A comparison of methods in feature extraction for the classification of algae

Eight common species of algae belonging to four categories were chosen and cultivated, and their three-dimensional fluorescence spectra were measured. The noise existed in spectra was eliminated by Empirical model decomposition technology. Three different ways of feature extraction were compared, among which, the first one was based on the wavelet analysis with the three-dimensional fluorescence spectra. The second was based on the distribution of the primary pigment, which simplified the raw spectra to a characteristic matrix with fewer data. The third one only used a discrete characteristic excitation spectrum with seven extraction wavelengths (405,435,470,490,530,555, and 590nm) to present the feature of each alga. By setting up the feature database with these three ways, the right recognition rates of pure test samples are 82.4%, 94.1% and 88.2%, and that in the mixture are 63.3%, 83.3% and 67.7% respectively. For the third feature-extracted method, after adding with another two new features: (a) the first component from principal component analysis; (b) the ratio of relative fluorescence intensity at 435nm and 470nm. The right recognition rate of the mixture could increase to 76.6%. Therefore, not only could the discrete characteristic excitation spectra satisfy with the basic algal identification requirements, but also could be more efficient.