[MIS07-P15] Recognition and classification of Martian chaos terrains using imagery machine learning
Keywords:Mars, Chaotic terrains, Machine learning
In the present study, we perform imagery machine learning for the chaos terrains on Mars with aims at for recognition and classification. We used the fine-tuning VGG19 model, which was constructed by Visual Geometry Group with convolutional layer depth of 19. We prepare total 3300 images of chaotic terrains on Mars, including 1100 visible images taken with the Context Camera, CTX, in grayscale, 1100 images of thermal inertia map taken with the Thermal Emission Imaging System, THEMIS, and colored 1100 images of Mars Orbiter Laser Altimeter (MOLA) High-Resolution Stereo Camera (HRSC) blended digital elevation model (DEM) taken by Mars Global Surveyor, MGS and Mars Express, MEX. Among the chaos terrains, we define two types of chaos terrains: Ones were probably formed through the discharge of groundwater based on the previous studies (e.g. Zegers et al., 2010; Roda et al., 2014; Rodriguez et al., 2005) (221 images) , and the others were probably generated in association with magma excavation and tectonic tiles based on the previous studies (Sharp, 1973; Bamberg et al., 2014) (151 images) . We also collect 200 images of non-chaotic features, such as valley networks and impact craters, on Mars to construct the classifier.
Our preliminary results show that using THEMIS images, we can classify chaotic terrains with 83% accuracy. The constructed classifier is useful to find new chaos terrains on Mars and other planets, such as Earth. However, the accuracy to recognize whether they formed through discharge of groundwater or magma excavation is low. This means that the formation mechanisms may not be able to suggest based on only the morphology.