| Abstract [eng] |
Remotely sensed image classification is probably one of the most important strategic notion in the remotely sensed data research. Constructed thematic maps have been widely used for urban areas, forests, glaciers change detection, as well as for crop or soil evaluation. In this work various spatial features have been analised in order to rate the impact on classification of remotely sensed images. Classical models are build based on the spectral values of an image. However it has been criticised due to lack of spatial information. There is a review of the variogram and its application to remotely sensed image classification at the beginning of the paper. Variogram fitting is one of the methods of extracting texture parameters. In this case, coefficients of a theoretical model set up the spatial features. Secondly, image texture can be described by simple parameters derived from empirical variogram values. The relevance of the proposed features for remotely sensed imgae classification has been evaluated using satellite images of Kaunas city. The classifiers are constructed by support vector machines models with a particular kernel function. After the comparison of classification by different features the conclusion is that texture information description by derived variogram parameters is easier and more accurate. In addition, the most effective remotely sensed image classification was achieved by combining spatial and spectral features. |
