[Outlier sample discriminating methods for building calibration model in melons quality detecting using NIR spectra].

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Spectroscopy

Context:

• Soluble solids content (SSC) is a key quality parameter for melons.
• Near-Infrared (NIR) spectroscopy is a rapid, non-destructive method for SSC measurement.
• Outlier samples in calibration datasets can negatively impact model precision.

Purpose:

• To investigate the influence of outlier samples on NIR calibration models for melon SSC.
• To evaluate different outlier detection methods (predicted concentration residual, Chauvenet, leverage and studentized residual tests).
• To develop an optimized calibration model by carefully reintroducing potentially valuable outlier samples.

Summary:

• Three outlier detection methods identified nine suspicious samples from an 85-sample calibration set.
• These nine samples were individually reassessed, with five beneficial samples reintegrated into the calibration set.
• The refined model achieved a higher correlation coefficient (r=0.889) and lower calibration errors (RMSEC=0.601 Brix) compared to models with all or no outliers removed.

Impact:

• The optimized model demonstrated superior predictive performance (RMSEP=0.854 Brix) compared to models excluding all outliers (RMSEP=1.19 Brix) or including all identified outliers (RMSEP=0.862 Brix).
• This approach leads to more representative and stable calibration models for SSC determination in melons.
• Highlights the importance of judicious outlier management in spectroscopic calibration for agricultural products.