Multiresolution based neutrosophic framework for building footprint identification using remote sensing images

Building identification using automatic segmentation techniques from remotely sensed images have been studied for a while and lately several supervised deep learning techniques have been applied to this task. However, supervised techniques demand dense pixel-level annotations and results are largely dependent on the characteristics of the dataset used for supervised learning. In this paper, we propose an unsupervised and fully automatic framework based on neutrosophic sets and multiresolution analysis which is able to robustly extract building from aerial images with the aid of an ensemble of spectral, textural and morphological features. We demonstrate the robustness of the proposed method by comparing with several state-of-the-art supervised techniques using imageries of different resolutions, building densities and backgrounds. We show that the proposed framework is not only able to extract shapes as close as possible to the ground truth (accuracy up to 94%) but also maintains homogenous semantics in the face of varying characteristics of the input aerial imagery.