De-noising And Feature Extraction Of Hyperspectral Remote Sensing Data Based On EMD And Wavelet Analysis

Hyperspectral remote sensing data has the characteristics of high spectral resolution and can provide almost continuous spectral curves for objects,which can fully reflect the detailed features of the objects.The following problems is that noises caused by the internal circuit of the instrument and the external environment of the sensor in the measurement process is more likely to change details,at same time,the data redundancy caused by a large number of bands also brings great difficulties to the feature extraction of the object.In this paper,the noise reduction and feature extraction of hyperspectral data are studied by empirical mode decomposition and wavelet analysis.Both methods can decompose the signal from low frequency to high frequency.The whiteboard noises of ASD FeiledSpec 4,with high frequency,are used as noise samples,which have high amplitude and decrease from both ends to the middle,The standard spectrum of Goethite plus whiteboard noise are denoised with empirical mode decomposition and wavelet analysis respectively.After that,signal-to-noise ratio,root mean square error,spectral similarity are used to evaluate the effect of noise reduction.In terms of feature extraction,three similar spectral curves including Andesine,Anorthite and Anhydrite are selected for experiments,and the corresponding bands of extreme points are extracted as features,so as to obtain the features of extreme positions at different scales and achieve the results of stratified features.The final result shows that for the whiteboard noise of ASD FeiledSpec 4,the first few high-frequency components of empirical mode decomposition are more sensitive to noise,and the obtained signal signal-to-noise ratio,root-mean-square error and spectral similarity are better than the wavelet decomposition results under the same level of component.For the feature extraction of the three similar spectral curves of Andesine,Anorthite and Anhydrite,which are characterized by the corresponding bands of extreme points,the stratified features of the low-frequency part of empirical mode decomposition are few and have higher discrimination degree.The number of low-frequency features of wavelet decomposition is larger and the discrimination degree is lower than that of the same level of empirical mode decomposition.