Gaussian process regression for prediction and confidence analysis of fruit traits by near-infrared spectroscopy

Abstract

Detection of fruit traits by using near-infrared (NIR) spectroscopy may encounter out-of-distribution (OOD) samples that exceed the generalization ability of a constructed calibration model. Therefore, need confidence analysis for a given prediction, but this cannot be done by using common calibration models of NIR spectroscopy. To handle this issue, this paper studied the Gaussian process regression (GPR) for fruit traits detection using NIR spectroscopy. The mean and variance of the GPR were used as the predicted value and confidence respectively. To show this, a real NIR data set related to dry matter content measurements in mango was used. As compared to the partial least squares regression (PLSR), GPR showed ~14% lower root mean squared error (RMSE) for the in-distribution (ID) test set. Compared with no confidence analysis, using the variance of GPR to remove abnormal samples made GPR and PLSR showed ~58% and ~10% lower RMSE on the mixed distribution test set respectively (when the type 1 error rate was set to 0.1). Compared with traditional one-class classification methods, the variance of the GPR can be used to effectively eliminate poorly predicted samples.