Denoising Hyperspectral Field Spectra of Vegetation with a Prosail-Fed Denoising Autoencoder

Hyperspectral field spectra are an essential foundation for the construction and evaluation of many vegetation indices and vegetation models. However, they are often contaminated by water vapor absorption and instrument noise. In this paper, a new denoising method based on a denoising autoencoder (DAE) is proposed, which was trained with simulated spectra generated by the radiative transfer model, PROSAIL. The performance of this new method was evaluated on both simulated and real-world spectra and was compared to that of several most used spectral denoising methods. Although limited by the soil spectra used for dataset generation, DAE looks remarkably promising in the denoising of vegetation field spectra, and works especially well in the removal of water vapor effects.