Noise in hyperspectral images

For clarity, we will utilize synthetic data to illustrate the problems and potential solutions. Images from a synthetic dataset with dimensions of 25 x 25 in the spatial domain and containing 100 channels in the spectral domain are shown in Fig. 4.5. Spectra at each pixel consist of a Gaussian peak at spectral channel 50. The area under the Gaussian change from 0 to 30 across the spatial domain in the patterns 

(b) Low frequency noise superimposed on (a), (c) High frequency noise superimposed on (a), (d) High and low frequency noise superimposed on (a). 

Perhaps, the simplest method for minimizing sample-induced spectral baseline fluctuations is by calculating derivative (first, second, etc.) spectra. In derivative spectra, broad baseline fluctuations have much smaller slopes and curvatures than vibrational absorption bands in the original spectra, and as a consequence, the resulting derivative spectra have diminished broader features. This effect is seen in the Fig. 4.6, 

First-derivative spectra Second-derivative spectra 

---