Elsevier

Signal Processing

Volume 124, July 2016, Pages 115-131
Signal Processing

A spectral-spatial based local summation anomaly detection method for hyperspectral images

https://doi.org/10.1016/j.sigpro.2015.09.037Get rights and content

Highlights

  • Anomaly detection technique has drawn the increasing interest in hyperspectral imagery processing fields.

  • This article proposes a novel local summation anomaly detection method.

  • The method combines the multiple local distributions from neighboring local windows with spectral-spatial feature integration.

Abstract

Anomaly detection is one of the most popular applications in hyperspectral remote sensing image analysis. Anomaly detection technique does not require any prior features or information of targets of interest and has draw the increasing interest in target detection domain for hyperspectral imagery (HSI) in the recent twenty years. From hyperspectral data, the approximately continuous spectral features which are attributed to the high spectral resolution of hyperspectral image can be achieved. Unfortunately, most conventional anomaly detectors merely take advantage of the spectral information in hyperspectral data and rarely give the consideration to spatial information within neighboring pixels. With the development of remote sensing technology, the high spatial resolution can also be acquired by the hyperspectral airborne/spaceborne sensors. Then, further improvement in algorithmic performance may be achieved if both the spectral and spatial information is combined. This article proposes a novel local summation anomaly detection method (LSAD) which combines the multiple local distributions from neighboring local windows surrounding the pixel under test (PUT) with spectral-spatial feature integration. Some other detection performance enhanced operations such as feature extraction and edge expansion are also used. The proposed local summation anomaly detection method makes allowance for exploiting more sufficient local spatial neighboring relationship of local background distribution around the test pixel considered in detection processing. Moreover, summated local background statistics can get better performance in suppressing background materials and extruding anomalies. Feature extraction enables LSAD with robust background feature statistics and edge expansion can ensure no loss of edge detection information. Experiments are implemented on a simulated PHI data and two real hyperspectral images. The experimental results demonstrate that the proposed anomaly detection strategy outperforms the other traditional anomaly detection methods.

Introduction

Recent years have witnessed the quick development of image processing theory [1], [2], [3], [4], [5], [6], [7], [8]. One newly developed imaging technique, IMAGING SPECTROSCOPY, which is also known as hyperspectral imaging, draws great interests of researchers in image processing field [9], [10], [11]. It is a type of remote sensing imaging technology concerned with the extraction of information about objects lying on the Earth ground surface, based on radiant signal acquired by spaceborne or airborne sensors [9], [10]. Hyperspectral imaging sensors are passive sensors that simultaneously attain images with hundreds of continuous and narrow regions of the electromagnetic spectrums [11]. Approximately continuous spectral characteristics of ground surface materials can be achieved from hyperspectral image because of its high spectral resolution which is usually less than 10 nm [12], [13], [14], [15]. This is the most different trait of hyperspectral images, compared with the traditional panchromatic and multispectral remote sensing images [16]. Then a hyperspectral image “cube” can be established by combining two spatial dimensions (width and height) and a spectral dimension together [17]. By means of approximately continuous spectral curves of ground surface materials, adequate spectral information can be provided for hyperspectral applications such as classification, target detection etc. [18]. For current hyperspectral practical application domains, spectral information is an important essential factor which is still mainly under consideration. Thus, significant attention in the field of hyperspectral remote sensing research is devoted to develop algorithmic techniques to detect, classify, identify, quantify, and characterize objects and features of interest in the captured data [19].

Two types of target detection techniques are usually interested in target detection domain: known and unknown with target spectra. If target spectra are known in the processing of target detectors, it is also named spectral matching detection algorithms refer detectors that require the spectral information about the targets of interest, and they try to identify pixels whose spectrum exhibits a high degree of correlation to the expected signature [20]. However, in most cases it is difficult to obtain the spectrum of ground surface objects in hyperspectral images. The absorption and scattering of the atmosphere, the subtle effects of illumination and the spectral response of the sensor must all be considered in measuring the spectral properties of a material through the atmosphere [21]. Moreover, spectral variability also needs to be addressed in this kind of target detection [22].

The multiplicity of possible spectra associated with the objects of interest and the complications of atmospheric compensation have led to the development and application of target detection without target spectra which is called anomaly detection. Anomaly detection algorithms are more practical in actual applications. Anomaly detection technique which does not require any prior features or information of targets of interest has been of great interest in hyperspectral imagery processing in recent twenty years [23], [24], [25], [26]. Hyperspectral anomaly detection technique has been successfully applied in many application domains such as mineral reconnaissance, border monitoring, search-and-rescue etc. in recent years [27].The goal of anomaly detection is to detect a small quantity of pixels in the hyperspectral image whose spectral characteristics differ significantly from those of a large proportion of pixels in the hyperspectral data cube, and then the hyperspectral image is segmented into two parts: anomaly targets and backgrounds [28]. Anomaly targets are usually inclined to be small man-made objects of minor quantity of pixels whose spectra are distinctly different with other major pixels in the images. For example, for those entrusted with search-and-rescue missions anomalies may be downed-aircraft or adventures lost in the wilderness. With national defense, anomalies may be tanks, aircraft, surface-to-air missile launchers, command bunkers, and other objects of military significance scattered across a battlefield. And for civilian applications, anomalies may be vehicles located in an urban or natural field.

Anomaly detection methods mainly extract the knowledge from backgrounds and use the differences between anomaly targets and the backgrounds to detect the anomalies [29]. According to the different ways of making use of information in hyperspectral images, anomaly detection methods are mainly classified into two kinds: spectral information based and spectral-spatial based. Some conventional anomaly detectors think hyperspectral image as a whole sample dataset and extract the background statistical features from it. These detectors can be called global anomaly detectors [30]. Global anomaly detection algorithms only utilize the spectral information in hyperspectral images. However, the global anomaly detector will not find an isolated local target whose signature is similar to that of global background material [31]. For example, consider a scene containing the isolated trees on a grass plain. Each separate tree may be seen as a local spectral anomaly even if the image contains a separate region with many pixels of trees. The global spectral anomaly detection algorithms are susceptible to this type of clutter and would generate false alarms. Moreover, although the spectral feature vector contains a lot of information about the spectral properties of the pixels, the spectral-vector-based HSI analysis just process each pixel independently, without considering the spatial relationship of neighboring pixels [32].

Most traditionally common used detectors or classifiers are mainly using the spectral information in the “hypercube”. They only deal with a set of first-order data as input, i.e., the vector representation, which is commonly used to represent the spectral feature of a certain pixel in HIS [33], [34], [35]. With the development of airborne/spaceborne remote sensing technology, spatial resolution of hyperspectral image increased gradually. Airborne/spaceborne hyperspectral sensors such as Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) [36] and Hyperspectral Digital Imagery Collection Experiment (HYDICE) [37] can reach a spatial resolution of 1–10 m. This level of spatial resolution can make ground surface materials in hyperspectral images contain a wealth of spatial information. And then if the spatial information is joined into hyperspectral applications, better performance can be potentially achieved. Some research on spectral-spatial integration has been implemented in hyperspectral image processing algorithms. Endmember extraction methods such as Abundance-Constrained Endmember Extraction (ACEE) [38] and Automatic Morphological Endmember Extraction (AMEE) [39] algorithms join spectral and spatial information and obtain improved performance than conventional methods. Some researchers also use spectral information as well as spatial information to achieve further improvement in classification [40], [41]. Anomaly detection algorithms mainly utilize two ways to integrate spectral and spatial information: model based methods and window filtering. Anomaly detectors such as Gaussian Markov Random Field (GMRFAD) based [42], [43], [44] and wavelet domain GARCH random field based [45] anomaly detection methods take on different spatial models to extract spatial features in hyperspectral image and detect anomalies that do not submit to these models. Nevertheless, most current spectral-spatial information based target detectors are more inclined to be the window filtering methods which extract adjacent local background spatial feature. For example, dual window has been employed by Kernel-RX [46] anomaly detector. Stochastic Spatial Filtering has been implemented in target detection (SSFTD) [47]. Other examples include Whitening Spatial Correlation Filtering has been utilized in WSCFAD [48] and Nested Spatial Window-based Anomaly Detector (NSWAD) [49]. Multiple-Window Anomaly Detector (MWAD) [50] also get significant performance by employing different windows statistic methods. Hence, spectral-spatial integration is a tendency in anomaly detection domain. The meaning of setting a chosen-size window in this kind of detection methods is to establish a profitable local background statistics estimation which utilizes adequate background information and meanwhile try to reduce the influence of the anomaly information. Then, the window size should not be too large or too small to obtain considerable background estimation. Most window filtering-based anomaly detection algorithms usually set only one sliding window to acquire neighborhood background statistic for the pixel under test (PUT). This kind of window detection strategy may contain anomalies in the test pixel centered window so that local background statistic will possibly be contaminated by some anomaly targets seriously. If the local distributions of some windows are most occupied by anomaly pixels then it is difficult to detect this kind of multi-pixel anomaly target because the background statistic will be contaminated seriously by anomaly pixels. Moreover, it is not convincing to set only one window to represent the local background information for the tested pixel concentrated in the center of window because the quantity of pixels in window filter is often so small that they can’t represent the local background distribution. Supposing that the multiple windows containing pixels under test are chosen to represent the local background distribution, it will be more effective to detect multi-pixel anomaly target. This paper proposes a novel anomaly detection method based on the multiple adjacent windows which also consider about the local summation strategy to integrate both the spectral and spatial information.

The rest of the paper is organized as follows: In Section 2, problems in conventional spectral-spatial integrated methods will be presented. Section 3 provides implementation of local summation anomaly detector. Section 4 demonstrates the experiments of the proposed algorithm compared with some traditional anomaly detection algorithms. Finally, Section 5 draws our conclusions.

Section snippets

Spectral-spatial integration

Some anomaly detection algorithms not only use spectral information but also spatial information to achieve further improvement in detection performance. In the following, Section 2.1 takes local-RX anomaly detection algorithm as an example to demonstrate how spectral and spatial information can be integrated in anomaly detection domain and Section 2.2 demonstrates some problems in window-based spectral-spatial integrated anomaly detection algorithms.

Local summation anomaly detector

The objective of hyperspectral anomaly detection is to identify the anomalous ground surface objects with significantly different spectra relative to the major spectral features in hyperspectral dataset. Then anomaly target pixels are usually seen as outliers in the manifold of hyperspectral data. Fig. 4 illustrates data distribution for anomaly detection in the hyperspectral manifold with a simplified 3-bands feature space. Blue and red dots represent the major background samples and anomaly

Experiments and analysis

In this section, we evaluate the detection performance of the Global-RX (GRX), Local-RX (LRX), subspace-based RX (SSRX) [57], blocked adaptive computationally efficient outlier nominator (BACON) [58] and the proposed LSAD algorithms on a simulated PHI data and real AVIRIS and HYDICE hyperspectral images. In order to quantitatively evaluate detection performance of these algorithms, receiver operating characteristic (ROC) curves [59] are used to compare the different algorithms. Based on the

Conclusion

This paper proposes an LSAD algorithm for hyperspectral anomaly detection. This method exploits a local summation anomaly detection strategy which implements a series of the local windows which contain the pixel under test to obtain various local background distributions to compute background feature statistics for the pixels under test. Some other operations such as edge expansion and subspace feature projection are included in this strategy to solve some existing problems in the conventional

References (68)

  • W. Liu et al.

    Multiview Hessian discriminative sparse coding for image annotation

    Comput. Vis. Image Understand.

    (2014)
  • W. Liu et al.

    Multiview Hessian regularized logistic regression for action recognition

    Signal Process.

    (2015)
  • N. Billor et al.

    BACON: blocked adaptive computationally efficient outlier nominators

    Comput. Stat. Data Anal.

    (2000)
  • W. Liu et al.

    Multiview Hessian regularization for image annotation

    IEEE Trans. Image Process.

    (2013)
  • D. Tao et al.

    General tensor discriminant analysis and gabor features for gait recognition

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2007)
  • D. Tao et al.

    Asymmetric bagging and random subspace for support vector machines-based relevance feedback in image retrieval

    IEEE Trans. Pattern Anal. Mac. Intell.

    (2006)
  • D. Tao et al.

    Geometric mean for subspace selection

    IEEE Trans. Pattern Analy. Mach. Intell.

    (2009)
  • C. Xu et al.

    Multi-view Intact Space Learning

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2015)
  • C. Xu et al.

    Large-Margin multi-view information bottleneck

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2014)
  • Gustavo Camps
    (2011)
  • John A. Richards

    Remote sensing digital image analysis: an introduction

    (2013)
  • S.M. Schweizer et al.

    Hyperspectral imagery: clutter adaptation in anomaly detection

    IEEE Trans. Inform. Theory

    (2000)
  • A. Plaza et al.

    Recent advances in techniques for hyperspectral image processing

    Remote Sens. Environ.

    (2009)
  • G. Shaw et al.

    Signal processing for hyperspectral image exploitation

    IEEE Signal Process. Mag.

    (2002)
  • D. Landgrebe

    Hyperspectral image data analysis

    IEEE Signal Process. Mag.

    (2002)
  • A. Plaza et al.

    Parallel hyperspectral image and signal processing

    IEEE Signal Process. Mag.

    (2011)
  • Prudhvi Gurram

    Algorithms for multispectral and hyperspectral image analysis

    J. Electr. Comput. Eng.

    (2013)
  • I.C. Chein et al.

    Real-time processing algorithms for target detection and classification in hyperspectral imagery

    IEEE Trans. Geosci. Remote Sens.

    (2001)
  • I.C. Chein et al.

    Anomaly detection and classification for hyperspectral imagery

    IEEE Trans. Geosci. Remote Sens.

    (2002)
  • M.T. Eismann et al.

    Automated hyperspectral cueing for civilian search and rescue.

    Proc. the IEEE

    (2009)
  • D. Manolakis

    Taxonomy of detection algorithms for hyperspectral imaging applications

    Opt. Eng.

    (2005)
  • D. Manolakis et al.

    Hyperspectral image processing for automatic target detection applications

    Linc. Lab. J.

    (2003)
  • N. Keshava et al.

    Spectral unmixing

    IEEE Signal Process. Mag.

    (2002)
  • A. Banerjee et al.

    A support vector method for anomaly detection in hyperspectral imagery

    IEEE Trans. Geosci. Remote Sens.

    (2006)
  • L. Edisanter et al.

    Hyperspectral anomaly detection based on minimum generalized variance method

    SPIE Process.

    (2008)
  • G. Yanfeng et al.

    A Selective KPCA algorithm based on high-order statistics for anomaly detection in hyperspectral imagery

    IEEE Geosci. Remote Sens. Lett.

    (2008)
  • P. Gurram et al.

    Sparse kernel-based hyperspectral anomaly detection

    IEEE Geosci. Remote Sens. Lett.

    (2012)
  • S. Matteoli et al.

    A tutorial overview of anomaly detection in hyperspectral images

    IEEE Aerosp. Electron. Syst. Mag.

    (2010)
  • D.W.J. Stein et al.

    Anomaly detection from hyperspectral imagery

    IEEE Signal Process. Mag.

    (2002)
  • D. Bo et al.

    Random-selection-based anomaly detector for hyperspectral imagery

    IEEE Trans. Geosci. Remote Sens.

    (2011)
  • D. Manolakis et al.

    Detection algorithms for hyperspectral imaging applications

    IEEE Signal Process. Mag.

    (2002)
  • P. Goovaerts, G. Jacquez, et al., Detection of local anomalies in high resolution hyperspectral imagery using...
  • Liangpei Zhang et al.

    Tensor discriminative locality alignment for hyperspectral image spectral-spatial feature extraction

    IEEE Trans. Geosci. Remote Sens.

    (2013)
  • M.J. Carlotto

    A cluster-based approach for detecting man-made objects and changes in imagery

    IEEE Trans. Geosci. Remote Sens.

    (2005)
  • Cited by (81)

    • Hyperspectral anomaly detection via fractional Fourier transform and deep belief networks

      2022, Infrared Physics and Technology
      Citation Excerpt :

      The pixels with spectral or spatial information clearly distinguished from the background are identified as the anomaly targets [5]. In application, it is not easy to get the spectral information of the ground objects, and AD, as a more practical category, is widely used [6]. In recent decades, lots of AD algorithms have appeared based on HSI.

    • Multi-level graph learning network for hyperspectral image classification

      2022, Pattern Recognition
      Citation Excerpt :

      However, the similarity occurring in the spectral bands among different land covers makes the classification task challenging. Confronted with this circumstance, spatial context is incorporated to generate discriminative spectral-spatial features [5], such as spatial-based filtering techniques [6], which generally employ handcrafted features conjuncted with a classifier with predefined hyperparameters, and thereby requiring massive experts’ experience [7]. Recently, deep learning methods, which act automatically to generate more robust and expressive feature representations than the handcrafted ones, have demonstrated their potentials in modeling the spectral-spatial features of HSI [8].

    View all citing articles on Scopus

    This work was supported in part by the National Basic Research Program of China (973 Program) under Grants 2011CB707105 and 2012CB719905, the National Natural Science Foundation of China under Grants 61471274 and 41431175, the Natural Science Foundation of Hubei Province, China under Grants 2014CFB193 and the Fundamental Research Funds for the Central Universities under 211-274175.

    View full text