Reducing boundary effects in a kernel-based classifier

Abstract

In this letter, a method is presented to reduce misclassification at the boundaries between different types of land covers. Errors are caused by use of spatial features extracted from pixel neighbourhoods. Simple thresholding and region-growing techniques are used to reduce such errors. The method is included as an extension to a frequency-based contextual classifier (FBC).

Some experiments have been carried out to evaluate both the original classifier, FBC, and the extended version using an image acquired with a Compact Airborne Spectrographic Imager (CASI). Experimental results show that overall classification accuracies with the CASI image are 84-80 per cent and 89-37 per cent obtained using the FBC and the extended FBC, respectively.