Detection of pesticide residues in lettuce based on fluorescence spectra

Abstract: Lettuce is a kind of important food for people all over the world, which contains a lot of protein, carbohydrate, vitamin and some nutrient elements. However, pesticide residue in lettuce leaves is becoming one of the most urgent problems and it needs to be solved as soon as possible in the world. Therefore, fast and efficient nondestructive detection of pesticide residues with different concentration in lettuce leaves plays a key role in the growth of lettuce and also improves food safety supervision. In this paper, we aimed to identify pesticide residues with different concentration in lettuce leaves in a novel and rapid nondestructive way by using fluorescence spectra technology. The fluorescence spectral data of 180 lettuce leaf samples with pesticide residues of 3 concentrations (the ratio of pesticides to water was 1:500, 1:1000 and 1:1200) were obtained using Cary Eclipse fluorescence spectrometer (at an excitation wavelength of 245 nm, an emission wavelength in the range of 300-510 nm and discharge voltage of 600 V). Besides, the actual concentrations of pesticide sprayed were 12.77, 2.55 and 1.28 mg/L, respectively. Moreover, Savitzky-Golay (SG), Standard Normalized Variable (SNV), Standard Normalized Variable detrending (SNV detrending), Savitzky-Golay coupled with Standard Normalized Variable (SG-SNV), Savitzky-Golay coupled with Standard Normalized Variable detrending (SG-SNV detrending) were used to preprocess the raw spectra, respectively. Finally, support vector machine (SVM) classification models were established based on full spectra (FS), fluorescence characteristic peak (FCP) and wavelet feature (WF), respectively. The Monte Carlo Cross Validation algorithm (MCCV) was used to select a total of 120 lettuce leaves samples as the training set, the remaining 60 lettuce leaves samples were taken as the prediction set, and the number of MCCV cycle was set as 1000. Moreover, the sample selection would be completed when the training set and the prediction set accuracy were less than 2%. Besides, FCP mainly contained the band including 371.07, 424, 440, 460 and 486.96 nm. Furthermore, the WF was selected by wavelet transform using db4, db6, sym5 and sym7 as wavelet basis functions, respectively. The results showed that the prediction set identification rate of SVM model based on WF obtained the best results compared with the SVM models based on the FS and FCP. SG-SNV detrending - FS-SVM model obtained the best performance among all FS-SVM models, with an identification rate of 100% in the calibration set and 63.33% in the prediction set. Furthermore, SG-SNV detrending -FCP-SVM model obtained the best performance among all FCP-SVM models with an identification rate of 95% in the calibration set and 70% in the prediction set. Moreover, SG-SNV detrending -WF-SVM model obtained the best performance among all SVM models with an identification rate of 98.33% in the calibration set and 93.33% in the prediction set, in which sym5 was used as wavelet basis function and the optimal wavelet decomposition level was 4. The results indicated that it was feasible to use fluorescence spectra technology to identify different concentration of pesticide residues in lettuce leaves.