Abstract: In deep space exploration missions, craters are widely recognized as crucial terrain landmarks for visual autonomous navigation due to their abundance, wide distribution and stable features. To address the challenges posed by existing crater identification methods, including high computational complexity, poor environmental adaptability and extensive data requirements, a crater localization method based on a composite channel-mixed neural network was proposed. Channel fusion input data were utilized, and a Convolution-Stem Vision Transformer (CS-ViT) was designed to optimize spatial feature modeling capabilities. Additionally, data generation and augmentation strategies were introduced, including random positional deviations, false positives and false negatives, to enhance model robustness and generalization. Experimental results demonstrate that the proposed CS-ViT model was significantly superior to traditional methods in identification accuracy, resource efficiency and environmental adaptability, and it achieved higher recognition accuracy and robustness under complex interference conditions. This study provides an efficient and reliable technical solution for visual autonomous navigation in deep space exploration missions.