Hyperspectral anomaly detection method based on deep self-representation learning framework

Abstract

The application provides a hyperspectral anomaly detection method based on a deep self-representation learning framework, and mainly solves the problem of poor detection effect of existing methods. The scheme comprises the following steps: 1) extracting multi-scale spatial features of a hyperspectral image; 2) constructing a subspace recovery autoencoder (SRAE) network model for each superpixel block, and designing a joint loss function for the model; 3) training the model by using an alternating optimization method based on a robust principal component analysis method, and obtaining a multi-scale reconstructed hyperspectral image; and 4) evaluating the abnormal score of the multi-scale reconstructed image by using an RX algorithm, obtaining initial detection results of different scales, and performing two-stage pixel-level addition and pixel-level multiplication operation fusion on the initial detection results, and obtaining detection results. The application integrates reliable prior knowledge and local spatial information into the autoencoder model, improves the representation ability of the background and weakens the representation ability of the anomaly, and effectively improves the precision of the hyperspectral image anomaly detection.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and machine learning technology, and further relates to hyperspectral anomaly detection, specifically a hyperspectral anomaly detection method based on a deep self-representation learning framework, which can be used for environmental monitoring, disaster prediction, and precision agriculture. Background Technology

[0002] Hyperspectral anomaly detection is a task in hyperspectral image processing, defined as capturing pixels with significant spectral differences from surrounding pixels in a hyperspectral image without prior knowledge. Its characteristics have been widely applied in military, agricultural, geological, and marine fields. In recent years, deep learning models have attracted the attention of researchers in the hyperspectral field, and their application in hyperspectral anomaly detection has achieved good detection accuracy.

[0003] After reviewing and analyzing literature such as [Xiang P, Ali S, Jung SK, et al. Hyperspectral anomaly detection with guided autoencoder[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-18.], it was found that while existing hyperspectral anomaly detection methods based on autoencoders (AEs) have improved detection performance to some extent, there are still areas that urgently need improvement:

[0004] 1) The training of the autoencoder is affected by the purity of the prior background samples. More importantly, background selection and AE training are performed separately, resulting in non-end-to-end training and the problem of local optima. Therefore, these factors can affect detection performance, especially in complex scenes.

[0005] 2) Some autoencoder models optimize the model by integrating prior information. However, during the training phase, prior knowledge sometimes cannot be continuously learned and updated. For example, the GAED model [Xiang P, Ali S, Jung SK, et al. Hyperspectral anomaly detection with guided autoencoder[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-18.] uses a guided map to constrain the learning of each pixel. Since the guided map is obtained through curvature filtering, it is fixed, which affects the robustness of the model. In addition, the inherent properties of low rank (background) and sparsity (anomalies and noise) of hyperspectral images are reliable prior knowledge. How to reasonably utilize this characteristic for autoencoder models requires further exploration.

[0006] 3) Autoencoder models in hyperspectral imaging typically utilize global samples for training, which is not very helpful in distinguishing between local and background anomalies, and may lead to better reconstruction of anomalies. This is because the degree of anomaly for each pixel is determined by its surrounding pixels, and global methods ignore the importance of local spatial information. On the other hand, when faced with complex terrain features, due to insufficient utilization of spatial information, some background pixels are easily misdetected as anomalies, leading to an increase in false alarms. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention proposes an anomaly detection method based on a deep self-representation learning framework. By integrating reliable prior knowledge and local spatial information into the autoencoder model, it effectively improves the ability to represent the background and weakens the ability to represent anomalies, thereby further optimizing the accuracy of anomaly detection in hyperspectral images.

[0008] The approach to this problem is as follows: First, multi-scale spatial feature extraction is performed. This involves converting the 3D hyperspectral image into a 2D image and using principal component analysis (PCA) to obtain the first principal component features. A simple linear iterative clustering algorithm with multiple different numbers of superpixels is then used to obtain multi-scale superpixel segmentation features. This spatial information is indexed into the 3D hyperspectral image, thus obtaining multi-scale spatial features. Next, a subspace restoration autoencoder network is constructed, trained, and tested. The subspace restoration layer is introduced into the autoencoder network, and a joint loss function is designed for the model. Alternating optimization based on robust PCA is used for training, and the trained network is tested to obtain reconstructed hyperspectral images at multiple scales. Finally, the multi-scale reconstruction results are evaluated using the near-anomaly score of the RX algorithm to obtain initial detection results at different scales. These results are then fused using a two-stage pixel-level addition and pixel-level multiplication operation to obtain the final detection result.

[0009] To achieve the above objectives, the technical solution of the present invention is implemented through the following steps:

[0010] (1) Extract the multi-scale spatial features of the hyperspectral image X to obtain K superpixel blocks at W scales; where W is a positive integer greater than or equal to 2;

[0011] (2) Construct and train the subspace restoration autoencoder network SRAE:

[0012] (2.1) For each superpixel block, construct a subspace restoration autoencoder (SRAE) network model consisting of an encoder, a subspace restoration layer, and a decoder:

[0013]

[0014]

[0015]

[0016] Where ω = 1, 2, ..., W; k = 1, 2, ..., K; The features of the k-th superpixel block at the ω-scale extracted by the encoder. For the subspace restoration layer of the k-th superpixel block at the ω scale Acquired features; and Represent the k-th superpixel block at scale ω. The encoder and decoder; This indicates that the code finally passes through the decoder. Acquired features;

[0017] (2.2) Calculate the joint loss function L of the SRAE network model according to the following formula.total :

[0018]

[0019] Among them, L Rec (·) represents the reconstruction loss; L SRL (·) represents the subspace repair loss;

[0020] (2.3) Based on the robust principal component analysis method, an alternating optimization approach is adopted, utilizing the joint loss function L. total The model parameters for each pixel block are iteratively optimized, and the corresponding detection model is obtained through training.

[0021] (3) Each pixel is input into the detection model corresponding to that pixel block. The test was conducted to generate multi-scale reconstruction results.

[0022] (4) Introduce the RX algorithm based on Mahalanobis distance to process the multi-scale reconstruction results. Anomaly score evaluation is performed, and the evaluation results are obtained as R = {R1, R2, ..., R...} ω , ..., R W Then, the results are fused in two stages, as follows:

[0023]

[0024] in, and These represent pixel-level addition and pixel-level multiplication, respectively; Z add and Z mul These are the pixel-level addition fusion result and the pixel-level multiplication fusion result of the first stage, respectively, N. max-min This indicates max-min normalization, O final This is the result of the second-stage fusion, i.e., the final detection result.

[0025] Compared with the prior art, the advantages of this invention are:

[0026] First, the hyperspectral anomaly detection method proposed in this invention is based on a deep self-representation learning framework, which constructs an end-to-end model and adopts a novel alternating optimization strategy to optimize the background representation by adaptively removing abnormal pixels and purifying background pixels.

[0027] Second, since the present invention introduces a subspace restoration layer in the autoencoder network, it helps to bring background pixels closer to their initial positions and move abnormal pixels away from their initial positions, thereby enhancing the reconstruction of the background while weakening the reconstruction of abnormal pixels. In order to effectively constrain the autoencoder network of subspace restoration, a joint loss function including reconstruction loss and subspace restoration layer loss is introduced.

[0028] Third, considering the utilization of spatial information, this invention attempts a training method different from existing hyperspectral anomaly detection methods. The autoencoder network for each subspace reconstruction is trained in a local homogeneous region, and the reconstruction of the entire hyperspectral image is achieved by using autoencoder networks for multiple subspace reconstructions. This method can improve the network's sensitivity to local anomalies. In addition, the proposed multi-scale fusion based on different spatial features further highlights anomalies and suppresses the background, thereby optimizing the detection effect. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the overall implementation of the present invention.

[0030] Figure 2 This is a schematic diagram illustrating the application process of hyperspectral image detection using the method of the present invention.

[0031] Figure 3 This is a comparison chart of the detection effects of the present invention and existing methods. Detailed Implementation

[0032] To make the objectives and advantages of the present invention clearer, the technical content of the present invention will be described in detail below with reference to the accompanying drawings.

[0033] Example 1: Refer to Figure 1 and Figure 2 The hyperspectral anomaly detection method based on a deep self-representation learning framework proposed in this invention specifically includes the following steps:

[0034] Step 1. Extract multi-scale spatial features from the hyperspectral image X to obtain K superpixel blocks at W scales; where W is a positive integer greater than or equal to 2; the implementation steps are as follows:

[0035] (1.1) Transform the three-dimensional hyperspectral image Transform into a two-dimensional matrix Where H, W and B represent the height, width and band of the hyperspectral image, respectively, and N represents the number of pixels, N = H × W;

[0036] (1.2) Principal Component Analysis (PCA) is used to obtain the first principal component features. Then, the Simple Linear Iterative Clustering (SLIC) algorithm is used to perform superpixel segmentation on the first principal component features to obtain K superpixel blocks {X1, X2, ..., X...}. k , ..., XK},in Represents the k-th superpixel block; n k The number of pixels in the k-th superpixel block;

[0037] (1.3) Multi-scale superpixel segmentation is adopted, and the k-th superpixel block at the ω-scale is represented as... in Indicates the number of superpixels at the ω scale;

[0038] (1.4) Take ω = 2, 3, ..., W to obtain all superpixel blocks at each scale, i.e., the multi-scale spatial features of the hyperspectral image.

[0039] Step 2. Construct and train the subspace restoration autoencoder network SRAE:

[0040] (2.1) For each superpixel block, construct a subspace restoration autoencoder (SRAE) network model consisting of an encoder, a subspace restoration layer, and a decoder:

[0041]

[0042]

[0043]

[0044] Where ω = 1, 2, ..., W; k = 1, 2, ..., K; The features of the k-th superpixel block at the ω-scale extracted by the encoder. For the subspace restoration layer of the k-th superpixel block at the ω scale Acquired features; and Represent the k-th superpixel block at scale ω. The encoder and decoder; This indicates that the code finally passes through the decoder. Acquired features;

[0045] (2.2) Calculate the joint loss function L of the SRAE network model according to the following formula. total :

[0046]

[0047] Among them, L Rec (·) represents the reconstruction loss; L SRL (·) represents the subspace repair loss; α and β are hyperparameters, set to 1; x i I represents a pixel; t It is a 60×60 identity matrix; z represents The feature of a certain pixel, x i For a single pixel in a hyperspectral image, x i T x represents i transpose, This indicates that the reconstructed features are obtained after the subspace-based reconstructive autoencoder.

[0048] (2.3) Based on the robust principal component analysis method, an alternating optimization approach is adopted, utilizing the joint loss function L. total The model parameters for each pixel block are iteratively optimized, and the corresponding detection model is obtained through training. This step yields a total of W×K trained SRAE models.

[0049] The method employs robust principal component analysis for alternating optimization. Specifically, an image is divided into two parts: a background part L and an abnormal noise part S. Then, the two parts L and S are alternately optimized according to the following formula:

[0050]

[0051] Among them, ||·|| 2,1 L represents 2,1 Norm; λ is the harmonic coefficient; E k A k D k L k These represent the encoder, subspace restoration layer, decoder, and background portion of the k-th superpixel block, respectively.

[0052] The optimization process is terminated under the following conditions:

[0053] a) The preset number of iterations has been reached;

[0054] b) One of the following conditions must be met:

[0055]

[0056] Where c1 and c2 are evaluation parameters for testing convergence; ε is the convergence parameter.

[0057] The process of alternately optimizing L and S is as follows:

[0058] (2.3.1) Optimizing L with a fixed S: First, through The background portion of the hyperspectral image is obtained, and the background portion at the ω-scale is indexed using the spatial information from superpixel segmentation. The background portion of each pixel block at the ω scale is input into its detection model to obtain the optimized background portion at the ω scale.

[0059] (2.3.2) Optimize S with fixed L: using Generate anomalous noise components in hyperspectral images; via L 2,1 The near-end method is used for S-optimization to obtain the optimized abnormal noise component. The L 2,1 The proximal method is to use L 2,1 Norm conversion to block-by-block soft thresholding function:

[0060]

[0061] Where i and j represent the i-th row and j-th column of the hyperspectral space, S i,j S represents the i-th row and j-th column of the abnormal noise part. i,: This represents the entire i-th row in the abnormal noise section.

[0062] (2.3.3) Repeat steps (3.1) and (3.2) above until the termination condition is met, to obtain the reconstruction result at the ω scale, i.e.

[0063] (2.3.4) For hyperspectral images containing different spatial information at multiple scales, following steps (2.3.1)-(2.3.3) above, the multi-scale reconstruction result is obtained, i.e.

[0064] Step 3. Each pixel is input into the detection model corresponding to that pixel block. The test was conducted to generate multi-scale reconstruction results.

[0065] Step 4. Introduce the RX algorithm based on Mahalanobis distance to process the multi-scale reconstruction results. Anomaly score evaluation is performed, and the evaluation results are obtained as R = {R1, R2, ..., R...} ω , ..., R W Then, the results are fused in two stages, as follows:

[0066]

[0067] in, and These represent pixel-level addition and pixel-level multiplication, respectively; Z add and Z mul These are the pixel-level addition fusion result and the pixel-level multiplication fusion result of the first stage, respectively, N. max-min This indicates max-min normalization, O final This is the result of the second-stage fusion, i.e., the final detection result.

[0068] Example 2: The overall implementation steps of the detection method proposed in this example are the same as in Example 1. Now, taking two scales as examples, the specific parameter settings are given to further describe the implementation process of this invention in detail:

[0069] Step A. Multi-scale spatial feature extraction:

[0070] 1.1) Transform three-dimensional hyperspectral images Transform into a two-dimensional matrix Where H, W and B represent the height, width and band of the hyperspectral image, respectively, and N represents the number of pixels, N = H × W;

[0071] 1.2) Principal Component Analysis (PCA) is used to obtain the first principal component features. The Simple Linear Iterative Clustering (SLIC) algorithm is then used to perform superpixel segmentation on these features. A superpixel block is represented as... Moreover {X1, X k , ..., X K} = X. Where, n k Let K represent the number of pixels in a superpixel block. To further utilize spatial information, multi-scale superpixel segmentation is employed, resulting in superpixel blocks at two scales, denoted as... and Then all superpixel blocks at a certain scale are Where ω represents the scale number; This represents the number of superpixels at a certain scale, and its value is set between 1 and 30.

[0072] Step B. Construction of the subspace-based autoencoder model:

[0073] 2.1) Regarding the model architecture, the model consists of three parts: an encoder (E), a subspace recovery layer (SRL), and a decoder (D). The network is built upon fully connected layers. Furthermore, for each superpixel block... They are constructed separately. This is represented as:

[0074]

[0075] in, This represents the features extracted by the encoder; This indicates the subspace restoration layer of the k-th superpixel block at the ω scale. The acquired features For a matrix of size t′×T′, in this embodiment, t′ and T′ are set to 60 and 50 respectively; Finally, the features are obtained through the decoder. The specific network design is shown in the table below:

[0076] Table 1. Subspace-based autoencoder network settings

[0077]

[0078] The optimizer for this network is Adam, and the learning rate is set to 0.001.

[0079] Furthermore, by employing a subspace restoration autoencoder model for hyperspectral images with spatial features at different scales, reconstruction results at multiple scales can be obtained, namely:

[0080] 2.2) Joint Loss Function (L) Based on Subspace Restoration Autoencoder Network total It consists of two parts: reconstruction loss (L) Rec ) and subspace repair loss (L SRL It can be represented as:

[0081]

[0082] Among them, L Rec (·) represents the reconstruction loss; L SRL (·) represents the subspace repair loss; α and β are hyperparameters, set to 1; x i I represents a pixel; t It is a 60×60 identity matrix; z represents The feature of a certain pixel, x i For a single pixel in a hyperspectral image, x i T x represents i transpose, This indicates that the reconstructed features are obtained after the subspace-based reconstructive autoencoder.

[0083] Step C. Training and testing of alternating optimization based on robust principal component analysis:

[0084] To better characterize background features, an alternating optimization strategy is adopted. This strategy is based on robust principal component analysis, which divides an image into two parts: X = L + S. Here, L represents the background part, and S represents the part containing anomalies and noise. Therefore, the L and S parts can be optimized alternately. The optimization can be expressed as:

[0085]

[0086] Among them, ||·|| 2,1 Represented as L 2,1 Norm; λ is the harmonic coefficient. This optimization process consists of two phases. Its termination conditions include the following two:

[0087] a) When the set number of iterations is reached;

[0088] b) One of the following conditions must be met:

[0089]

[0090] Where c1 and c2 are evaluation parameters for testing convergence; ε is the convergence parameter, set to 1×10. -5 ; and Let L and S represent the iteratively optimized values, respectively. The following optimization is performed for spatial information at a specific scale.

[0091] 3.1) Fixed S, optimized L: First, through The background portion of the hyperspectral image is obtained, and the spatial information of superpixel segmentation is used to index L = {L1, L...} k ,…,L K}; Then, based on each superpixel block L K Build an SRAE k Model; finally, sample L K Input to SRAE k The model was optimized. And using L total Optimizing SRAE k parameter.

[0092] 3.2) Fixed L optimization S: Utilizing The generated hyperspectral image contains anomalous and noisy components; through L 2,1 The proximal method performs S optimization, which can improve L. 2,1 The norm is converted to a block-by-block soft thresholding function, i.e.:

[0093]

[0094] Where λ is the harmonic coefficient, with a value range of [1, 1000]. Through this process, the optimized [value] is obtained.

[0095] 3.3) Repeat steps 3.1) and 3.2) above until the iteration termination condition is met, and obtain the final reconstruction, i.e.

[0096] 3.4) For hyperspectral images containing different spatial information at two scales, repeat the processes described in 3.1)-3.3) above to obtain the reconstruction results at the two scales. and

[0097] Step D. Anomaly Detection: Introduce the RX algorithm to analyze the reconstruction results at both scales. and Anomaly score evaluation is performed to obtain R1 and R2. Then, a two-stage fusion of R1 and R2 is performed, as follows:

[0098]

[0099]

[0100]

[0101] Among them, Z add and Z mul This is the result of the first phase of integration. and For pixel-level addition and pixel-level multiplication; O final For the results of the second phase of fusion, N max-min This is a maximum-minimum normalization process.

[0102] The advantages of this invention are summarized as follows: 1. By combining background sample purification with model training, the deep self-representation learning network becomes an end-to-end approach, solving the local optima problem caused by non-end-to-end methods, thereby improving detection performance. 2. To better characterize the background of hyperspectral images, a subspace restoration layer is introduced into the autoencoder network, proposing a subspace restoration-based autoencoder network that can strengthen background reconstruction and weaken anomalous reconstruction. 3. To better constrain the network, a joint loss function including reconstruction loss and subspace restoration loss is proposed, which improves detection accuracy. 4. For the first time, alternating optimization based on robust principal component analysis is introduced, effectively removing anomalous samples and purifying background samples, enabling the subspace restoration-based autoencoder network to better characterize the background of hyperspectral images. 5. By incorporating spatial information into the subspace restoration-based autoencoder network, a local deep learning network is constructed, forming a new training paradigm that improves the model's sensitivity to local anomalies and its ability to distinguish background. 6. In order to further utilize spatial information, a multi-scale strategy was adopted, which fused the detection results of multiple scales in two stages, which not only improved the detection rate but also reduced the false alarm rate.

[0103] The advantages of this invention are further demonstrated through qualitative and quantitative comparisons below:

[0104] Reference Figure 3The subjective qualitative comparison results of this invention with seven popular algorithms on three publicly available hyperspectral anomaly detection datasets show that this invention captures anomalies more effectively. The publicly available hyperspectral anomaly detection datasets used in the experiments included Gulfport, Gainesville, and Salinas. The seven popular algorithms compared were: RX [Reference: I.S. Reed, and X. Yu, “Adaptive multiple-band CFAR detection of an optical pattern with unknown spectral distribution,” IEEE Trans. Acoust. Speech Signal Process., vol. 38, no. 10, pp. 1760-1770, Oct. 1990.], CRD [Reference: W. Li and Q. Du, “Collaborative representation for hyperspectral anomaly detection,” IEEE Trans. Geosci. Remote Sens., vol. 53, no. 3, pp. 1463-1474, Mar. 2015.], and NJCR [S. Chang and P. Ghamisi, “Nonnegative-Constrained Joint Collaborative Representation With Union Dictionary for Hyperspectral Anomaly Detection,” in IEEE Transactions on Geoscience]. and Remote Sensing, vol.60, pp.1-13, 2022, Artno.5534913, doi:10.1109 / TGRS.2022.3195339.], KIFD [Reference: S.Li, K.Zhang, P.Duan, and X.Kang, "Hyperspectral anomaly detection with kernel isolation forest," IEEE Trans.Geosci.Remote Sens., vol.58, no.1, pp.319-329, Jan.2020.], GAED[P.Xiang,S.Ali,SKJung et al., "Hyperspectral anomaly detection with guided autoencoder," IEEE Trans.Geosci.Remote Sens., vol.60, pp.1-18, Sep., 2022.Artno.5538818.], RGAE [Reference: Fan G, Ma Y, Mei X, et al. Hyperspectral anomaly detection with robust graph autoencoders[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60: 1-14.], Auto-AD [S.Wang, no.5503314.]. .

[0105] The following is a quantitative comparison between this invention and seven popular algorithms, using the following rating metric: Detection Rate (AUC(P)). d ,P f ) and false alarm rate AUC(P f ,τ). Among them, the detection rate AUC(P) d ,P f The larger the value of AUC(P), the better the anomaly detection performance; the false alarm rate (AUC(P)) is. f The smaller the value of τ, the better the performance in detecting false anomalies. See Tables 2 and 3 below:

[0106] Table 2 AUC(P) d ,P f Indicator Comparison Table

[0107]

[0108] Table 3 AUC(P) f Comparison table of indicators (τ)

[0109]

[0110] As shown in the AUC values ​​in the table above, this invention not only achieves the best performance in terms of detection rate but also in terms of false alarm rate. Therefore, this demonstrates the superiority of this invention over mainstream algorithms.

[0111] This invention incorporates the inherent properties of hyperspectral images, namely the low rank of the background and the sparsity of anomalies and noise, and fully utilizes the spatial and spectral information of hyperspectral images to construct a novel end-to-end network based on a deep self-representation learning framework. This solves the problems of low detection rate and high false alarm rate in existing encoder-based networks. Furthermore, qualitative and quantitative experimental comparisons verify the advantages of this invention in terms of detection rate and false detection rate.

[0112] The parts of this invention not described in detail are common knowledge to those skilled in the art.

[0113] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, those skilled in the art, after understanding the content and principle of the present invention, may make various modifications and changes in form and detail without departing from the principle and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A hyperspectral anomaly detection method based on a deep self-representation learning framework, characterized in that, Includes the following steps: (1) Extract the multi-scale spatial features of the hyperspectral image X to obtain K superpixel blocks at W scales; where W is a positive integer greater than or equal to 2; (2) Construct and train the subspace restoration autoencoder network SRAE: (2.1) For each superpixel block, construct a subspace restoration autoencoder (SRAE) network model consisting of an encoder, a subspace restoration layer, and a decoder: Where ω = 1, 2, ..., W; k = 1, 2, ..., K; The features of the k-th superpixel block at the ω-scale extracted by the encoder. For the subspace restoration layer of the k-th superpixel block at the ω scale Acquired features; and Represent the k-th superpixel block at scale ω. The encoder and decoder; This indicates that the code finally passes through the decoder. Acquired features; (2.2) Calculate the joint loss function L of the SRAE network model according to the following formula. total : Among them, L Rec (·) represents the reconstruction loss; L SRL (·) represents the subspace repair loss; (2.3) Based on the robust principal component analysis method, an alternating optimization approach is adopted, utilizing the joint loss function L. total The model parameters for each pixel block are iteratively optimized, and the corresponding detection model is obtained through training. (3) Each pixel is input into the detection model corresponding to that pixel block. The test was conducted to generate multi-scale reconstruction results. (4) Introduce the RX algorithm based on Mahalanobis distance to process the multi-scale reconstruction results. Anomaly score evaluation is performed, and the evaluation results are obtained as R = {R1, R2, ..., R...} ω , ..., R W Then, the results are fused in two stages, as follows: in, and These represent pixel-level addition and pixel-level multiplication, respectively; Z add and Z mul These are the pixel-level addition fusion result and the pixel-level multiplication fusion result of the first stage, respectively, N. max-min This indicates max-min normalization, O final This is the result of the second-stage fusion, i.e., the final detection result.

2. The method according to claim 1, characterized in that: Step (1) involves extracting multi-scale spatial features from the hyperspectral image. The steps are as follows: (1.1) Transform the three-dimensional hyperspectral image Transform into a two-dimensional matrix Where H, W and B represent the height, width and band of the hyperspectral image, respectively, and N represents the number of pixels, N = H × W; (1.2) Principal Component Analysis (PCA) is used to obtain the first principal component features. Then, the Simple Linear Iterative Clustering (SLIC) algorithm is used to perform superpixel segmentation on the first principal component features to obtain K superpixel blocks {X1, X2, ..., X...}. k , ..., X K },in Represents the k-th superpixel block; n k The number of pixels in the k-th superpixel block; (1.3) Multi-scale superpixel segmentation is adopted, and the k-th superpixel block at the ω-scale is represented as... in Indicates the number of superpixels at the ω scale; (1.4) Take ω = 2, 3, ..., W to obtain all superpixel blocks at each scale, i.e., the multi-scale spatial features of the hyperspectral image.

3. The method according to claim 1, characterized in that: Step (2.2) The joint loss L of the model total Specifically, it is calculated as follows: Where α and β are hyperparameters, their values ​​are set to 1; x i I represents a pixel; t Let z be the identity matrix of t′×t′, where t′ ranges from 10 to 100; z represents The feature of a certain pixel, x i For a single pixel in a hyperspectral image, x i T x represents i transpose, This indicates that the reconstructed features are obtained after the subspace-based reconstructive autoencoder.

4. The method according to claim 1, characterized in that: Step (2.3) employs an alternating optimization method based on robust principal component analysis, which involves dividing an image into two parts: the background part L and the abnormal noise part S, and then alternately optimizing L and S according to the following formula: Among them, ||·|| 2,1 L represents 2,1 Norm; λ is the harmonic coefficient; E k A k D k L k These represent the encoder, subspace restoration layer, decoder, and background portion of the k-th superpixel block, respectively. The optimization process is terminated under the following conditions: a) The preset number of iterations has been reached; b) One of the following conditions must be met: Where c1 and c2 are evaluation parameters for testing convergence; ε is the convergence parameter.

5. The method according to claim 4, characterized in that: In step (2.3), the L and S parts are optimized alternately, as follows: (2.3.1) Optimizing L with a fixed S: First, through The background portion of the hyperspectral image is obtained, and the background portion at the ω-scale is indexed using the spatial information from superpixel segmentation. The background portion of each pixel block at the ω scale is input into its detection model to obtain the optimized background portion at the ω scale. (2.3.2) Optimize S with fixed L: using Generate anomalous noise components in hyperspectral images; via L 2,1 The near-end method is used for S-optimization to obtain the optimized abnormal noise component. (2.3.3) Repeat steps (3.1) and (3.2) above until the termination condition is met, to obtain the reconstruction result at the ω scale, i.e. (2.3.4) For hyperspectral images containing different spatial information at multiple scales, following steps (2.3.1)-(2.3.3) above, the multi-scale reconstruction result is obtained, i.e.

6. The method according to claim 5, characterized in that: Step (2.3.2) L 2,1 The proximal method is to use L 2,1 Norm conversion to block-by-block soft thresholding function: Where i and j represent the i-th row and j-th column of the hyperspectral space, S i,j S represents the i-th row and j-th column of the abnormal noise part. i,: This represents the entire i-th row in the abnormal noise section.

Images