Energy-Based Generative Models with Morphological Attention Networks for Hyperspectral Image Classification: a Unified Framework

This paper presents an innovative framework that synergistically integrates energy-based generative models with morphological attention networks for robust hyperspectral image classification. Despite recent advances in data-driven approaches, existing methods struggle to model complex data distributions while preserving geometric invariance and maintaining stable representations under limited supervision. To address these challenges, we propose a novel architecture that combines three key components: (1) a morphological attention module that fuses learned structural elements with multi-head attention through parallel dilation and erosion pathways, (2) an uncertainty-aware energy-based learning paradigm that employs adaptive margin constraints and Langevin dynamics sampling, and (3) a multi-scale feature extraction block with parallel convolutional branches for dynamic spectral-spatial feature integration. Extensive experimental validation on benchmark datasets demonstrates our framework’s superior ability over state-of-the-art methods.