Building extraction and change detection from remotely sensed imagery based on layered architecture

Abstract

The processing and analysis of remotely sensed imagery (RSI) is a research hotspot in the information field, and building extraction and change detection are some of the difficult problems. In order to make the maximum use of the effective characteristics and design independently the algorithm of feature extraction, an approach to building extraction and change detection from RSI based on layered architecture containing pixel layer, object layer and configuration layer is proposed. In the pixel layer, the input image is over-segmented and under-segmented, respectively, by a quantity-controllable algorithm using super-pixel segmentation to obtain the segmentation object sets, with which the input image is decomposed into shadow layer, homogeneity layer and edge layer, where the buildings are extracted based on the spatial relationship between the feature areas and segmentation objects. In the object layer, for preserving the accurate contour of the buildings, a new segmentation method based on the traditional graph-cut theory and mathematical morphology is introduced, and then, the buildings extracted from each layer are merged. Finally, in the configuration layer, the change information is detected using spatial relationship of buildings between the old image and the new one. The experimental results reveal that the building contour is extracted accurately, and three types of change including the newly built, the demolished and the reconstructed buildings can be detected; in addition, there is no strict requirement for registration accuracy. For the test images, the overall performance F1 of the building extraction is over 85, and the precision and recall of the change detection are both higher than 90%.