Denoising hyperspectral images using spectral domain statistics

Hyperspectral imaging has proven useful in a diverse range of applications in agriculture, diagnostic medicine, and surveillance to name a few. However, conventional hyperspectral images (HSIs) tend to be noisy due to limited light in individual bands; thus making denoising necessary. Previous methods for HSI de-noising have viewed the entire HSI as a general 3D volume or focused on processing the spatial domain. However, past findings suggest that spectral distributions exhibit less variation than spatial patterns. Hence it would be fruitful to take specific advantage of the more predictable behavior of spectral domain data for denoising. In this paper, we present a two-stage de-noising framework that first emphasizes denoising in the spectral domain and then uses spatial information to further improve spectral domain denoising. Our results indicate that specifically leveraging the spectral domain for denoising can provide state-of-the-art performance even from a relatively simple approach.