Super-resolution mapping of hyperspectral data using Artificial Neural Network and wavelet

Abstract

Hyperspectral images are characterized by high spectral but moderate spatial resolution and are therefore contains mixed pixels. Hyperspectral is capable of detecting the type of material represented by endmembers by utilizing the unique absorption dip in the reflected wavelength spectra. The quantification of present endmembers in mixed pixel has been investigated widely in the recent times. This quantification is known as Hyperspectral unmixing which also refers to the decomposition of these mixed pixels into the set of endmembers and abundance fractions. Estimation of abundance fraction has often been solved based on the linear mixing model because of their simplicity while on many occasions non-linear spectral unmixing has been utilized to enhance the simulation of spectral mixing. In this study, we have proposed a novel approach (cum algorithm) to super resolved the fractional images (generated through spectral unmixing). We have used Artificial Neural Network (ANN) coupled with wavelet, that decomposes the images into different level of frequencies. These frequency components are further fed to ANN to learn the network. A learned network is then used to predict coarser and finer features separately thereby preserving the smaller features and produces a better representation of higher resolution map. The accuracy of the super-resolved map was compared using a high resolution image. Object characterization is adopted for metric accuracy measure and positional accuracy for spatial consistency. We also performed metric comparison for few features in original image, high resolution image (obtained from Google Earth image at best resolution) and super-resolved image. We found the algorithm is performing very well for comparatively larger features and satisfactory for others. We also found the very good metric accuracy and positional accuracy to represent the features in super-resolved images. The results shows, features are better represented in super-resolved image than original image when compared their dimensions with high resolution Google Earth image. Since we have degraded the original image to collect the training data so this approach may not be suitable for small features however performing very well for relatively bigger features.