Mars Imagery Classification: A Galactic Battle between Knowledge Transfer Networks and their Dual-Attention Armed Adversaries

The development of intelligent and reliable autonomous systems will determine the course of space exploration in the future. One of the most important challenges in space exploration is automating the analysis of a planet’s environment, in search of the space rover’s instruments and potential environmental threats to the rover. Leveraging artificial intelligence for this task can aid exploration missions in guiding autonomous vehicles and saving the data transmission costs. In this work, we compare the performance of dual attention networks, interplanetary transfer learning methods and Vision transformers in classifying objects in images collected by the Mars Science Laboratory Curiosity rover from August 2012 to July 2015. Our EfficientNet knowledge transfer model outperforms the state of the art model by 17.14% on test set and 16.11% on validation set. The successful interplanetary knowledge transfer demonstrated in this work bodes well for the future of deep neural networks in space imagery classification.