Anomaly Detection for Camouflage Targets via Hierarchical Attentional Fusion in UAV Hyperspectral Imagery

Hyperspectral imaging (HSI) provides abundant spectral information for fine-grained target analysis, yet anomaly detection of camouflaged objects in UAV-based HSI remains challenging due to strong background similarity, spectral redundancy, and complex spatial variations across different observation scales. To address these issues, this paper proposes a Hierarchical Attentional Fusion (HAF) framework for hyperspectral camouflage target anomaly detection. The proposed model integrates a spectral–spatial self-attention mechanism to capture long-range dependencies among highly correlated spectral bands and spatial contexts, together with a hierarchical multi-scale feature fusion module to enhance robustness against target scale variations caused by different UAV flight altitudes. Specifically, the self attention module sequentially models spectral correlation and spatial dependency, while a triple-branch convolutional structure with 1×1, 3×3, and 5×5 kernels adaptively aggregates multi-scale contextual information. A composite loss function combining re construction loss, orthogonal constraint, and attention regularization is further employed to stabilize feature learning and improve anomaly separability. Experiments are conducted on real-scene UAV hyperspectral datasets containing four types of military standard camouflage nets acquired at multiple flight heights. The results demonstrate that the proposed HAF framework consistently outperforms state-of-the-art hyperspectral anomaly detection methods, achieving superior AUC performance and robust generalization under complex background conditions. These findings indicate that HAF provides an effective and practical solution for UAV-based hyperspectral camouflage target detection.