Texture-based segmentation for sand and rock in Mars images

The exploration of Martian’s surface is one of the most important aspects of understanding Martian environment. Investigating features such as sand dunes can provide valuable insights into Mars past environmental dynamics and geological history. However, self-driving in uncontrolled and unsupervised environments, like Mars, is a very challenging problem. To date, six Mars rovers have been successfully sent and operated on the surface, the latest of which is NASA’s Perseverance Rover and the next is planned to be ESA’s Rosalind Franklin rover. These cutting-edge spacecraft are driven by classical machine vision systems, which could cause some limitations to the safety, reliability, and productivity of these missions. This research aims to develop new technologies in Feature Extraction, Image Processing and Deep Learning for the identification of Mars Terrain to help future self-driving rovers navigate. Specifically, in this work, we focus on the identification of rocks and sands by utilising both linear (Gabor Filter), and nonlinear (Polynomial Bilateral Filter) feature extraction techniques. To combine the benefits of both filters we also investigate the effect of using image fusion technique. Our feature extraction response achieves promising results when integrated with deep learning methods with the highest accuracy achieved at 89.84%. However, to improve the accuracy further, we analysed some of the images contributing to erroneous classifications. Most of these images feature low-contrast terrain, primarily due to dust or poor lighting conditions. A framework is introduced which includes a contrast investigation stage to determine the required level of image enhancement when processing the images. This resulted in an improved accuracy of 93.70%. Finally, some suggestions for future improvements are included in this paper.