Mark M. Baumback

A classification approach and comparison to other object identification algorithms for hyperspectral imagery

This study adapts a variety of techniques derived from multi-spectral image classification to find objects amid cluttered backgrounds in hyperspectral imagery. This study quantitatively compares the algorithms against a standard object search, the matched filter (MF) and recently developed object detector, Adaptive Cosine Estimator (ACE). These object searches require calculating the Mahalanobis distance between the average object spectral signature and the test pixel spectrum and needs the computation of a covariance matrix. The covariance matrix is generated using the entire image (Whitened Euclidean Distance, WED) or using pixels associated with the object (Maximum Likelihood Classifier, MLC). The latter computation requires a relatively large number of pixels to generate a non-singular, accurate covariance matrix. Regularizing object pixels via optimally mixing (likelihood maximization) diagonal, object, and entire image covariance matrices to generate the object covariance matrix estimate. This approximation is called the Regularized Maximum Likelihood Classifier (RMLC). The object searches MF, ACE, WED, MLC, and RMLC were applied to visible/near IR data collected from forest and desert environments. This study searched for objects using object signatures and covariance matrices taken directly from the scene and from statistically transformed object signatures and covariance matrices from another time. This study found a substantial reduction in the number of false alarms (factor of 10 to 1000) using WED, ACE, RMLC relative to MF searches for the two independent data collects. The regularization of in-scene and transformed covariance matrices substantially reduced false alarms relative to using unprocessed covariance matrices. This study adds simple, high performing algorithms to the object search arsenal.

Fusion and normalization to enhance anomaly detection

This study examines normalizing the imagery and the optimization metrics to enhance anomaly and change detection, respectively. The RX algorithm, the standard anomaly detector for hyperspectral imagery, more successfully extracts bright rather than dark man-made objects when applied to visible hyperspectral imagery. However, normalizing the imagery prior to applying the anomaly detector can help detect some of the problematic dark objects, but can also miss some bright objects. This study jointly fuses images of RX applied to normalized and unnormalized imagery and has a single decision surface. The technique was tested using imagery of commercial vehicles in urban environment gathered by a hyperspectral visible/near IR sensor mounted in an airborne platform. Combining detections first requires converting the detector output to a target probability. The observed anomaly detections were fitted with a linear combination of chi square distributions and these weights were used to help compute the target probability. Receiver Operator Characteristic (ROC) quantitatively assessed the target detection performance. The target detection performance is highly variable depending on the relative number of candidate bright and dark targets and false alarms and controlled in this study by using vegetation and street line masks. The joint Boolean OR and AND operations also generate variable performance depending on the scene. The joint SUM operation provides a reasonable compromise between OR and AND operations and has good target detection performance. In addition, new transforms based on normalizing correlation coefficient and least squares generate new transforms related to canonical correlation analysis (CCA) and a normalized image regression (NIR). Transforms based on CCA and NIR performed better than the standard approaches. Only RX detection of the unnormalized of the difference imagery in change detection provides adequate change detection performance.

Robustness tests for object identification algorithms in hyperspectral imagery

A previous study adapted a variety of techniques derived from multi-spectral image classification to find objects amid cluttered backgrounds in hyperspectral imagery. That study quantitatively compared the adapted algorithms against a standard object search, the matched filter (MF) and a recently developed object detector, Adaptive Cosine Estimator (ACE) and found substantial reduction in false alarm rates for a given target detection probability. One adapted object search, Regularized Maximum Likelihood Classifier (RMLC), requires calculating the covariance matrix involving the average object spectral signature and the target pixels. The object covariance matrix requires a relatively large number of pixels to generate a non-singular, accurate covariance matrix. This study examines the robustness of the RMLC algorithm on number of training pixels, the optimal mixing covariance matrices, and choice of object subspaces for the ACE algorithm. The tests were applied to visible/near IR data collected from forest and desert environments. This study finds that high detection performance standards for RMLC are invariant for pixel number for homogenous targets, down to two pixels. Regularization is relatively unaffected by the choice of areas to optimize the object covariance matrix although targets that mix background appear to be more sensitive to choice of covariance matrix. Reducing the object subspace dimensions by using the average target signature or choosing the first principle component enhances ACE performance relative to using the entire object space.