Neighborhood methods for image classification

Nonparametric classifiers predicated on spectral neighborhoods have not been widely used for pixel assignment in remote sensing. This is probably because their computational requirements make them too slow for practical classification of large images. Regardless, when properly specified, the nearest neighbor classifier is a Bayesian classifier, and does not require conditions of multivariate normality as a prerequisite for good pixel assignment. In this study, six nearest neighbor classifiers and four parametric classifiers are applied to several Landsat images embracing a broad variety of land cover. The accuracy of the classifiers is compared. In a majority of hybrid experiments, classifiers predicated on spectral neighborhoods were significantly superior to parametric classifiers when training sample proportions matched the true population proportions. A nonparametric classifier that weighted neighbor “votes” inversely by distance proved particularly accurate in classifying land cover