Research

4D-EARTH

The hyperspectral game

HymapCC_small
Color composite of bands at 650 nm (red), 543 nm (green) and 452 nm (blue).

Preamble

The geology community, in particular those involved in mining applications, has been one of the first to use hyperspectral remote sensing. The commonly used hyperspectral end-products describing surface geology are mineral maps. A wealth of spectral matching and spectral unmixing techniques are now available for producing such maps. There is very little standardization and protocols underlying ‘mineral mapping from hyperspectral data’ nor objective methods for quantification, verification and testing of repeatability of methods.

The idea

Our idea is to make an inventory of existing techniques that are used to produce mineral maps from hyperspectral data. By applying these to a unique hyperspectral data set we can contribute to the notion of standardization and create awareness of the range of approaches that are available. The hyperspectral game is an experiment to reach a sensitivity analysis of mapping methods. As we limit the prior information, it is aimed to be objective and purely a spectral geology exercise.

The hyperspectral  game

A subset of a HyMAP scene is made available to all participants of this experiment. Alongside we provide four regions of interest for four key minerals (alunite, kaolinite, illite, calcite) that are found in the imaged area and a spectral library with image spectra. Participants may also choose to use spectra from a standard spectral library. We ask participants to produce mineral maps for the four target minerals/spectra. These can be hard classifications of pixels (pixels labeled as classes) or rule images (probability or proportion images or the like). You can use a standard algorithm or your own algorithm as long as it is properly documented and the parameter settings properly described.

Purpose of the game

We hope many people will participate with a range of different algorithms. We will gather all the different results and compare them in a qualitative way. That means we want to do a sensitivity analysis (spectral, spatial), highlight similarities and differences in classification results, indicate which areas (spatial) were easy/difficult to map, which classes (spectral) were confused. It is not our purpose to link the results to ground truth or to a validation data set, it is also not our purpose to conclude that method A outperforms method B.

The outcome

The outcome of the hyperspectral game will be an overview of spectral mineral mapping methods and a sensitivity analysis for the four classes foreseen. We intend to publish the results in an ISI (remote sensing) journal. All participants will become co-author on this paper provided they share their results and provide a reference to a journal paper describing their algorithm.

Interested in participating?

Please send an email to Freek van der Meer (f.d.vandermeer@utwente.nl) expressing your interest to participate indicating with which algorithm. Our tentative timeline is to compile a list of participants during April-May, participants perform the mineral mapping and return results end of June, initial comparative analysis and feedback in July, paper writing September-October with submission aimed for in November 2011. We will make this a group effort, so during the comparative stage and paper writing we will regularly update and consult all participants. All participants may withdraw at any time prior to paper submission.

The data

Click here to download the data (.zip, 19.7 MB)

Hyperspectral_image: the HyMAP image of the area
Image_spectra: a spectral library of input spectra for the four endmembers
ROI_image_spectra: regions of interest for the four image endmembers