XAttDnet: Explainable Deep Convolutional Network for Crater Detection and Pose Estimation During Lunar Landing

Intelligent algorithms through the recent developments of Artificial Intelligence (AI) are making enormous progress in providing intelligence to the decision-making process for space vehicles descending on the Moon. However, the trust in freely deploying them on space vehicles to increase their autonomy is still questionable by the most qualified scientists in the space domain. This paper therefore delves into explainable artificial intelligence (XAI) for vision-based crater detection and navigation, to offer clear and comprehensible predictions to facilitate intelligent lunar landings. We incorporate an attention mechanism into the backbone architecture during the network construction phase, delivering an attention-enhanced Darknet53 (XAttDnet) for explainable feature extraction. Further, we replaced the XAttDnet with the backbone of YOLOv3 and developed the XAttDnet-LSTM models for crater detection and pose estimation tasks. Synthetic datasets generated by PANGU software are used to train and evaluate the proposed networks, showing competitive accuracy in detecting craters and estimating relative poses with RGB images-driven descent scenarios. The roles played by the different convolutional layers in concentrating and extracting features for two tasks are suggested by Visualising the attention maps of each trained convolutional layer. A quantitative assessment of the explainability is attempted by employing the Pearson correlation coefficient (PCC), which further reveals the evolution of the relationship among the feature extraction layers and inputs/outputs.