Multi-sensor feature fusion for very high spatial resolution built-up area extraction in temporary settlements

Detailed and up-to-date knowledge on the situation in temporary settlements of forced migrants plays an important role for effective humanitarian assistance. These settlements emerge as planned or spontaneous camps or camp-like structures, characterized by a small-scale physical morphology and high dynamics. Information on the built-up area (BUA; i.e. areas occupied by buildings) in these settlements provides important evidence on the local situation. The objective of this work is to present a generic procedure for the detailed extraction of BUA in complex temporary settlements from very high spatial resolution satellite data collected by different sensor types. The proposed approach is embedded in the methodological framework of object-based image analysis and is compound of i) the computation of an exhaustive set of spectral-spatial features aggregated on multiple hierarchic segmentation scales, ii) filter based feature subset selection and iii) supervised classification using a Random Forest classifier. Experimental results are obtained based on Pleiades multispectral optical and TerraSAR-X Staring Spotlight Synthetic Aperture Radar satellite imagery for six distinct but representative test areas within the refugee camp Al Zaatari in Jordan. The experiments include a detailed assessment of classification accuracy for varying configurations of considered feature types and training data set sizes as well as an analysis of the feature selection (FS) outcomes. We observe that the classification accuracy can be improved by the use of multiple segmentation levels as well as the integration of multi-sensor information and different feature types. In addition, the results show the potential of the applied FS approach for the identification of most relevant features. Accuracy values beyond 80% in terms of kappa statistic and True Skill Statistic based on significantly reduced feature sets compared to the input underline the potential of the proposed method.