Adaptive Scale Selection for Multiscale Segmentation of Satellite Images

With dramatically increasing of the spatial resolution of satellite imaging sensors, object-based image analysis (OBIA) has been gaining prominence in remote sensing applications. Multiscale image segmentation is a prerequisite step that splits an image into hierarchical homogeneous segmented objects for OBIA. However, scale selection remains a challenge in multiscale segmentation. In this study, we presented an adaptive approach for defining and estimating the optimal scale in the multiscale segmentation process. Central to our method is the combined use of image features from segmented objects and prior knowledge from historical thematic maps in a top-down segmentation procedure. Specifically, the whole image was first split into segmented objects, with the largest scale in a presupposition segmentation scale sequence. Second, based on segmented object features and prior knowledge in the local region of thematic maps, we calculated complexity values for each segmented object. Third, if the complexity values of an object were large enough, this object would be further split into multiple segmented objects with a smaller scale in the scale sequence. Then, in the similar manner, complex segmented objects were split into the simplest objects iteratively. Finally, the final segmentation result was obtained and evaluated. We have applied this method on a GF-1 multispectral satellite image and a ZY-3 multispectral satellite image to produce multiscale segmentation maps and further classification maps, compared with the state-of-the-art and the traditional mean shift algorithm. The experimental results illustrate that the proposed method is practically helpful and efficient to produce the appropriate segmented image objects with optimal scales.