Deep Learning based Semantic Segmentation for Mars Rover Terrain Classification

Terrain classification is crucial for the successful execution of autonomous navigation and path planning during Mars rover missions. This study focuses on enhancing the rover's capability to traverse the Martian surface by investigating the integration of advanced semantic segmentation models based on deep learning. The aim is to identify the most effective deep learning model from recent advancements and establish efficient training approaches.The study selected the state-of-the-art U-Net and DeepLabV3+ models for further assessment and evaluation, utilizing both the AI4Mars and ESA's LabelMars datasets. Techniques such as preprocessing, augmentation, and various loss functions were investigated to improve model performance and class imbalance issues are tackled. To mitigate overfitting, regularization techniques like weight decay and early stopping were applied, ensuring robust model training. Additionally, to further enhance the model's performance, especially in recognizing rare classes, the study investigated the use of state-of-the-art GAN models for generating new images and expand training sets.Our findings reveal that excluding the background class from training and testing significantly improves model performance. Using early stopping regularization reduces the training time drastically while giving high model performance. Notably, the DeepLabV3+ model surpasses the performance reported in existing literature, achieving a maximum segmentation accuracy and Mean Intersection over Union (mIoU) of 99% and 87% on the AI4Mars dataset, and 87% and 72% on the LabelMars dataset, respectively. The integration of GAN-generated images into training further improved rare class performance by up to 2%. These advancements in deep learning models for terrain classification promise to significantly enhance the capabilities of Mars rovers in autonomous navigation and path planning.