Research project S0/00/011 (Research action S0)
Project summary:
This proposal aims to extract all different kinds of information over urban and peri-urban environment using hyperspectral APEX data. This proposal is in conjunction with the Stereo-project "Improving spatial information extraction for local and regional authorities using very-high-resolution data". Ground truth and derived products generated from the mentioned project will serve as validation data in this research on hyperspectral data. As test site is the southern part of the city of Ghent selected. This proposal aims to extract information out of hyperspectral data on urban and semi urban environments in two ways: firstly geometrical information and secondly thematic information.
The geometrical information (x, y and z) will be extracted using stereoscopic flight strips with a lateral overlap of 60%. This means that the flight scheme has to be prepared carefully with the flight management. From this stereoscopic hyperspectral data an envelope DEM and an orthophoto will be produced by aims of digital photogrammetrical methods, based on ground control points measured with a differential GPS. These products will be evaluated by using reference data based on very high-resolution aerial pictures (8 cm.).
The thematic analysis will be done in a visual and digital way. In a first step the hyperspectral images will be analysed and map by on screen digitalisation. Specific elements will be mapped, asphalt surfaces, different type of roofs, vegetation etc. A statistical analysis will be performed on this different land cover/land use classes. This analysis will serve as one of the options for band reduction. Besides this approach to reduce the data, also principle component analysis (PCA), Discriminant Analysis Feature Extraction (DAFE) and Decision Boundary Feature Extraction (DBFE) methodologies will be tested out. The image classification of the hyperspectral data will be investigated with different approaches: 1) by classification on the first components of the PCA, 2) by conventional maximum likelihood classification, 3) by neural network analysis and 4) by a spatial-spectral variant of the maximum-likelihood classifier. The classifier, known as ECHO, proceeds by first segmenting the scene on a multi-variant basis into statistically homogeneous objects using spatial information, and then classifies these objects using a distribution to distribution comparison. Using the same class descriptions as a pixel classifier, it usually achieves higher accuracy and requires less computation time.
The tested approaches, as well the methodologies for data reduction and image classification will be compared and evaluated in order to select the best methodology to extract thematic information on urban and peri-urban environments using hyperspectral data.
