Improving Spectral Resolution of Multispectral Data Using Convolutional Neural Network

Hyperspectral data, despite of possessing high spectral resolution, suffers from narrow swath, which hinders its wider applications. Till now, there are many fusion methods to improve the resolution of data. However, most of the fusion methods are focused on enhancement of the overlapping area and cannot extend the swath of the hyperspectral data. Thus, in this paper, a multispectral image spectral resolution improving method using convolutional neural network (CNN) is proposed, which is based on the hypothesis that there exists learnable nonlinear mapping between hyperspectral data and multispectral data, and when the land cover is the same, the mapping between data from the overlapped area is the same as that from the non-overlapped area. By training the network using the data of the overlapped area, the method can predict the hyperspectral data of nonoverlapping area, or in other words, extend the swath of hyperspectral data. The architecture used in this method is a relatively simple three-layered structure yet powerful to extend the swath of hyperspectral data as the experimental results shows.