MINERALOGICAL CHARACTERIZATION USING NEURAL NETWORKS: COMPOSITION OF MAFIC MINERALS IN MARTIAN METEORITES FROM THEIR SPECTRA A. Rozel 1 , H. Cl´ enet 2 , S. Dout´ e 3 and C. Quantin 4 1 ETH Zurich, Institute of Geophysics, Switzerland, 2 EPFL, Earth and Planetary Science Laboratory, Swizerland, 3 Institut de Plan´ etologie et d’Astrophysique de Grenoble, IPAG, CNRS UJF, France, 4 UCBL / ENS Lyon, Laboratoire de G´ eologie de Lyon, UMR CNRS 5276, France. ABSTRACT In this study, we test the ability of neural networks to deter- mine the composition of magmatic rocks from their labora- tory spectra. We first describe the structure and behaviour of the multi- layer perceptron that we implement and train for quantitative characterization. For that purpose, reference laboratory spec- tra of mafic minerals from both natural and synthetic samples are used. As their composition in terms of the three mafic minerals, olivine (OL), orthopyroxene (OPX) and clinopyrox- ene (CPX) are known, those spectra are given as inputs during the learning phase of the neural network. In the analysis phase, we use the neural network to pro- cess spectra acquired on SNCs (Shergottites, Nakhlites, Chas- signites) meteorite samples that are considered to be repre- sentative of Mars surface. The network outputs mineralogical compositions very quickly, performing only explicit opera- tions. Our preliminary results show that neural networks are able to quantify mafic minerals, especially in the case of complex mixtures, with much improved computer efficiency and comparable accuracy compared to usual methods. This is very promising regarding future analysis of huge datasets. Index Terms— Neural, Network, Hyperspectral, Detec- tion, Meteorite 1. INTRODUCTION Mafic minerals are key components when trying to under- stand the geological history of planetary bodies like Mars. Indeed, their presence in igneous rocks is directly related to mantle properties and crystallization conditions. They also partially control the nature of the alteration products which could be formed subsequently. In this respect detection of olivine and pyroxenes, and characterization of their respective composition, is an important step that must be done carefully. Support was provided by the Marie Curie Initial Training Network TOPOMOD and the ERC project iGEO. Reflectance spectroscopy, hyperspectral remote sensing in visible/near-infrared is a very powerful tool to achieve this objective. Indeed, olivine and pyroxenes have characteristics broad absorption features near 1 and 2 μm [1, 2] due to the Fe 2+ electronic transition. During the last decade, imaging spectrometers onboard spacecraft have acquired huge amount of such data and it is actually challenging to process them both quickly and efficiently. Several techniques (e.g. linear unmixing [3], radiative transfer modeling [4]) aimed at deconvolving absorption bands in terms of mineralogy. Modified Gaussian Model [5] can also be used to quantitatively estimate the chemical composition of each mineral in a rock. Such approach has been successfully applied on OMEGA data [6]. However, the results obtained with those techniques are basically not accu- rate enough, these methods are time consuming and efforts are still to be done to develop new algorithms. This is why we test in this study the ability of neural networks to determine the composition of minerals from their mafic signatures on laboratory spectra. 2. NEURAL NETWORK IMPLEMENTATION 2.1. Structure of a Neural Network A neural network is a learning machine which uses a set of scalars as input (the data to analyse) and produces another set of numbers, carefully chosen [7, 8]. In our case, the input vector is the spectrum and the output vector is the modal or the chemical composition (percentage of olivine, clinopyrox- ene and orthopyroxene, cf. figure 2). Various schemes can be used to connect the input to the output set. We use the mul- tilayer perceptron method in which information propagates forward in the analysis phase and backward in the learning phase through adjacent layers. Each adjacent layers are linked by non-linear weighted connections, as represented in figure 1. At each node, the contributions of all elements of the pre- vious layer are summed up and a threshold is applied. It has been shown that the use of the threshold has a fundamental importance in the learning capacities of the neural network