A quasi-qualitative strategy for FT-NIR discriminant prediction: Case study on rapid detection of soil organic matter

Abstract

Fourier transform near infrared (FT-NIR) is a technology to provide direct and rapid quantitative determinations of soil organic matter (SOM). In this paper, a new discriminant method is proposed for quasi-qualitative determination by combining the interval search principal component analysis algorithm with logistic regression (iPCA-LR). We firstly predict the SOM content of soil samples based on the partial least square (PLS) regression. To build up a quasi-qualitative analytical strategy, we design various fault-tolerant thresholds. Discriminate the sample marks as accurate or non-accurate according to the predicted values from priori PLS and the thresholds. The quantitative calibration model is thereby transformed into a quasi-qualitative discriminant model. We then leverage iPCA-LR to select informative FT-NIR wavebands with parameter optimization, according to the optimal discriminant accuracy. Results show that the FT-NIR quasi-qualitative discriminant predictive accuracy varies significantly with thresholds varying, but fortunately that the optimal accuracy climbed to above 74%. Furthermore, the test of different informative wavebands outputs the optimal calibration models with an accuracy above 88%. In the SOM content prediction of FT-NIR, iPCA-LR converts the quantitative problem into the quasi-qualitative discriminant issue when combined with the threshold-transformed PLS results. The quasi-qualitative strategy helps to overcome the over-idealistic modeling in PLS quantitative analysis. It is more beneficial for the real-time application of spectroscopy technology.