Simultaneous qualitative and quantitative analyses of pesticide residues on fruit peels with ATR-FTIR spectroscopy and multi-task learning

Abstract

With the increased awareness of food safety, rapid, accurate, and non-destructive detection of pesticide residues on fruit peels has attracted widespread attention. In this work, we utilize the attenuated total reflectance-Fourier transform infrared spectroscopy (ATR-FTIR) to directly detect multiple pesticide residues (including carbendazim, thiophanate-methyl, and thiabendazole) on the surface of the apple peels. To further improve the efficiency of detection and meet the practical application needs, a multi-task learning (MTL) model based on multi-task neural networks is introduced to perform qualitative and quantitative analysis of three pesticides, simultaneously. The optimal results in the testing set demonstrate an average accuracy of 100 % for the qualitative task, while the average R² of 0.9415 and root mean square error (RMSE) of 2.567 μg/cm² can be achieved in the quantitative task. The limit of detection (LOD) of carbendazim, thiophanate-methyl, and thiabendazole were determined as 7.308 μg/cm², 1.595 μg/cm² and 0.159 μg/cm², respectively. Compared with the traditional single-task model, our work greatly simplifies the complexity of pesticide detection while ensuring prediction accuracy, which offers an alternative approach for further deployment and operation of the on-site system.