A Geostatistically Weighted k-NN Classifier for Remotely Sensed Imagery

This study aims to increase the accuracy with which remotely sensed data can be used to generate thematic maps of land cover classes. It explores the use of geostatistical models to characterize the inherent spatial variation between different land covers (woodland, rough grassland, managed grassland, and built land) and integrates these into a supervised, nonparametric, k-nearest neighbor (k-NN) per-pixel classifier. The study defines three geographical weighting methods, two of which are based on geostatistical functions. These produce a geographical weighting that is incorporated into two pure k-NN classifiers (inverse distance weighted and difference distance weighted) using a scheme that allows the weights for the information from feature space and geographical space to be varied. The relative merits of the enhanced approach are explored using a spatially and spectrally variable IKONOS subscene. Compared with the original k-NN classifications, which use only the information in the spectral response of pixels treated independently, a statistically significant increase in the overall accuracy was achieved, particularly for land cover classes with considerable within-class variation and between-class confusion.