MDL and RMSEP assessment of spectral pretreatments by adding different noises in calibration/validation datasets

In multivariate calibration, the optimization of pretreatment methods is usually according to the prediction error and there is a lack of robustness evaluation. This study investigated the robustness of pretreatment methods by adding different simulate noises to validation dataset, calibration and validation datasets, respectively. The root mean squared error of prediction (RMSEP) and multivariate detection limits (MDL) were simultaneously calculated to assess the robustness of different pretreatment methods. The result with two different near-infrared (NIR) datasets illustrated that Multiplicative Scatter Correction (MSC) and Standard normal variate (SNV) were substantially more robust to additive noise with smaller REMSP and MDL value.