- Email: [email protected]
Fast detection of fenthion on fruit and vegetable peel using dynamic surface-enhanced Raman spectroscopy and random forests with variable selection
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
Contents lists available at ScienceDirect
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy journal homepage: www.elsevier.com/locate/saa
Fast detection of fenthion on fruit and vegetable peel using dynamic surface-enhanced Raman spectroscopy and random forests with variable selection Shizhuang Weng a,⁎, Mengqing Qiu a, Ronglu Dong b, Fang Wang a, Linsheng Huang a, Dongyan Zhang a, Jinling Zhao a,⁎ a b
Anhui Engineering Laboratory of Agro-Ecological Big Data, Anhui University, Hefei 230601, China Hefei Institute of Physical Science, Chinese Academy of Sciences, Hefei 230031, China
a r t i c l e
i n f o
Article history: Received 14 December 2017 Received in revised form 1 April 2018 Accepted 8 April 2018 Available online 10 April 2018 Keywords: Dynamic surface-enhanced Raman spectroscopy Fenthion Random forests Fruit and vegetable peel Fast detection
a b s t r a c t Dynamic surface-enhanced Raman spectroscopy (D-SERS) based on the state change of the substrate not only significantly enhances but also provides a highly reproducible Raman signal. Hence, we develop a fast and accurate method for the detection of fenthion on fruit and vegetable peel using D-SERS and random forests (RF) with variable selection. With uniform Ag nanoparticles, the dynamic spectra of fenthion solution at different concentrations were obtained using D-SERS, and fenthion solution greater than or equal to 0.05 mg/L can be detected. Then, the quantitative analysis models of fenthion were developed by RF with variable selection for spectra of different range. The model of best performance is developed by RF and spectra of characteristic range with higher RF importance (top 40%), and the root mean square error of cross-validation is 0.0101 mg/L. Moreover, the fenthion residue of tomato, pear, and cabbage peel were extracted by a swab dipped in ethanol and analyzed using the above method to further validate the practical effect. Compared to gas chromatography, the maximal relative deviation is below 12.5%, and the predicted recovery is between 87.5% and 112.5%. Accordingly, D-SERS and RF with variable selection can realize the fast, simple, ultrasensitive, and accurate analysis of fenthion residue on fruit and vegetable peel. © 2018 Elsevier B.V. All rights reserved.
1. Introduction Fenthion is widely used for the control of Carposina niponensis Walshingham, caterpillar, and aphid in fruits and vegetables [1,2]. However, its overuse and inappropriate application lead to residue. Excessive fenthion residue can induce intermediate syndrome with long-lasting and reiterative damage. The accurate detection of residue is a key approach for avoiding the above problem. The detection process must be rapid for large-scale residue. Moreover, the conventional method as gas chromatography (GC) is incapable for meeting the above requirement [3,4]. Surface-enhanced Raman spectroscopy (SERS), whose signal is largely enhanced based on the electromagnetic and chemical enhancement of nanoscale rough noble metal, allows highly sensitive detection. Furthermore, with the simple pretreatment and rapid spectrummeasuring procedure, SERS has been applied to the sensitive and fast detection of pesticides [5–7], drugs [8,9], food additives [10], and biomacromolecules [11]. Many researchers have detected fenthion ⁎ Corresponding authors. E-mail address: [email protected]. (S. Weng).
https://doi.org/10.1016/j.saa.2018.04.012 1386-1425/© 2018 Elsevier B.V. All rights reserved.
using SERS [4,12]. However, the conventional SERS measurement is performed on substrates of dry or wet state, and simultaneously ensuring good sensitivity and high reproducibility is highly difficult [13]. Recently, our group developed a novel method called dynamic SERS (D-SERS) [14–16], which originates from the two conventional methods as the dry film-based and solution-based methods. Dry film-based SERS detection: place the general colloidal nanoparticles on a solid substrate (silicon, glass wafer), dry the sample on the substrate, and obtain the SERS spectra. Generally, it gains the enormous SERS response accompanied with weak reproducibility and stability for fabrication of complicated substrates and damage of laser. Solution-based detection involves that analyte is mixed with colloidal nanoparticles and hot spots are generated through inducing particles aggregation prior to spectra measurement. Due to the hot spots of random distribution and low density, the high sensitivity is difficult to ensure. Measurement of D-SERS depends on the translation-based Raman substrate state from wet to dry. Under the critical state of wet to dry, the nanostructures can self-assemble to form hot spots driven by solvent capillary forces. Meanwhile, the aggregated nanostructures can provide a bigger capture space for the analyte molecules. The larger signal enhancement is obtained. In addition, self-assembled hot spots are of good uniformity,
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
and the residual liquid at the critical state can prevent from damage of laser. Repeatability of D-SERS is also excellent. Previous results have proven that D-SERS can provide reproducible, stable, and sensitive SERS signals [12]. Accordingly, D-SERS is used for the fast and accurate detection of fenthion. When spectra are obtained using D-SERS, the qualitative and quantitative analysis of analytes should be performed by professionals. Artificial intervention renders the fast detection difficult to achieve. However, the intelligent identification model obtained by chemometric methods can achieve automatic and rapid spectral analysis without experts [17–20]. Frequently used methods include artificial neural networks [21], support vector machine [18,19], and random forests (RF) [20]. Among them, RF possesses significant advantages of robustness and simple parameter optimization [22]. Moreover, RF can estimate the importance of variables during model construction, which is useful for variable selection [23]. In the present study, RF with variable selection is applied to analyze D-SERS spectra and intelligently obtain information on analytes. With D-SERS and RF, we aimed to explore the ultrasensitive and fast detection of fenthion on fruit and vegetable peel. Firstly, the dynamic spectra of fenthion solution at different concentrations were measured using D-SERS with uniform Ag nanoparticles (AgNPs). Then, regression models for the quantitative analysis of fenthion were developed by combining RF with variable selection. Moreover, fenthion residue on peel of tomato, pear, and cabbage was extracted by wiping with a swab dipped in ethanol, then spectra were also measured using DSERS. The established model and the obtained spectra were used to predict the residue concentration on peel. 2. Experiments and Methods 2.1. Reagents and Chemicals Silver nitrate (AgNO3, 99.9%), sodium citrate (99%), ethanol (99.9%), and fenthion powder (99.9%) were purchased from Beijing Century Aoke Biological Technology Co. Ltd. (Beijing, China). Tomato, pear, and cabbage were purchased from Hefei Hongfu market (Anhui, China). 2.2. SERS Substrate AgNPs sol solution was synthesized using chemical reduction [24]. Firstly, 1 mL 0.1 M silver nitrate was added into 99 mL deionized water and heated to slight boiling. Then, 4 mL 1% sodium citrate was
21
added, and kept boiling for 1 h after the color change. Finally, the solution was centrifuged for 10 min at 7500–8000 rpm, and dark green and transparent colloid was obtained. Ultraviolet and visible (UV–vis) absorption spectrum was shown in Fig. 1, and the inset is the corresponding SEM image of silver colloid nanoparticles. According to Fig. 1, the absorption peak of the silver colloid was about 425 nm with a full width at half maximum (FWHM) of 115 nm. The size of silver particle was uniform, and the average diameter was about 40–50 nm. 2.3. Sample Preparation and Measurement D-SERS measurement: fenthion powder was dissolved in ethanol to obtain concentrations of 50, 10, 5, 1, 0.5, 0.1, 0.05, and 0.01 mg/L. Samples of fruit and vegetable peel: tomato, pear, and cabbage were sprayed with 1 mL of 50, 10, 5, 1, 0.5, 0.1 mg/L fenthion solution on the labeled 3 × 3 cm2 region. After the solution dried, the labeled region was wiped by a swab dipped in ethanol, and the swab was vortexed in ethanol to release the pesticide [25]. The total amount of ethanol was 1 mL, and ten samples were collected for each fruit and vegetable. The testing solution and AgNPs colloid solution (V/V = 1:1) were mixed adequately and dropped on the silicon chip, and the dropping volume of the mixture was 2 μL. During the wet-to-dry state of the mixed solutions, 18 spectra were collected on the Raman spectrometer (LabRAM HR800) with a 5 s integration time and a 5 s interval time. During interval time of 5 s, the focus was corrected on the droplets. For the LabRAM HR800, a 532 nm laser was used as the excitation source with a measured power of 3.5 mW at center of mixing droplet surface with a focal spot of about 1 μm in diameter. And temperature and relative humidity were kept at 25 °C and 40% to reduce the fluctuation of signals. Two spectra were selected from the obtained 18 spectra, which were regarded as dynamic spectra for the subsequent analysis. Samples that underwent the same treatment were also analyzed by GC to validate the results. GC measurement: the obtained ethanol solution was removed into centrifuge tube of 5 mL and evaporated with a termovap sample concentrator (GIPP-AUTO-12S, Jipu, Shanghai). The extractive was eluted using 1 mL ethyl acetate and used for GC measurement. A GC instrument (ThermoFinnigan Trace GC 2000, USA) was used to measure fethion residue with a flame photometric detector. The oven was heated at 110 °C for 1 min, then raised to 220 °C at 40 °C min−1, and then ascended to 222 °C at 1 °C min−1. Sample of 1 mL was injected into the instrument using high purity nitrogen as a carrier gas. The gas flow rate was 3.0 mL min−1. Interface temperature was 230 °C, and detector temperature was 250 °C. The collision gas was high purity nitrogen. 2.4. RF
Fig. 1. UV–vis absorption spectrum of silver colloid, and the inset is SEM image of silver colloid nanoparticles.
RF is a powerful algorithm for solving classification and regression problems. In specific, RF takes advantage of multiple classification and regression tree (CART) to construct the recognition model, which was first developed by Leo Breiman [26] and Adele Cutler [27]. When the RF model identifies the unknown objects based on some properties, every CART provides its own conclusion, and the final output of the model is the most selected option (classification) or the mean value of all results (regression). Furthermore, the main parameters of RF [28], such as the number of CART (ntree) and the number of features (mtry), significantly affect the performance of the RF model. According to experience, ntree was set to traverse between 400 and 1000 with an interval of 100, and mtry was set to the round number of the square root of feature dimension (dimension of spectrum). The prediction accuracy of the models was quantitatively evaluated using root mean square error of crossvalidation (RMSECV). Additionally, the dual fold cross-validated method was adopted in this study to verify the effect of variable selection based on the estimation of variable importance (Fig. 2). The dataset
22
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
Fig. 2. Structure of dual fold cross-validated method.
was selectively divided into the training and testing subsets (N−1:1). The training subset (N−1) was used for variable selection, and the testing subset was used for model evaluation (N = 5). All computations and chemometric methods were implemented in MATLAB 2011b (Mathworks Inc., Natick, MA, USA). 3. Results and Discussion 3.1. Dynamic Spectra of Fenthion The measuring method for D-SERS is to drop 2 μL mixture droplet of testing solution and AgNPs colloid solution (V/V = 1/1) on the silicon wafer, and acquire the spectra until the droplet become dried. The obtained spectra of 10 mg/L fenthion solution during the entire D-SERS measurement are shown in Fig. 3(A), and the sensitivity and reproducibility of the spectra are clearly different during measurement process. With solvent evaporation, the sensitivity and reproducibility of the
Raman signal improve until the solvent evaporates completely. The distance between the Ag nanoparticles shortens and the solvent evaporates, the hot spots of high uniformity and excellent capture capability begin to form [12], and the sensitivity and reproducibility of the Raman signal gradually improve. Furthermore, the optimal effect occurs at the critical state between wet and dry. However, when the substrates become dry, the hot spots disappear, and the Raman signal deteriorates quickly. From the intensity change of peak at 1045 cm−1 (Fig. 3B), the critical point of SERS substrate can be easily determined. So the spectra at the critical state can be selected based on the intensity of characteristic peaks. And the spectra of 50, 5, 1, 0.5, 0.1 and 0.05 mg/L fenthion during entire D-SERS measurement were shown in Fig. S1 and also obtained the similar results. Spectra of five testing samples of 10 mg/L fenthion obtained at the critical state are shown in Fig. 4(A). and spectra were moved vertically for direct observation. The spectra of the different samples are of good uniformity (Fig. S2), which validates the high reproducibility. For subsequent analysis, only the spectra at the critical state were selected and regarded as dynamic spectra. Then, the dynamic spectra of 10 mg/L fenthion, whose baseline was deducted using the polynomial fitting method, are shown in Fig. 4(B). From the figure, the characteristic peaks at approximately 637, 712, 755, 930, 1045, 1225, 1364, 1566, and 1647 cm−1 can clearly be observed. These characteristic peaks carry the vibration information of fenthion molecule [4,12]: peaks at 637 and 1647 cm−1 are attributed to the stretching vibration of P_S, and the Raman bands at 712, 755, 930, 1045, 1225, 1364, and 1566 cm−1 are assigned to the stretching vibration of C\\S, vibration of aryl, stretching vibration of P\\O, deformation vibration of aryl, stretching vibration of P\\O\\C, rocking vibration of P\\O\\CH3, symmetric deformation vibration of CH3, and stretching vibration of aryl, respectively. Dynamic spectra of fenthion solution at different concentrations are shown in Fig. 5. As shown in the figure, the spectral characteristic peaks are uniform when fenthion solutions are greater than or equal to 0.05 mg/L, but the spectra of 0.01 mg/L fenthion are the same as those of the blank sample. Therefore, the limit of detection of fenthion using D-SERS with AgNPs is approximately 0.05 mg/L, the national maximum residue limit of China. Relationship between intensity of characteristic peaks and residue concentrations was of positive correlation (Fig. S3A). Furthermore, to achieve the fast and accurate analysis of fenthion spectra without the involvement of professionals, RF was used for developing the intelligent recognition model. According to the spectral assignment of fenthion, the spectra of 615–659, 688–793, 905–944, 1013–1075, 1198–1248, 1339–1389, 1541–1577, and 1613–1664 cm−1 were selected as the spectra of characteristic range for its physical significance.
Fig. 3. Obtained spectra of 10 mg/L fenthion during entire D-SERS measurement (A), and intensity change of spectral peak at 1045 cm−1 of 10 mg/L fenthion (B).
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
23
Fig. 4. Spectra at the critical state of 10 mg/L fenthion obtained from five samples (A), and dynamic spectra of 10 mg/L fenthion after baseline subtraction (B).
3.2. Spectral Analysis Using RF With Variable Selection The obtained spectra of 50–0.05 mg/L fenthion were utilized to develop the quantitative model using RF, which includes the spectra of full range (600–1800 cm−1) and characteristic range (615–659, 688–793, 905–944, 1013–1075, 1198–1248, 1339–1389, 1541–1577, and 1613–1664 cm−1). Five samples were measured for each fenthion concentration, and 20 spectra were collected for each sample; therefore, 100 spectra were used to construct the RF model for fenthion at each concentration. Meanwhile, to obtain the RF model of higher accuracy, the variable selection was tested based on the estimation of variable importance. The prediction accuracy of the models was quantitatively evaluated by RMSECV and dual fold cross-validated method (Fig. 2). The analysis results of the RF models with different spectral data are shown in Table 1. The table shows that the lowest prediction error of the RF model with spectra of full range is 0.1147 mg/L, and the detection limit of D-SERS can reach 0.05 mg/L; therefore, the prediction accuracy is inadequate. Then, for the models obtained with the spectra of characteristic range, the prediction accuracy improves significantly, and the lowest RMSECV value decreases to 0.0135 mg/L. This finding is mainly because the spectra of characteristic range provide almost all the
molecular information of fenthion and remove the irrelevant information. Furthermore, the spectra of characteristic range were selected based on the estimation of variable importance obtained through RF, and the spectra of higher importance (top 40%) were used to build the new models. The prediction error of the new model decreases, and the lowest RMSECV is 0.0101 mg/L, which are the reasons for further retaining the information of positive effect and eliminating the negative information. Hereafter, the spectra used for the following analysis are the spectra of characteristic range with higher RF importance (top 40%). During the experiments, the RF models always maintain high prediction accuracy when the parameters undergo large-scale change which indicates that RF possesses excellent robustness. Considering the simplicity of application, we set ntree to 1000 and mtry to the round number of square root of spectrum dimension in the subsequent experiments. Accordingly, D-SERS and RF with variable selection can provide fast, automatic, accurate, and ultrasensitive detection of fenthion. Basing on the results, we hope that the method can be expanded to the actual detection of fenthion on fruit and vegetable peel. 3.3. Detection of Fenthion on Fruit and Vegetable Peel The labeled regions (3 × 3 cm2) of tomato, pear, and cabbage were sprayed with 1 mL of 50, 10, 5, 1, 0.5, 0.1 mg/L fenthion solution. After the solution dried, the labeled region was wiped by a swab dipped in ethanol, and the swab was vortexed in ethanol to release the pesticide. The total amount of ethanol was 1 mL, and ten samples were collected for each fruit and vegetable. And the wiping times of swab and time of vortex should be set to 3 and 10 s for extracting the residue using swab dipped in ethonal (Fig. S4). The testing solution and AgNPs colloid solution (V/V = 1:1) were mixed adequately. The spectra of all mixed solutions were obtained using D-SERS. The spectra of the fenthion residue of tomato, pear, and cabbage peel and the relative standard deviation (RSD) of spectral characteristic peaks are shown in Fig. 6. As
Table 1 Prediction results of the RF models with different spectral data (RMSECV: mg/L).
Fig. 5. Dynamic spectra of 50, 10, 5, 1, 0.5, 0.1, 0.05, 0.01 mg/L fenthion and blank sample (AgNPs colloid solution).
Value of ntree
Spectra of full range
Spectra of characteristic range
Spectra of characteristic range with higher RF importance
400 500 600 700 800 900 1000
0.1319 0.1613 0.1597 0.1913 0.1745 0.1147 0.1462
0.0249 0.0161 0.0292 0.0189 0.0196 0.0177 0.0135
0.0105 0.0131 0.0119 0.0122 0.0132 0.0121 0.0101
mtry is set to the round number of square root of feature dimension.
24
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
Fig. 6. Spectra of fenthion residue on tomato, pear, and cabbage peel (A), and relative standard deviation (RSD) of spectral characteristic peaks of the same species and concentration (B).
Fig. 7. Spectra of AgNPs and extraction solutions from peel of tomato, pear, cabbage without spraying fenthion (A), and spectra of fenthion residue at different concentrations on tomato peel (B).
shown in Fig. 6, the characteristic spectral peaks are highly consistent, and the RSD of all the spectral characteristic peaks is less than 8.02%. As seen in Fig. S3(B)–(D), it can be known that intensity of characteristic peaks is proportional to concentrations of fenthion residue on peel, which demonstrates the feasibility of quantitative detection of fenthion using D-SERS. Spectra of AgNPs and extraction solutions from peel of tomato, pear, cabbage by the swab dipped in ethanol are highly consistent (Fig. 7A), we can know that there is no interference coming from peels using the proposed extraction method. From the spectra of fenthion residue on tomato peel (Fig. 7B), all the residue of spraying fenthion can be detected, and spectral intensity is positively related to the concentration of residue (Fig. S3B). The amount of residue was predicted based on the obtained spectra and established RF model, and the fruit and vegetable samples sprayed with fenthion were also detected using GC (Table 2). The table shows that the detecting value between GC and D-SERS with RF is highly agreeable, the maximal relative deviation of detecting value is below 12.5%, and the predicted recovery is between 87.5% and 112.5%. Meanwhile, the swab dipping with ethanol can extract the fenthion residue by wiping the peel. The extraction efficiency depends on the
Table 2 Detection results of fenthion residue on tomato, pear, and cabbage peel. Samples Spraying fenthion (mg/L)
Detecting value by D-SERS and RF with variable selection (mg/L)
Detecting value by GC (mg/L)
Relative deviation (%)
Predicted recovery (%)
Tomato
48.52 9.25 4.97 0.91 0.45 0.09 44.44 8.53 4.65 0.89 0.44 0.08 43.38 8.41 4.57 0.87 0.44 0.07
48.71 9.41 4.91 0.95 0.48 0.08 48.23 9.35 4.89 0.94 0.48 0.08 48.51 9.37 4.88 0.95 0.47 0.08
0.39 1.75 1.22 4.21 6.25 12.5 8.23 8.77 4.90 5.32 8.33 0 10.58 10.25 6.35 8.42 6.38 12.5
99.61 98.30 101.22 95.79 93.75 112.5 92.14 91.23 95.09 94.68 91.67 100 89.42 89.75 93.65 91.58 93.62 87.5
50 10 5 1 0.5 0.1 Pear 50 10 5 1 0.5 0.1 Cabbage 50 10 5 1 0.5 0.1
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
smoothness of the peel, which is the reason why the detection effect on tomato peel is superior to that on the other two objects. Additionally, the extraction of peel residue can be completed in 1 min using the swab dipped with ethanol, the dynamic spectra can be obtained in 2 min, and the quantitative results of unknown samples is achieved in a few seconds based on the established RF model. Therefore, D-SERS and RF with variable selection can realize the accurate and fast detection of fenthion on fruit and vegetable peel. In the future, the presented method has a brilliant application prospect for applying to the detection of various pesticides and objects.
4. Conclusion A fast and accurate method for detection of fenthion on fruit and vegetable peel was developed using D-SERS and RF with variable selection. First, with uniform AgNPs, the dynamic spectra of fenthion solution at 50, 10, 5, 1, 0.5, 0.1, 0.05, and 0.01 mg/L were obtained using D-SERS. Then, the regression models for the automatic and accurate analysis of fenthion concentration were developed by RF with spectra of different range, and variable selection was realized based on the estimation of variable importance for obtaining models of higher accuracy. Comparing the spectra of AgNPs and fenthion solution at different concentrations, the Raman signal of fenthion can be detected when the concentration is greater than or equal to 0.05 mg/L, which can reach the national maximum residue limit of China and review the ultra-sensitiveness of D-SERS. Furthermore, DSERS with excellent reproducibility for the spectra of the same concentration are highly uniform. Meanwhile, the spectral intensity is positively related to the concentration of fenthion which provides the possibility of quantitative analysis using D-SERS. For the automatic analysis, the model of best performance is developed by RF and the spectra of characteristic range with higher RF importance (top 40%), and the lowest RMSECV is 0.0101 mg/L. Then, the above method is expanded to detect fenthion residue on tomato, pear, and cabbage peel. The obtained spectra are still highly consistent. And the quantitative analysis of spectra is executed based on the established model. Compared to the results of GC, the maximal relative deviation is below 12.5%, and the predicted recovery is between 87.5% and 112.5%. The results demonstrate that D-SERS and RF with variable selection can accurately detect the fenthion residue on the peel. Additionally, the extraction of residue, dynamic spectra measurement, and automatic quantitative analysis can be realized in 3 min, and only 2 μL sample volume is needed for spectra measurement. Therefore, D-SERS and RF with variable selection can achieve the fast, ultrasensitive, and accurate analysis of fenthion residue on fruit and vegetable peel. In the future, we believe the presented method will play an important role in the safety inspection of agricultural products.
Acknowledgement This study is supported by the Natural Science Foundation of Anhui Province (No. 1708085QF134), Natural Science Research Project of Anhui Provincial Education Department (No. KJ2017A006), Anhui Provincial Major Scientific and Technological Special Project (17030701062), National Natural Science Foundation of China (Nos. 31401285, 61475163, and 61672032), and National Key Research and Development Program (Nos. 4014YFD0800904). Appendix A. Supplementary Data Supplementary data to this article can be found online at https://doi. org/10.1016/j.saa.2018.04.012.
25
References [1] W. Guo, B.J. Engelman, T.L. Haywood, N.B. Blok, D.S. Beaudoin, S.O. Obare, Dual fluorescence and electrochemical detection of the organophosphorus pesticides—ethion, malathion and fenthion, Talanta 87 (2011) 276–283. [2] M.T. Jafari, M. Saraji, H. Sherafatm, Polypyrrole/montmorillonite nanocomposite as a new solid phase microextraction fiber combined with gas chromatography–corona discharge ion mobility spectrometry for the simultaneous determination of diazinon and fenthion organophosphorus pesticides, Anal. Chim. Acta 814 (2014) 69–78. [3] E. Meyer, D. Borrey, W. Lambert, C. Van-Peteghem, M. Piette, A. De-Leenheer, Analysis of fenthion in postmortem samples by HPLC with diode-array detection and GC-MS using solid-phase extraction, J. Anal. Toxicol. 22 (3) (1998) 248–252. [4] X. Li, S. Zhang, Z. Yu, T. Yang, Surface-enhanced Raman spectroscopic analysis of phorate and fenthion pesticide in apple skin using silver nanoparticles, Appl. Spectrosc. 68 (4) (2014) 483–487. [5] P. Guo, D. Sikdar, X. Huang, K.J. Si, W. Xiong, S. Gong, L.W. Yap, M. Premaratne, W. Cheng, Plasmonic core–shell nanoparticles for SERS detection of the pesticide thiram: size-and shape-dependent Raman enhancement, Nano 7 (7) (2015) 2862–2868. [6] S. Huang, J. Hu, P. Guo, M. Liu, R. Wu, Rapid detection of chlorpyriphos residue in rice by surface-enhanced Raman scattering, Anal. Methods 7 (10) (2015) 4334–4339. [7] J. Kubackova, G. Fabriciova, P. Miskovsky, D. Jancura, S. Sanchez-Cortes, Sensitive surface-enhanced Raman spectroscopy (SERS) detection of organochlorine pesticides by alkyl dithiol-functionalized metal nanoparticles-induced plasmonic hot spots, Anal. Chem. 87 (1) (2014) 663–669. [8] F. Inscore, C. Shende, A. Sengupta, H. Huang, S. Farquharson, Detection of drugs of abuse in saliva by surface-enhanced Raman spectroscopy (SERS), Appl. Spectrosc. 65 (9) (2011) 1004–1008. [9] R. Dong, S. Weng, L. Yang, J. Liu, Detection and direct readout of drugs in human urine using dynamic surface-enhanced Raman spectroscopy and support vector machines, Anal. Chem. 87 (5) (2015) 2937–2944. [10] J.F. Betz, Y. Cheng, G.W. Rubloff, Direct SERS detection of contaminants in a complex mixture: rapid, single step screening for melamine in liquid infant formula, Analyst 137 (4) (2012) 826–828. [11] Y. Zhao, L. Liu, H. Kuang, L. Wang, C. Xu, SERS-active Ag@ Au core–shell NP assemblies for DNA detection, RSC Adv. 4 (99) (2014) 56052–56056. [12] Y. Teng, W. Liu, J. Liu, Y. Nie, P. Li, C. He, Surface-enhanced Raman scattering sensing of trace fenthion coupled with stable silver colloids and oh stretching band of water as an internal standard, J. Anal. Chem. 71 (9) (2016) 891–895. [13] L. Yang, P. Li, H. Liu, X. Tang, J. Liu, A dynamic surface enhanced Raman spectroscopy method for ultra-sensitive detection: from the wet state to the dry state, Chem. Soc. Rev. 44 (10) (2015) 2837–2848. [14] X. Yan, P. Li, B. Zhou, X. Tang, X. Li, S. Weng, L. Yang, J. Liu, Optimal hotspots of dynamic surfaced-enhanced Raman spectroscopy for drugs quantitative detection, Anal. Chem. 89 (9) (2017) 4875–4881. [15] H. Liu, Z. Yang, L. Meng, Y. Sun, J. Wang, L. Yang, J. Liu, Z. Tian, Three-dimensional and time-ordered surface-enhanced Raman scattering hotspot matrix, J. Am. Chem. Soc. 136 (14) (2014) 5332–5341. [16] Z. Han, H. Liu, B. Wang, S. Weng, L. Yang, J. Liu, Three-dimensional surface-enhanced Raman scattering hotspots in spherical colloidal superstructure for identification and detection of drugs in human urine, Anal. Chem. 87 (9) (2015) 4821–4828. [17] C.V. Di Anibal, L.F. Marsal, M.P. Callao, I. Ruisanchez, Surface enhanced Raman spectroscopy (SERS) and multivariate analysis as a screening tool for detecting Sudan I dye in culinary spices, Spectrochim. Acta A 87 (2012) 135–141. [18] A. Marz, B. Monch, P. Rosch, M. Kiehntopf, T. Henkel, J. Popp, Detection of thiopurine methyltransferase activity in lysed red blood cells by means of lab-on-a-chip surface enhanced Raman spectroscopy (LOC-SERS), Anal. Bioanal. Chem. 400 (9) (2011) 2755–2761. [19] C.S. Deroo, R.A. Armitage, Direct identification of dyes in textiles by direct analysis in real time-time of flight mass spectrometry, Anal. Chem. 83 (18) (2011) 6924–6928. [20] M. Liu, M. Wang, J. Wang, D. Li, Comparison of random forest, support vector machine and back propagation neural network for electronic tongue data classification: application to the recognition of orange beverage and Chinese vinegar, Sensors Actuators B Chem. 177 (2013) 970–980. [21] B. Ozbalci, I.H. Boyaci, A. Topcu, C. Kadilar, U. Tamer, Rapid analysis of sugars in honey by processing Raman spectrum using chemometric methods and artificial neural networks, Food Chem. 136 (3) (2013) 1444–1452. [22] T.F. Cootes, M.C. Ionita, C. Lindner, P. Sauer, Robust and Accurate Shape Model Fitting Using Random Forest Regression Voting, European Conference on Computer Vision, 2012 278–291. [23] A.L. Boulesteix, S. Janitza, J. Kruppa, I.R. Konig, Overview of random forest methodology and practical guidance with emphasis on computational biology and bioinformatics, WIRES 2 (6) (2012) 493–507. [24] A. Michota, J. Bukowska, Surface-enhanced Raman scattering (SERS) of 4mercaptobenzoic acid on silver and gold substrates, J. Raman Spectrosc. 34 (1) (2003) 21–25. [25] L.L. He, T. Chen, T.P. Labuza, Recovery and quantitative detection of thiabendazole on apples using a surface swab capture method followed by surface-enhanced Raman spectroscopy, Food Chem. 148 (2014) 42–46. [26] L. Breiman, Random forests, Mach. Learn. 45 (1) (2001) 5–32. [27] D.R. Cutler, T.C. Edwards, K.H. Beard, A. Cutler, K.T. Hess, Random forests for classification in ecology, Ecology 88 (11) (2007) 2783–2792. [28] A. Verikas, A. Gelzinis, M. Bacauskiene, Mining data with random forests: a survey and results of new tests, Pattern Recogn. 44 (2) (2011) 330–349.
Contents lists available at ScienceDirect
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy journal homepage: www.elsevier.com/locate/saa
Fast detection of fenthion on fruit and vegetable peel using dynamic surface-enhanced Raman spectroscopy and random forests with variable selection Shizhuang Weng a,⁎, Mengqing Qiu a, Ronglu Dong b, Fang Wang a, Linsheng Huang a, Dongyan Zhang a, Jinling Zhao a,⁎ a b
Anhui Engineering Laboratory of Agro-Ecological Big Data, Anhui University, Hefei 230601, China Hefei Institute of Physical Science, Chinese Academy of Sciences, Hefei 230031, China
a r t i c l e
i n f o
Article history: Received 14 December 2017 Received in revised form 1 April 2018 Accepted 8 April 2018 Available online 10 April 2018 Keywords: Dynamic surface-enhanced Raman spectroscopy Fenthion Random forests Fruit and vegetable peel Fast detection
a b s t r a c t Dynamic surface-enhanced Raman spectroscopy (D-SERS) based on the state change of the substrate not only significantly enhances but also provides a highly reproducible Raman signal. Hence, we develop a fast and accurate method for the detection of fenthion on fruit and vegetable peel using D-SERS and random forests (RF) with variable selection. With uniform Ag nanoparticles, the dynamic spectra of fenthion solution at different concentrations were obtained using D-SERS, and fenthion solution greater than or equal to 0.05 mg/L can be detected. Then, the quantitative analysis models of fenthion were developed by RF with variable selection for spectra of different range. The model of best performance is developed by RF and spectra of characteristic range with higher RF importance (top 40%), and the root mean square error of cross-validation is 0.0101 mg/L. Moreover, the fenthion residue of tomato, pear, and cabbage peel were extracted by a swab dipped in ethanol and analyzed using the above method to further validate the practical effect. Compared to gas chromatography, the maximal relative deviation is below 12.5%, and the predicted recovery is between 87.5% and 112.5%. Accordingly, D-SERS and RF with variable selection can realize the fast, simple, ultrasensitive, and accurate analysis of fenthion residue on fruit and vegetable peel. © 2018 Elsevier B.V. All rights reserved.
1. Introduction Fenthion is widely used for the control of Carposina niponensis Walshingham, caterpillar, and aphid in fruits and vegetables [1,2]. However, its overuse and inappropriate application lead to residue. Excessive fenthion residue can induce intermediate syndrome with long-lasting and reiterative damage. The accurate detection of residue is a key approach for avoiding the above problem. The detection process must be rapid for large-scale residue. Moreover, the conventional method as gas chromatography (GC) is incapable for meeting the above requirement [3,4]. Surface-enhanced Raman spectroscopy (SERS), whose signal is largely enhanced based on the electromagnetic and chemical enhancement of nanoscale rough noble metal, allows highly sensitive detection. Furthermore, with the simple pretreatment and rapid spectrummeasuring procedure, SERS has been applied to the sensitive and fast detection of pesticides [5–7], drugs [8,9], food additives [10], and biomacromolecules [11]. Many researchers have detected fenthion ⁎ Corresponding authors. E-mail address: [email protected]. (S. Weng).
https://doi.org/10.1016/j.saa.2018.04.012 1386-1425/© 2018 Elsevier B.V. All rights reserved.
using SERS [4,12]. However, the conventional SERS measurement is performed on substrates of dry or wet state, and simultaneously ensuring good sensitivity and high reproducibility is highly difficult [13]. Recently, our group developed a novel method called dynamic SERS (D-SERS) [14–16], which originates from the two conventional methods as the dry film-based and solution-based methods. Dry film-based SERS detection: place the general colloidal nanoparticles on a solid substrate (silicon, glass wafer), dry the sample on the substrate, and obtain the SERS spectra. Generally, it gains the enormous SERS response accompanied with weak reproducibility and stability for fabrication of complicated substrates and damage of laser. Solution-based detection involves that analyte is mixed with colloidal nanoparticles and hot spots are generated through inducing particles aggregation prior to spectra measurement. Due to the hot spots of random distribution and low density, the high sensitivity is difficult to ensure. Measurement of D-SERS depends on the translation-based Raman substrate state from wet to dry. Under the critical state of wet to dry, the nanostructures can self-assemble to form hot spots driven by solvent capillary forces. Meanwhile, the aggregated nanostructures can provide a bigger capture space for the analyte molecules. The larger signal enhancement is obtained. In addition, self-assembled hot spots are of good uniformity,
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
and the residual liquid at the critical state can prevent from damage of laser. Repeatability of D-SERS is also excellent. Previous results have proven that D-SERS can provide reproducible, stable, and sensitive SERS signals [12]. Accordingly, D-SERS is used for the fast and accurate detection of fenthion. When spectra are obtained using D-SERS, the qualitative and quantitative analysis of analytes should be performed by professionals. Artificial intervention renders the fast detection difficult to achieve. However, the intelligent identification model obtained by chemometric methods can achieve automatic and rapid spectral analysis without experts [17–20]. Frequently used methods include artificial neural networks [21], support vector machine [18,19], and random forests (RF) [20]. Among them, RF possesses significant advantages of robustness and simple parameter optimization [22]. Moreover, RF can estimate the importance of variables during model construction, which is useful for variable selection [23]. In the present study, RF with variable selection is applied to analyze D-SERS spectra and intelligently obtain information on analytes. With D-SERS and RF, we aimed to explore the ultrasensitive and fast detection of fenthion on fruit and vegetable peel. Firstly, the dynamic spectra of fenthion solution at different concentrations were measured using D-SERS with uniform Ag nanoparticles (AgNPs). Then, regression models for the quantitative analysis of fenthion were developed by combining RF with variable selection. Moreover, fenthion residue on peel of tomato, pear, and cabbage was extracted by wiping with a swab dipped in ethanol, then spectra were also measured using DSERS. The established model and the obtained spectra were used to predict the residue concentration on peel. 2. Experiments and Methods 2.1. Reagents and Chemicals Silver nitrate (AgNO3, 99.9%), sodium citrate (99%), ethanol (99.9%), and fenthion powder (99.9%) were purchased from Beijing Century Aoke Biological Technology Co. Ltd. (Beijing, China). Tomato, pear, and cabbage were purchased from Hefei Hongfu market (Anhui, China). 2.2. SERS Substrate AgNPs sol solution was synthesized using chemical reduction [24]. Firstly, 1 mL 0.1 M silver nitrate was added into 99 mL deionized water and heated to slight boiling. Then, 4 mL 1% sodium citrate was
21
added, and kept boiling for 1 h after the color change. Finally, the solution was centrifuged for 10 min at 7500–8000 rpm, and dark green and transparent colloid was obtained. Ultraviolet and visible (UV–vis) absorption spectrum was shown in Fig. 1, and the inset is the corresponding SEM image of silver colloid nanoparticles. According to Fig. 1, the absorption peak of the silver colloid was about 425 nm with a full width at half maximum (FWHM) of 115 nm. The size of silver particle was uniform, and the average diameter was about 40–50 nm. 2.3. Sample Preparation and Measurement D-SERS measurement: fenthion powder was dissolved in ethanol to obtain concentrations of 50, 10, 5, 1, 0.5, 0.1, 0.05, and 0.01 mg/L. Samples of fruit and vegetable peel: tomato, pear, and cabbage were sprayed with 1 mL of 50, 10, 5, 1, 0.5, 0.1 mg/L fenthion solution on the labeled 3 × 3 cm2 region. After the solution dried, the labeled region was wiped by a swab dipped in ethanol, and the swab was vortexed in ethanol to release the pesticide [25]. The total amount of ethanol was 1 mL, and ten samples were collected for each fruit and vegetable. The testing solution and AgNPs colloid solution (V/V = 1:1) were mixed adequately and dropped on the silicon chip, and the dropping volume of the mixture was 2 μL. During the wet-to-dry state of the mixed solutions, 18 spectra were collected on the Raman spectrometer (LabRAM HR800) with a 5 s integration time and a 5 s interval time. During interval time of 5 s, the focus was corrected on the droplets. For the LabRAM HR800, a 532 nm laser was used as the excitation source with a measured power of 3.5 mW at center of mixing droplet surface with a focal spot of about 1 μm in diameter. And temperature and relative humidity were kept at 25 °C and 40% to reduce the fluctuation of signals. Two spectra were selected from the obtained 18 spectra, which were regarded as dynamic spectra for the subsequent analysis. Samples that underwent the same treatment were also analyzed by GC to validate the results. GC measurement: the obtained ethanol solution was removed into centrifuge tube of 5 mL and evaporated with a termovap sample concentrator (GIPP-AUTO-12S, Jipu, Shanghai). The extractive was eluted using 1 mL ethyl acetate and used for GC measurement. A GC instrument (ThermoFinnigan Trace GC 2000, USA) was used to measure fethion residue with a flame photometric detector. The oven was heated at 110 °C for 1 min, then raised to 220 °C at 40 °C min−1, and then ascended to 222 °C at 1 °C min−1. Sample of 1 mL was injected into the instrument using high purity nitrogen as a carrier gas. The gas flow rate was 3.0 mL min−1. Interface temperature was 230 °C, and detector temperature was 250 °C. The collision gas was high purity nitrogen. 2.4. RF
Fig. 1. UV–vis absorption spectrum of silver colloid, and the inset is SEM image of silver colloid nanoparticles.
RF is a powerful algorithm for solving classification and regression problems. In specific, RF takes advantage of multiple classification and regression tree (CART) to construct the recognition model, which was first developed by Leo Breiman [26] and Adele Cutler [27]. When the RF model identifies the unknown objects based on some properties, every CART provides its own conclusion, and the final output of the model is the most selected option (classification) or the mean value of all results (regression). Furthermore, the main parameters of RF [28], such as the number of CART (ntree) and the number of features (mtry), significantly affect the performance of the RF model. According to experience, ntree was set to traverse between 400 and 1000 with an interval of 100, and mtry was set to the round number of the square root of feature dimension (dimension of spectrum). The prediction accuracy of the models was quantitatively evaluated using root mean square error of crossvalidation (RMSECV). Additionally, the dual fold cross-validated method was adopted in this study to verify the effect of variable selection based on the estimation of variable importance (Fig. 2). The dataset
22
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
Fig. 2. Structure of dual fold cross-validated method.
was selectively divided into the training and testing subsets (N−1:1). The training subset (N−1) was used for variable selection, and the testing subset was used for model evaluation (N = 5). All computations and chemometric methods were implemented in MATLAB 2011b (Mathworks Inc., Natick, MA, USA). 3. Results and Discussion 3.1. Dynamic Spectra of Fenthion The measuring method for D-SERS is to drop 2 μL mixture droplet of testing solution and AgNPs colloid solution (V/V = 1/1) on the silicon wafer, and acquire the spectra until the droplet become dried. The obtained spectra of 10 mg/L fenthion solution during the entire D-SERS measurement are shown in Fig. 3(A), and the sensitivity and reproducibility of the spectra are clearly different during measurement process. With solvent evaporation, the sensitivity and reproducibility of the
Raman signal improve until the solvent evaporates completely. The distance between the Ag nanoparticles shortens and the solvent evaporates, the hot spots of high uniformity and excellent capture capability begin to form [12], and the sensitivity and reproducibility of the Raman signal gradually improve. Furthermore, the optimal effect occurs at the critical state between wet and dry. However, when the substrates become dry, the hot spots disappear, and the Raman signal deteriorates quickly. From the intensity change of peak at 1045 cm−1 (Fig. 3B), the critical point of SERS substrate can be easily determined. So the spectra at the critical state can be selected based on the intensity of characteristic peaks. And the spectra of 50, 5, 1, 0.5, 0.1 and 0.05 mg/L fenthion during entire D-SERS measurement were shown in Fig. S1 and also obtained the similar results. Spectra of five testing samples of 10 mg/L fenthion obtained at the critical state are shown in Fig. 4(A). and spectra were moved vertically for direct observation. The spectra of the different samples are of good uniformity (Fig. S2), which validates the high reproducibility. For subsequent analysis, only the spectra at the critical state were selected and regarded as dynamic spectra. Then, the dynamic spectra of 10 mg/L fenthion, whose baseline was deducted using the polynomial fitting method, are shown in Fig. 4(B). From the figure, the characteristic peaks at approximately 637, 712, 755, 930, 1045, 1225, 1364, 1566, and 1647 cm−1 can clearly be observed. These characteristic peaks carry the vibration information of fenthion molecule [4,12]: peaks at 637 and 1647 cm−1 are attributed to the stretching vibration of P_S, and the Raman bands at 712, 755, 930, 1045, 1225, 1364, and 1566 cm−1 are assigned to the stretching vibration of C\\S, vibration of aryl, stretching vibration of P\\O, deformation vibration of aryl, stretching vibration of P\\O\\C, rocking vibration of P\\O\\CH3, symmetric deformation vibration of CH3, and stretching vibration of aryl, respectively. Dynamic spectra of fenthion solution at different concentrations are shown in Fig. 5. As shown in the figure, the spectral characteristic peaks are uniform when fenthion solutions are greater than or equal to 0.05 mg/L, but the spectra of 0.01 mg/L fenthion are the same as those of the blank sample. Therefore, the limit of detection of fenthion using D-SERS with AgNPs is approximately 0.05 mg/L, the national maximum residue limit of China. Relationship between intensity of characteristic peaks and residue concentrations was of positive correlation (Fig. S3A). Furthermore, to achieve the fast and accurate analysis of fenthion spectra without the involvement of professionals, RF was used for developing the intelligent recognition model. According to the spectral assignment of fenthion, the spectra of 615–659, 688–793, 905–944, 1013–1075, 1198–1248, 1339–1389, 1541–1577, and 1613–1664 cm−1 were selected as the spectra of characteristic range for its physical significance.
Fig. 3. Obtained spectra of 10 mg/L fenthion during entire D-SERS measurement (A), and intensity change of spectral peak at 1045 cm−1 of 10 mg/L fenthion (B).
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
23
Fig. 4. Spectra at the critical state of 10 mg/L fenthion obtained from five samples (A), and dynamic spectra of 10 mg/L fenthion after baseline subtraction (B).
3.2. Spectral Analysis Using RF With Variable Selection The obtained spectra of 50–0.05 mg/L fenthion were utilized to develop the quantitative model using RF, which includes the spectra of full range (600–1800 cm−1) and characteristic range (615–659, 688–793, 905–944, 1013–1075, 1198–1248, 1339–1389, 1541–1577, and 1613–1664 cm−1). Five samples were measured for each fenthion concentration, and 20 spectra were collected for each sample; therefore, 100 spectra were used to construct the RF model for fenthion at each concentration. Meanwhile, to obtain the RF model of higher accuracy, the variable selection was tested based on the estimation of variable importance. The prediction accuracy of the models was quantitatively evaluated by RMSECV and dual fold cross-validated method (Fig. 2). The analysis results of the RF models with different spectral data are shown in Table 1. The table shows that the lowest prediction error of the RF model with spectra of full range is 0.1147 mg/L, and the detection limit of D-SERS can reach 0.05 mg/L; therefore, the prediction accuracy is inadequate. Then, for the models obtained with the spectra of characteristic range, the prediction accuracy improves significantly, and the lowest RMSECV value decreases to 0.0135 mg/L. This finding is mainly because the spectra of characteristic range provide almost all the
molecular information of fenthion and remove the irrelevant information. Furthermore, the spectra of characteristic range were selected based on the estimation of variable importance obtained through RF, and the spectra of higher importance (top 40%) were used to build the new models. The prediction error of the new model decreases, and the lowest RMSECV is 0.0101 mg/L, which are the reasons for further retaining the information of positive effect and eliminating the negative information. Hereafter, the spectra used for the following analysis are the spectra of characteristic range with higher RF importance (top 40%). During the experiments, the RF models always maintain high prediction accuracy when the parameters undergo large-scale change which indicates that RF possesses excellent robustness. Considering the simplicity of application, we set ntree to 1000 and mtry to the round number of square root of spectrum dimension in the subsequent experiments. Accordingly, D-SERS and RF with variable selection can provide fast, automatic, accurate, and ultrasensitive detection of fenthion. Basing on the results, we hope that the method can be expanded to the actual detection of fenthion on fruit and vegetable peel. 3.3. Detection of Fenthion on Fruit and Vegetable Peel The labeled regions (3 × 3 cm2) of tomato, pear, and cabbage were sprayed with 1 mL of 50, 10, 5, 1, 0.5, 0.1 mg/L fenthion solution. After the solution dried, the labeled region was wiped by a swab dipped in ethanol, and the swab was vortexed in ethanol to release the pesticide. The total amount of ethanol was 1 mL, and ten samples were collected for each fruit and vegetable. And the wiping times of swab and time of vortex should be set to 3 and 10 s for extracting the residue using swab dipped in ethonal (Fig. S4). The testing solution and AgNPs colloid solution (V/V = 1:1) were mixed adequately. The spectra of all mixed solutions were obtained using D-SERS. The spectra of the fenthion residue of tomato, pear, and cabbage peel and the relative standard deviation (RSD) of spectral characteristic peaks are shown in Fig. 6. As
Table 1 Prediction results of the RF models with different spectral data (RMSECV: mg/L).
Fig. 5. Dynamic spectra of 50, 10, 5, 1, 0.5, 0.1, 0.05, 0.01 mg/L fenthion and blank sample (AgNPs colloid solution).
Value of ntree
Spectra of full range
Spectra of characteristic range
Spectra of characteristic range with higher RF importance
400 500 600 700 800 900 1000
0.1319 0.1613 0.1597 0.1913 0.1745 0.1147 0.1462
0.0249 0.0161 0.0292 0.0189 0.0196 0.0177 0.0135
0.0105 0.0131 0.0119 0.0122 0.0132 0.0121 0.0101
mtry is set to the round number of square root of feature dimension.
24
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
Fig. 6. Spectra of fenthion residue on tomato, pear, and cabbage peel (A), and relative standard deviation (RSD) of spectral characteristic peaks of the same species and concentration (B).
Fig. 7. Spectra of AgNPs and extraction solutions from peel of tomato, pear, cabbage without spraying fenthion (A), and spectra of fenthion residue at different concentrations on tomato peel (B).
shown in Fig. 6, the characteristic spectral peaks are highly consistent, and the RSD of all the spectral characteristic peaks is less than 8.02%. As seen in Fig. S3(B)–(D), it can be known that intensity of characteristic peaks is proportional to concentrations of fenthion residue on peel, which demonstrates the feasibility of quantitative detection of fenthion using D-SERS. Spectra of AgNPs and extraction solutions from peel of tomato, pear, cabbage by the swab dipped in ethanol are highly consistent (Fig. 7A), we can know that there is no interference coming from peels using the proposed extraction method. From the spectra of fenthion residue on tomato peel (Fig. 7B), all the residue of spraying fenthion can be detected, and spectral intensity is positively related to the concentration of residue (Fig. S3B). The amount of residue was predicted based on the obtained spectra and established RF model, and the fruit and vegetable samples sprayed with fenthion were also detected using GC (Table 2). The table shows that the detecting value between GC and D-SERS with RF is highly agreeable, the maximal relative deviation of detecting value is below 12.5%, and the predicted recovery is between 87.5% and 112.5%. Meanwhile, the swab dipping with ethanol can extract the fenthion residue by wiping the peel. The extraction efficiency depends on the
Table 2 Detection results of fenthion residue on tomato, pear, and cabbage peel. Samples Spraying fenthion (mg/L)
Detecting value by D-SERS and RF with variable selection (mg/L)
Detecting value by GC (mg/L)
Relative deviation (%)
Predicted recovery (%)
Tomato
48.52 9.25 4.97 0.91 0.45 0.09 44.44 8.53 4.65 0.89 0.44 0.08 43.38 8.41 4.57 0.87 0.44 0.07
48.71 9.41 4.91 0.95 0.48 0.08 48.23 9.35 4.89 0.94 0.48 0.08 48.51 9.37 4.88 0.95 0.47 0.08
0.39 1.75 1.22 4.21 6.25 12.5 8.23 8.77 4.90 5.32 8.33 0 10.58 10.25 6.35 8.42 6.38 12.5
99.61 98.30 101.22 95.79 93.75 112.5 92.14 91.23 95.09 94.68 91.67 100 89.42 89.75 93.65 91.58 93.62 87.5
50 10 5 1 0.5 0.1 Pear 50 10 5 1 0.5 0.1 Cabbage 50 10 5 1 0.5 0.1
S. Weng et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 200 (2018) 20–25
smoothness of the peel, which is the reason why the detection effect on tomato peel is superior to that on the other two objects. Additionally, the extraction of peel residue can be completed in 1 min using the swab dipped with ethanol, the dynamic spectra can be obtained in 2 min, and the quantitative results of unknown samples is achieved in a few seconds based on the established RF model. Therefore, D-SERS and RF with variable selection can realize the accurate and fast detection of fenthion on fruit and vegetable peel. In the future, the presented method has a brilliant application prospect for applying to the detection of various pesticides and objects.
4. Conclusion A fast and accurate method for detection of fenthion on fruit and vegetable peel was developed using D-SERS and RF with variable selection. First, with uniform AgNPs, the dynamic spectra of fenthion solution at 50, 10, 5, 1, 0.5, 0.1, 0.05, and 0.01 mg/L were obtained using D-SERS. Then, the regression models for the automatic and accurate analysis of fenthion concentration were developed by RF with spectra of different range, and variable selection was realized based on the estimation of variable importance for obtaining models of higher accuracy. Comparing the spectra of AgNPs and fenthion solution at different concentrations, the Raman signal of fenthion can be detected when the concentration is greater than or equal to 0.05 mg/L, which can reach the national maximum residue limit of China and review the ultra-sensitiveness of D-SERS. Furthermore, DSERS with excellent reproducibility for the spectra of the same concentration are highly uniform. Meanwhile, the spectral intensity is positively related to the concentration of fenthion which provides the possibility of quantitative analysis using D-SERS. For the automatic analysis, the model of best performance is developed by RF and the spectra of characteristic range with higher RF importance (top 40%), and the lowest RMSECV is 0.0101 mg/L. Then, the above method is expanded to detect fenthion residue on tomato, pear, and cabbage peel. The obtained spectra are still highly consistent. And the quantitative analysis of spectra is executed based on the established model. Compared to the results of GC, the maximal relative deviation is below 12.5%, and the predicted recovery is between 87.5% and 112.5%. The results demonstrate that D-SERS and RF with variable selection can accurately detect the fenthion residue on the peel. Additionally, the extraction of residue, dynamic spectra measurement, and automatic quantitative analysis can be realized in 3 min, and only 2 μL sample volume is needed for spectra measurement. Therefore, D-SERS and RF with variable selection can achieve the fast, ultrasensitive, and accurate analysis of fenthion residue on fruit and vegetable peel. In the future, we believe the presented method will play an important role in the safety inspection of agricultural products.
Acknowledgement This study is supported by the Natural Science Foundation of Anhui Province (No. 1708085QF134), Natural Science Research Project of Anhui Provincial Education Department (No. KJ2017A006), Anhui Provincial Major Scientific and Technological Special Project (17030701062), National Natural Science Foundation of China (Nos. 31401285, 61475163, and 61672032), and National Key Research and Development Program (Nos. 4014YFD0800904). Appendix A. Supplementary Data Supplementary data to this article can be found online at https://doi. org/10.1016/j.saa.2018.04.012.
25
References [1] W. Guo, B.J. Engelman, T.L. Haywood, N.B. Blok, D.S. Beaudoin, S.O. Obare, Dual fluorescence and electrochemical detection of the organophosphorus pesticides—ethion, malathion and fenthion, Talanta 87 (2011) 276–283. [2] M.T. Jafari, M. Saraji, H. Sherafatm, Polypyrrole/montmorillonite nanocomposite as a new solid phase microextraction fiber combined with gas chromatography–corona discharge ion mobility spectrometry for the simultaneous determination of diazinon and fenthion organophosphorus pesticides, Anal. Chim. Acta 814 (2014) 69–78. [3] E. Meyer, D. Borrey, W. Lambert, C. Van-Peteghem, M. Piette, A. De-Leenheer, Analysis of fenthion in postmortem samples by HPLC with diode-array detection and GC-MS using solid-phase extraction, J. Anal. Toxicol. 22 (3) (1998) 248–252. [4] X. Li, S. Zhang, Z. Yu, T. Yang, Surface-enhanced Raman spectroscopic analysis of phorate and fenthion pesticide in apple skin using silver nanoparticles, Appl. Spectrosc. 68 (4) (2014) 483–487. [5] P. Guo, D. Sikdar, X. Huang, K.J. Si, W. Xiong, S. Gong, L.W. Yap, M. Premaratne, W. Cheng, Plasmonic core–shell nanoparticles for SERS detection of the pesticide thiram: size-and shape-dependent Raman enhancement, Nano 7 (7) (2015) 2862–2868. [6] S. Huang, J. Hu, P. Guo, M. Liu, R. Wu, Rapid detection of chlorpyriphos residue in rice by surface-enhanced Raman scattering, Anal. Methods 7 (10) (2015) 4334–4339. [7] J. Kubackova, G. Fabriciova, P. Miskovsky, D. Jancura, S. Sanchez-Cortes, Sensitive surface-enhanced Raman spectroscopy (SERS) detection of organochlorine pesticides by alkyl dithiol-functionalized metal nanoparticles-induced plasmonic hot spots, Anal. Chem. 87 (1) (2014) 663–669. [8] F. Inscore, C. Shende, A. Sengupta, H. Huang, S. Farquharson, Detection of drugs of abuse in saliva by surface-enhanced Raman spectroscopy (SERS), Appl. Spectrosc. 65 (9) (2011) 1004–1008. [9] R. Dong, S. Weng, L. Yang, J. Liu, Detection and direct readout of drugs in human urine using dynamic surface-enhanced Raman spectroscopy and support vector machines, Anal. Chem. 87 (5) (2015) 2937–2944. [10] J.F. Betz, Y. Cheng, G.W. Rubloff, Direct SERS detection of contaminants in a complex mixture: rapid, single step screening for melamine in liquid infant formula, Analyst 137 (4) (2012) 826–828. [11] Y. Zhao, L. Liu, H. Kuang, L. Wang, C. Xu, SERS-active Ag@ Au core–shell NP assemblies for DNA detection, RSC Adv. 4 (99) (2014) 56052–56056. [12] Y. Teng, W. Liu, J. Liu, Y. Nie, P. Li, C. He, Surface-enhanced Raman scattering sensing of trace fenthion coupled with stable silver colloids and oh stretching band of water as an internal standard, J. Anal. Chem. 71 (9) (2016) 891–895. [13] L. Yang, P. Li, H. Liu, X. Tang, J. Liu, A dynamic surface enhanced Raman spectroscopy method for ultra-sensitive detection: from the wet state to the dry state, Chem. Soc. Rev. 44 (10) (2015) 2837–2848. [14] X. Yan, P. Li, B. Zhou, X. Tang, X. Li, S. Weng, L. Yang, J. Liu, Optimal hotspots of dynamic surfaced-enhanced Raman spectroscopy for drugs quantitative detection, Anal. Chem. 89 (9) (2017) 4875–4881. [15] H. Liu, Z. Yang, L. Meng, Y. Sun, J. Wang, L. Yang, J. Liu, Z. Tian, Three-dimensional and time-ordered surface-enhanced Raman scattering hotspot matrix, J. Am. Chem. Soc. 136 (14) (2014) 5332–5341. [16] Z. Han, H. Liu, B. Wang, S. Weng, L. Yang, J. Liu, Three-dimensional surface-enhanced Raman scattering hotspots in spherical colloidal superstructure for identification and detection of drugs in human urine, Anal. Chem. 87 (9) (2015) 4821–4828. [17] C.V. Di Anibal, L.F. Marsal, M.P. Callao, I. Ruisanchez, Surface enhanced Raman spectroscopy (SERS) and multivariate analysis as a screening tool for detecting Sudan I dye in culinary spices, Spectrochim. Acta A 87 (2012) 135–141. [18] A. Marz, B. Monch, P. Rosch, M. Kiehntopf, T. Henkel, J. Popp, Detection of thiopurine methyltransferase activity in lysed red blood cells by means of lab-on-a-chip surface enhanced Raman spectroscopy (LOC-SERS), Anal. Bioanal. Chem. 400 (9) (2011) 2755–2761. [19] C.S. Deroo, R.A. Armitage, Direct identification of dyes in textiles by direct analysis in real time-time of flight mass spectrometry, Anal. Chem. 83 (18) (2011) 6924–6928. [20] M. Liu, M. Wang, J. Wang, D. Li, Comparison of random forest, support vector machine and back propagation neural network for electronic tongue data classification: application to the recognition of orange beverage and Chinese vinegar, Sensors Actuators B Chem. 177 (2013) 970–980. [21] B. Ozbalci, I.H. Boyaci, A. Topcu, C. Kadilar, U. Tamer, Rapid analysis of sugars in honey by processing Raman spectrum using chemometric methods and artificial neural networks, Food Chem. 136 (3) (2013) 1444–1452. [22] T.F. Cootes, M.C. Ionita, C. Lindner, P. Sauer, Robust and Accurate Shape Model Fitting Using Random Forest Regression Voting, European Conference on Computer Vision, 2012 278–291. [23] A.L. Boulesteix, S. Janitza, J. Kruppa, I.R. Konig, Overview of random forest methodology and practical guidance with emphasis on computational biology and bioinformatics, WIRES 2 (6) (2012) 493–507. [24] A. Michota, J. Bukowska, Surface-enhanced Raman scattering (SERS) of 4mercaptobenzoic acid on silver and gold substrates, J. Raman Spectrosc. 34 (1) (2003) 21–25. [25] L.L. He, T. Chen, T.P. Labuza, Recovery and quantitative detection of thiabendazole on apples using a surface swab capture method followed by surface-enhanced Raman spectroscopy, Food Chem. 148 (2014) 42–46. [26] L. Breiman, Random forests, Mach. Learn. 45 (1) (2001) 5–32. [27] D.R. Cutler, T.C. Edwards, K.H. Beard, A. Cutler, K.T. Hess, Random forests for classification in ecology, Ecology 88 (11) (2007) 2783–2792. [28] A. Verikas, A. Gelzinis, M. Bacauskiene, Mining data with random forests: a survey and results of new tests, Pattern Recogn. 44 (2) (2011) 330–349.