Constraining the thermal inertia of Mars utilizing machine learning techniques

Machine learning techniques, showing high automation and efficiency in handling large amounts of observation data, have been applied to predict the thermal inertia of Mars from surface kinetic temperatures. We created a large data set from well-established thermal models. Using this data set, we trained random forest (RF) models using surface kinetic temperatures, time of day, and other five accessible parameters as inputs to the model. The model performances for different local times were analysed and the characteristics of derived thermal inertia in typical regions on Mars were discussed. It is found that it is feasible and reliable to predict the thermal inertia of Mars using the well-trained RF. The RF predictions reflect the thermal signatures of Mars and show good agreement with previous retrievals. When using the nighttime data to make predictions, the RF model shows the best performance compared with those at other times of day. We also classified thermal inertia into four units: low, intermediate, relatively large, and large thermal inertia, and the RF model works for all four units. The predictive ability of the RF is also demonstrated for five representative regions on Mars, where the RF predictions are in good agreement with the bolometric nighttime thermal inertia from the thermal emission spectrometer. More importantly, the RF model provides a rapid retrieval of thermal inertia and speeds up the thermal analysis in upcoming Mars exploration missions with substantial data.