Transfer learning for hierarchical Conditional Random Fields for the classification of urban aerial and satellite images

• Applicant: Professor Dr.-Ing. Christian Heipke
• Subject Area: Geodesy, Photogrammetry, Remote Sensing, Geoinformatics, Cartography
• Term: from 2013 to 2017
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 246463617

It is the goal of the proposed project to develop a methodology for the supervised context-based classification of aerial and high-resolution satellite images of urban areas. The main scientific contribution is the development and application of methods for transfer learning for determining the parameters of the classification model in order to reduce the amount of training data required for such a model. In order to overcome problems of methods based on local context only, a hierarchical model is proposed. The classification is based on a digital surface model and a true orthophoto, the mathematical framework is provided by Conditional Random Fields (CRF). Recent work on CRF-based classification has shown that simple models of local context may lead to over-smoothing of the results. More complex models can lead to better results, but they require a considerably larger amount of training data. Furthermore, CRF are known to have problems in modelling long-range interactions between objects in a scene. In order to tackle these problems, we suggest a new CRF-based classification technique using more complex context models than comparable methods. In order to reduce the amount of training data required for learning the parameters of these models, it is our goal to develop models that are suitable for transfer learning, so that training data acquired at another time and/or for another place can be transferred to a new scene to be classified. The suggested project constitutes the first application of the principles of transfer learning in the context of graph-based classification methods in image analysis. Furthermore, in order to be able to model long-range interactions in the probabilistic model with a realistic computational effort, scale space is considered explicitly by building a hierarchical model. The new methodology is evaluated on real data with a reference that was generated manually. The methodology developed in both projects, which deal with different aspects of transfer learning, will be exchanged and compared.

• DFG Programme: Research Grants
• Participating Person: Professor Dr.-Ing. Franz Rottensteiner