摘要
The dry-extract system for (near) infrared (DESIR) technique was implemented using reflectance near-infrared spec-troscopy in context of detection of contact pesticide residues on fruit. Based on chemical structure, spectra features and regression statistics for PLSR models, a product containing metiram and pyraclostrobin was chosen from six pesticides for further consideration. Regression models based on spectra of dry extracts of aqueous solutions and either acetone or water washes of contaminated fruit were encouraging (RMSECV of approximately 0.03 - 0.06 mg a.i.). This level of analytical performance would support the use of the technique as a rapid screening tool, with suspect samples then subject to the reference GC-MS analysis method. However, the PLSR model performance was poor across populations of fruit, suggesting that matrix changes in the solvent wash between sets of fruit is problematic. Further work is required to establish whether sufficient variation can be built into a calibration set to overcome this issue, without degrading model performance to the point where it loses practical application.
The dry-extract system for (near) infrared (DESIR) technique was implemented using reflectance near-infrared spec-troscopy in context of detection of contact pesticide residues on fruit. Based on chemical structure, spectra features and regression statistics for PLSR models, a product containing metiram and pyraclostrobin was chosen from six pesticides for further consideration. Regression models based on spectra of dry extracts of aqueous solutions and either acetone or water washes of contaminated fruit were encouraging (RMSECV of approximately 0.03 - 0.06 mg a.i.). This level of analytical performance would support the use of the technique as a rapid screening tool, with suspect samples then subject to the reference GC-MS analysis method. However, the PLSR model performance was poor across populations of fruit, suggesting that matrix changes in the solvent wash between sets of fruit is problematic. Further work is required to establish whether sufficient variation can be built into a calibration set to overcome this issue, without degrading model performance to the point where it loses practical application.
