Hyperspectral Detection of Pesticide Residues in Black Vegetable Based on Multi-Classifier Entropy Weight Method

This study addresses the issue of rapid, non-destructive detection of pesticide residues in black vegetable (Brassica narinosa L.H. Bailey) by proposing a multi-classifier entropy weighted method algorithm that combines hyperspectral technology and the entropy weight method. 10 black vegetable samples were sprayed with each of the four different pesticides (trichlorfon, propargite, cypermethrin, and imidacloprid) at concentrations of 0.10, 2.00, 0.20, and 2.00 mg/kg, respectively. Hyperspectral data were acquired using a hyperspectral imaging system. Savitzky-Golay convolutional smoothing was used for spectral preprocessing. Three dimensionality reduction techniques, competitive adaptive reweighted sampling, random frog leaping, and successive projections algorithm, were compared. Models were built using eXtreme gradient boosting, random forest, and support vector machine algorithms. The entropy weight method was then used to optimize model weights, developing the multi-classifier entropy weighted method algorithm to improve detection accuracy and robustness. Experimental results show that competitive adaptive reweighted sampling significantly improved overall classification accuracy and Kappa coefficient after dimensionality reduction. The multi-classifier entropy weighted method algorithm outperformed both single classifiers and other fusion algorithms in pesticide residue detection, achieving an overall accuracy of 94.20% and a Kappa coefficient of 0.93. Thus, integrating hyperspectral technology, machine learning, and the entropy weight method effectively leverages the strengths of each classifier, overcoming the “curse of dimensionality” and improving detection accuracy. This method offers an effective solution for the rapid detection of pesticide residues in black vegetable.