A large-scale Martian rock dataset and a spatial-frequency dual-domain Mamba for Martian rock segmentation

Rocks, as typical exposed geological units on the Martian surface, hold a significant role in both planetary research and engineering safety. Existing rock research faces two major limitations. First, the lack of a benchmark dataset hinders the advancement of Martian rock research. Second, high background similarity and weak texture features make rock segmentation methods often fail to achieve accurate segmentation. For the first limitation, we construct CUGB-M 2 Rock, a large-scale multi-source Martian rock dataset, containing over 600,000 annotated instances. For the second limitation, we propose an ultra-lightweight boundary-aware network (UBNet). Specifically, UBNet incorporates the frequency-assisted enhanced Mamba (FAEM), which integrates global spatial modeling and wavelet-based frequency transform to establish a dual domain learning mechanism. The former leverages a selective scanning mechanism to achieve efficient global modeling of rock sequences, while the latter utilizes a wavelet transform branch to capture discriminative frequency information across multiple scales. Furthermore, the four-layer parallel architecture significantly reduces model complexity. In addition, the multi-scale feature enhancer (MSFE) is designed to improve cross-scale feature discriminability through dual-attention mechanisms and adaptive feature selection. Simultaneously, the boundary awareness auxiliary branch (BAAB) is devised to capture fine-grained details and elevate boundary precision by explicitly modeling edge semantics. Comprehensive experiments on the CUGB-M 2 Rock and SynMars datasets demonstrate that our UBNet outperforms the other 18 methods, and achieves superior performance with low computational overhead. The CUGB-M 2 Rock dataset is available at https://github.com/ffushiyang/CUGB-M2Rock .