HYPERWAVE - Validation of classification techniques developed in Hypercrunch project under various conditions

Context and objectives

The objective of this proposal is to apply and validate the algorithms developed in the frame of the Hypercrunch project. In this earlier STEREO 1 project (SR/00/05), state of the art algorithms to analyse Hyperspectral datasets have been developed and validated via stress detection in orchards. In this context Hyperspectral data cubes were acquired for both stressed and non-stressed orchard plots. One of the goals of Hypercrunch was to develop generic data reduction algorithms, i.e. algorithms that are application independent to the extent possible. This to allow us to implement and automate the generic part in operational data processing chains such as the APEX-chain. A lot of effort has been put in the development of a prototype toolbox to apply the algorithms. To be able to validate and fully test the toolbox, it has to be confronted with as many different data cubes dealing with as many diverse remote sensing applications as possible and feasible. Data cubes and expertise from previous projects within the STEREO envelope and from different projects performed at Vito can be fully re-used for this purpose, without additional cost.

Project outcome

Expected scientific results

For the dune vegetation mapping application, a 4-step classification framework is proposed, that works directly on the posterior class probabilities and includes a coupling of binary classifier procedure, spatial smoothing and unmixing. A large scale classification experiment was set up, for which a large amount of ground reference data and airborne hyperspectral data of the entire Belgian coastline were collected. A total of 23 classes was distinguished, within the vegetation groups salty vegetation, marram dune, moss dune, grassland, scrub and woodland, some mixed vegetation and 4 non-vegetation classes.
For the aquatic application, a model is fitted to reflectance spectra, using simulating annealing for optimising the means square error of the reflectance over all spectra.