Identification of martian surface minerals in crism imagery using a deep neural network

Introduction: Visible and Near-Infrared spectroscopy has been used by the Observatoire pour la Minéralogie, l’Eau, les Glaces et l’Activité (OMEGA) instrument[1] and the Compact Reconnaissance Imaging Spectrometer for Mars (CRISM)[2] instrument on the Mars Reconnaissance Orbiter to elucidate the composition of the Martian surface through the detection of diagnostic infrared absorptions attributed to specific mineral compositions [3,4]. Aided by the spectral parameters formulated by [5], these instruments have observed a wide array of minerals indicative of formation in ancient Martian environments [6,7,8,9]. Typically, a CRISM image RGB composite is produced by combining three parameters, which allows for visual representation of the presence of specific mineral groups. Determining the presence of specific minerals currently involves a tedious process of generating spectral ratios, and visually comparing them to infrared spectra from mineral libraries. While this type of manual inspection is suitable for a small subset of observations, it is untenable for a global or even regional large-scale spectroscopic survey of the surface. Automated processing of CRISM data has been previously employed in the detection and mapping of specific mineral phases. [10-13]. Implementing the use of a Deep Neural Network (DNN) to the entire CRISM dataset, a large regionalscale study, or the new MRDR map tile mosaics [14] would simplify the automation process and enhance the versatility of surface mineral research. A previous neural network approach was successfully used to determine the temperature and single scattering albedo for each pixel in a group of CRISM scenes from 13.8μm [15]. In a previous iteration of this study I attempted to utilize a Convolutional Neural Network to identify significant outcrops of mineral parameters from [5] but given the random shapes of outcrops and the lack of distinct patterns in parameter values, the results were not optimal [16]. For this study I construct a DNN that, unlike [16], will examine the entire VNIR range covered by CRISM (.436-3.89μm) of each pixel in a CRISM image or set of images and illuminate contributions of end-member minerals to the final spectra. Methods: A DNN, like most machine learning applications, is an algorithm that is trained on large datasets for the specific purpose of complex pattern recognition. The DNN is trained on grouped endmember mineral reference spectra from the CRISM targeted United States Geological Survey and Minerals Identified in CRISM Analysis (MICA) libraries. The DNN is then evaluated on MapProjected Targeted Record CRISM images. The endmember spectra present in these mineral spectra libraries have been pre-configured to account for absorption of the Martian atmosphere and are specific to Martian chemistry. The DNN is trained to classify the 31 minerals found in the MICA spectral library, in addition to 1 null class corresponding to ignored values. Because the spectra of each pixel will contain contributions from multiple end-member minerals, the DNN will not function properly as a standard classifier, which attempts to identify one mineral that is a “best-fit” for the spectra. The DNN would need to be modified to identify the contributions of each mineral to the observed spectrum. To accomplish this, we implement a multinomial regression (Softmax) algorithm for the class evaluation which returns a decimal confidence value between 0 (No confidence) and 1 (100% confidence) for each class [17]. Confidence values of greater than .1% are considered significant contributions to spectral signature, and smaller confidence values are considered insignificant. The CRISM image data are preprocessed for the DNN using the same method as [16]. The DNN is developed using Tensorflow in the Python