Hyperspectral redundancy detection and modeling with local Hurst exponent

Hyperspectral reflectance means a curve in a range of certain wavelength, the complex dynamic structure of which reflects the rich information of an object at different bands, which is often used as various modeling inputs. However, the potential redundancy associating with the information mentioned above will have serious impacts for the accurate extraction of spectral features. Thus, detecting information redundancy is a critical processing for the spectral analysis. By using the local detrended fluctuation analysis, we propose a new method detecting the redundant bands, which focuses on the spectral auto-correlation represented by local Hurst exponent in the moving windows, and the redundant bands can be defined through comparing the auto-correlation between two adjacent windows. Finally, with the fractal feature of the removing redundant bands as the augment, the rapeseed oleic acid prediction model based on the random decision forest is constructed to test our method. For the purpose of comparing, the same feature as the original spectrum is also employed as the augment for the model. The testing result shows that the feature obtained by removing the redundant bands has better performance over the feature of the original spectrum.