Skip to content
Open access

FAD-Net: Frequency Alignment Dual-Branch Network for Hyperspectral Image Super-Resolution

Aug 2026 · Remote Sensing · Vol 18, pp. 2631 · 0 citations · 38 references

TL;DR

This work proposes a Frequency Alignment Dual-Branch Network (FAD-Net) for HSI SR, a hybrid methods combining convolutional neural networks and Transformers that achieves the best Spectral Angle Mapper in all nine settings while maintaining competitive spatial performance.

Abstract

Hyperspectral image super-resolution (HSI SR) aims to recover high-resolution hyperspectral images from low-resolution observations while preserving spatial details and spectral fidelity. Accurate spectral preservation is a key distinction between HSI SR and natural image SR. Recent hybrid methods combining convolutional neural networks (CNNs) and Transformers have substantially improved spatial reconstruction performance, but spectral fidelity remains insufficiently explored. Our frequency perturbation analysis reveals that CNNs and Transformers exhibit complementary frequency-response characteristics. To explicitly exploit this complementarity, we propose a Frequency Alignment Dual-Branch Network (FAD-Net) for HSI SR. Specifically, the Multi-Scale Frequency Refinement Branch (MFRB) restores local high-frequency details through wavelet decomposition, while the Spatial–Spectral Frequency Interaction Branch (S2FIB) models long-range spatial and spectral dependencies via window-based attention. The two branches are fused by a Frequency Alignment Block (FAB). During training, Hyperspectral Frequency Loss and Spectral Angle Mapper loss are further introduced to constrain the reconstructed results toward the ground truth. Experiments on Chikusei, Botswana, and Pavia Center at ×2, ×3, and ×4 scales show that FAD-Net achieves the best Spectral Angle Mapper (SAM) in all nine settings while maintaining competitive spatial performance.

Read PDF

Similar papers

Open access Nov 2026

Hyperspectral image super-resolution using spatial-spectral attention and spectral transformer

Hyperspectral image super-resolution (HSI-SR) aims to reconstruct high-resolution hyperspectral images (HR-HSI) from low-resolution observations (LR-HSI). Existing convolutional neural network (CNN)-based methods excel at extracting local spatial features but struggle to capture long-range spectral dependencies, while...

Thanh-Van Nguyen, Duy-Thanh Phan, Quoc-Dung Nguyen · 0 citations
Oct 2026

Single Hyperspectral Image Super-Resolution Using Multi-scale Cross-spectral Transformer Network

Hyperspectral image super-resolution seeks to produce images with finer spatial detail while keeping spectral information consistent, which supports more reliable analysis in hyperspectral imaging applications. However, existing methods mainly rely on 2D/3D convolutions, which are limited to modeling local spatial-spec...

Hai-Jun Wang, Yao-Wei Wang, Ya-Lin Nie et al. · 0 citations
Sep 2026

Position-Aware Dual-Domain Hybrid Transformer for Single Hyperspectral Image Super-Resolution

A Position-Aware Attention Module (PAM) that employs MLP-based positional encoding to inject spatial structure information, thereby strengthening positional representation and a Query Interaction Mechanism (QIM) to enable complementary fusion of cross-domain features, where outputs of spatial and spectral branches serv...

Hai-Jun Wang, Hao-Yu Hu, Ya-Lin Nie et al. · 0 citations
Open access 2026

Residual Learning and Feature Weighing-Based Mamba Network for Hyperspectral Image Super Resolution

The hyperspectral image super-resolution (HSI-SR) enhances spatial resolution while maintaining the information-rich spectral bands in HSI. However, managing the computational overload of spectral bands and capturing the long-range spatial dependencies remains a challenging task. To address these issues, this letter pr...

J. Sandana, S. Deivalakshmi · 0 citations
Nov 2026

Preserving spectral fidelity in hyperspectral super-resolution with gradient-weighted loss

Deep learning has substantially improved the spatial enhancement of hyperspectral imagery, yet training objectives remain dominated by band-agnostic pixel-wise fidelity terms that treat every pixel as equally informative, regardless of how much diagnostic spectral structure its spectrum contains. A super-resolved image...

Aysha Alharam, Paul Murray, Annalisa Riccardi et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.