The geographical weighted K-NN classifiers in land cover classification from remote sensing image: A case study of a subregion of Xi'an, China

The classification of land cover is one of the most important objectives of remote sensing. Class-conditional probability plot has been presented to documentation classification. In this paper, we try to incorporate two geostatistical models (Exponential model and Gaussian model) into a supervised k-nearest neighbor (k-NN) classifier to improve the accuracy of land cover classification. A subregion of Xi'an city (multispectral quickbird satellite image, 2.4m spatial resolution) is taken as an example to illustrate the validation of these land cover classification methods. The geographical weighting k-NN classifiers have been demonstrated that the accuracy of classification of land cover is very high, which is up to 91.58 percent. In addition, this classifier has eliminated the salt-and-pepper effect of the remote sensing image to some degree.