An object-based convolutional neural network for urban building semantic classification

Building classification (including identifying geometrical shapes and their semantic labels) in aerial and satellite images is one of the most important issues in urban planning, environment management, and disaster recovery. With the fast development of remote sensing technology, both the spectral and spatial resolution of remote sensing image have been significantly improved. In this regard, remote sensing imagery may contain hundreds of spectral bands whilst the spatial resolution can be as high as sub-meter for each pixel. The difficulties of building identification have significantly increased as the quickly evolving spatial and spectral variations. Moreover, due to the limitation of remotely sensed data, it is almost impossible to acquire the semantic information of ground-based buildings.