Quantile BEAST Attacks the False-Sample Problem in Near-Infrared Reflectance Analysis

Abstract

The multiple linear regression approach typically used in near-infrared calibration yields equations in which any amount of reflectance at the analytical wavelengths leads to a corresponding composition value. As a result, when the sample contains a component not present in the training set, erroneous composition values can arise without any indication of error. The Quantile BEAST (Bootstrap Error-Adjusted Single-sample Technique) is described here as a method of detecting one or more “false” samples. The BEAST constructs a multidimensional form in space using the reflectance values of each training-set sample at a number of wavelengths. New samples are then projected into this space, and a confidence test is executed to determine whether the new sample is part of the training-set form. The method is more robust than other procedures because it relies on few assumptions about the structure of the data; therefore, deviations from assumptions do not affect the results of the confidence test.