Recently, it was shown that a detailed reconstruction of urban height maps is possible from single very high resolution (VHR) synthetic aperture radar (SAR) images with deep convolutional neural networks. Being merely a proof-of-concept so far, the potential of this approach has not been fully exploited yet. With this work, we present an optimized deep learning model for height estimation from single VHR SAR images, which incorporates sensor knowledge into the estimation. We embed this model into a SAR-specific processing chain that allows the generation of seamless georeferenced digital surface models (DSMs) with geodetically defined heights in an orthometric map coordinate system. Extensive experiments are carried out with a custom-generated dataset including over 50 TerraSAR-X images from 8 different cities. They confirm that our workflow generalizes well across different locations while being robust to different properties of the input data. Thus, our workflow provides the unique ability to produce elevation models of urban areas quickly, regardless of weather, around the clock, and at low cost. This can be of immense benefit when time is critical, e.g. in disaster response scenarios or in the context of reconnaissance activities.
«Recently, it was shown that a detailed reconstruction of urban height maps is possible from single very high resolution (VHR) synthetic aperture radar (SAR) images with deep convolutional neural networks. Being merely a proof-of-concept so far, the potential of this approach has not been fully exploited yet. With this work, we present an optimized deep learning model for height estimation from single VHR SAR images, which incorporates sensor knowledge into the estimation. We embed this model int...
»