MarsUnet: An Adapter-Enhanced Attention Network for Martian Terrain Semantic Segmentation

Abstract

Autonomous navigation of Mars rovers requires accurate terrain perception, as surface properties directly influence traversability and wheel slip. However, semantic segmentation of Martian terrain remains challenging due to the high visual similarity and low contrast between surface classes, particularly sand and rocks. In addition, the scarcity of annotated Martian data necessitates transfer learning from Earthpretrained models, leading to significant domain-shift issues. To address these challenges, this paper proposes MarsUnet, a semantic segmentation network based on a ResNet-UNet architecture. The proposed method introduces parameterefficient parallel adapters into the encoder to improve adaptation from terrestrial to Martian imagery while preserving pretrained representations. Furthermore, a Spatial and Channel Squeezeand-Excitation (SCSE) module is incorporated into the decoder to enhance feature discrimination and improve the segmentation of low-contrast rock boundaries. Experiments conducted on the AI4MARS dataset demonstrate that MarsUnet outperforms conventional fine-tuning approaches in both Balanced Accuracy and Jaccard Score. In particular, the proposed network achieves a 12% improvement in mean IoU and a 3.8-fold improvement in obstacle-rock detection sensitivity, indicating its effectiveness for robust terrain perception and autonomous navigation in Martian environments.