Object-based sub-pixel mapping of buildings incorporating the prior shape information from remotely sensed imagery

Sub-pixel mapping (SPM) is a promising method to predict the spatial locations of land cover classes at the sub-pixel scale for remotely sensed imagery, using the fraction images generated by soft classification as input. At present, SPM treats all sub-pixels of different land cover classes in the same strategy by maximizing their spatial dependence. Although the maximal spatial dependence is a simple method to describe the spatial pattern of land cover classes and has been proved to be an effective principle for SPM, it does not reflect real-world situations. Given that spatial patterns are land cover class- or object-specific, each land cover class or object should be designated its own specific spatial pattern description when SPM is applied. In this paper, a novel object-based sub-pixel mapping (OBSPM) method was proposed to map buildings at the sub-pixel scale. On the basis of the prior information of the building shape (i.e., the building boundaries are parallel or perpendicular to the main orientation), a novel anisotropic spatial dependence model is adopted in the SPM procedure. The proposed OBSPM model includes three main steps: building segmentation, building feature extraction, and anisotropic SPM of buildings. The proposed model is evaluated with a simulated synthetic image and an actual AVIRIS image. The results show that OBSPM obtains more accurate building maps than do conventional SPM models, and the accuracy of fraction images and the spatial resolutions of remotely sensed images are two crucial factors that influence the OBSPM results. Furthermore, extending the OBSPM model to more land cover classes to incorporate more specific prior information is a promising method in enhancing the applicability of SPM to practical situations.

► An object-based sub-pixel mapping method was proposed to map buildings at the sub-pixel scale.
► An anisotropic spatial dependence model is adopted to incorporate the prior information.
► Three steps are included: segmentation, feature extraction, and anisotropic SPM.
► Fraction image errors and the image spatial resolutions are two crucial factors.