Semantic Segmentation of Remote-Sensing Images Through Fully Convolutional Neural Networks and Hierarchical Probabilistic Graphical Models

Deep learning (DL) is currently the dominant approach to image classification and segmentation, but performances of DL methods are remarkably influenced by quantity quality ground truth (GT) used for training. In this article, a method presented deal with semantic segmentation very-high-resolution (VHR) remote-sensing data in case scarce GT. The main idea combine specific type deep convolutional neural networks (CNNs), namely fully (FCNs), probabilistic graphical models (PGMs). Our takes advantage intrinsic multiscale behavior FCNs representations connect them hierarchical Markov model (e.g., making use quadtree). As consequence, spatial information present better exploited, allowing reduced sensitivity GT incompleteness be obtained. marginal posterior mode (MPM) criterion inference proposed framework. To assess capabilities method, experimental validation conducted ISPRS 2D Semantic Labeling Challenge datasets on cities Vaihingen Potsdam, some modifications simulate spatially sparse GTs that common real applications. results quite significant, as exhibits higher producer accuracy than standard considered especially mitigates impact minority classes small details.