ANOMALY DETECTION IN COMPLEX ENVIRONMENTS: EVALUATION OF THE INTER- AND INTRA-METHOD CONSISTENCY D. Borghys, E. Truyen, M. Shimoni Royal Military Academy Signal & Image Centre Brussels, Belgium C. Perneel Royal Military Academy Dept. of Appl. Mathematics Brussels, Belgium ABSTRACT Many anomaly detection methods, depending on various pa- rameters, have been proposed in literature. Given the diversity of available anomaly detectors, from an operational viewpoint it is interesting to determine an efficient strategy to find the best suited detector for a given application. This is not obvi- ous, especially in scenes with a highly structured background. The work presented here proposes a generic approach to the problem by examining the following questions: How differ- ent are the results of the various anomaly detectors ? Are the parameters influencing the results significantly ? Are there classes of methods sufficiently similar so that one can test only one of each class and see which results are most ade- quate for a given application ? What are the spectral/spatial characteristics of the differences between methods ? Can one predict which detector will give the best results for a given application ? The current paper tries to answer the first three questions by comparing results of different types of anomaly detectors applied to different complex (urban, industrial and harbor) scenes. The comparison is not in absolute terms be- cause it does not rely on a priori ground truth. In stead the detectors are compared relative to one another, the aim be- ing to evaluate the similarities between the performance of the detectors and the dependency of their results on the used parameters, i.e. the inter- and intra method consistency. Index Terms— Anomaly detection, hyperspectral, clus- tering, segmentation 1. INTRODUCTION Anomaly detection in hyperspectral data has received a lot of attention for various applications. The aim of anomaly detec- tion is to detect pixels in the hyperspectral datacube whose spectra differ significantly from the background spectra. In anomaly detection, in contrast to target detection, no a pri- ori knowledge about the target is assumed [1]. Anomaly de- tection methods in general estimate the spectra of the back- ground (locally or globally) and then detect anomalies as pix- els with a large spectral distance w.r.t. the determined back- ground spectra. Many types of anomaly detectors have been proposed in literature, each depending on several parameters. The Reed Xiaoli (RX) algorithm [2], which is the benchmark anomaly detector for hyperspectral imagery, models the local background by a multi-variate normal distribution. Sub-space detectors form a very interesting family of anomaly detectors. In [3] several linear and non-linear (kernelized) sub-space detectors are compared and the Kernel Principal Component Analysis (KPCA) [4] based detector was found to be very promising for the investigated scenes. All of these methods detect anomalies by considering the spectral difference be- tween the current pixel and its immediate surroundings. An alternative approach for anomaly detection consists in apply- ing a scene segmentation prior to the actual anomaly detec- tion. Several types of segmentation-based anomaly detection (SBAD) methods have been proposed [5]-[8]. These meth- ods seem promising for anomaly detection in complex envi- ronments [5, 8] because they estimate a set of background spectra globally over the image. In this paper the similari- ties between different anomaly detectors and the dependence of their results on their parameters. Three classes of detec- tors are examined. The first class is specifically designed for complex environments and is based on a two-level clus- tering scheme. The second class consists of global image segmentation-based methods. The third class is based on lo- cal statistics. From each class several detectors were selected for the comparison. The test dataset consists of six datacubes, acquired by four different airborne sensors, showing scenes of diverse complexity. 2. DATASET The presented analysis was performed on a set of 6 hyper- cubes of scenes with various complexity, acquired by 4 dif- ferent airborne sensors. Table 1 presents the main character- istics of the dataset. The first column is the name by which the scenes will be referred further in this paper. 3. ANOMALY DETECTION METHODS The three classes of anomaly detectors examined in this pa- per are briefly described below. The main parameters of each