MINERAL IDENTIFICATION On MARTIAN SURFACE USING SUPERVISED LEARNING APPROACH FROM CRISM HYPERSPECTRAL DATA

Abstract

The availability of spectral libraries for CRISM (Compact Reconnaissance Imaging Spectrometer for Mars) data through NASA’s Planetary Data System has revolutionized the study of the surface mineralogy of Mars. However, building supervised learning models for mineral mapping remains a challenge due to the scarcity of ground-truth training data. In this thesis, an innovative framework is presented that leverages supervised learning to classify spectra within CRISM hyperspectral images. To overcome the data limitation, an augmentation approach is employed that creates the training data by augmenting the minerals available in the MICA spectral library, preserving key absorption signatures of each mineral class while introducing adequate variability. The framework includes a comprehensive pre-processing pipeline, featuring a novel feature extraction method to capture distinctive absorption patterns in the spectra. The approach is validated using CRISM images from diverse Martian locations and interactive mineral maps are also provided for the detected dominant minerals. While this initial framework ensures acceptable accuracy, utilizing more sophisticated learning models and advanced preprocessing techniques can enhance the performance of the framework. Spectra in remotely sensed hyperspectral images are often affected by the presence of continuum, which changes the global curvature of the spectra, although the key absorption signatures are present. The continuum removal process, one of the critical preprocessing steps in the pipeline, is modified from the traditional approach to a novel method named Segmented Curve Fitting, which can identify more absorption shoulder points in a spectrum and thus can detect the absorption features in it more distinctively. Lastly, the thesis introduces MICAnet, a specialized Deep Convolutional Neural Network (DCNN) architecture tailored for mineral identification using CRISM hyperspectral data. Inspired by Inception-V3 and InceptionResnet-V1 architectures, MICAnet leverages 1-dimensional convolutions for processing spectra at the pixel level. This innovative architecture represents a significant contribution, being the first solely dedicated to this objective. The performance of the mineral mapping framework is assessed using both simulated data of varying complexity and a real CRISM TRDR/MTRDR hyperspectral dataset. In conclusion, this study advances the field of planetary science and remote sensing by providing automated approaches for mineral identification and mapping on Mars, also, enhances the understanding of Martian surface mineralogy, offering valuable insights into the planet’s geological history and habitability.