ResGAT-F: a novel graph neural network-based approach for evaluating landing suitability in the lunar southern polar region

Landing suitability evaluation in the Lunar Southern Polar region is critical for future exploration, it requires integrating various environmental factors to balance safety and scientific value. This study proposes a Residual Connection Graph Attention Forest (ResGAT-F) model, which systematically integrates multi-source spatial data to extract regional features and environmental relationships, enabling a quantitative assessment of landing suitability that addressing safety and multi-disciplinary scientific exploration scenarios. Results show all ResGAT sub-models achieve over 96% accuracy based on pixel-wise labels generated by the adapted Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) from multiple small-sample regions. Ensemble ResGAT-F attains AUC above 0.92, outperforming the baseline model Attn-CNN (accuracy: 93%, AUC: 0.86). A 256 m resolution suitability map between 80°S to 90°S was generated and been scored (scoring ≥8 means highly suitable), which can evaluate site suitability across the region. Only 7.81% of the area meets safety requirements and potential of multi-target scientific exploration. Suitability evaluations for Artemis III candidate landing zones, such as Malapert Massif, indicates 28% of this area meets the requirements. The ResGAT-F in handling complex, multi-dimensional lunar data shows potential for supporting future landing missions and improving lunar exploration planning.