Feasibility for rapid on-site screening of pesticide residues in fresh produce using machine learning-assisted diffuse reflectance spectroscopy

Abstract

Ensuring food safety by minimizing pesticide residues in crops is crucial to safeguarding human health. However, farmers often fail to follow the minimal time interval known as the Pre-Harvest Interval (PHI) between pesticide application and crop harvest. This lapse leads to high pesticide residue levels, often exceeding the recommended maximum residue levels (MRLs). Conventional methods for detecting pesticide residues are laboratory-based and rely on complex wet chemistry procedures. Although highly accurate, these techniques are destructive, expensive, and time-consuming. In addition, they take several hours before establishing results, making them unsuitable for analyzing large volumes of perishable fresh produce, particularly in resource-constrained settings. In this study, we investigated the feasibility of utilizing machine learning-assisted diffuse reflectance spectroscopy (DRS), a direct, rapid, and cost-effective approach for conducting an on-site preliminary assessment of pesticide residue levels in fresh produce, referred to as “screening.” This method facilitates the validation of fresh produce compliance with crop-specific PHI requirements after pesticide application, indicating adherence to MRLs. We used a portable Vis/NIR spectrometer to acquire DRS measurements from tamarillo fruits that had just been sprayed with mancozeb at the cultivation site. In total, 646 measurements were collected from treated and control (untreated) fruits over five alternating days, covering the duration between pesticide application and PHI lapse. We used spectral data pre-processing techniques and principal component analysis (PCA) to differentiate the fruits as the residue levels decreased, approaching the recommended PHI and demonstrating their suitability for harvesting. Furthermore, using two principal components that explained 88% of the total variance, we developed a support vector machine (SVM) classification model. The model successfully classified both the treated and control samples, as well as the samples based on their PHI status, distinguishing between ‘before PHI’ (days 1 - 7), ‘after PHI’ (day 9), and ‘Control’ (no residues), all with 100% accuracy. Given the quick turnaround time for results, typically within five minutes, we conclude that DRS combined with machine learning holds the potential for rapid and cost-effective on-site screening of large volumes of fresh produce in a market setting. This approach allows for the rapid identification of samples that require additional investigation in a food safety laboratory, thereby improving the efficient use of laboratory resources.