A Framework for Satellite Image Classification in the Context of Crisis Mapping using Markov Random Fields

Abstract

In this contribution a general framework for classification of very high resolution optical satellite images is proposed and evaluated. This approach is designed in order to cope with the specific conditions accompanied by crisis mapping applications and is moreover well suited for several other applications. Multiscale image information (data model) as well as hierarchical and spatial context information (prior model) is incorporated into the classification process using a hybrid Markov model which combines a hierarchical directed as well as a planar lattice-based Markov Random Field (MRF). The modelling of arbitrary semantic classes in different scales enables the definition of a hierarchical semantic network representing the dependencies and relations between the classes in adjacent scales. Classification is carried out using non-iterative hierarchical maximum a-posteriori (MAP) or mode of posterior marginal (MPM) inference as well as a subsequent optimization step using a planar MRF. Additionally, a modified MAP inference which is able to outperform the original inference methods is proposed. The impact of incorporation of image data from multiple scales is evaluated in this contribution. Furthermore, the dependency between the quantity of the training data and the classification accuracy is analyzed.