Hyperspectral anomaly detection via low-rank and sparse decomposition with cluster subspace accumulation Authors: Baozhi Cheng, Yan Gao Published: 2024 ABSTRACT AbstractAnomaly detection (AD) has emerged as a prominent area of research in hyperspectral imagery (HSI) processing. Traditional algorithms, such as low-rank and sparse matrix decomposition (LRaSMD), often struggle to effectively address challenges related to background interference, anomaly targets, and noise. To overcome these limitations, we propose a novel method that leverages both spatial and spectral features in HSI. Initially, the original HSI is segmented into several subspaces using the k-means method, which reduces redundancy among HSI bands. Subsequently, the fractional Fourier transform (FrFT) is applied within each subspace, enhancing the distinction between background and anomaly target information while simultaneously suppressing noise. To further improve the stability and discriminative power of the HSI, LRaSMD is employed. Finally, the modified Reed–Xiaoli (RX) detector is utilized to identify anomalies within each subspace. The results from these detections are then aggregated to produce a comprehensive final outcome. Experiments conducted on five real HSI data sets yield an average area under the curve (AUC) of 0.9761 with a standard deviation of 0.0156 for the proposed algorithm. These results indicate that our method is highly competitive in the field of anomaly detection.Keywords: Hyperspectral imagery (HSI), Anomaly target detection, Low-rank and sparse decomposition, Cluster subspaceSubject terms: Imaging and sensing, Electrical and electronic engineering Introduction Compared to multispectral images, HSI offers superior capabilities for recognizing and analyzing surface materials1. HSI has found extensive applications in various domains, including image quality evaluation, image correction, classification, unmixing, and target detection2,3. Notably, AD has gained traction across multiple fields due to its independence from prior target information, AD has a wider practical application4. This characteristic makes it particularly valuable for detecting marine oil spills, accurately identifying battlefield targets, and inspecting the quality of agricultural products5,6. In recent decades, research on AD has focused on both algorithm implementation and application. AD algorithms are generally categorized into two types: statistics-based and representation-based methods7. Among the statistical approaches, the RX detector stands out as a classical anomaly detection method. It simplifies the background model using a Gaussian multivariate distribution, allowing for the detection of anomaly targets that deviate from the background feature distribution through the Mahalanobis distance8. Beyond the RX detector, several extensions have been developed, including the fractional Fourier transform-based tensor RX detector9, as well as the local RX detector (LRXD)10, global RX detector (GRXD)11, weighted RX12, and RX-segmented (RXSD)13. However, the performance of the RX algorithm is constrained by two primary issues: (1) the conventional RX distance measure fails to account for higher-order relationships among spectral bands at different wavelengths, and (2) the local Gaussian assumption and homogeneity of background distributions may not hold in certain scenarios. Given the nonlinear characteristics of anomaly targets and background information in HSI5, nonlinear methods have been employed to enhance AD performance. For instance, the kernel RX (KRX) detector utilizes kernel mapping to transform the original HSI into a high-dimensional feature space, thereby improving anomaly detection capabilities14. Various KRX-based enhancements have emerged, such as the modified kernel-RX (MKRX) detector15, cluster kernel RX (CKRX) detector16, and fast real-time kernel RX (LRT-KRX) detector17. Additionally, Banerjee et al. developed the Support Vector Data Description (SVDD) algorithm for anomaly target detection, which circumvents the need to solve high-dimensional inverse covariance matrices18. Moreover, some real HSI data do not conform to Gaussian distributions. To mitigate false alarms, sparse representation techniques have been proposed for anomaly detection, including the sparse representation for target detection (SR-TD) detector19 and the local sparsity divergence (LSD) detector20. These methods operate on the principle that HSI can be projected into low-dimensional subspaces, allowing target and background pixels to be expressed as sparse linear combinations19. Recently, low-rank techniques have also been recommended for anomaly detection21,22. The LRaSMD method has proven effective in HSI anomaly detection23,24. This approach posits that HSI consist of background, anomaly targets, and noise. Given the strong interband correlations in HSI, the background spectral signatures can be modeled as a low-rank matrix, expressed as a linear combination of several endmembers. In contrast, anomaly targets, being fewer in number, exhibit sparse properties. The noise in HSI typically comprises stripe noise, Gaussian white noise, and photoelectric noise25. This analysis has led to the development of advanced algorithms, such as the Low-rank and Sparse Decomposition with Mixture of Gaussian (LSDM-MoG) detector26, which enhances the detection and identification of anomaly targets using a mixture of Gaussian distributions. Additionally, the orthogonal subspace projection LRaSMD (OSP-LRaSMD) detector effectively addresses anomaly detection challenges through orthogonal subspace projection27. The LRaSMD struggles to determine the optimal rank and sparsity, and it lacks robustness to noise and outliers, resulting in a decrease in detection performance. Thus, we propose a novel LRaSMD-based algorithm that incorporates subspace accumulation through k-means clustering and fractional Fourier entropy (KFRLSMD) for anomaly detection. Initially, we reduce the dimensions of the HSI data using a clustering method, decomposing the HSI into several clustering subspaces via k-means clustering. Subsequently, we apply FrFT to the image data within each subspace7,28. The LRaSMD method is then employed to decompose the HSI matrix into low-rank and sparse matrices. Finally, we construct the covariance matrix from the low-rank matrix and obtain anomaly detection results using a modified RX approach. Briefly, the contributions of related research work on KFRLSMD are summarized as follows: HSI is characterized by their rich spectral data and superior spatial features. Traditional detectors often struggle with incomplete matching between their background models and HSI in real-world scenarios. Consequently, analytical approaches that depend on specific background model assumptions have been set aside. This paper introduces hyperspectral anomaly detection method that leverages subspace theory and FrFT, grounded in low-rank and sparse matrix decomposition. By optimizing both the spectral and spatial attributes of HSI, this approach effectively accentuates the abstract differences between anomaly and backgrounds, thereby enhancing anomaly distinguish ability. HSI is characterized by their rich spectral data and superior spatial features. Traditional detectors often struggle with incomplete matching between their background models and HSI in real-world scenarios. Consequently, analytical approaches that depend on specific background model assumptions have been set aside. This paper introduces hyperspectral anomaly detection method that leverages subspace theory and FrFT, grounded in low-rank and sparse matrix decomposition. By optimizing both the spectral and spatial attributes of HSI, this approach effectively accentuates the abstract differences between anomaly and backgrounds, thereby enhancing anomaly distinguish ability. Considering the intricate distribution of land covers in real HSI, the proposed method integrates K-means clustering, FrFT, and an modified RX anomaly detection operator. Anomaly detection is conducted within individual subspaces using the FrFT processing and the modified RX algorithm, achieving optimal results in each subspace. These techniques not only address the challenges of determining optimal rank and sparsity in LRaSMD and its susceptibility to noise and outliers but also enhance the accuracy and robustness of hyperspectral anomaly detection while minimizing computational complexity. Considering the intricate distribution of land covers in real HSI, the proposed method integrates K-means clustering, FrFT, and an modified RX anomaly detection operator. Anomaly detection is conducted within individual subspaces using the FrFT processing and the modified RX algorithm, achieving optimal results in each subspace. These techniques not only address the challenges of determining optimal rank and sparsity in LRaSMD and its susceptibility to noise and outliers but also enhance the accuracy and robustness of hyperspectral anomaly detection while minimizing computational complexity. The KFRLSMD method developed in this study is applied to real hyperspectral image scenarios, with experimental simulations conducted on five data sets: San_Diego, Abu-Airport-2, HYDICE_Urba, Airport and Abu-Urban-3. This method effectively tackles the variability, high dimensionality, and large data volumes inherent in real HSI scenes, achieving superior detection accuracy and computational efficiency. It not only addresses the limitations of the LRaSMD method but also surpasses several classic and recently proposed anomaly detection methods, demonstrating robust performance and low computational complexity in practical applications. The KFRLSMD method developed in this study is applied to real hyperspectral image scenarios, with experimental simulations conducted on five data sets: San_Diego, Abu-Airport-2, HYDICE_Urba, Airport and Abu-Urban-3. This method effectively tackles the variability, high dimensionality, and large data volumes inherent in real HSI scenes, achieving superior detection accuracy and computational efficiency. It not only addresses the limitations of the LRaSMD method but also surpasses several classic and recently proposed anomaly detection methods, demonstrating robust performance and low computational complexity in practical applications. The remainder of this paper is organized as follows: Section II reviews the k-means clustering, fractional Fourier transform, and modified RX detector. Section III provides a detailed introduction to the materials and methods employed. Section IV presents the experimental results and analysis, while Section V concludes with a summary of our findings. Methods K-means clutering K-means clustering is an unsupervised learning method that employs an iterative algorithm to categorize data into a predetermined number of clusters. This clustering technique is particularly useful for analyzing high-dimensional data29, such as HSI. The process consists of several key steps: (A)The algorithm begins by defining HSI matrix , the number of clusters , the cluster centers , and the normalization factor . The algorithm begins by defining HSI matrix , the number of clusters , the cluster centers , and the normalization factor . (B)Next, it randomly selects samples from the HSI data set to serve as the initial cluster centers. Next, it randomly selects samples from the HSI data set to serve as the initial cluster centers. (C)For each sample yi (i = 1, 2,⋯,B; j = 1, 2,⋯,ε) in the HSI data set, the algorithm calculates the distance from to each cluster center. The formula used for this distance calculation is as follows:1 For each sample yi (i = 1, 2,⋯,B; j = 1, 2,⋯,ε) in the HSI data set, the algorithm calculates the distance from to each cluster center. The formula used for this distance calculation is as follows:1 where is the dimension of the sample. (D)The distance from each sample to the cluster center is calculated. The sample is then assigned to the corresponding cluster based on the minimum distance identified. The distance from each sample to the cluster center is calculated. The sample is then assigned to the corresponding cluster based on the minimum distance identified. (E)The cluster center is recalculated and updated according to Eq. (2). Finally, the result of the objective function is computed using Eq. (3).23 The cluster center is recalculated and updated according to Eq. (2). Finally, the result of the objective function is computed using Eq. (3).23 The algorithm evaluates the clustering center and the objective function. It terminates when the specified requirements are met; otherwise, it proceeds to step (B). The results of K-means clustering for HSI are illustrated in Fig. 1. Specifically, Fig. 1a presents a false-color image of a mining area in Nevada, USA, utilizing bands 2, 70, and 159 as RGB channels. This area is characterized by a rich mineral presence and a limited amount of vegetation. Additionally, Fig. 1b displays the clustering results for the first ten bands of HSI. To streamline data processing, only these ten bands were selected and categorized into two distinct groups. Cuprite_ref HSI and the clustering result. (a) False-color image of the cuprite_ref imagery scene, (b) Clustering result map. Fractional fourier fransform The FrFT is more effective than the traditional Fourier transform in processing nonstationary signals28. This capability makes it particularly suitable for anomaly detection in HSI. Consider an HSI matrix , where represents the number of pixels and denotes the number of spectral bands. For each pixel , its representation in the FrFT domain can be described as follows28. With where where and represent indices, while is an integer. denotes the fractional order of the FrFT. Additionally, when is present, is determined; similarly, when ϑ = (2n+1) π is present, is established. indicates the rotation angle associated with . The results for vary, as illustrated in Fig. 2. These variations confirm that the FrFT effectively suppresses noise in the original HSI. Illustration with tenth hyperspectral band and different for the FrFT domain. (a) the original HSI, (b) , (c) , (d) , (e) , (f) . Low-rank and sparse matrix decomposition model HSI typically involves issues of relevance and redundancy in a low-dimensional subspace, with anomaly targets being sparse. Therefore, in LRaSMD, the HSI matrix data can be modeled as follows: In HSI, the background is represented by , which is a low-rank matrix. The anomalies are denoted by , which constitutes a sparse matrix. Additionally, represents the noise matrix present in HSI. To address the components of low-rank and sparse matrices, we employ the Go Decomposition (GoDec) algorithm30. This algorithm formulates the anomaly detection in HSI as a minimization problem, as outlined below: where, and represent the upper bounds for the rank of matrix and the cardinality of , respectively, while denotes the Frobenius norm. As demonstrated in Eq. (9), the optimization problem outlined in Eq. (8) can be addressed by alternatively solving two subproblems. To solve for and , we update using singular value hard thresholding of , as follows: updating via entry-wise hard thresholding of , as follows: where Ω denotes the nonzero subset of the first largest entries of , while indicates the projection of a matrix onto the entry set 1. The GoDec algorithm is summarized in Algorithm 1, as follows: Godec algorithm for LRaSMD. Modified RX detector The detection results of the GoDec algorithm do not satisfy the requirements for anomaly detection. Therefore, we need to further process the low-rank components and , using the RX detector for anomaly detection. In the classical RX detector8,9, the binary hypothesis of the RX algorithm is defined as follows: In this study, we consider two competing hypotheses characterized by the same background covariance but differing means. The target signature, denoted as , and the background covariance, represented as , are assumed to be unknown. The background clutter noise is modeled by the vector , while denotes the spectral signature of the anomaly target. Under the framework , the data representing background clutter is modeled as , and under framework , it is modeled as . The RX algorithm employs a threshold test based on the Mahalanobis distance between the test pixel and its background. It is assumed that this background follows a multivariate normal distribution. The background covariance matrix, denoted as , is estimated from the reference data of background clutters. Additionally, represents the estimated sample mean of these background clutters. We define as the anomaly detection threshold. This framework facilitates a decision between two competing hypotheses: if is true, we adopt , which posits the absence of a target. Conversely, if is not true, we utilize , indicating the presence of an anomaly target. Given that the low-rank matrix derived from the Godec algorithm contains some anomalous components, we propose a novel method to enhance the classical RX approach. Initially, we employ the low-rank matrix to construct the background covariance matrix, denoted as , as follows: is the number of pixels, where Eq. (13) can be represented as follows: The anomaly detection results obtained from Eq. (16) will be utilized to construct a new background. Specifically, the first smallest entries from these results will be selected for this purpose, as indicated in reference16. The background will then be used to construct a covariance matrix, employing Eqs. (14) and (15). Ultimately, this process will lead to the final anomaly detection results for HSI. Proposed detection method The anomaly detection (AD) process utilizes a representation method that leverages the similarity relationship between background pixels in HSI. This approach effectively incorporates local spatial information and background spectral data surrounding the tested pixel, resulting in successful anomaly detection outcomes. As illustrated in Fig. 3, the proposed KFRLSMD algorithm integrates k-means clustering, FrFT, LRaSMD, and a modified Reed-Xiaoli (RX) detector. By combining representation-based and statistics-based methods, the algorithm capitalizes on their respective strengths to identify anomalous targets. Specifically, k-means clustering partitions the HSI into several subspaces. Subsequently, feature extraction is performed using the FrFT, LRaSMD, and modified RX detectors within each subspace. The detection results from all subspaces are then aggregated to produce the final outcome. The detailed procedure of the KFRLSMD algorithm is outlined in Algorithm 2. Flowchart of KFRLSMD algorithm. The proposed KFRLSMD algorithm. Experiments Data set The effectiveness and superiority of the KFRLSMD method are demonstrated using five HSI data sets. All experiments were conducted using MATLAB on a Windows 10 system equipped with an Intel i7-11800H processor (2.30 GHz) and 16 GB of RAM. The data sets utilized are as follows: San_Diego: This HSI was acquired by the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) in San Diego, USA. It contains various anomaly components in specific areas, making it suitable for simulating anomaly target detection. The original image consists of 224 bands; however, after removing bands severely affected by moisture and noise, 189 bands remain. For the experiment, an area of 100 × 100 pixels was selected, resulting in an HSI of size 100 × 100 × 189. San_Diego: This HSI was acquired by the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) in San Diego, USA. It contains various anomaly components in specific areas, making it suitable for simulating anomaly target detection. The original image consists of 224 bands; however, after removing bands severely affected by moisture and noise, 189 bands remain. For the experiment, an area of 100 × 100 pixels was selected, resulting in an HSI of size 100 × 100 × 189. Abu-Airport-2: This data set was also obtained by AVIRIS, this time from Los Angeles, USA. The selected experimental area measures 100 × 100 pixels. After removing bands impacted by moisture and noise, 205 bands are retained. The HSI has a size of 100 × 100 × 205, with a spatial resolution of 7.1 m and a wavelength range of 400-2500 nm. Abu-Airport-2: This data set was also obtained by AVIRIS, this time from Los Angeles, USA. The selected experimental area measures 100 × 100 pixels. After removing bands impacted by moisture and noise, 205 bands are retained. The HSI has a size of 100 × 100 × 205, with a spatial resolution of 7.1 m and a wavelength range of 400-2500 nm. HYDICE_Urban: Acquired from an urban area in the USA using the Hyperspectral Digital Imagery Collection Experiment (HYDICE) sensor, this data set consists of 80 × 100 pixels. It includes 175 bands covering wavelengths from 400 to 2500 nm and contains 21 anomaly pixels representing cars and rooftops. The HSI is of size 80 × 100 × 175. HYDICE_Urban: Acquired from an urban area in the USA using the Hyperspectral Digital Imagery Collection Experiment (HYDICE) sensor, this data set consists of 80 × 100 pixels. It includes 175 bands covering wavelengths from 400 to 2500 nm and contains 21 anomaly pixels representing cars and rooftops. The HSI is of size 80 × 100 × 175. Airport: This HSI was captured by AVIRIS in Gulfport, Mississippi, USA. The experimental area is 100 × 100 pixels, with 191 bands included in the data set. The HSI measures 100 × 100 × 191, featuring a spatial resolution of 3.4 m and a wavelength range of 400-2500 nm. Airport: This HSI was captured by AVIRIS in Gulfport, Mississippi, USA. The experimental area is 100 × 100 pixels, with 191 bands included in the data set. The HSI measures 100 × 100 × 191, featuring a spatial resolution of 3.4 m and a wavelength range of 400-2500 nm. Abu-Urban-3: This data set was obtained by AVIRIS from Gainesville, Florida, USA. Notably, Abu Urban 3 contains a higher number of anomaly targets within the same background area compared to the other data sets. The HSI is of size 100 × 100 × 191, with a spatial resolution of 3.5 m. Abu-Urban-3: This data set was obtained by AVIRIS from Gainesville, Florida, USA. Notably, Abu Urban 3 contains a higher number of anomaly targets within the same background area compared to the other data sets. The HSI is of size 100 × 100 × 191, with a spatial resolution of 3.5 m. Experimental results and discussion In this section, the anomaly detection performance of the proposed KFRLSMD is evaluated. To assess the effectiveness of the KFRLSMD algorithm, we compare it with several other anomaly detection methods, including the RX detector (RXD)8, RX segmented (RXSD)13, Low-Rank and Sparse Representation (LRASR)23, Local Sparsity Divergence (LSD)20, LRaSMD-based Mahalanobis distance (LSMAD)25, Two-Step the Generalized Likelihood Ratio Test(2SGLRT)31, and Anomaly Detection Based on Tree Topology (TTAD)32. All the detectors are implemented with optimal parameters. For RXSD, HSI is segmented by a K-means type algorithm; we then calculate a covariance matrix based on the pixels in the segment. For LRASR, is a parameter to trade off low rankness and sparsity, is used to balance the effects of the two parts, stable detection performance is obtained when and is set 0.1. For LSD, the outer window is 15 × 15 and the inner window is 5 × 5. For LSMAD, stable detection performance is obtained when the rank value is set to 28. For 2SGLRT, the outer window is 5 × 5 and the inner window is 3 × 3. For TTAD, TTAD involves two key parameters, and , which indicate the number of parallel topological spaces and the subsampling percentage corresponding to the framework required for a single topology, respectively. the corresponding optimal parameter combination is and . Image quality evaluation We utilize the receiver operating characteristic (ROC) curve33 and the area under the curve (AUC) to evaluate the performance of these algorithms5. The ROC curve illustrates the relationship between detection probability and the false alarm rate 4. By establishing a threshold, we can generate ROC curves that provide all coordinates relevant to the image scene. A pixel is classified as a target if it falls within the coordinate ranges designated for target regions. In this context, we define the axes and as follows: where represents the number of detected true target pixels, denotes the total number of target pixels in HSI, indicates the number of false alarms, and signifies the total number of pixels in HSI. The area under the receiver operating characteristic (ROC) curve, known as the AUC value, serves as a performance metric for the detector. A larger AUC value indicates better detector performance at the same false alarm rate. Discussion for parameters The paper examines the impact of the number of clusters on anomaly detection results. As illustrated in Fig. 4, the anomaly detection maps for the airport data set reveal that when the number of clusters is set to 3, the anomaly targets are more distinguishable, and the false alarm rate is lower. Additionally, Fig. 5 presents the ROC curves for the airport data set across different cluster numbers. The results indicate that the probability of detection is superior when the number of clusters is 3 compared to other configurations. Furthermore, Table 1 provides a quantitative evaluation of the AUC values, demonstrating that the largest AUC occurs with three clusters. Overall, these findings confirm that using three clusters yields stable and effective anomaly detection results. Consequently, the original HSI is clustered using the proposed KFRLSMD algorithm with three clusters. Comparison anomaly detection results with different cluster numbers for the Airport data set. (a) Pseudocolor image, (b) Ground-truth map, (c) Cluster numbers are 2, (d) Cluster numbers are 3 , (e) Cluster numbers are 4 , (f) Cluster numbers are 5 , (g) Cluster numbers are 6. ROC curves using different clustering numbers for the Airport data set. Table 1Performance evaluation through AUC values based on different cluster numbers.Cluster numbers23456AUC values0.98400.99230.98320.98020.9775Significant values are in bold. Performance evaluation through AUC values based on different cluster numbers. Significant values are in bold. Additionally, the selection of the fractional order parameter value for FrFT is thoroughly analyzed in Ref.28, concluding that when the fractional order ranges between 0.6 and 0.9, there is an increase in FrFE values. This increase indicates enhanced discrimination between anomalies and backgrounds, leading to improved detection performance as measured by AUC. Consequently, the KFRLSMD method is validated using experimental data sets. Based on the experimental results, the fractional order parameter value of FrFT is set to 0.7 for all five real hyperspectral image data sets. Detection performance First, the effectiveness of the proposed KFRLSMD algorithm for anomaly detection is assessed using five HSI data sets, Fig. 6 through Fig. 10 illustrate the results. Specifically, Figs. 6a, 7a, 8a, 9a and 10a present the 10th band image scenes of the original HSI data. In contrast, Figs. 6b, 7b, 8b, 9b and 10b display the ground-truth maps of the anomaly results. The anomaly detection outcomes from various algorithms are shown in Figs. 6c–j, 7c–j, 8c–j, 9c–j and 10c–j. To enhance the anomaly detection performance, optimal parameters were selected for all algorithms. For the KFRLSMD algorithm, the number of clusters is set to three, and the fractional order of FrFT is specified. The detection result maps indicate that the KFRLSMD algorithm outperforms other algorithms across the five HSI data sets, demonstrating superior distinguish ability and robustness. In contrast, the LRASR, RXD, and LSD methods, as depicted in Figs. 6c–j, 7c–j, 8c–j, 9c–j and 10c–j, struggle to effectively identify anomaly objects. These methods also exhibit issues with false detections, mistakenly identifying background information as anomaly. Although the LSMAD algorithm addresses some of the challenges associated with hyperspectral anomaly detection and produces effective results, its robustness is inferior to that of the KFRLSMD algorithm, resulting in lower detection accuracy. In the experimental simulation, the TTAD method outperforms 2SGLRT in anomaly detection, yet both are less effective than the KFRLSMD method. Comparison anomaly detection results of the proposed KFRLSMD with other algorithms for the San_Diego data set. (a) 10th band image scene of original image; (b) Ground-truth map; (c) LRASR; (d) SGLRT; (e) LSD; (f) LSMAD; (g)RXD; (h)RXSD; (i) TTAD; (j) the proposed KFRLSMD. Comparison anomaly detection results of the proposed KFRLSMD with other algorithms for the Abu-urban-3 data set. (a) 10th band image scene of original image; (b) Ground-truth map; (c) LRASR; (d) SGLRT; (e) LSD; (f) LSMAD; (g)RXD; (h) RXSD; (i) TTAD; (j) the proposed KFRLSMD. Comparison anomaly detection results of the proposed KFRLSMD with other algorithms for the abu-airport-2 data set. (a) 10th band image scene of original image; (b) Ground-truth map; (c) LRASR; (d) SGLRT; (e) LSD; (f) LSMAD; (g)RXD; (h) RXSD; (i) TTAD; (j) the proposed KFRLSMD. Comparison anomaly detection results of the proposed KFRLSMD with other algorithms for the HYDICE_urban data set. (a) 10th band image scene of original image; (b) Ground-truth map; (c) LRASR; (d) SGLRT; (e) LSD; (f) LSMAD; (g)RXD; (h)RXSD; (i) TTAD; (j) the proposed KFRLSMD. Comparison anomaly detection results of the proposed KFRLSMD with other algorithms for the Airport data set. (a) 10th band image scene of original image; (b) Ground-truth map; (c) LRASR; (d) SGLRT; (e) LSD; (f) LSMAD; (g) RXD; (h) RXSD; (i) TTAD; (j) the proposed KFRLSMD. Second, the ROC curve serves as an effective evaluation metric for HSI anomaly target detection methods. This is illustrated in Figs. 11, 12, 13, 14 through 15, which present experimental simulations using five data sets: San_Diego, Abu-airport-2, HYDICE_urban, Airport, and Abu-urban-3. In the initial false alarm rate (FAR) range of 0 to 0.05, the KFRLSMD method exhibits excellent detection performance across the first four data sets. However, its performance in Abu-urban-3 is slightly inferior to that of the LSD, 2SGLRT, LSMAD, and LRASR methods. As the FAR increases to between 0.05 and 0.15, KFRLSMD maintains strong detection capabilities in the Abu-airport-2, HYDICE_urban, Airport, and Abu-urban-3 data sets, though it underperforms compared to the TTAD algorithm in the San_Diego data set. Notably, when the FAR surpasses 0.15, the KFRLSMD method significantly outperforms the other algorithms. ROC curves obtained by different methods for the San_Diego data set. ROC curves obtained by different methods for the abu-airport-2 data set. ROC curves obtained by different methods for the HYDICE-urban data set. ROC curves obtained by different methods for the Airport data set. ROC curves obtained by different methods for the Abu-urban-3 data set. Third, the AUC is a standard metric for evaluating the quality of anomaly detection. Table 2 presents the AUC values for eight anomaly detection algorithms across five HSI data sets: San_Diego, ABU-airport-2, HYDICE_urban, Airport, and Abu-urban-3. The average AUC values for these data sets are as follows: 0.8260, 0.9339, 0.9479, 0.8353, 0.9277, 0.9668, 0.9218 and 0.9761, respectively. Correspondingly, the standard deviations for these AUC values are 0.1046, 0.0550, 0.0212, 0.1540, 0.0435, 0.0268, 0.0486, and 0.0156. Notably, the KFRLSMD algorithm exhibits the highest average AUC value among all algorithms, indicating superior performance. Additionally, its standard deviation is the lowest, suggesting that the KFRLSMD algorithm maintains a relatively stable performance across the data sets. Therefore, it is evident that KFRLSMD outperforms the other algorithms in the experimental HSI data sets. Table 2AUC values and standard deviation of the algorithms for five data sets.Data setsLSDRXDTTAD2SGLRTRXSDLSMADLRASRKFRLSMDSan_Diego0.81530.94020.98000.59310.94760.96880.86810.9801Abu-airport-20.82530.84030.93900.88720.90270.92400.87510.9551HYDICE_urban0.66540.98550.92190.98290.97720.99130.97540.9924Airport0.95110.95250.94810.78320.94480.98730.95990.9877Abu-urban-30.87270.95120.95060.93020.86600.96260.93040.9651Average values0.82600.93390.94790.83530.92770.96680.92180.9761Standard deviation0.10460.05500.02120.15400.04350.02680.04860.0156Significant values are in bold. AUC values and standard deviation of the algorithms for five data sets. Significant values are in bold. Finally, Table 3 summarizes the computational complexity of the anomaly detection methods, with all average run times measured in seconds. The KFRLSMD method offers significant advantages in terms of execution efficiency and robustness, primarily because it eliminates the need for training time. This characteristic makes it particularly practical for scenarios that demand high time efficiency in anomaly detection. As detailed in Table 3, the execution times of various methods are compared, revealing that among the five experimental simulation data sets, the RXD method exhibits the shortest computation time, whereas the LSD method requires the longest. Notably, the KFRLSMD method strikes an optimal balance between computation time and detection accuracy. This balance results in superior performance across key metrics, including detection accuracy, robustness, and computation time, when compared to other methods. Table 3Execution time of various anomaly detectors using different data sets.Data setsLSDRXDTTAD2SGLRTRXSDLSMADLRASRKFRLSMDSan_Diego154.860.9714.7772.264.7713.5441.2216.33Abu-airport-2241.861.0516.6583.735.9614.8541.4217.84HYDICE_urban251.220.8411.0950.074.1610.2532.0712.38Airport850.481.1214.7373.354.3713.5238.7316.51Abu-urban-3711.991.0216.3775.236.0514.5841.5217.76 Execution time of various anomaly detectors using different data sets. Conclusions This paper proposes a novel algorithm, KFRLSMD, to enhance the performance of anomaly detection (AD) in HSI by leveraging both spatial and spectral features. The KFRLSMD algorithm incorporates the GoDec method for LRaSMD. While LRaSMD-based anomaly detection has been addressed in previous literature, existing algorithms often yield only rank candidates and may not achieve satisfactory detection performance. Given that HSI is characterized as big data imagery with inherent redundancy among bands, we employ the k-means clustering method to partition HSI into several subspaces. Processing is conducted within each subspace, where features are extracted using the FrFT method. This approach enhances the distinguishability of background and anomaly target information while partially suppressing noise. Subsequently, the GoDec algorithm is applied, followed by the modified RX detector for anomaly target detection. The performance of the KFRLSMD algorithm is validated using five HSI data sets. Experimental results demonstrate that the KFRLSMD algorithm achieves a high target detection rate, a low false alarm rate, and notable robustness compared to other methods. Additionally, we discuss the parameters of the KFRLSMD algorithm, which contribute to its improved performance in detecting anomaly targets. Overall, the KFRLSMD algorithm exhibits strong anomaly detection capabilities and robustness. Acknowledgements All the works were supported by The Natural Science Foundation of the Jiangsu Higher Education Institutions of China (Grant Nos. 23KJA510002), and the National Social Science Fund of China (Grant Nos. 24BTJ051), the authors also would like to thank the handling editors and the reviewers for providing valuable comments. Author contributions Baozhi Cheng wrote the main manuscript text and Yan Gao prepared Figs. 1, 3, Tables 1, 2. All authors reviewed the manuscript. Data availability The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.