Remotely sensed image intelligent interpretation based on robust segmentation and GIS system

The continuous improving spatial resolution of remote sensing sensors sets new demand for applications utilizing this information. For mining useful information from remote sensing data, many classifiers based on the spectral analysis of individual pixels have been proposed and significant progress has been achieved. However, these approaches have their limitations, usually they produce “salt and pepper” noisy results. To conquer such problems, object-oriented image analysis method based on multi-resolution segmentation technique was put forward and it has been utilized for various different application purposes successfully. In this study, an efficient remotely sensed image intelligent interpretation method based on image segmentation and geographical information system (GIS) was proposed. First, segmentation based on mean shift was employed to gain the initial segments from remote sensing images. Then, apply vectorization (Raster to Vector Convertor) method to generate polygons from the segmented image and feature attributions such as spectral, shape, texture etc. are extracted by zonal analysis based on original raster and polygons. Finally, creating training sample and supervised classification are implemented. Nearly all steps are achieved in geo-information system except image segmentation. On the basis of the study, we developed a software system of remotely sensed image analysis. Compared with the interpretation approach of a commercial software eCognition, the proposed one was feasible and efficient when applied to the Quickbird remotely sensed images.