Mark Schmalz (University of Florida), Gary Key (Frontier Technology, Inc.)
Keywords: NROC, Non-resolved Object Characterization
Abstract:
Accurate spectral signature classification is key to the nonimaging detection and recognition of spaceborne objects. In classical hyperspectral recognition applications, signature classification accuracy depends on accurate spectral endmember determination [1]. However, in selected target recognition (ATR) applications, it is possible to circumvent the endmember detection problem by employing a Bayesian classifier. Previous approaches to Bayesian classification of spectral signatures have been rule- based, or predicated on a priori parameterized information obtained from offline training, as in the case of neural networks [1,2]. Unfortunately, class separation and classifier refinement results in these methods tends to be suboptimal, and the number of signatures that can be accurately classified often depends linearly on the number of inputs. This can lead to potentially significant classification errors in the presence of noise or densely interleaved signatures.
In this paper, we present an emerging technology for nonimaging spectral signature classfication based on a highly accurate but computationally efficient search engine called Tabular Nearest Neighbor Encoding (TNE) [3]. Based on prior results, TNE can optimize its classifier performance to track input nonergodicities, as well as yield measures of confidence or caution for evaluation of classification results. Unlike neural networks, TNE does not have a hidden intermediate data structure (e.g., the neural net weight matrix). Instead, TNE generates and exploits a user-accessible data structure called the agreement map (AM), which can be manipulated by Boolean logic operations to effect accurate classifier refinement algorithms. This allows the TNE programmer or user to determine parameters for classification accuracy, and to mathematically analyze the signatures for which TNE did not obtain classification matches. This dual approach to analysis (i.e., correct vs. incorrect classification) has been shown to significantly strengthen analysis of classifier performance in support of classifier optimization.
We show that AM-based classification can be modified to include dynamic tracking of input statistical changes, to achieve accurate signature classification in the presence of noise, closely spaced or interleaved signatures, and simulated optical distortions. In particular, we examine two critical cases: (1) classification of multiple closely spaced signatures that are difficult to separate using distance measures, and (2) classification of materials in simulated hyperspectral images of spaceborne satellites. In each case, test data are derived from a NASA database of space material signatures. Additional analysis pertains to computational complexity and noise sensitivity, which are superior to Bayesian techniques based on classical neural networks.
%Z Winter, M.E., 1999, Fast autonomous spectral end-member determination in hyperspectral data, Proceedings of the 13th International Conference On Applied Geologic Remote Sensing, Vancouver, B.C., Canada, pp. 337-44
N. Keshava, 2003, A survey of spectral unmixing algorithms, Lincoln Laboratory Journal 14:55-78
Key, G., Schmalz, M.S., Caimi, F.M., & Ritter, G.X., 1999, Performance analysis of tabular nearest neighbor encoding algorithm for joint compression and ATR, Proceedings SPIE 3814:115-126
Date of Conference: September 12-15, 2007
Track: Non-resolved Object Characterization