Sensitivity of multiresolution segmentation to spatial extent

Abstract Spatial extent (i.e. the size of the study area) is acknowledged as an important component of scale, together with grain (i.e. cell size). While the influence of grain on multiresolution segmentation has been evaluated, the impact of spatial extent is still poorly understood. The main goal of our study was to evaluate how changing the extent affects multiresolution segmentation, in respect to the geometric accuracy of the resulting image objects. The experiments were carried out on very-high resolution optical images in four study areas: the City of Manchester (UK), the region of Normandy (France), the City of Tampa (Florida), USA, and the province of Flevoland (the Netherlands). Data sets of various extents were created by partitioning each image into regular tiles with eCognition® Server. The smallest tile size was 100 × 100 pixels, which doubled iteratively, until no further partition was possible, so that the image was processed at its full extent. Each tile was segmented with the Estimation of Scale Parameters (ESP2) algorithm and for each of the three generated levels the degree of overlap between the image objects and the reference polygons representing buildings and crop fields was checked. Segmentation accuracy was performed with the following metrics: Area Fit Index, Under-Segmentation, Over-Segmentation, D- index, and Quality Rate. The results show that the geometric accuracy improved by 8–19% in Quality Rate when multiresolution image segmentation was performed in the smallest extent (100 × 100 pixels), as compared to the segmentation of whole images. These findings challenge previous assumptions and findings that partitioning an image into regularly-sized tiles can bias segmentation, and are relevant to guiding the setup of tile size in a distributed computing framework.