Noise Robustness Comparison for near Infrared Prediction Models

The effect of six artificial noises added to near infrared spectra on the predictive robustness of ten calibration algorithms was investigated for the prediction of whole corn moisture. White noise, multiplicative noise, baseline shift, wavelength shift, spectral stretch/ shrink and stray light were independently added to a set of whole corn near infrared spectra used to validate various regression models based on partial least squares regression, locally weighted regression, variable selection, artificial neural networks and least-squares support vector machines. The quantitative increase in the root mean square error of prediction relative to a unit of noise was used as the robustness criterion. The effectiveness of standard normal variate (SNV) was also tested to remove artificial noises for all models. All models were highly sensitive to white noise and stray light (0.92% pt increase in error for 0.1% of white noise; 3.01% pt increase in error for 0.1% of stray light added to the data, with partial least squares regression). Model robustness was improved by the use of SNV. Variable selection techniques were particularly sensitive to wavelength shift, white noise and stray light. However, SNV pre-treatment before modelling improved the results.