Fuzzy Linear Regression Analysis of Land-cover Classification Using Remotely Sensed Data

Abstract

In this study developed is a new method for land-cover classification of remotely sensed data. The proposed method bases its modeling structure on so-called category decomposition principle, estimating categorical proportions of a mixel rather than assigning only one category to it. Classification algorithm is established in the framework of fuzzy linear regression analysis so that observed radiance data of the subject mixel may be contained, at least to a predetermined degree, in the estimated fuzzy set of radiance which is accomplished by the weighted sum of category specific spectra expressed by fuzzy numbers. Validity tests using Landsat TM data show that the proposed method substantially outperforms three existing methods compared, the quadratic programming method based on category decomposition idea, the linear discriminant analysis method and the maximum likelihood method, by up to 40% in terms of root mean square error of estimation and also the proposed method has a kind of robustness in that it can make relatively accurate estimates compared to the existing methods even though inexact categorical characteristics of spectra are input to the model.