Optimizing multiscale segmentation with local spectral heterogeneity measure for high resolution remote sensing images

Image segmentation is a vital and fundamental step in Geographic Object-Based Image Analysis (GEOBIA). Many multiscale segmentation algorithms have been widely used in high resolution (HR) remote sensing images. These segmentation algorithms need a preset parameter, named scale parameter, to control the average size of each object. However, due to the spatial variation, single scale parameter can hardly describe the boundaries of regions with different land covers. To overcome this limitation, in this study, an adaptive parameter optimization method is proposed for multiscale segmentation. To find the optimal scale of objects, a local spectral heterogeneity measure is applied by calculating the spectral angle between inter and intra objects. Different from selecting a global optimal scale parameter, this study aims to directly search the optimal objects from results of all different scales and combine them into final segmentation results. In experiments, a multi-resolution segmentation is used to generate segmentation results of different scales and the QuickBird-2 images are used as test data. Optimization results over four HR test images reveal that the proposed method provides better segmentation performance than single scale segmentation result.