Hyperspectral data scarcity problem from a super resolution perspective: data augmentation analysis and scheme

Hyperspectral Single Image Super Resolution is an important field of research due to the low spatial resolution of Hyperspectral Images (HSI) that limits their usability. Deep Convolutional Neural Networks (DCNNs) have been commonly used for SISR tasks, however, a large dataset is typically needed for training. Because of HSI data scarcity, training DCNNs for HSI-SISR becomes a challenging task. In this study, HSI data scarcity problem is tackled from an SISR perspective via Data Augmentation (DA). Several DA techniques are reviewed in this context, and a new DA technique called CutMixBlur is introduced. The best techniques are decided based on maximizing Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Index Measurement (SSIM) while minimizing Spectral Angle Mapper (SAM). The aim is to enhance HSI spatially without distorting their unique spectral signature. Experiments on Pavia University and Indian Pines datasets show that CutMixBlur boosts all quality metrics. Additionally, applying DA techniques by randomly stacking their effect significantly enhances the performance of SISR DCNNs, particularly the 3D-SRCNN.