Artificial Intelligence-Enhanced Detection of Biogenicity Using Laboratory Specimens of Biologically and Microbially Induced Sedimentary Structures in a Controlled Experiment

Abstract

The search for traces of life can be based on the detection of specific signatures produced by microorganisms on sedimentary rocks. Microbially induced sedimentary structures (MISSs) develop under specific physicochemical conditions that are likely to have potentially existed on Mars during the Noachian period. We designed an experiment under controlled laboratory conditions to explore the wide range variability in biogeomorphological responses of clay-sand substrates to the development of biological mats-including microbial mats-of different strains and biomasses, and an abiotic control. A 3D picture dataset based on the experiment was built using multi-image photogrammetry. Visual observations were combined with multivariate statistics on computed topographical variables to interpret the diversity in the resulting biotic and abiotic mud cracks. Finally, an artificial intelligence (AI) classifier based on convolutional neural networks was trained with the data. The resulting model predicted accurately not only the biotic-abiotic differences but also the differences between strains and biomasses of biotic treatments. Its results outperformed the blind human classification, even using only grayscale pictures. Class Activation Maps showed that AI followed several decision paths, not always like those of the human expert. Next steps are proposed for application of these models to ex situ biogeomorphological structures (fossil and modern MISS) on Earth's surface, to ultimately transpose them to a martian context.

Keywords: Artificial intelligence; Biogeomorphology; Biological mats; Classification; Controlled laboratory experiment; Convolutional neural networks; Mars; Microbially induced sedimentary structures; Multivariate statistics.