Buildings Detection in VHR SAR Images Using Fully Convolution Neural Networks

Abstract

This paper addresses the highly challenging prob-lem of automatically detecting man-made structures especiallybuildings in very high-resolution (VHR) synthetic apertureradar (SAR) images. In this context, this paper has two majorcontributions. First, it presents a novel and generic work-flow that initially classifies the spaceborne SAR tomography(TomoSAR) point clouds - generated by processing VHR SARimage stacks using advanced interferometric techniques knownas TomoSAR - into buildings and nonbuildings with the aid ofauxiliary information (i.e., either using openly available 2-Dbuilding footprints or adopting an optical image classificationscheme) and later back project the extracted building pointsonto the SAR imaging coordinates to produce automatic large-scale benchmark labeled (buildings/nonbuildings) SAR data sets.Second, these labeled data sets (i.e., building masks) have beenutilized to construct and train the state-of-the-art deep fullyconvolution neural networks with an additional conditionalrandom field represented as a recurrent neural network to detectbuilding regions in a single VHR SAR image. Such a cascadedformation has been successfully employed in computer vision andremote sensing fields for optical image classification but, to ourknowledge, has not been applied to SAR images. The resultsof the building detection are illustrated and validated over aTerraSAR-X VHR spotlight SAR image covering approximately39 km2 - almost the whole city of Berlin - with the mean pixelaccuracies of around 93.84%.