Abstract

Hyperspectral (HS) images are critical in fields like remote sensing, where high spectral resolution is essential for specific detection tasks. Recent advancements in spectral reconstruction (SR) from RGB or multispectral (MS) images have been driven by deep convolutional networks (CNNs). Rather than introducing a new architecture, this work proposes a novel spectral-aware perceptual loss function that enhances the spectral fidelity of existing learning-based SR methods. Experiments show that incorporating this loss function improves spectral reconstruction quality. Additionally, a new dataset from Landsat-8/OLI and EO-1/Hyperion images is introduced, offering a valuable resource for SR research in remote sensing. The method successfully reconstructs HS spectra from MS data, providing a cost-effective solution for remote-sensing applications where HS data is scarce.