MRT-43K: A Large-Scale Benchmark for Data-Centric Mars Terrain Classification

Data-centric artificial intelligence is essential for autonomous Mars exploration, where long communication delays prevent real-time human intervention and require reliable onboard perception. In this setting, the performance of Mars terrain classification—critical for safe rover navigation—is primarily limited by training data quality and diversity rather than model architecture. However, existing Martian terrain datasets are constrained by low resolution, missing color information, non-terrain artifacts, and insufficient coverage of complex mixed terrains, limiting robust data-driven learning. We introduce MRT-43K, a large-scale benchmark for data-centric Mars terrain classification, comprising 43,756 high-resolution RGB images curated from the raw archives of NASA’s Curiosity and Perseverance rovers. To support scalable dataset construction from large unlabeled planetary imagery, we propose MRT-AL, a single-pass active learning framework that leverages Transformer self-attention to quantify local feature richness and select informative samples without iterative retraining. Using MRT-AL, we further construct MRT-challenge, a challenging evaluation set featuring mixed terrains and diverse occlusions to assess model robustness. Extensive experiments demonstrate that MRT-43K consistently improves terrain classification performance across diverse architectures, while evaluations on MRT-challenge reveal substantial robustness gaps in existing methods. These results confirm that data quality and scale are the primary bottlenecks in Mars terrain classification, establishing MRT-43K as a foundational benchmark for data-centric autonomous exploration.