Evaluation of a robust regression method (RoBoost-PLSR) to predict biochemical variables for agronomic applications: Case study of grape berry maturity monitoring

Visible and near infrared spectroscopy (VIS-NIR) is increasingly being transferred from laboratory to industry for in-line and portable applications in various domains. By intensively using VIS-NIR spectroscopy, some abnormal observations may certainly arise. It is then important to properly handle outliers to elaborate effective prediction models. The objective of this study is to investigate the potential of using a robust method called Roboost-PLSR to improve prediction model performances for a viticulture application. This work focuses on a case study to predict sugar content in grape berries of three different grape varieties of Vitis Vinifera in a maturity monitoring context. Hyperspectral images were acquired of grape berries of Syrah, Fer-Servadou and Mauzac varieties. Reference measurements of sugar levels were made in the laboratory by densimetric baths. Performances of RoBoost-PLSR models were compared to performances of reference models using Partial Least Square Regression (PLSR). Reference prediction criteria using PLSR were obtained for all varieties with these following values: Syrah (R(2)p = 0.971; RMSEp = 5.36 g/L), Fer-servadou (R(2)p = 0.788; RMSEp = 11.69 g/L) and Mauzac (R(2)p = 0.690; RMSEp = 15.61 g/L). Prediction qualities are improved with RoBoost-PLSR: Syrah (R(2)p = 0.990; RMSEp = 3.14 g/L), FerServadou (R(2)p = 0.848; RMSEp = 10.20 g/L) and Mauzac (R(2)p = 0.927; RMSEp = 7.58 g/L). Results confirm that Roboost-PLSR method allows a better consideration of outliers within the calibration set.

Automated summary

This study investigates the potential of using the robust method Roboost-PLSR to improve prediction model performances for predicting sugar content in grape berries. The results show that Roboost-PLSR method allows for better consideration of outliers, resulting in improved prediction qualities.