Effect of denoising methods for hyperspectral images classification: DnCNN, NGM, CSF, BM3D and Wiener

Abstract

Hyperspectral images contain a large amount of information and are composed of several spectral bands. These data are used to classify pixels for land use/cover analysis, which is a popular subject, particularly in remote sensing. Because of noise caused by systematic and random errors, these data cannot be presented to end users most of the time. In this paper, Pavia university’s hyperspectral dataset with Gaussian, salt & pepper, poisson, and speckle noise were denoised using DnCNN, NGM, CSF, BM3D, and Wiener. Then, denoised data was classified using the k-nearest neighbor method. Following that, statistical and visual performance comparisons were performed for classified data. The BM3D method was statistically the most successful, whereas the NGM method was the least successful. The validated instruments not provide effective results when it came to denoising salt & pepper noise, but they managed to produce outstanding results when it came to denoising poisson noise.