Hyperspectral Image Super Resolution using Deep Learning

Abstract

Hyperspectral images can easily discriminate different materials due to theirfine spectral resolution. At any condition, obtaining a hyper-spectral image (HSI) with the high level configured spatial resolution is still a challenge as we are limited by the hardware present. Super-resolution (SR) is basically concerns on the different methods which can improve the spatial resolution of theimage. Viewing the current configured methods in Convolutional neural network (CNN) extended in super-resolution with respect to natural (RGB) images, we have optimized a model for spatialsuper-resolution in hyperspectral images which is three-dimensional complex CNN(3D-FCNN). Herethree-dimensional(3D) convolution is used so that the spatial content of the image issuper-resolved without losing the spectral information. 3D-FCNN which we have used has a sensor-specific mode so that while training the model we do not require images from target scene. However, fine-tuning using a few images from the target scene can definitely improve the performance of the model. First, we ran our model using Pavia center dataset and as wehad taken a few images from AVIRIS-NG sensor we tested our model on that Dataset too.