Hierarchical dynamic spatio-temporal graph neural network for aerial group target intention prediction

Predicting the intention of aerial group targets is central to battlefield situation awareness: the aim is to forecast, from a sequence of historical observations, how the future tactical intentions of heterogeneous formations will evolve. Building this capability faces three major challenges: the lack of hierarchical multi-granularity representations at the individual, interaction, and formation levels; the standard graph attention is scale-sensitive to formations and ignores platform heterogeneity; a single time scale cannot capture multi-scale tactical patterns simultaneously. This paper proposes HD-STGNN, which has three core modules: HMR uses three-layer encoding and cross-granularity bidirectional attention for multi-level feature fusion; SA-DHAT ensures the consistent modeling of heterogeneous formations relying on type-specific projection and adaptive temperature scaling; IEA-TM captures multi-scale tactical dynamics using multi-scale parallel temporal convolution and intention transition detection mechanisms. On the datasets covering air-to-air, air-to-ground and air-to-sea scenarios, HD-STGNN achieves accuracies of 83.82%, 89.67% and 90.07% respectively, which are 7.55%, 5.33% and 5.92% higher than the second-best baseline respectively. Ablation study and statistical significance analysis confirm the effectiveness and reliability of the proposed method.