ATN-Hybrid: a hybrid attention network with deterministic-probabilistic mechanism for hyperspectral image classification

Hyperspectral image classification faces significant challenges due to the high dimensionality of spectral data and complex feature relationships between bands. To address the limitations of existing methods, we propose a novel hybrid attention network (ATN-Hybrid) based on a deterministic-probabilistic mechanism for hyperspectral image classification. The network employs a dual-branch structure: the deterministic branch generates stable feature weights through explicit channel mapping and normalization, while the probabilistic branch implements temperature-guided adaptive feature selection through a soft attention mechanism. Additionally, we design an adaptive feature fusion strategy with learnable fusion coefficients for dynamic feature optimization. Extensive experimental validation is conducted on four benchmark datasets: Indian Pines, Pavia University, Kennedy Space Center, and Salinas Valley. Experimental results demonstrate that ATN-Hybrid achieves superior performance compared with existing state-of-the-art methods, attaining overall accuracies of 96.01%, 95.23%, 97.37%, and 95.99% on the four datasets, respectively. The method shows particular advantages in both distinguishing classes with similar spectral features and handling limited training samples, achieving favorable results using only 3–9% of the data for training. The hybrid attention mechanism demonstrates strong generalization capabilities in both agricultural and urban scene classification while maintaining stable performance across different spatial resolutions and spectral characteristics.