Hybrid-domain feature fusion network with neural ordinary differential equations and multiview perception for hyperspectral image classification

Abstract

Hyperspectral image (HSI) classification often suffers from insufficient local detail, limited global semantic correlation, and inadequate frequency information, which impede effective multi-dimensional feature integration. To address these challenges, this paper proposes a novel Hybrid-domain Feature Fusion Network (HFFN) that integrates Neural Ordinary Differential Equations (NODEs) with a multi-view perception mechanism to jointly model spatial, spectral, and frequency information for robust HSI classification. The proposed framework comprises three key modules. First, a Dual-Phase Local Feature Extractor (DP-LFE) leverages NODEs to model spectral features from a continuous perspective. By formulating spectral feature extraction as a differential dynamic process, DP-LFE effectively captures smooth spectral variations inherent in hyperspectral data, mitigating the limitations of conventional discrete-layer networks in modeling spectral continuity while avoiding excessive parameter expansion. Second, a Global Perception-driven Multiview Feature Extraction (GPMFE) module explicitly encodes global semantic relationships using Euclidean distance-based similarity and a multi-view strategy, enhancing spatial feature representation while reducing redundancy. Third, a Wavelet Spectral Enhancer (WSE) selectively boosts high- and low-frequency components to suppress redundant frequency information and refine discriminative spectral representations across different bands. Experimental results on three public HSI datasets with limited training samples demonstrate that the proposed HFFN consistently outperforms state-of-the-art methods in terms of classification accuracy and robustness, particularly for complex land-cover distributions with subtle spectral differences.