Rapid detection of pesticide residues on tomato surfaces using hyperspectral imaging and machine learning

Rapid and non-destructive detection of pesticide residues is of great importance for ensuring food safety and quality control of fresh agricultural products. In this study, hyperspectral imaging (HSI) combined with machine learning methods was employed for the classification and quantitative analysis of pesticide residues on tomato surfaces. A total of 90 tomato samples were prepared in this study, from which 1800 spectral samples were extracted. Among them, 800 spectral samples were used for pesticide identification, including the control, difenoconazole, imidacloprid, and kresoxim-methyl groups. The remaining 1000 spectral samples were used for imidacloprid analysis at five concentration levels: 0, 0.25, 0.5, 0.75, and 1 mg/kg. Based on average spectral information, models were built combining preprocessing methods (MSC, MSC-D1) with machine learning models (SVM, MLP, PLSR, SVR, 1D-CNN). Results demonstrated that the MSC-D1-PCA-SVM model optimally classified pesticide types with 96.67% accuracy. For imidacloprid concentration classification, the MSC-D1-MLP model performed best with 93% accuracy. In quantitative analysis, MSC-D1-MLP was also most effective, yielding a prediction R² of 0.9230, RMSEP of 0.0981, and RPD of 3.6027. These findings indicate that HSI, combined with appropriate preprocessing and feature learning, effectively captures subtle spectral differences induced by surface pesticides, enabling rapid identification and prediction of residue types and concentrations. The proposed method provides an effective strategy for rapid and non-destructive monitoring of pesticide residues in tomatoes and shows considerable potential for food safety inspection and quality assessment of fresh agricultural products.