Selection and evaluation of optimal segmentation scale for high-resolution remote sensing images based on prior thematic maps and image features

Abstract

Multiscale segmentation is the premise and key step of geographic object-based image analysis (GEOBIA), but scale selection remains a challenge in multiscale segmentation. Over the years, scale selection and evaluation in image segmentation has been extensively explored and many methods have been developed. In these methods, when a scale is chosen or evaluated, all the features are generally extracted from images. In addition, an optimal scale is generally selected based on the pre-estimation of the statistical variance in remote sensing images or determined based on the postsegmentation evaluation of segmented results. In this study, a method was proposed to identify the optimal scale of each segmented object during the segmentation through combining the a priori thematic map knowledge with image features. First, 25 image segmentations were obtained using multiresolution segmentation algorithm of Definiens Professional 9.0 with different scales. A global score (GS) value was assigned to each segmentation based on the calculation results of the weighted variance and global Moran’s I and the single-scale optimal segmentation result was determined according to each GS of 25 segmentation scales. Second, the image feature complexity information and the a priori thematic map complexity information of each segmentation object were extracted to calculate complexity values for each object. Third, the optimal scale of each segmentation object was determined through the iterative calculation with the multithreshold method. Finally, the segmentation results of the proposed method were evaluated. The proposed method was applied to process Gaofen-2 (GF-2), GF-1, Korea Multipurpose Satellite (KOMPSAT-2), IKONOS, QuickBird and WorldView-2 high resolution satellite images to obtain the segmentation results and classification results, compared with results obtained of the optimal single-scale segmentation and the unsupervised evaluation method. The experimental results of GEOBIA showed that the method was helpful for generating the segmentation object with the optimal scale.