Hyperspectral and Multispectral Image Fusion Based on NMF-Unmixing Method and Sparse Unmixing to Tackle Spectral Variability

Introduction

Hyperspectral images (HSIs) are of great interest for urban monitoring and environmental applications, as they provide contiguous spectral coverage spanning the visible to infrared wavelength domains, enabling accurate discrimination and classification of materials. However, HSIs have inherently low spatial resolution, which limits their use in applications requiring fine spatial detail. Hypersharpening aims at generating a high-spatial-resolution hyperspectral image by fusing a low-spatial-resolution HSI with a high-spatial-resolution multispectral image (HR-MSI). Most existing fusion methods fail to account for spectral variability (SV), a phenomenon that significantly impacts hyperspectral data analysis, particularly in urban environments where illumination and material composition changes are prevalent. Addressing SV is therefore critical to producing fusion products with accurate spectral and spatial fidelity.

Methods

A novel hypersharpening method is proposed. First, the Inertia-Constrained Pixel-by-Pixel Nonnegative Matrix Factorization (IP-NMF) algorithm is employed to construct a spectral library that explicitly incorporates SV by estimating pixel-dependent endmember spectra organized into classes. Second, a pruning strategy is applied to reduce the library size and improve computational efficiency. Third, the Sparse Unmixing by Variable Splitting and Augmented Lagrangian (SUnSAL) technique is used to extract high-spatial-resolution abundance fractions from the HR-MSI using the pruned spectral library. Finally, the sharpened hyperspectral image is generated by recombining the estimated spectral and spatial information.

Results

The proposed approach is evaluated on synthetic datasets using standard spectral and spatial performance criteria, including SAM, NMSEλ, NMSEs, PSNR, UIQI, and ERGAS. Experimental results demonstrate that the proposed method outperforms state-of-the-art methods, including CNMF, HySure, FuVar, and HMF-IPNMF, across all considered metrics.

Conclusions

The proposed fusion technique provides an effective and efficient solution for hyperspectral and multispectral image fusion under SV. The high quality of the fused products makes this approach particularly suitable for urban scene classification and time-series monitoring applications.