A texture classification experiment for SAR radar images

This paper presents the first results obtained in a study on synthetic aperture radar (SAR) image processing in the framework of using the SAR data for conservation site and monitoring. A comparative study of median filtering, adaptive Lee filtering, and lineament enhancement was done with emphasis on the applicability to speckle elimination and texture edges preserving. The result shows that the adaptive Lee filtering preserves the texture edges best followed by the median filtering. A comparative study of the texture features based on a concept of texture units, the texture features based on the co-occurrence matrix, and the texture features based on the use of local statistics was also done with emphasis on the applicability to object differentiation. Random sampling, clustering, and minimum distance classification techniques were used for the image segmentation. At this stage of experiment, it could be recommended that the geometric symmetry and the degree of direction features of texture units, the contrast and the inverse difference moment features of the co-occurrence matrix, the local mean and the local ratio between the standard deviation and the mean value are potential features for SAR image classification. A combination of filtered image tone and texture features in some cases may improve the classification accuracy.