Correcting sample selection and elimination of singular sample is very important for the quantitative and qualitative modeling of near infrared spectroscopy. However, methods for identification of singular sample available are generally based on data center estimates which require an experience decision threshold, this largely limit its recognition accuracy and practicability. Aiming at the low accuracy of the existing methods of singular sample recognition problem, this paper improves the existing metric - Leverage value and presents a new algorithm for near infrared singular sample identification based on strong influence degree. This metric reduces the dependence on the data center to a certain extent, so that the normal samples become more aggregation, and the distance between the singular samples and the normal samples is opened; at the same time, in order to avoid artificial setting threshold unreasonably according to experience, this paper introduces the concept of the jump degree in the field of statistics, and proposes an automatic threshold setting method to distinguish singular samples. In order to verify the validity of our algorithm, abnormal samples of 200 representative samples were eliminated in the calibration set with using Mahalanobis distance, Leverage- Spectral residual method and the algorithm presented in this paper respectively; then through partial least squares (PLS), the rest of the calibration samples were made quantitative modelings (took Nicotine as index), and the results of quantitative modelings were made a comparative analysis; besides, 60 representative testing samples were made a prediction through the modelings; at last, all the algorithms above were made a comparison with took Root Mean Square Error of Cross Validation (RMSECV), Correlation Coefficient (r) and Root Mean Square Error of Prediction (RMSEP) as evaluation Index. The experimental results demonstrate that the algorithm for near infrared singular sample identification based on strong influence degree significantly improves the accuracy of singular sample identification over existing methods. With lower RMSECV (0. 104), RMSEP (0. 112) and higher r (0. 983), it also contribute to boost the stability and prediction ability of the model.
