Semi-automatic Hyperspectral Image Classification of Urb- an Areas Using Logistic Regression

Urban areas are among the most dynamic regions on earth, continuously and rapidly changing. For monitoring these changes, remote sensing has proven over the years to be a reliable source. Current airborne hyperspectral systems with spatial resolution of a few meters, combined with very high spectral resolution, facilitate the urban scene analysis by allowing to distinguish small details in the urban environment. This paper presents part of a project aiming to classify man-made objects using hyperspectral images and to investigate the complementarity between hyperspectral and SAR data. The intention is to develop methods that are able to quickly obtain an overview of the current situation and require as little human intervention as possible. This is very important for various applications related to disasters, e.g. emergency cartography, disaster monitoring, damage assessment, mission planning, etc. The paper describes a new method for classifying the main classes in an urban environment using hyperspectral data. The method is based on logistic regression (LR), which is a supervised multi-variate statistical tool that finds an optimal combination of the input channels for distinguishing one class from all the others. LR thus results in detection images per class that can then be combined into a classification image. The LR uses a stepwise method that implicitly performs a channel selection. The method is supervised in the sense that existing digital maps are used for learning. However, the method does not require the laboratory spectra or extensive ground truth. The method is applied on HyMAP data of an urban area in the South of Germany. The results of the proposed approach are compared to classical methods. Furthermore, a sensitivity analysis is presented, which investigates the robustness of the detection of the different classes against various influences and in particular the influence of channel width and pre-processing level. The classification results are better than those obtained by a classical method. The sensitivity analysis shows that the pre-processing level applied to the hyperspectral data does not influence the classification results significantly for this application. Furthermore, reducing the number of channels results in a drop of performance for some classes only when less the number of channels becomes inferior to 40.