Spatial Super Resolution of Hyperspectral Images: Novel Approaches for Learning Deep Spatial-Spectral Prior

Hyperspectral imaging (HSI) has become a crucial tool in remote sensing, providing rich spectral information. This study explores the application of deep learning techniques like convolution, image generation, vector quantization etc. and deep learning models such as autoencoders(AE), convolution neural networks(CNN) and generative adversarial networks(GAN) for enhancing the spatial super resolution of hyperspectral images. The proposed architectures have been specifically designed for hyperspectral imagery by using 3D convolution, conditioning, bilinear/bicubic interpolation techniques and employing loss functions like Root Mean Square Error(RMSE), Spectral Angle Mapper (SAM) and Adversarial loss for stricter training to preserve the spectral fidelity and refine the spatial resolution. Results demonstrate that the proposed models show the best performance among the other state-of-the-art models when compared using performance metrics RMSE, SAM and Peak-Signal-to-Noise ratio(PSNR).