Value Focused Thinking Applications to Supervised Pattern Classification With Extensions to Hyperspectral Anomaly Detection Algorithms.

摘要

Hyperspectral imaging (HSI) is an emerging analytical tool with flexible applications in different target detection and classification environments, including Military Intelligence, environmental conservation, etc. Algorithms are being developed at a rapid rate, solving various related detection problems under certain assumptions. At the core of these algorithms is the concept of supervised pattern classification, which trains an algorithm to data with enough generalizability that it can be applied to multiple instances of data. It is necessary to develop a logical methodology that can weigh responses and provide an output value that can help determine an optimum algorithm. This research focuses on the comparison of supervised learning classification algorithms through the development of a value focused thinking (VFT) hierarchy. This hierarchy represents a fusion of qualitative/ quantitative parameter values developed with Subject Matter Expert a priori information. Parameters include a fusion of bias/variance values decomposed from quadratic and zero/one loss functions, and a comparison of cross-validation methodologies and resulting error. This methodology is utilized to compare the aforementioned classifiers as applied to hyperspectral imaging data. Conclusions reached include a proof of concept of the credibility and applicability of the value focused thinking process to determine an optimal algorithm in various conditions.