| Summary: | Nonparametric nearest neighbor classification and a post-classification contextual correction can be used successfully to classify multispectral images. Accuracy is similar to that of parametric quadratic discriminant classifiers if the training set is well-defined and much better if the training set is not well-defined. Before 1-NNR classification, training set is redefined by selecting a reduced and representative subset. After classification, a contextual correction is performed in order to to get homogeneous spatial classes, improving the accuracy and credibility of classification. The proposed methodology is tested on a Landsat-5 TM image of the Ymer Ø region (Greenland, Denmark). 1 Introduction Classification of remote sensed images is an important topic in digital image processing as it has important economic implications and numerous practical applications. Basically, a remote sensed image consist of measurements of refectance in several spectral bands so we have a vector X as.
|
|---|