Research On Fourier Transform Infrared Ultraspectral Data Classification

The Fourier transform infrared(FTIR) ultraspectral spectrometers based on Michelson interferometey have the advantages of full range cover of infrared bands, higher spectral resolution(0.1 nm) and shorter full-range scanning time(0.01 s) compared with the dispersive spectrometers. It makes the applications such as high speed object recognition and atmospheric composition fine analysis become possible. The ultraspectral waveform pattern classification is the fundamental of these new applications, and these new applications raise new challenges. The academia and industry have not yet started to research on the classification method for the FTIR ultraspectral data, and they apply the hyperspectral classification method to the FTIR ultraspectral applications. Since the high dimension of ultraspectral data, the traditional hyperspectral classification method cannot meet the high real-time and high accuracy needs.The dimensions of data that the traditional hyperspectral classification methods dealed with are about hundreds, and the dimension of FTIR ultraspectral spectrometer can reach up to about ten thousands. Thus the time consumption for classification increase dramatically and make it difficult to the real-time applications. Aimed to this problem, this research conducts on enhancing both of algorithm efficiency and hardware performance. Firstly,this thesis proposes a space hierarchical spectral histogram concept and calculation method.Base on it, the SPM algorithm has been firstly introduced to the ultraspectral data classification and used for feature extraction and matching, which make the time consumption of the similarity calculation significantly reduced, and thereby improves the real-time performance of the system. Secondly, a FTIR ultraspectral signal processing system prototype is designed, developed and tested through a field experiment. It provides a hardware platform for the real-time hyperspectral remote sensing applications.In the ultraspectral spectrometers, the signal-to-noise ratio(SNR) is in inversely proportion to spectral resolution. i.e., when the spectral resolution improves, the SNR decreases, and finally the profile of the spectrum changes. Experiments show that with the SNR decreases, the classification accuracy decreases exponentially. Aimed to this problem, we proposed a restricted Boltzmann machine(RBM) based FTIR ultraspectral classification method. The experimental results show the proposed method can effectively address the SNR decrease issue.In conclusion, this thesis works on the theoretical research of the FTIR ultraspectral classification method and experimental demonstrations and solve the key problems in the applications of the FTIR ultraspectral technology. It brings strong technical support for the applications such as high speed object recognition and atmospheric composition fine analysis and can help to broaden the fields of FTIR ultraspectral applications.