dc.description.abstract | There is an urgent need for accurate, non-destructive, rapid, and affordable techniques for
screening pesticide residues in fresh fruits to assess whether regulatory standards are met
by not surpassing the allowed maximum residue limits (MRL). Although conventional
methods such as chromatography and mass spectroscopy are accurate, they are expensive,
destructive, and require tedious wet lab sample preparations. This study aimed to assess
the utility of machine learning techniques for rapid and non-destructive assessment of
residues in fruits based on diffuse reflectance spectroscopy (DRS) measurements in the
near-infrared region. Towards this goal, tree tomatoes fruits were spiked with Mancozeb
and thiocyclam hydrogen oxalate (THO) in varying concentrations, and near-infrared
spectra data (900-2500nm) were collected in DRS geometry. Another dataset was collected
from the field in the 200 nm to 1050 nm range on control and treated tree tomatoes for
11 consecutive days for qualitative analysis. All the measurements were transformed into
absorbance using log10(1/R) before preprocessing using the smoothing, normalization,
and multiplicative scatter correction techniques. Principal component analysis of the field
data showed distinct clusters for the control and treated fruits in the 800 nm to 850 nm
third overtone region with PC1 and PC2 accounting for 82% and 5.9% of the variance,
respectively. The combination region (1900-2500 nm) was optimal in discriminating the
samples with varying pesticide concentrations for the laboratory data, with PC1 and
PC2 accounting for 93% and 3.8% of the variance, respectively. The first four PCs,
which explained 98% of the cumulative variation of the data, were extracted from the
laboratory data and used as inputs to the support vector machine (SVM), artificial neural
networks (ANN), and Random forest (RF) machine learning models. All the developed
models had R2 values greater than 92% and RMSEP values of less than 0.06 ppm for
Mancozeb models and not more than 0.08 ppm for THO models. Limits of detection
and quantification were also determined using a pseudo-univariate approach. The models
were tested on a new dataset of samples collected from four local markets. From the
results obtained, all predicted values were below the acceptable MRL values (0.5 ppm and
0.3 ppm for Mancozeb and THO, respectively). One-way Tukey ANOVA analysis of the
predictions of the market samples showed that ANN and SVR models were more reliable
than the RF model. Therefore, it was concluded that the combination of diffuse reflectance
spectroscopy with machine learning techniques has potential for rapid, non-destructive,
in-situ assessment of pesticide residues in fruits and vegetables. | en_US |