A Dimension Reduction Method Applied in Spectrum Analysis

It is the premise of establishing stable and accurate model to extract useful information from spectrum data in Vis/NIR spectrum analysis technology. ISOMAP is a dimension reduction method, and can effectively extract the intrinsic low dimension from high dimensional data, but is sensitive to noise and neighborhood parameter. In this paper, an improved ISOMAP algorithm, called supervised dimension reduction, is proposed. It guides the construction of the neighborhood graph using correlation owned by spectrum data, and reduces sensitivity to noise and neighborhood parameter. The algorithm was applied to two datasets, and then PLS models were established. The experiment results indicated that the improved algorithm was less sensitive to the neighborhood size and more robust and more topologically stable. In addition, smaller dimension was extracted, and the model precision was improved at the same time.