Item

Abstract

The advantage of hyperspectral data sets with its ability for diagnostic analysis is widely recognized in different fields of application. Real time spectra extraction in combination with high spatial information can significantly improve the accuracy of mineral identification. Insufficient signal to noise ratios as well as mixed pixel/signatures decrease the accuracy to diagnostic mineral identification. Thus, it is often difficult to link extracted spectral signatures to distinct minerals. In this context a software for Neural Evaluation of Spectral Signatures (NESSI) was developed. The advantage of the neural network technique in combination with spectral libraries has been proven by noise folded spectra of well defined mineral mixtures. The technique was then applied to natural rock types and GERIS hyperspectral image data of the Maktesh Ramon test-site located in Israel. The network guaranteed a relative precise identification of the most matching signature. The average accuracy for a correct class decision was 92% within the SWIR-network and 88% within the VNIR/SWIR-network. About 25% of the image could be assigned to minerals of the spectral library. Misclassifications can mostly be traced to the poor SNR of the GER data and the missing of mineral mixtures in the spectral library.