Science Service System

Summary of Proposal LAN1627

TitleClassification of Urban Structure Types based on high-resolution InSAR data - the case of Munich
Investigator Novack, Tessio - Technische Universitaet Muenchen, Photogrammetry and Remote Sensing
Team Member
M.Sc. Novack, Tessio - Technische Universitaet Muenchen, Photogrammetry and Remote Sensing
Prof. Dr. Ing. Stilla, Uwe - Technische Universitaet Muenchen, Photogrammetry and Remote Sensing

InSAR data provided by theTerraSAR-X satellite operating in High-Resolution Spotlight Mode have a spatial resolution that enables the structural characterization of urban areas. The objective of this PhD research financed by the Deutscher Akademischer Austausch Dienst (DAAD) is to develop a framework for the probabilistic classification of Urban Structure Types (USTs) based on space-borne InSAR data. USTs can be understood as the different types of urban settlements regarding their building densities, impervious and green open spaces. USTs are a very important spatial indicator of physical, functional and energetic factors of the urban fabric which in turn guide many urban planning interventions. The most important attributes of the USTs are the presence of vegetation and the number and volume of the buildings. Our assumption is that this attributes can be estimated by jointly using all the information content of InSAR datasets generated by the TerraSAR-X satellite operating on High-Resolution Spotlight Mode. We also assume that the best way to characterize the urban fabric regarding is different UTSs is through a classification approach that enables the consideration of potentially many input variables and the modelling of complex contextual relations between these variables and the possible classes. Conditional Random Fields (CRFs) are able to model the posterior distribution of all the possible classifications of the scene given the observed variables of each of its analysis units (i.e.individual pixels or image segments). It is well known that the automatic learning of the CRF structure and its parameters based on sampled data leads to models that perform better than manually defined ones. Besides, the automatic learning may reveal important contextual relations between the possible classes of an image segment and the observed variables and possible classes of its neighbouring segments. A significant problem though is still the very high computational costs of these learning methods, due to the necessity of running inference over the current CRF network at each step of the optimization process. However, two state-of-the-art objective functions have been proposed in the last years, which when combined can significantly reduce computational costs and yet find the global optimal solution. Besides proposing the joint use of these two objectives and profiting from their advantages, we propose a framework for the insertion of human knowledge in the learning process. This knowledge refers to our capacity of visually interpreting images (be it optical or radar imagery). The model learned through this approach will be then applied for predicting the most probable USTs classification of an InSAR scene from Munich (Germany). The application of this CRF learning/classification framework involves the following steps: (1) segmentation of the scene into image objects,(2) sample collection, (3) optimization of the objective function in order to learn the best model structure and set of parameters and (4) application of the model over the whole data, which means performing inference in order to obtain the most probable UST classification.

In order to test this framework and apply it for the classification of USTs, two High-Resolution Spotlight InSAR acquisitions (of5 km x 5 km each) are necessary. These two required scenes cover most of Munich’s urban area and its different USTs. The deliverables are (1) a tractable framework for learning CRFs models, which in theory can be appliedfor any remote sensing imagery classification task, (2) a specific CRFs classification model that detects and differentiate USTs based on high-resolution InSAR data and (3) a relevant geoinformation product for several urban applications, namely the USTs map of Munich. More than that, the most important outcome of this research is the transferable methodology for automatically mapping UTSs based on space-borne InSAR data.

Back to list of proposals

© DLR 2004-2016