HDP-Net: a hyperspectral deterministic-probabilistic dual attention network for hyperspectral image classification

Hyperspectral image classification faces significant challenges in the simultaneous processing of deterministic and uncertain features, leading to suboptimal performance in complex scenarios involving mixed pixels and spectral confusion issues. This paper proposes a Hyperspectral Deterministic-Probabilistic Dual Attention Network (HDP-Net) that addresses these limitations through three key innovations: (1) a deterministic attention branch using Sigmoid activation for independent feature weight assignment, (2) a probabilistic attention branch employing temperature-regulated Softmax for controllable uncertainty modelling, and (3) an adaptive feature fusion module that dynamically balances dual-pathway contributions based on data characteristics. The deterministic branch excels at capturing discriminative spectral-spatial patterns, while the probabilistic branch models uncertainty through the temperature parameter τ, enabling continuous adjustment between deterministic decisions and probabilistic exploration. Comprehensive experiments on three benchmark datasets demonstrate HDP-Net’s superior performance: 96.94% overall accuracy (OA) on Indian Pines (0.62% improvement over best baseline with 17% error reduction), 98.39% OA on Pavia University (superior performance in challenging mixed-pixel scenarios), and 97.08% OA on Salinas Valley (1.28% improvement representing 30.5% error reduction). The method achieves consistent top-tier performance across diverse scenarios, with average accuracies of 94.38%, 96.92%, and 98.34%, and Kappa coefficients of 96.51%, 97.87%, and 96.75%, respectively. Notably, HDP-Net demonstrates exceptional performance in challenging conditions: perfect classification (100%) for small-sample classes, 83.50% accuracy for mixed-pixel ‘Gravel’ class (vs. 1.05%–66.05% for competing methods), and robust handling of spectrally similar categories. The uncertainty quantification mechanism provides interpretable confidence estimates alongside classification results, enabling automated decision-making systems with reliability assessment for precision agriculture, environmental monitoring, and urban planning applications. These results establish HDP-Net as an effective solution for uncertainty-aware hyperspectral classification, contributing to a unified framework that advances both classification accuracy and practical deployment capabilities in complex remote sensing scenarios.