A Remote Sensing Image Segmentation Method Based on Spectral and Texture Information Fusion

Since inadequacy information from spectral characteristics for very high resolution remote sensing multispectral imagery segmentation/classification, we propose the combination of spectral feature which was extracted by a variable mean shift clustering algorithm and spatial features by Gabor filter banks and support vector machine is employed to achieve feature fusion and classification. Some issues on feature dimension reduction and parameter estimation in mean shift procedure are discussed. The whole algorithm is evaluated on the synthetic texture image, which are cropped from a QuickBird image with typical land-cover types. The result show an improvement in classification of land cover classes having similar spectral surface.