A hierarchical geo-aware cross-modal transformer for multiscale lunar crater detection

Impact craters are fundamental geomorphological features on the lunar surface and play a crucial role in planetary chronology, geological evolution analysis, and mission planning. Accurate and automated detection of multiscale craters remains challenging due to extreme diameter variation, degraded rim structures, overlapping crater fields, and heterogeneous terrain conditions. Although recent deep learning–based methods have significantly improved detection performance, most existing approaches rely on single-modality inputs, anchor-based detection mechanisms, or convolution-dominated architectures that inadequately model long-range spatial dependencies and geomorphological priors. This study proposes a Hierarchical Geo-Aware Cross-Modal Transformer (HGC-Transformer) for multiscale lunar crater detection. The framework integrates Kaguya Terrain camera optical imagery and SLDEM2015-derived terrain features through early multi-modal patch embedding. A hierarchical Transformer encoder that incorporates a geo-attention mechanism (with gradient prior) is used for long-range context dependency modelling and attention mechanisms. In addition to enabling the integration of terrain gradient prior knowledge into the attention mechanism, a cross-scale transformer feature pyramid is introduced to improve the consistency of representations at multiple scales. The geo-shape aware refinement module is designed to improve the accuracy of the boundaries and radial symmetry of objects by using the spatial characteristics of each object. Finally, an anchor‑free centre‑based detection head is added to enable flexible location estimation over continuous crater diameter ranges and to avoid discretising the scale. Experiments conducted on the multimodal lunar crater dataset demonstrate that the proposed method achieves a precision of 92.84%, recall of 95.17%, F1-score of 93.99%, and IoU of 91.43%, outperforming lightweight CNN backbones, hybrid CNN–Transformer models, and state-of-the-art multimodal approaches. The proposed HGC-Transformer establishes a robust and scalable framework for high-precision multiscale lunar crater detection and provides enhanced support for automated crater cataloguing and planetary geological analysis.