Semi-supervised method for land cover classification of remotely sensed image considering spatial arrangement of the pixels

We can recognize the objects in the remotely sensed image by classifying the pixels into several categories. The pixels are usually classified based on spectral characteristics of objects on the land surface. The supervised method is often used for classification when objective categories are prescribed and adequate training data are available. However, it is often difficult to provide sufficient training data which are indispensable for accurate classification by the supervised method. This situation often causes considerable errors in classification results. We propose here a semi-supervised method in which the unsupervised clustering is employed to enhance the training data for the supervised classification. We extract several clusters out of the image considering both of the spectral and the spatial information. Then we extract additional training data from several clusters and use them for the supervised classification. The accuracy of classification was confirmed to be improved by using the semi-supervised method.