Mark A. Friend

Further extensions to robust parameter design: three factor interactions with an application to hyperspectral imagery

Hyperspectral imagery (HSI) provides opportunities for locating anomalous objects through the use of multivariate statistics. Global anomaly detectors, such as the autonomous global anomaly detector (AutoGAD), require the user to provide various parameters/thresholds to analyse an image. These user-defined settings can be thought of as control variables and properties of the imagery can be employed as noise variables. The presence of these factors suggests the use of robust parameter design (RPD) to locate the best settings for the algorithm. Mindrup et al. (2012) showed that the standard RPD model might not be sufficient for use with more complex data and extended the model to include noise by noise interactions. This paper extends the model to include control by noise by noise and noise by control by control interactions. These new models are then applied to AutoGAD output and the Lin and Tu MSE method is employed to locate optimal settings.

Clustering hyperspectral imagery for improved adaptive matched filter performance

Abstract This paper offers improvements to adaptive matched filter (AMF) performance by addressing correlation and non-homogeneity problems inherent to hyperspectral imagery (HSI). The estimation of the mean vector and covariance matrix of the background should be calculated using “target-free” data. This statement reflects the difficulty that including target data in estimates of the mean vector and covariance matrix of the background could entail. This data could act as statistical outliers and severely contaminate the estimators. This fact serves as the impetus for a 2-stage process: First, attempt to remove the target data from the background by way of the employment of anomaly detectors. Next, with remaining data being relatively “target-free” the way is cleared for signature matching. Relative to the first stage, we were able to test seven different anomaly detectors, some of which are designed specifically to deal with the spatial correlation of HSI data and/or the presence of anomalous pixels in local or global mean and covariance estimators. Relative to the second stage, we investigated the use of cluster analytic methods to boost AMF performance. The research shows that accounting for spatial correlation effects in the detector yields nearly “target-free” data for use in an AMF that is greatly benefitted through the use of cluster analysis methods.

Extending robust parameter design to noise by noise interactions with an application to hyperspectral imagery

Anomaly detection algorithms for hyperspectral imagery (HSI) are an important first step in the analysis chain. Improvements to anomaly detection algorithm effectiveness directly reduce the amount of data that must go through time intensive processing before final analysis. The effectiveness of most anomaly detection algorithms is a function of user selected algorithm parameters, or controls, and uncontrollable noise factors which introduce additional variance into the detection process. In the case of HSI, these noise factors are embedded in the image under consideration. Robust parameter design (RPD) offers a method to reduce the impact of these noise variables on anomaly detection algorithms by identifying control settings robust to the noise factors found in HSI. This paper extends standard RPD modelling by removing the assumption that squared noise terms and noise by noise interactions are negligible. A mean squared error approach proposed by Lin and Tu is applied which employs expected value and var...

An Entropy-based Scheme for Automatic Target Recognition:

Often the performance of a classification system is reported in terms of classification accuracy. In an environment with objects unknown to the classification system, classification accuracy may provide unrealistic expectations. In this paper we contrast classification and label accuracy in a challenging classification environment. A statistical-based method is used to identify records not represented in the template library used by the classifier and three different information theory-based methods are used to identify label records likely to be misidentified. These methods are applied to an automatic target recognition (ATR) problem, using features drawn from high-range resolution profiles generated from synthetic aperture radar (SAR) data. An optimization framework is used to select the optimal classification system choices based on the measurement of evaluation. The choices selected by the framework when classification or label accuracy is the optimization focus are contrasted.

Small sample training and test selection method for optimized anomaly detection algorithms in hyperspectral imagery

Abstract. There are numerous anomaly detection algorithms proposed for hyperspectral imagery. Robust parameter design (RPD) techniques provide an avenue to select robust settings capable of operating consistently across a large variety of image scenes. Many researchers in this area are faced with a paucity of data. Unfortunately, there are no data splitting methods for model validation of datasets with small sample sizes. Typically, training and test sets of hyperspectral images are chosen randomly. Previous research has developed a framework for optimizing anomaly detection in HSI by considering specific image characteristics as noise variables within the context of RPD; these characteristics include the Fisher’s score, ratio of target pixels and number of clusters. We have developed method for selecting hyperspectral image training and test subsets that yields consistent RPD results based on these noise features. These subsets are not necessarily orthogonal, but still provide improvements over random training and test subset assignments by maximizing the volume and average distance between image noise characteristics. The small sample training and test selection method is contrasted with randomly selected training sets as well as training sets chosen from the CADEX and DUPLEX algorithms for the well known Reed-Xiaoli anomaly detector.

The effectiveness of using diversity to select multiple classifier systems with varying classification thresholds

In classification applications, the goal of fusion techniques is to exploit complementary approaches and merge the information provided by these methods to provide a solution superior than any single method. Associated with choosing a methodology to fuse pattern recognition algorithms is the choice of algorithm or algorithms to fuse. Historically, classifier ensemble accuracy has been used to select which pattern recognition algorithms are included in a multiple classifier system. More recently, research has focused on creating and evaluating diversity metrics to more effectively select ensemble members. Using a wide range of classification data sets, methodologies, and fusion techniques, current diversity research is extended by expanding classifier domains before employing fusion methodologies. The expansion is made possible with a unique classification score algorithm developed for this purpose. Correlation and linear regression techniques reveal that the relationship between diversity metrics and accuracy is tenuous and optimal ensemble selection should be based on ensemble accuracy. The strengths and weaknesses of popular diversity metrics are examined in the context of the information they provide with respect to changing classification thresholds and accuracies.

Forecasting effects of Miso actions: an ABM methodology

Agent-based models (ABM) have been used successfully in the field of generative social science to discover parsimonious sets of factors that generate social behavior. This methodology provides an avenue to explore the spread of anti-government sentiment in populations and to compare the effects of potential Military Information Support Operations (MISO) actions. We develop an ABM to investigate factors that affect the growth of rebel uprisings in a notional population. Our ABM expands the civil violence model developed by Epstein by enabling communication between agents through a genetic algorithm and by adding the ability of agents to form friendships based on shared beliefs. We examine the distribution of opinion and size of sub-populations of rebel and imprisoned civilians, and compare two counter-propaganda strategies. Analysis identifies several factors with effects that can explain some real-world observations, and provides a methodology for MISO operators to compare the effectiveness of potential actions.

Selecting training and test images for optimized anomaly detection algorithms in hyperspectral imagery through robust parameter design

There are numerous anomaly detection algorithms proposed for hyperspectral imagery. Robust parameter design (RPD) techniques have been applied to some of these algorithms in an attempt to choose robust settings capable of operating consistently across a large variety of image scenes. Typically, training and test sets of hyperspectral images are chosen randomly. Previous research developed a frameworkfor optimizing anomaly detection in HSI by considering specific image characteristics as noise variables within the context of RPD; these characteristics include the Fishers score, ratio of target pixels and number of clusters. This paper describes a method for selecting hyperspectral image training and test subsets yielding consistent RPD results based on these noise features. These subsets are not necessarily orthogonal, but still provide improvements over random training and test subset assignments by maximizing the volume and average distance between image noise characteristics. Several different mathematical models representing the value of a training and test set based on such measures as the D-optimal score and various distance norms are tested in a simulation experiment.