Objectives  Traversability analysis is one of the necessary parts for rovers on extraterrestrial surface to explore unknown environment.

  Methods  In this paper, we propose a data-driven method for traversability analysis for rovers on extraterrestrial surface. Based on the inputting multi-dimensional terrain information, the proposed method models traversability analysis as a semantic segmentation problem, which can explicitly compute a traversability map of this circumstances for a specific rover. Meanwhile, we provide an algorithm for generating training dataset for the rover. We first run the rover in the field to collect directed traversability results at certain positions, and then fulfill the undirected traversability map with these results by converting this problem into a global optimization problem, since undirected traversability map is more intuitive and straightforward for path planning. We can get the dataset for a specific rover by linking this map with the data of environment. In order to get the data more efficiently, we design an algorithm to generate virtual extraterrestrial terrains randomly and to simulate the running of a specific rover.

  Results  We generate a set of visible multi-dimensional terrain information and perform traversing test in virtual environment, which is used for generating traversibility labels in the optimization method. Based on the terrain information and labels, we train a U-Net-like network for predicting labels according to the given multi-dimensional information, and the network performs well on test dataset with the accuracy of 93.8% on average.

  Conclusions  The proposed data⁃driven method for traversability analysis is effective in virtual environment.