution of features in each channel to a similar range, eliminating training oscillations caused by data scale differences and significantly improving model convergence speed. Finally, a nonlinear transformation using the ReLU activation function is applied to suppress invalid feature responses, enhance the expressive power of effective features, and mitigate the risk of gradient vanishing, making the initially extracted features more discriminative, thus obtaining the initial extracted features for the first phase. .

[0080] Subsequently, the preliminary extracted features of the first and second time phases are input into the frequency-space interaction attention module and the multi-channel time-domain difference correction module, respectively.

[0081] In some embodiments, the architecture diagram of the frequency-space interaction attention module is as follows: Figure 3 As shown. In this embodiment, the frequency-space interaction attention module includes a spatial domain encoder, a frequency domain encoder, and a frequency-space interaction attention mechanism. Figure 3 In the figure, (a) is a schematic diagram of the overall architecture of the frequency-space interaction attention module, (b) is the structure of the spatial domain encoder, and (c) is the structure of the frequency domain encoder.

[0082] In the frequency-space interaction attention module, the first phase initially extracts features. Preliminary feature extraction in the second phase The inputs are processed by the spatial domain encoder and the frequency domain encoder respectively. Finally, the first temporal-space-frequency collaborative features are obtained based on the frequency-space interactive attention mechanism. Second phase frequency-space co-operation characteristics The following section describes the preliminary feature extraction based on the first time phase. Obtain the first phase frequency-space cooperative features For example, combined with Figure 3 Describe the specific implementation process.

[0083] like Figure 3 As shown in Figure (b), the spatial domain encoder uses a dual-path feature extraction process, as follows:

[0084] Preliminary feature extraction for the first time phase The first path processing result is obtained by sequentially performing 1×1 convolution and multi-scale convolution. The result of multi-scale convolution is then concatenated and processed again with 1×1 convolution to obtain the first path processing result for the first time phase. Specifically, the 1×1 convolution first extracts features from the input, compresses the channel dimension, and aggregates core information, preserving key features while controlling computational load. The multi-scale convolution achieves parallel processing through convolution kernels of varying gradient sizes. 3×3 convolution kernels accurately capture fine-grained details such as ground object edges and textures, while large (2k+1)×(2k+1) series convolution kernels cover a wider range, uncovering macro- and mid-scale spatial relationships such as ground object distribution patterns and regional outlines, achieving comprehensive capture of features at different scales. The subsequent concatenation operation integrates features at each scale according to the channel dimension, preserving complementary information across multiple scales. Finally, the 1×1 convolution further fuses the concatenated features, eliminating redundant interference caused by scale differences and enhancing feature consistency.

[0085] Preliminary feature extraction for the first time phase The execution layer performs normalization processing, and the processed result is fed into three independent fully connected layers. Through linear transformation, the features are mapped to different vector spaces to obtain the query vector for the first phase of self-attention. Key vector Sum value vector ;Query vector based on first-phase self-attention and key vector Obtain the self-attention matrix :

[0086] ;in, This is the scaling factor.

[0087] Then, the self-attention matrix Sum value vector Multiply them to obtain the second path processing result of the first time phase.

[0088] Finally, the first path processing result and the second path processing result of the first time phase are concatenated and passed through a fully connected layer to obtain the spatial domain features of the first time phase, which are then labeled as follows. .

[0089] Preliminary feature extraction in the second phase Spatial domain encoding is performed using the same method described above to obtain the second temporal spatial domain features.

[0090] like Figure 3 As shown in Figure (c), the specific process of frequency domain encoding by the frequency domain encoder is as follows:

[0091] Preliminary feature extraction for the first time phase The frequency domain results of the first time phase are obtained by sequentially performing 3×3 convolution and two-dimensional fast Fourier transform. Then, the amplitude A and phase P of the first time phase are separated from the frequency domain results, and convolution is performed on the amplitude and phase to enhance the spectral feature representation. Finally, a two-dimensional inverse fast Fourier transform is performed on the convolutioned amplitude and phase to obtain the frequency domain features of the first time phase, labeled as follows. .

[0092] Preliminary feature extraction in the second phase Frequency domain coding is performed using the same method described above to obtain the second temporal spatial domain features.

[0093] like Figure 3 As shown in Figure (a), the frequency-space interactive attention mechanism incorporates the first-phase frequency domain features. The query vector for the first temporal-frequency-space interactive attention is generated through linear transformation. and key vector The first phase spatial domain features The value vector of the first temporal-frequency-space interactive attention is generated through linear transformation. By calculating the query vector and key vector The product of the transposes of the vectors is divided by a scaling factor, and the result is normalized using the softmax function to obtain the frequency-space interaction attention matrix between the frequency domain and spatial domain features. Finally, this attention matrix is ​​compared with the value vector. Weighted multiplication yields frequency-space collaborative features, completing cross-domain feature fusion.

[0094] In some embodiments, the architecture diagram of the multi-channel time-domain differential correction module is as follows: Figure 4 As shown. In this embodiment, the multi-channel time-domain differential correction module includes a differential feature extraction branch and a correction feature extraction branch, which utilizes the dual branches to accurately model cross-temporal feature differences.

[0095] like Figure 4 As shown, in the differential feature extraction branch, the features of the first time phase are initially extracted. Preliminary feature extraction in the second phase Channel-by-channel summation and channel-by-channel differencing are performed separately. Then, the query vector for channel fusion attention is obtained based on the summation results. and key vector The value vector of channel fusion attention is obtained based on the difference results. Then, based on the query vector of channel fusion attention... and key vector Obtain the channel fusion attention matrix : ;in, The scaling factor is then used to fuse the channels into an attention matrix. Value vector of attention fused with channels Multiply to obtain the difference features .

[0096] In the correction feature extraction branch, features are initially extracted based on the second time phase. Query vector for obtaining uniphase attention Key vector Sum value vector Then, based on the query vector of uniphase attention... and key vector Obtain the single-phase attention matrix : ;in, This is the scaling factor. Then, the attention matrix is ​​fused through a difference operation to fuse the channels. and single-phase attention matrix Remove non-changing information, highlight the true temporal variation features, and obtain the standard attention matrix. Then apply the standard attention matrix. Value vector of uniphase attention Multiply to obtain the correction feature .

[0097] In some other embodiments, the correction feature extraction branch may also employ preliminary feature extraction from the first time phase. The corrected features are obtained through processing.

[0098] Finally, the differential features extracted by the differential feature extraction branch are... Correction features extracted by the correction feature extraction branch Add them together to obtain the final corrected time-domain features. .

[0099] See further Figure 2 Overall architecture diagram, correcting time-domain features The features are fused with the first and second temporal-space co-location features via residual connections to obtain the first and second temporal-calibrated features. These two-phase calibrated features are first processed by a random deactivation module to reduce feature redundancy and prevent overfitting during training. They are then input into a feedforward neural network, where a combination of two fully connected layers and a nonlinear activation function enhances the discriminative power of the features. Simultaneously, residual connections integrate information from before and after the feedforward neural network, ensuring the integrity of the feature representation. After these processes, pixel-by-pixel absolute difference operations are performed on the optimized two-phase calibrated features to generate a difference feature map. Finally, this difference feature map is input into a multilayer perceptron, which consists of two fully connected layers. The hidden layer includes ReLU activation and random deactivation regularization, and the output layer is activated by Softmax to obtain the probability distribution of change and invariance classes for each pixel. Finally, a binarized pixel-level change detection result is output based on the maximum probability.

[0100] For detection network models, pre-training using a training set is required. In some embodiments, when training the detection network model using the training set, the cross-entropy loss function is used as the optimization objective. By measuring the difference between the model's predicted distribution of changed regions and the actual distribution of changed regions, the correct classification probability is maximized, guiding the update of model parameters. In other embodiments, the Adam optimizer is also used to iteratively optimize the model parameters during model training. To prevent overfitting, a weight decay term is set in the optimizer to impose regularization constraints on the model parameter complexity. The specific implementation process uses existing technology, which will not be described in detail in this embodiment.

[0101] In other embodiments, to enhance the intuitiveness and readability of the detection results, test set data is input into the trained detection network model to obtain test detection results, which are the change prediction results for each pixel. These test detection results are then visualized. Visualization not only presents the spatial distribution of the changed areas intuitively but also allows for comparison and analysis of the model's detection performance in complex scenes with real change maps, providing data support and decision-making basis for practical applications. Furthermore, performance metrics can be set and the test detection results used to verify the performance of the trained detection network model. These performance metrics include, but are not limited to, the number of missed pixels, the number of false positives, the overall accuracy, and the total number of errors.

[0102] The technical effectiveness of the hyperspectral image change detection method proposed in this invention will be further illustrated below with two simulation experiments.

[0103] Hyperspectral image data of a certain location captured in a relevant database was used as the first image dataset. This dataset contains hyperspectral images of land, trees, rivers, and farmland, with changes primarily concentrated in the soil and rivers. The hyperspectral image change detection method based on frequency-space interaction and temporal differential correction (hereinafter referred to as the Ours method) described in the above embodiment was applied to this first image dataset to output change detection results. Furthermore, existing methods such as CVA, SST-Former, CSANet, CAMixer, and HyGSTAN were also used to detect changes in this first image dataset, obtaining corresponding change detection results.

[0104] Figure 5 The diagram shows the actual changes in the first image dataset and the visualization results of the change detection results obtained by applying six methods to the first image dataset. Specifically: (a) The top image is the actual change map, and the bottom image is a magnified view of the boxed portion of the top image; (b) The top image is the change detection result obtained using the CVA method, and the bottom image is a magnified view of the boxed portion of the top image; (c) The top image is the change detection result obtained using the SST-Former method, and the bottom image is a magnified view of the boxed portion of the top image; (d) The top image is the change detection result obtained using the CSANet method, and the bottom image is a magnified view of the boxed portion of the top image; (e) The top image is the change detection result obtained using the CAMixer method, and the bottom image is a magnified view of the boxed portion of the top image; (f) The top image is the change detection result obtained using the HyGSTAN method, and the bottom image is a magnified view of the boxed portion of the top image; (g) The top image is the change detection result obtained using the Ours method, and the bottom image is a magnified view of the boxed portion of the top image.

[0105] from Figure 5As can be seen, on the first image dataset, the change detection results generated by the method of this invention have the highest degree of agreement between the changed areas and the real labels. It not only accurately delineates the boundaries between farmland and buildings and the exposed areas of river beaches, but also effectively suppresses background noise interference. The pixels inside the changed areas are evenly distributed, with no obvious false detections or missed detections. Even subtle change areas can be clearly identified, fully demonstrating the ability of the frequency-space interaction attention module to capture detailed features and the anti-interference advantages of the multi-channel temporal domain difference correction module.

[0106] In addition to visualizing the results, multiple performance metrics were used to analyze and compare the detection results of six different methods. The comparison results are shown in Table 1 below.

[0107] Table 1. Comparison of performance metrics of various methods on the first image dataset.

[0108]

[0109] In Table 1 above, FP represents the number of missed pixels, indicating the number of pixels in multi-temporal hyperspectral images that were misclassified as having changed but actually had not; FN represents the number of falsely detected pixels, indicating the number of pixels in multi-temporal hyperspectral images that actually changed but were not detected by the model; OE represents the total number of errors, indicating the total number of pixels misclassified; PCC represents the overall accuracy, measuring the proportion of correctly predicted pixels out of the total number of pixels; and KC represents the Kappa coefficient, used to measure the consistency between the model's generated results and the actual situation. As can be seen from Table 1, on the first image dataset, the method of this invention exhibits significant advantages in all performance metrics. Specifically, the PCC of the method of this invention reaches 98.38%, and the KC is as high as 96.09%, ranking first among all methods, which fully demonstrates that the classification results of the method of this invention have higher consistency and reliability with the real labels. The FN of this invention is only 585, which is the lowest value among all the methods compared, indicating that the method of this invention can accurately capture subtle changes in soil and river areas and avoid omission of change areas due to insufficient feature extraction. The FP of this invention is controlled at 433, and the OE is 1018, which is also at the optimal level, indicating that the method of this invention effectively filters out the spurious change interference caused by changes in illumination and atmospheric disturbances, verifies the strong suppression effect of the multi-channel temporal difference correction module on spurious change interference, and ensures the purity of the detection results.

[0110] Hyperspectral image data of an agricultural area taken in a relevant database was used as the second image dataset. Change detection was performed on this second image dataset using the CVA method, SST-Former method, CSANet method, CAMixer method, HyGSTAN method, and Ours method mentioned above, and the corresponding change detection results were obtained.

[0111] Figure 6 The diagram shows the actual changes in the second image dataset and visualizations of the change detection results obtained by applying six different methods to the dataset. Specifically: (a) The top image is the actual change map, and the bottom image is a magnified view of the area within the box in the top image; (b) The top image is the change detection result obtained using the CVA method, and the bottom image is a magnified view of the area within the box in the top image; (c) The top image is the change detection result obtained using the SST-Former method, and the bottom image is a magnified view of the area within the box in the top image; (d) The top image is the change detection result obtained using the CSANet method, and the bottom image is a magnified view of the area within the box in the top image; (e) The top image is the change detection result obtained using the CAMixer method, and the bottom image is a magnified view of the area within the box in the top image; (f) The top image is the change detection result obtained using the HyGSTAN method, and the bottom image is a magnified view of the area within the box in the top image; (g) The top image is the change detection result obtained using the Ours method, and the bottom image is a magnified view of the area within the box in the top image.

[0112] from Figure 6 As can be seen, on the second image dataset, the change detection results generated by the method of the present invention exhibit better accuracy and completeness. It not only accurately distinguishes between real changes caused by crop rotation and pseudo changes caused by differences in soil moisture, but also accurately locates small change areas such as irrigation canal branches and field ridge boundaries. The segmentation between change areas and non-change areas is clear, without over-expansion or omission. This directly demonstrates the strong adaptability and high detection accuracy of the method of the present invention in complex agricultural scenarios.

[0113] In addition to visualizing the results, six performance metrics—FP, FN, OE, PCC, and KC—were used to analyze and compare the detection results of the six different methods. The comparison results are shown in Table 2 below.

[0114] Table 2 Comparison of performance metrics of various methods on the second image dataset.

[0115]

[0116] As shown in Table 2, the method of this invention also demonstrates significant advantages in various performance indicators on the second image dataset. The PCC reaches 96.45%, and the KC reaches 89.66%, maintaining a leading position even in complex agricultural scenarios and exhibiting excellent generalization ability. The FN is 927, which can accurately identify subtle changes such as crop planting structure adjustment and land use type conversion. The OE is only 2627, which shows strong anti-interference ability and robustness in the processing of hyperspectral data with long time intervals. Even when faced with spectral information drift and complex background noise, it can effectively distinguish between real changes and pseudo changes, fully demonstrating the synergistic advantages of the frequency-space interaction attention module and the multi-channel temporal difference correction module.

[0117] Other embodiments of the present invention also provide a computer storage medium storing a computer program. When the computer program is executed by a processor, it implements the hyperspectral image change detection method based on frequency-space interaction and time-domain differential correction of the above embodiments, and achieves the technical effects of the corresponding embodiments.

[0118] The aforementioned computer storage media can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The computer storage media can be any available storage medium accessible to general-purpose or special-purpose computers.

[0119] In some embodiments, a computer storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in application-specific integrated circuits (ASICs). Of course, the processor and storage medium can also exist as discrete components in the device.

[0120] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0121] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions claimed by the present invention.

Claims

1. A method for detecting hyperspectral image changes based on frequency-space interaction and temporal difference correction, characterized in that, The method includes: Two temporal hyperspectral image data are acquired, and data preprocessing is performed on each to obtain the first temporal input data and the second temporal input data. The first and second time-phase input data are input into the trained detection network model for the following detection process to obtain the detection results: Preliminary feature extraction is performed on the first phase input data and the second phase input data to obtain preliminary extracted features for the first phase and preliminary extracted features for the second phase. Spatial domain encoding and frequency domain encoding are performed on the first and second phase preliminary extracted features, respectively, to obtain first phase spatial domain features, second phase spatial domain features, first phase frequency domain features, and second phase frequency domain features. Based on the first phase spatial domain features, a value vector for the first phase frequency-space interaction attention is obtained; based on the first phase frequency domain features, a query vector and a key vector for the first phase frequency-space interaction attention are obtained; based on the value vector, query vector, and key vector of the first phase frequency-space interaction attention, a first phase frequency-space collaborative feature is obtained. Based on the second phase spatial domain features, a value vector for the second phase frequency-space interaction attention is obtained; based on the second phase frequency domain features, a query vector and a key vector for the second phase frequency-space interaction attention are obtained; based on the value vector, query vector, and key vector of the second phase frequency-space interaction attention, a second phase frequency-space collaborative feature is obtained. The first and second phase preliminary extracted features are added channel by channel, and the query vector and key vector of the channel fusion attention are obtained based on the addition result. The first and second phase preliminary extracted features are then differencing channel by channel, and the value vector of the channel fusion attention is obtained based on the difference result. Differential features are obtained based on the query vector, key vector, and value vector of the channel fusion attention. The query vector, key vector, and value vector of the single-phase attention are obtained based on the first or second phase preliminary extracted features, and the correction features are obtained based on the query vector, key vector, and value vector of the single-phase attention. The differential features and the correction features are then added to obtain the corrected temporal domain features. The first time-phase frequency-space coordination feature and the second time-phase frequency-space coordination feature are added to the corrected time-domain feature to obtain the first time-phase calibrated feature and the second time-phase calibrated feature; the first time-phase calibrated feature and the second time-phase calibrated feature are differentially processed to obtain a differential feature map; The differential feature map is input into a multilayer perceptron for classification to obtain the detection result.

2. The hyperspectral image change detection method based on frequency-space interaction and temporal differential correction according to claim 1, characterized in that, Spatial domain encoding is performed on the preliminary extracted features of the first time phase and the preliminary extracted features of the second time phase to obtain the spatial domain features of the first time phase and the spatial domain features of the second time phase, including: The features initially extracted in the first time phase are processed sequentially using 1×1 convolution and multi-scale convolution. The results of the multi-scale convolution are then concatenated and processed again using 1×1 convolution to obtain the first path processing result for the first time phase. Layer normalization is then performed on the features initially extracted in the first time phase. The processed results are then processed using a self-attention mechanism to obtain the query vector, key vector, and value vector of the self-attention for the first time phase. Based on the query vector, key vector, and value vector of the self-attention for the first time phase, the second path processing result for the first time phase is obtained. The first path processing result and the second path processing result for the first time phase are then concatenated to obtain the spatial domain features of the first time phase. The features initially extracted in the second time phase are processed sequentially using 1×1 convolution and multi-scale convolution. The results of the multi-scale convolution are then concatenated and processed again using 1×1 convolution to obtain the first path processing result of the second time phase. Layer normalization is then performed on the features initially extracted in the second time phase. The processed results are then processed using a self-attention mechanism to obtain the query vector, key vector, and value vector of the self-attention in the second time phase. Based on the query vector, key vector, and value vector of the self-attention in the second time phase, the second path processing result of the second time phase is obtained. The first path processing result of the second time phase and the second path processing result of the second time phase are then concatenated to obtain the spatial domain features of the second time phase.

3. The hyperspectral image change detection method based on frequency-space interaction and temporal differential correction according to claim 1, characterized in that, Frequency domain encoding is performed on the first and second phase preliminary extracted features to obtain the first and second phase frequency domain features, including: The features of the first time phase are initially extracted and then subjected to 3×3 convolution and two-dimensional fast Fourier transform to obtain the frequency domain results of the first time phase. The amplitude and phase of the first time phase are separated from the frequency domain results of the first time phase. The amplitude and phase of the first time phase are then subjected to convolution and two-dimensional inverse fast Fourier transform to obtain the frequency domain features of the first time phase. The features of the second time phase are initially extracted and then subjected to 3×3 convolution and two-dimensional fast Fourier transform to obtain the frequency domain results of the second time phase. The amplitude and phase of the second time phase are separated from the frequency domain results of the second time phase. The amplitude and phase of the second time phase are then subjected to convolution and two-dimensional inverse fast Fourier transform to obtain the frequency domain features of the second time phase.

4. The hyperspectral image change detection method based on frequency-space interaction and temporal differential correction according to claim 1, characterized in that, The differential features are obtained based on the query vector, key vector, and value vector of the channel fusion attention, including: Based on the query vector and key vector of the channel fusion attention, the channel fusion attention matrix is ​​obtained; The differential feature is obtained by multiplying the channel fusion attention matrix with the value vector of the channel fusion attention.

5. The hyperspectral image change detection method based on frequency-space interaction and temporal differential correction according to claim 4, characterized in that, The correction features are obtained based on the query vector, key vector, and value vector of the uniphase attention, including: Based on the query vector and key vector of the uniphase attention, obtain the uniphase attention matrix; The standard attention matrix is ​​obtained by fusing the channel fusion attention matrix and the single-phase attention matrix through differential operations; The correction feature is obtained by multiplying the standard attention matrix with the value vector of the uniphase attention.

6. The hyperspectral image change detection method based on frequency-space interaction and temporal differential correction according to claim 1, characterized in that, Preliminary feature extraction is performed on the first time-phase input data and the second time-phase input data to obtain preliminary extracted features for the first time phase and preliminary extracted features for the second time phase, including: The first phase input data is sequentially processed by 3×3 convolution feature extraction, batch normalization, and ReLU activation function transformation to obtain the preliminary extracted features of the first phase. The second phase input data is sequentially processed by 3×3 convolution feature extraction, batch normalization, and ReLU activation function transformation to obtain the preliminary extracted features of the second phase.

7. The hyperspectral image change detection method based on frequency-space interaction and temporal differential correction according to claim 1, characterized in that, Two temporal hyperspectral image data sets were acquired, and data preprocessing was performed on each to obtain the first temporal input data and the second temporal input data, including: The two temporal hyperspectral image data are sequentially subjected to dimensionality reduction and standardization processing to obtain standardized data; By sliding a sliding window over the standardized data, image patch samples are constructed to obtain the first phase input data and the second phase input data, respectively.

8. The hyperspectral image change detection method based on frequency-space interaction and temporal differential correction according to any one of claims 1 to 7, characterized in that, The method further includes: When training the detection network model using the training set, the cross-entropy loss function is used as the optimization objective, and the Adam optimizer is used to iteratively optimize the model parameters.

9. The hyperspectral image change detection method based on frequency-space interaction and temporal differential correction according to any one of claims 1 to 7, characterized in that, The method further includes: The test set data is input into the trained detection network model to obtain the test detection results, and the test detection results are then visualized. The trained detection network model is validated using the set performance metrics and the test detection results.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the hyperspectral image change detection method based on frequency-space interaction and time-domain differential correction as described in any one of claims 1 to 9.

Images