Mineralogical Classification of CRISM Hyperspectral Data Under Uncertainty With Hybrid Neural Networks

Orbital remote sensing observations are a lynchpin of planetary science research. Hyperspectral infrared spectroscopy in particular is key for planetary mineralogical exploration, for example, CRISM for Mars, as this underpins our understanding of the distribution of specific lithologies and the geological process leading to their formation. Yet routine analysis workflows involving summary parameters have significant limitations and are highly time‐consuming. This work presents a novel methodology and framework for the analysis and classification of CRISM SWIR reflectance spectroscopy, leveraging Machine Learning (ML). We train a model to classify 37 minerals previously manually identified on the planet. We show this model is highly performant, with test data across Mars and a case study within Jezero crater, where ML results match previous manual analyses and rover observations. We also adapt Expected Cost (EC) to remote sensing data for use in geological context for the first time. We demonstrate that EC can be used to dynamically weight misclassification penalties based on geological context, as a rigorous measure of automated classification methods. We envision this model to make analysis of CRISM data more accessible to the planetary science community, allowing rapid searches for a vast range of minerals across a global/regional scale. As a result, areas of interest for further satellite or rover exploration can be quickly identified, leading to greater understanding of geological processes on Mars.