Change detection in periurban areas based on contextual classification

Abstract

This paper presents a methodology for change detection in peri-urban areas using high spatial resolution image and lidar data, founded on object-based image classification and a comparison of the classification results from two epochs. The definition of the objects is based on cadastral boundaries obtained from a geospatial database. An exhaustive set of descriptive features is computed, characterising each object for both epochs regarding spectral, texture, geometrical, and three-dimensional (3D) aspects. In addition, contextual features describing the object at two levels are defined. Internal context features describe the relations between different land cover elements within the object, whereas external context features describe each object considering the common properties of neighbouring objects, usually coinciding in urban areas with an urban block. Both the classification and the change detection process are thoroughly evaluated, and the specific contribution of 3D features to the accuracy of the processes is analysed. The results show that 3D information enables to improve the classification results, remarkably increasing the accuracy values of certain classes, and allowing for an enhanced discrimination of building typologies. Moreover, the change detection efficiency is notably improved by a significant reduction of both commission and omission errors.