Nondestructive detection of thiophanate-methyl pesticide content in cherry tomato based on handheld spectrometer and SpecTransformer algorithm

To address the drawbacks of traditional pesticide detection methods such as sample disruption and procedural complexity, a rapid, non-destructive spectral detection system was developed in this paper. This system consists of a handheld spectrometer, detection algorithm, cloud computing, and an app that enables real-time detection of thiophanate-methyl content in cherry tomatoes. As the key of the system and the focus of this study, a novel deep learning algorithm called SpecTransformer was proposed to drive the spectrometer for spectral feature extraction and model detection. The algorithm was designed as a module architecture including input layer, spectral preprocessing layer, Block1, Block2 and output layer, which could achieve better detection performance than other current spectral algorithms. The results showed that the determination coefficient (R²) of the spectrometer for thiophanate-methyl detection was 0.91, with a root mean square error (RMSE) of 1.05. The effective detection range was between 1:100 and 1:5000 dilutions, with a limit of detection (LOD) of 1:5000 dilution (0.2 g/L). The spectrometer is compact, user-friendly, and has strong scalability. It can be expanded to detect multiple pesticide residues in the future, which provides new insights into rapid and accurate measurement of pesticide residues on agricultural produce.