Title: A Data-Efficient Complex Feature Fusion Network for Hyperspectral Image Classification

Abstract

This study introduces DE-CFFN, a data-efficient adaptation of the Attention-Based Dual-Branch Complex Feature Fusion Network (CFFN) designed for hyperspectral image classification. The architecture maintains the original dual-stream framework, wherein the Real-Valued Neural Network (RVNN) analyzes standard hyperspectral patches, and the Complex-Valued Neural Network (CVNN) processes their Fourier-transformed versions. The primary innovation of this research focuses on optimizing the feature extraction methodology and refining the network structure. To achieve superior latent feature representation compared to Principal Component Analysis, the model employs Factor Analysis for dimensionality reduction. Furthermore, architectural complexity is minimized by progressively reducing the filter count in the 3D convolutional layers across both the RVNN and CVNN branches. The resulting outputs from both streams are combined and processed through a Squeeze and Excitation (SE) block to strengthen the joint feature representation. Testing on the Pavia University and Salinas datasets demonstrates that DE-CFFN delivers classification accuracy on par with the original CFFN. Crucially, it achieves this with a substantially smaller model footprint, lower memory usage, and faster inference speeds, thereby rendering it well-suited for real-time hyperspectral imaging tasks.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC