Gas-mixture IR absorption spectra denoising using deep learning

Abstract

A preliminary experimental absorption spectrum denoising improves the further analysis. The majority of popular filters causes not only a nose decreasing but also a spectrum shape distortion. An approach to a gas-mixture IR absorption spectrum denoising using a multilayered perceptron deep neural network (MLP DNN) with autoencoder architecture was suggested. This deep neural network was trained and tested on gas mixtures related closely to the ground atmosphere. Absorption spectra of the latter were calculated in the 0.9− 11μm spectral range using data from the HITRAN database, then a random noise was added. The results of the MLP DNN filter application were compared with the standard Gaussian filter. In common, MLP DNN filter provided more effective noise decreasing and less initial spectrum shape distortion compared to the Gaussian filter.