Machine learning methods in near infrared spectroscopy for predicting sensory traits in sweetpotatoes

It has been established that near infrared (NIR) spectroscopy has the potential of estimating sensory traits given the direct spectral responses that these properties have in the near infrared (NIR) region.In sweetpotato, sensory traits are key for improving acceptability of the crop for food security and nutrition. Studies have statistically modelled the levels of near infrared (NIR) spectroscopy sensory characteristics using Partial Least Squares (PLS) regression methods. To improve prediction accuracy, there are many advanced modelling techniques, particularly, which could be helpful when handling fresh (wet and un-processed) samples or where modelling may involve nonlinear dependence relationships. Performance of different quantitative prediction models for sensory traits developed using different machine learning methods were compared. Overall, results show that linear methods; linear support vector machine (L-SVM), principal component regression (PCR) and PLS performed better than other statistical methods. For all the 27 sensory traits, calibration models using L-SVM and PCR has slightly higher overall R2 (x = 0.33) compared to PLS (x̄ ̅= 0.32) and radial based SVM (NL-SVM; x ̅= 0.30). The levels of orange color intensity were the best predicted by all the calibration models (R2 = 0.87 – 0.89). The elastic net linear regression (ENR) and tree-based methods; extreme gradient boost (XGBoost) and random forest (RF) performed worse than would be expected but could possibly be improved with increased sample size. Lower average R2 were observed for calibration models of ENR (x ̅ = 0.26), XGBOOST (x ̅ = 0.26) and RF (x ̅ = 0.22). The overall RMSE in calibration, models was lower in PCR models (X = 0.82) compared to L-SVM (x= 0.86) and PLS (x= 0.90). ENR, XGboost and RF also had higher RMSE (0.90 -0.92). Effective wavelengths selection using the interval partial least-squares regression (iPLS), improved the robustness of the models but did not perform as good as the PLS. SNV pre-treatment was useful in improving model robustness.