A Novel Method for Denoising Remotely Sensed Hyperspectral Image Using Autoencoder Technique

摘要

Denoising of Hyper Spectral Imagery (HSI) is an essential stage prior to further processing such as dimensionality reduction, classification and object identification. The traditional noise reduction filtering techniques may not preserve the information across the spectral bands. Hence the information loss leads to performance degradation in advanced processing stages. This paper proposes denoising of HSI using stacked autoencoders, in which concept of deep neural network is used. Particularly, this model is experimented to determine the degree of reconstruction of HSI image in which, it improves the perception quality of the inherent noisy bands and also reconstructs the normal bands with negligible changes. The proposed method has been used to experiment the robustness of the model against input images with various quantity of noise. The comprehensive evaluation of the model is extended by computing the statistical parameters Peak Signal to Noise Ratio (PSNR), Image quality Index (IQI), Mutual Information (MI) and entropy. The performance of the model is analysed by signal analysis, visual inspection of different bands, comparison of spectral signature of pixels. The proposed algorithm demonstrated that stacked autoencoder is better solution for noise reduction in Hyperspectral images with different noise densities.