HYPERCRUNCH - Data analysis in hyperspectral remote sensing

Context and objectives

The overall objective of the HyperCrunch project is to improve the information extraction (data mining) from hyperspectral datacubes.

The high spectral and high spatial resolution, intrinsic to hyperspectral remote sensing, result in huge quantities of data. The selection of a limited number of relevant bands without loss of essential information for a given application is a first subject of this research project. The second subject is the development of acquisition protocols to improve classification performance, both in view of spectral unmixing and the classification procedures themselves. The mathematical algorithms will be developed as application-independent as possible, such that they can be implemented and automated in operational data processing chains such as the APEX-chain. This way a higher level end-product becomes available to the scientific community.

The data reduction techniques and mathematical algorithms which will be the end-products of this research, will be tested and validated. The application envisaged here is precision agriculture, more precisely the monitoring of stress in orchards.

Project outcome

Expected scientific results

Two types of end products were envisaged. The first is an operational method for hyperspectral data extraction that can be used for future project. The method must be ready to implement in operational chains such as the APEX Processing and Archiving Facility (PAF). This objective is reached with success. All programs implementing the algorithms are written in C++ and are ready for operational use. They are compiled with the open source GNU compiler in a Linux environment (distribution Red Hat 9.0). No commercial libraries or other software is needed.
Another end product is the classification map of the hyperspectral data cube, acquired over our test plot Gorsem. Early in the project, it was obvious that this objective would be difficult to reach. In particular, the limited number of relevant pixels in the image put a constraint to its use. The pixels could clearly not be used as training data . We have made an attempt to classify the image by using the upscaled (simulated) leaf spectra as training data with moderate success (70% of the pixels are detected correctly). In order to increase this number, the classifier must be trained with real instead of simulated data. Appropriate training is feasible only with orchards of hundreds of trees (per class). In that case, results are expected with an accuracy of more than 80%, which is more acceptable for end users.