Multispectral anomaly detection method and system based on physical constraint spectral decoupling gan

Abstract

This invention discloses a multispectral anomaly detection method and system based on a physically constrained spatial-spectral decoupling GAN, belonging to the field of spectral detection technology. The method includes: acquiring and preprocessing a multispectral image to obtain a standardized tensor block; inputting this block into a dual-stream generator to output an ideal background reconstruction image; inputting this reconstruction image and a real background image into a multi-scale block discriminator for adversarial discrimination, and sending the discrimination result into a model training and optimization module; calculating the gradient using a composite loss function integrating physical constraints, synchronously updating the parameters of the dual-stream generator and the multi-scale block discriminator to obtain a multi-scale dual-stream GAN model; inputting the multispectral image to be tested into the trained model to obtain the ideal background reconstruction image, calculating the anomaly score map between the reconstruction image and the original image, and outputting the anomaly detection binarized result after threshold segmentation. This application solves the problem of difficulty in simultaneously considering texture and spectral features in multispectral data, significantly improves the model's sensitivity to anomalous targets, and reduces the false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of spectral detection technology, specifically to a multispectral anomaly detection method and system based on a physically constrained spatial-spectral decoupling GAN. Background Technology

[0002] With the rapid development of Earth observation technology, multispectral remote sensing imagery has become the main data source for geological prospecting and environmental monitoring due to its low acquisition cost, wide coverage, and short revisit cycle. Anomaly detection, as one of the core tasks of remote sensing data analysis, aims to extract sparse targets from massive background data whose spectra or textures are significantly different from the surrounding environment. Currently, existing technologies in this field mainly include: First, traditional statistical methods, such as the classic RX operator (Reed-Xiaoli Detector) and its variants (KRX, LRX). These methods typically assume that background features follow a multivariate Gaussian distribution and determine anomalies by calculating Mahalanobis distance. Second, deep learning-based reconstruction methods. In recent years, generative adversarial networks (GANs) and autoencoders (AEs) have been widely used. For example, patent application number 2024102319208 discloses a hyperspectral image anomaly detection method based on generative adversarial networks, proposing to use OPTICS clustering to preprocess the data and combine it with the Transformer architecture to process hyperspectral data, and to mine spectral sequence features through a self-attention mechanism. In addition, in other fields of remote sensing image processing, there are also technologies based on the "dual-stream" architecture, such as the existing technology "Remote Sensing Image Colorization Method Based on Dual-Stream Generator Deep Convolutional Adversarial Generative Network". This method uses a dual-stream network at the original scale and the macro scale (downsampling) to solve the problem of inconsistent color space.

[0003] While existing technologies have made progress in specific scenarios, their direct application to multispectral anomaly detection suffers from the following significant drawbacks: redundant network architecture and low computational efficiency. The patent application number 2024102319208, primarily designed for data with hundreds of bands, employs computationally intensive architectures such as Transformer to process long spectral sequences and relies on time-consuming OPTICS clustering as a preprocessing step. For multispectral data with fewer bands (typically 10-13) but higher spatial resolution, this architecture not only results in a huge waste of computational power but also fails to meet the timeliness requirements for rapid screening of large-scale images. The imbalance between spatial and spectral feature mining is a significant issue. Existing CNN or GAN models often treat multispectral data as ordinary RGB image processing, neglecting the implicit nonlinear physical relationships between spectral bands (such as band ratio features). Alternatively, they may overemphasize spectral dimension reconstruction, leading to numerous false alarms in complex textured backgrounds (such as fragmented terrain or urban-rural fringe areas) due to blurred spatial texture reconstruction. Overfitting during reconstruction also causes detection failures: ordinary GAN models, during unsupervised training, often not only learn to reconstruct normal backgrounds but also, through strong fitting capabilities, "incidentally" reconstruct anomalous targets (i.e., overfitting). This results in insignificant differences in the reconstruction residuals between anomalous and background regions, leading to missed detections. The limitations of existing "dual-stream" colorization methods stem from their fundamental fusion of image scales (large and small images) rather than decoupling of feature attributes (spatial and spectral). Their "downsampling" operation leads to the loss of small anomalous targets, and the simple "additive fusion" strategy lacks physical constraints such as spectral angle (SAM), failing to guarantee the physical authenticity of ground object spectral curves. Therefore, they are unsuitable for quantitative geological anomaly detection tasks. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multispectral anomaly detection method and system based on physically constrained spatial-spectral decoupling GAN.

[0005] The objective of this invention is achieved through the following technical solution: Firstly, this application discloses a multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GANs, comprising the following steps: S1. Acquire the multispectral image to be tested and preprocess it to obtain a standardized tensor block, and construct a pure background training sample set; S2. Input the standardized tensor block obtained in step S1 into a dual-stream generator with spatial-spectral feature decoupling capability, and sequentially extract, fuse and reconstruct its features to output an ideal background reconstruction image. S3. Input the real background image and the ideal background reconstruction image obtained in step S2 into the multi-scale block discriminator to perform real vs. fake adversarial judgment, and input the judgment result into the model training and optimization module. S4. The model training and optimization module calculates the gradient of the entire network through a composite loss function that integrates physical constraints. It uses the Adam optimizer to synchronously update the network parameters of the dual-stream generator and the multi-scale block discriminator until the network parameters converge, thus obtaining the trained multi-scale dual-stream GAN model. S5. Input the multispectral image to be tested into the trained multi-scale dual-stream GAN model to obtain the ideal background reconstruction image of the image to be tested. Calculate the anomaly score map between the multispectral image to be tested and the ideal background reconstruction image. Perform threshold segmentation on the anomaly score map and output the anomaly detection binarization result.

[0006] Based on the first aspect, step S1 specifically includes the following sub-steps: S11. To address the issue of inconsistent reflectance dimensions across different bands in the multispectral image under test, a linear mapping method is used to uniformly scale it, i.e. ,in This represents the original reflectance of the multispectral image under test. This represents the minimum reflectance value of the corresponding band under the current image. This represents the maximum reflectance value of the corresponding band under the current image. This represents the normalized tensor mapped to the interval [−1, 1]. S12. Using a sliding window with size W × H and step size S, the linearly mapped multispectral image to be tested is cut into local feature blocks with size W × H × C, where C represents the total number of sensor bands, to construct a pure background training sample set for unsupervised learning.

[0007] Based on the first aspect, the dual-stream generator described in step S2 includes a parallel spatial feature stream, a spectral feature stream, and a spatiotemporal spectrum fusion module, specifically including the following steps: S21. The spatial feature flow is composed of four concatenated sets of convolutional units. Each set of convolutional units includes an instance normalization layer, a LeakyReLU activation function, and a 3×3 deep convolutional kernel. The spatial feature flow widens the receptive field through the stacking of 3×3 convolutional kernels and performs downsampling using a stride of 2 to extract the spatial flow features of the multispectral image under test. ; S22. The spectral feature flow employs a 1×1 convolutional layer combined with an improved channel attention module. While maintaining spatial position, the 1×1 convolutional layer reconstructs cross-band features. The channel attention module compresses spatial information through global pooling, extracts nonlinear physical constraints between bands via a two-layer fully connected network, generates channel weight vectors, and applies them to the feature map of the normalized tensor block to enhance the features of key spectral components, thereby obtaining the spectral feature flow. ; S23. The spatiotemporal spectrum fusion module, at the bottleneck layer of the decoder, integrates spatial flow features. With spectral characteristic flow Cascaded along the channel dimension to obtain high-dimensional features, these features are then subjected to deep dimensionality reduction and fusion via 1×1 convolutional layers. Finally, spatial resolution is restored through three sets of transposed convolutional layers with skip connections, outputting an ideal background reconstruction image. .

[0008] Based on the first aspect, the multi-scale block discriminator in step S3 includes a fine-scale path and a coarse-scale path, specifically including: the fine-scale path at a resolution of... The reconstruction block is used as input to determine the fidelity of local features, which include building edges and road curves. The coarse-scale path performs 2×2 average pooling dimensionality reduction on the input image and then determines the background coherence and spectral distribution pattern at a large spatial scale.

[0009] Based on the first aspect, step S4 specifically includes: guiding the model to approximate the real background in three dimensions—numerical, shape, and structure—through a composite loss function, wherein the composite loss function is... ,in, Indicating resistance to loss, The hyperparameter weights represent the pixel reconstruction loss. The weighting coefficients represent the hyperparameters of the spectral angular cosine loss. The weight coefficients of the hyperparameters representing the marginal gradient loss are... Represents pixel reconstruction loss, used to ensure ideal background reconstruction of the image. Pixels and real background images The difference between pixels is within a preset range, and its calculation formula is as follows: , Represents real background image Reconstructing images with an ideal background The mathematical expectation obtained from the joint data distribution. Describing the L1 norm, This represents the spectral angular cosine loss, used to ensure that the shape of the spectral curve is not distorted. Its calculation formula is: ,in Indicates the total number of pixels. Represents real background image spectral vector, Represents an image reconstructed from an ideal background. spectral vector, Represents the vector dot product. It is the L2 norm. i Indicates the pixel index; This represents the edge gradient loss, used to constrain the high-frequency spatial structure features of the generated image and prevent blurring of edge textures in the reconstructed image. Its calculation formula is: ,in Represents background image Sobel gradient operator, Represents an image reconstructed from an ideal background. The Sobel gradient operator.

[0010] Based on the first aspect, step S5 specifically includes: calculating the ideal background reconstruction image respectively. With real background images The intensity residual and the spectral angle residual are obtained through the formula The results are merged into anomaly score map A, where and Preset harmonic weighting coefficients; execute on outlier score graph A. Adaptive threshold segmentation identifies regions exceeding a preset threshold as abnormal and outputs a binarized target mask.

[0011] Secondly, this application discloses a multispectral anomaly detection system based on physically constrained spatial-spectral decoupling GANs, employing the aforementioned multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GANs, including: The data preprocessing module is used for quality cleaning, normalized range alignment, and high-dimensional tensor sliding window slicing of multispectral images. The multi-scale generative adversarial network module includes a two-stream generator with spatial-spectral feature decoupling capability and a multi-scale block discriminator; The model training optimization module integrates a composite loss function calculation engine with physical constraints, and drives network weight convergence through the backpropagation algorithm. The anomaly scoring and detection module is used to perform inference generation tasks, calculate anomaly score maps, and output the final detection results through adaptive segmentation.

[0012] The beneficial effects of this invention are: 1) The detection accuracy and robustness are significantly improved. This application explicitly decouples spatial and spectral features through a "dual-stream architecture", which not only preserves the rich texture details of multispectral images (reducing false alarms caused by terrain texture), but also makes full use of the physical correlation between bands (improving the detection rate of hidden spectral anomalies). At the same time, experiments show that on Sentinel-2 geological data, the detection AUC value of this method for weak alteration information is improved by more than 15% compared with the RX algorithm.

[0013] 2) The architecture is lightweight and computationally efficient. This application abandons the Transformer self-attention mechanism with quadratic growth in computational complexity and the time-consuming OPTICS clustering process. It adopts a fully convolutional network architecture, and the inference speed is 5-10 times faster than existing hyperspectral anomaly detection methods. It can meet the engineering processing needs of massive satellite data and support deployment on edge devices such as drones.

[0014] 3) High spectral fidelity and low false alarm rate: The spectral angle cosine loss (L_SAM) introduced in this application enables the model to strictly adhere to the spectral curve characteristics of ground objects during the reconstruction process, effectively avoiding color distortion caused by changes in lighting conditions or seasonal differences, thereby significantly reducing false alarms caused by environmental factors.

[0015] 4) Unsupervised learning and low application cost: This application does not require manual labeling of massive amounts of abnormal samples (which is often difficult to obtain in some fields), and only cheap and readily available normal background data is needed to complete the training, which greatly reduces the threshold for algorithm implementation and promotion. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the steps of the multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GAN in an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This application discloses a multispectral anomaly detection method and system based on a physically constrained spatial-spectral decoupling GAN. It proposes an end-to-end, lightweight, and high-precision anomaly detection framework that does not require pre-clustering. By designing a generator architecture with "spatial-spectral dual-stream decoupling" and a loss function with "spectral consistency constraints," it addresses the challenge of simultaneously considering texture and spectral features in multispectral data, significantly improving the model's sensitivity to anomalous targets and reducing the false alarm rate. A schematic diagram of the method's steps is shown below. Figure 1 As shown, the method includes: S1. Acquire the multispectral image to be tested and preprocess it to obtain a standardized tensor block, and construct a pure background training sample set; this step completes the morphological transformation from "raw reflectance" to "standardized tensor block"; S2. Input the standardized tensor block obtained in step S1 into a dual-stream generator with spatial-spectral feature decoupling capability, and sequentially extract, fuse and reconstruct its features to output an ideal background reconstruction image. S3. Input the real background image and the ideal background reconstruction image obtained in step S2 into the multi-scale block discriminator for real vs. fake adversarial judgment, and input the judgment result into the model training optimization module. S4. The model training and optimization module calculates the gradient of the entire network through a composite loss function that integrates physical constraints. It uses the Adam optimizer to synchronously update the network parameters of the dual-stream generator and the multi-scale block discriminator until the network parameters converge, thus obtaining the trained multi-scale dual-stream GAN model. S5. Input the multispectral image to be tested into the trained multi-scale dual-stream GAN model to obtain the ideal background reconstruction image of the image to be tested. Calculate the anomaly score map between the multispectral image to be tested and the ideal background reconstruction image. Perform threshold segmentation on the anomaly score map and output the anomaly detection binarization result.

[0019] For example, step S1 specifically includes the following sub-steps: S11. To address the issue of inconsistent reflectance dimensions across different bands in the multispectral image under test, a linear mapping method is used to uniformly scale it, i.e. ,in This represents the original reflectance of the multispectral image under test. This represents the minimum reflectance value of the corresponding band under the current image. This represents the maximum reflectance value of the corresponding band under the current image. This represents the normalized tensor mapped to the interval [−1, 1]. S12. Using a sliding window with size W × H and step size S, the linearly mapped multispectral image to be tested is cut into local feature blocks with size W × H × C, where C represents the total number of sensor bands (e.g., 13 bands for Sentinel-2). Since anomalous targets only account for a very small proportion in the whole image, most of these local feature blocks represent normal background, thus constructing a pure background training sample set for unsupervised learning.

[0020] For example, the dual-stream generator described in step S2 includes a parallel spatial feature stream, a parallel spectral feature stream, and a spatiotemporal spectral fusion module. Through two parallel feature streams, it explicitly decouples and extracts the spatial structure and spectral attributes of the image, specifically including the following steps: S21. The spatial feature flow is composed of four concatenated sets of convolutional units. Each set of convolutional units includes an instance normalization layer, a LeakyReLU activation function, and a 3×3 deep convolutional kernel. The spatial feature flow widens the receptive field through the stacking of 3×3 convolutional kernels and performs downsampling using a stride of 2 to extract the spatial flow features of the multispectral image under test. ; S22. The spectral feature flow employs a 1×1 convolutional layer combined with an improved channel attention module. While maintaining spatial position, the 1×1 convolutional layer reconstructs cross-band features. The channel attention module compresses spatial information through global pooling, extracts nonlinear physical constraints between bands (such as the spectral absorption peak characteristics of specific minerals) via a two-layer fully connected network, generates channel weight vectors, and applies them to the feature map of the normalized tensor block to enhance the features of key spectral components, thereby obtaining the spectral feature flow. ; S23. The spatiotemporal spectrum fusion module, at the bottleneck layer of the decoder, integrates spatial flow features. With spectral characteristic flow Cascaded along the channel dimension to obtain high-dimensional features, these features are then subjected to deep dimensionality reduction and fusion via 1×1 convolutional layers. Finally, spatial resolution is restored through three sets of transposed convolutional layers with skip connections, outputting an ideal background reconstruction image. .

[0021] For example, the multi-scale block discriminator in step S3 includes a fine-scale path and a coarse-scale path, specifically including: the fine-scale path at a resolution of The reconstruction block is used as input to determine the fidelity of local features, which include building edges and road curves, etc. The coarse-scale path performs 2×2 average pooling dimensionality reduction on the input image and then determines the background coherence and spectral distribution pattern at a large spatial scale.

[0022] For example, step S4 specifically includes: guiding the model to approximate the real background in three dimensions—numerical, shape, and structure—through a composite loss function, wherein the composite loss function is... This is used to comprehensively evaluate and guide the backpropagation and optimization of the two-stream generator network parameters, where... This represents the adversarial loss, used to measure the ability of the generator to deceive the discriminator, forcing the reconstructed image to approximate the real background image in the overall data distribution. The hyperparameter weights represent the pixel reconstruction loss. The weighting coefficients represent the hyperparameters of the spectral angular cosine loss. The hyperparameter weighting coefficients representing the edge gradient loss are used to dynamically adjust the relative importance of each sub-loss term in the total loss, in order to balance the ratio of spatial structure constraints to spectral morphology constraints. Represents pixel reconstruction loss, used to ensure ideal background reconstruction of the image. Pixels and real background images The difference between pixels is within a preset range, and its calculation formula is as follows: , Represents real background image Reconstructing images with an ideal background The mathematical expectation obtained from the joint data distribution (i.e., the average level of the computational error of all pairs of real and generated graphs in the training sample set (or the current training batch)). Describing the L1 norm, This represents the spectral angle cosine loss, used to ensure the shape of the spectral curve is not distorted. This term calculates the angle between the two in multidimensional space; the smaller the angle, the higher the material consistency. The calculation formula is as follows: ,in Indicates the total number of pixels. Represents real background image spectral vector, Represents an image reconstructed from an ideal background. spectral vector, Represents the vector dot product. It is the L2 norm. i Indicates the pixel index; This represents the edge gradient loss, used to constrain the high-frequency spatial structure features of the generated image and prevent blurring of edge textures in the reconstructed image. Its calculation formula is: The Sobel operator is used to extract edge gradient feature maps in the horizontal and vertical directions of the real and reconstructed images, respectively. The absolute difference (L1 distance) between these two sets of high-frequency feature maps is calculated, thereby forcing the generative network to retain realistic geometric texture and edge sharpness. Represents background image Sobel gradient operator, Represents an image reconstructed from an ideal background. The Sobel gradient operator is used to extract edge strength in the horizontal and vertical directions.

[0023] For example, through the strong intervention of spectral angle cosine loss, the model can perfectly model the spatial arrangement and spectral distribution of the "normal background" in the region. When the input image to be tested (which may contain anomalous targets) is given, the trained generator will force it to be "cleaned" and reconstructed into an ideal background image without anomalous targets. Step S5 specifically includes: calculating the ideal background reconstructed image respectively. With real background images The intensity residual and the spectral angle residual are obtained through the formula The results are merged into anomaly score map A, where and The pre-defined harmonic weighting coefficients are used, and the formula within parentheses represents a variation of cosine similarity, used to measure the degree of abrupt changes in spectral material; this is applied to anomaly score map A. Adaptive threshold segmentation identifies regions exceeding a preset threshold as abnormal and outputs a binarized target mask.

[0024] For example, this application discloses a multispectral anomaly detection system based on physically constrained spatial-spectral decoupling GANs. The multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GANs described above includes: The data preprocessing module is used for quality cleaning, normalized range alignment, and high-dimensional tensor sliding window slicing of multispectral images. The multi-scale generative adversarial network module includes a two-stream generator with spatial-spectral feature decoupling capability and a multi-scale block discriminator; The model training optimization module integrates a composite loss function calculation engine with physical constraints, and drives network weight convergence through the backpropagation algorithm. The anomaly scoring and detection module is used to perform inference generation tasks, calculate anomaly score maps, and output the final detection results through adaptive segmentation.

[0025] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GAN, characterized in that, Includes the following steps: S1. Acquire the multispectral image to be tested and preprocess it to obtain a standardized tensor block, and construct a pure background training sample set; S2. Input the standardized tensor block obtained in step S1 into a dual-stream generator with spatial-spectral feature decoupling capability, and sequentially extract, fuse and reconstruct its features to output an ideal background reconstruction image. S3. Input the real background image and the ideal background reconstruction image obtained in step S2 into the multi-scale block discriminator to perform real vs. fake adversarial judgment, and input the judgment result into the model training and optimization module. S4. The model training and optimization module calculates the gradient of the entire network through a composite loss function that integrates physical constraints. It uses the Adam optimizer to synchronously update the network parameters of the dual-stream generator and the multi-scale block discriminator until the network parameters converge, thus obtaining the trained multi-scale dual-stream GAN model. S5. Input the multispectral image to be tested into the trained multi-scale dual-stream GAN model to obtain the ideal background reconstruction image of the image to be tested. Calculate the anomaly score map between the multispectral image to be tested and the ideal background reconstruction image. Perform threshold segmentation on the anomaly score map and output the anomaly detection binarization result.

2. The multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GAN according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: S11. To address the issue of inconsistent reflectance dimensions across different bands in the multispectral image under test, a linear mapping method is used to uniformly scale it, i.e. ,in This represents the original reflectance of the multispectral image under test. This represents the minimum reflectance value of the corresponding band under the current image. This represents the maximum reflectance value of the corresponding band under the current image. This represents the normalized tensor mapped to the interval [−1, 1]. S12. Using a sliding window with size W × H and step size S, the linearly mapped multispectral image to be tested is cut into local feature blocks with size W × H × C, where C represents the total number of sensor bands, to construct a pure background training sample set for unsupervised learning.

3. The multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GAN according to claim 2, characterized in that, The dual-stream generator mentioned in step S2 includes a parallel spatial feature stream, a spectral feature stream, and a spatiotemporal spectrum fusion module, specifically including the following steps: S21. The spatial feature flow is composed of four concatenated sets of convolutional units. Each set of convolutional units includes an instance normalization layer, a LeakyReLU activation function, and a 3×3 deep convolutional kernel. The spatial feature flow widens the receptive field through the stacking of 3×3 convolutional kernels and performs downsampling using a stride of 2 to extract the spatial flow features of the multispectral image under test. ; S22. The spectral feature flow employs a 1×1 convolutional layer combined with an improved channel attention module. While maintaining spatial position, the 1×1 convolutional layer reconstructs cross-band features. The channel attention module compresses spatial information through global pooling, extracts nonlinear physical constraints between bands via a two-layer fully connected network, generates channel weight vectors, and applies them to the feature map of the normalized tensor block to enhance the features of key spectral components, thereby obtaining the spectral feature flow. ; S23. The spatiotemporal spectrum fusion module, at the bottleneck layer of the decoder, integrates spatial flow features. With spectral characteristic flow Cascaded along the channel dimension to obtain high-dimensional features, these features are then subjected to deep dimensionality reduction and fusion via 1×1 convolutional layers. Finally, spatial resolution is restored through three sets of transposed convolutional layers with skip connections, outputting an ideal background reconstruction image. .

4. The multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GAN according to claim 3, characterized in that, The multi-scale block discriminator mentioned in step S3 includes a fine-scale path and a coarse-scale path, specifically including: the fine-scale path at a resolution of The reconstructed block is used as input to determine the fidelity of local features, which include building edges and road curves. The coarse-scale path performs 2×2 average pooling dimensionality reduction on the input image and then determines the background coherence and spectral distribution pattern at a large spatial scale.

5. The multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GAN according to claim 4, characterized in that, Step S4 specifically includes: guiding the model to approximate the real background in three dimensions—numerical, shape, and structure—through a composite loss function, wherein the composite loss function is... ,in, Indicating resistance to loss, The hyperparameter weights represent the pixel reconstruction loss. The weighting coefficients represent the hyperparameters of the spectral angular cosine loss. The weight coefficients of the hyperparameters representing the marginal gradient loss are... Represents pixel reconstruction loss, used to ensure ideal background reconstruction of the image. Pixels and real background images The difference between pixels is within a preset range, and its calculation formula is as follows: , Represents real background image Reconstructing images with an ideal background The mathematical expectation obtained from the joint data distribution. Describing the L1 norm, This represents the spectral angular cosine loss, used to ensure that the shape of the spectral curve is not distorted. Its calculation formula is: ,in Indicates the total number of pixels. Represents real background image spectral vector, Represents an image reconstructed from an ideal background. spectral vector, Represents the vector dot product. It is the L2 norm. i Indicates the pixel index; This represents the edge gradient loss, used to constrain the high-frequency spatial structure features of the generated image and prevent blurring of edge textures in the reconstructed image. Its calculation formula is: ,in Represents background image Sobel gradient operator, Represents an image reconstructed from an ideal background. The Sobel gradient operator.

6. The multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GAN according to claim 5, characterized in that, Step S5 specifically includes: calculating the ideal background reconstruction image respectively. With real background images The intensity residual and the spectral angle residual are obtained through the formula The results are merged into anomaly score map A, where and Preset harmonic weighting coefficients; execute on outlier score graph A. Adaptive threshold segmentation identifies regions exceeding a preset threshold as abnormal and outputs a binarized target mask.

7. A multispectral anomaly detection system based on physically constrained spatial-spectral decoupling GANs, employing the multispectral anomaly detection method based on physically constrained spatial-spectral decoupling GANs as described in any one of claims 1-6, characterized in that, include: The data preprocessing module is used for quality cleaning, normalized range alignment, and high-dimensional tensor sliding window slicing of multispectral images. The multi-scale generative adversarial network module includes a two-stream generator with spatial-spectral feature decoupling capability and a multi-scale block discriminator; The model training optimization module integrates a composite loss function calculation engine with physical constraints, and drives network weight convergence through the backpropagation algorithm. The anomaly scoring and detection module is used to perform inference generation tasks, calculate anomaly score maps, and output the final detection results through adaptive segmentation.

Images