Prototype-enhanced contrastive learning for terrain classifier of extraterrestrial mobile robots

Terrain classification is critical for autonomous navigation and scientific exploration of extraterrestrial mobile robots. However, the performance of conventional segmentation models degrades when deployed to extraterrestrial environments characterized by monotonous visual textures, low inter-class variation, and extreme illumination conditions. To address these challenges, this paper proposes a Global Prototype-Enhanced Contrastive Segmentation (GPECS) framework. The framework introduces a novel two-stage prototype learning strategy to enhance feature discriminability. During training, the model is optimized via a combined loss that integrates a pixel-wise contrastive loss and a batch-wise prototypical contrastive loss to learn robust feature representations. Subsequently, a set of global prototype vectors is computed for each semantic category using the entire training set, which serve as category priors during the inference stage. The final segmentation is obtained by similarity scores calculated via the global prototype set. Experiments conducted on a multi-mission planetary image dataset comprising data from active rovers demonstrate the effectiveness of GPECS, which confirms our method can provide technical support for future missions.