IN11E-1800
Global Learning Spectral Archive- A new Way to deal with Unknown Urban Spectra –

Monday, 14 December 2015
Poster Hall (Moscone South)
Marianne Jilge1, Uta Heiden2, Martin Habermeyer2 and Carsten Jürgens1, (1)Ruhr University Bochum, Bochum, Germany, (2)German Aerospace Center DLR Oberpfaffenhofen, German Remote Sensing Data Center (DFD), Oberpfaffenhofen, Germany
Abstract:
Rapid urbanization processes and the need of identifying urban materials demand urban planners and the remote sensing community since years. Urban planners cannot overcome the issue of up-to-date information of urban materials due to time-intensive fieldwork. Hyperspectral remote sensing can facilitate this issue by interpreting spectral signals to provide information of occurring materials. However, the complexity of urban areas and the occurrence of diverse urban materials vary due to regional and cultural aspects as well as the size of a city, which makes identification of surface materials a challenging analysis task. For the various surface material identification approaches, spectral libraries containing pure material spectra are commonly used, which are derived from field, laboratory or the hyperspectral image itself. One of the requirements for successful image analysis is that all spectrally different surface materials are represented by the library. Currently, a universal library, applicable in every urban area worldwide and taking each spectral variability into account, is and will not be existent. In this study, the issue of unknown surface material spectra and the demand of an urban site-specific spectral library is tackled by the development of a learning spectral archive tool. Starting with an incomplete library of labelled image spectra from several German cities, surface materials of pure image pixels will be identified in a hyperspectral image based on a similarity measure (e.g. SID-SAM). Additionally, unknown image spectra of urban objects are identified based on an object- and spectral-based-rule set. The detected unknown surface material spectra are entered with additional metadata, such as regional occurrence into the existing spectral library and thus, are reusable for further studies. Our approach is suitable for pure surface material detection of urban hyperspectral images that is globally applicable by taking incompleteness into account. The generically development enables the implementation of different hyperspectral sensors.