A New Better-Fit Decision Features Selection Method for C5.0 Decision Tree

Abstract

C5.0 decision tree method has a disadvantage that it’s difficult to select better-fit decision features, so a better-fit decision features selection (abbreviated as BFDFS) methods are proposed in this paper. The procedure of BFDFS is as follows: (1) all decision features are pre-processed and then integrated into one image file; (2) interesting regions of typical kinds of land objects are chosen; (3) values of all features in interesting regions are computed according the "3σ" theory; (4) all ratios of different objects features’ values are computed and sorted to obtain the better-fit decision features; (5) using the better-fit decision features to build C5.0 decision tree rules. The BFDFS is tested by the use of classifying the rubber woods from high resolution remote sensing images, experimental area is selected in Guangba Farm, Dongfang City, Hainan Island, China. The results indicate that the producer accuracy, user accuracy, total accuracy of rubber woods, and the Kappa coefficient are 88%, 91.67%, 92%, and 0.89, respectively. All four indices are better than the other classification methods, proving the feasibility and efficiency of the BFDFS method.