A CNN-BiGRU-selfattention model combined with GAN reconstruction and Reverse Feature Fusion for apple pesticide residues detecting

Abstract

Pesticide residues in fruits, particularly apples, pose a significant threat to human health, highlighting the urgent need for precise and efficient detection methods. Although spectral detection techniques have emerged as a promising lossless solution, their effectiveness is often hindered by challenges such as imbalanced data distribution and redundant spectral features. To address these limitations, we propose a novel method that combines spectral reconstruction and key wavelength selection. This method introduces a CNN-BiGRU-Selfattention (CBS) model, which utilizes Generative Adversarial Networks (GAN) for spectral reconstruction and Reverse Feature Fusion (RFF) for selecting key wavelengths. By tackling these critical issues, the proposed approach achieves significant improvements in pesticide residue detection accuracy for apples. Experimental results on apple datasets demonstrated exceptional performance, with an accuracy of 96.31 %, recall of 96.31 %, precision of 0.9632, and an F1 score of 0.9630. These results underscore the model's robustness and effectiveness in extracting critical spectral features, ensuring reliable detection. By enhancing the accuracy and reliability of pesticide residue detection, this study offers a pioneering approach to safeguard food safety and lays a foundation for advancing future research in agricultural applications.