Effect of Linearly-Combined Variables in Pattern Discrimination and a Method of Variable Selection
Masanobu YOSHIKAWA, Nobuhito TAKAMURA, Sadao FUJIMURA, Ryuei Nishii, Izumi Miyamoto, Shojiro Tanaka · Transactions of the Society of Instrument and Control Engineers · 1999
A new feature extraction method for land-cover classification is proposed. New variables are produced by a simple linear combination of two variables in the original measurement vector. They correspond to the first and second principal componets. At first, an equivalent relation is proven to exist in coefficient matrices for the proposed variable sets. Variable selection is employed for representatives of the largely-reduced equivalent classes. The total number of selection is defined using the Stirling number of the second kind. Secondly, strategies for selecting quasi-optimal variables for discriminant analysis are described. Akaike's Information Criterion (AIC) for linear discriminant functions is utilized in selection. Results of experiments by simulation data and real remotely-sensed data show effective performance of the proposed method.