Abstract With fresh winter jujube from the southern region of Xinjiang as the object of study， the method for detecting pesticide residues on winter jujube surface by near??infrared spectroscopy （NIR） combined with successive projections algorithm （SPA） and partial least squares （PLS） was investigated. The absorbance information of winter jujube sample surface was obtained through NIR technology， for the building of a full??wave band PLS model of fresh winter jujube sample sprayed with different concentrations of pesticide （taking chlorpyrifos as an example） as well as an SPA??PLS model which was built with the characteristic wavelengths extracted with SPA as the input variables for PLS， and the prediction precision of the two kinds of models was compared. The model built with the five characteristic waveslengths extected by SPA method only used the variables 0.32% of all the variables in the full wave band， but its accuracy and precision were better than the model built with the full wave band. It is feasible to build a model for different concentrations of chlorpyrifos on fresh winter jujube surfarce by NIR technology combined with SPA and PLS， and SPA method could simplify the complexity of the model and improve the precision and stablity of the model.
　　Key words Near infrared spectroscopy； Pesticide residue； SPA； Winter jujube； PLS
　　Winter jujube is an important economic crop in southern region of Xinjiang， which is acknowledged as one of the fresh??eating jujube varieties with the best quality due to its good taste， high sugar content， high contents of 19 amino acids needed for human body， vitamins and microelements， and more medicinal components[1]. It has very high dietary therapy and various health??care effects， and is honored as "living vitamin pill"[2]. Due to large??area plantation and environmental factors such as climate， the diseases and pests on winter jujube trees are increasingly severe， which seriously influences fruit farmers?? interests. Fruit farmers have to apply farm chemicals， but unreasonable abuse would cause pesticide residues on winter jujube??s surface exceeding standards. With the improvement of people??s living standard and the increase of awareness of health and environmental protection， consumers pay more and more attention on the quality safety of fruits and vegetables， especially pesticide residues on fruit and vegetable surface[3]. Because the last time of pesticide application before the harvest of winter jujube has a short interval from the time of entering the market， the fresh jujube on the market contains different quantities of pesticide residues， while many consumers might eat the fruit without cleaning or thorough cleaning. It is very important to detect pesticide residues on fresh winter jujube sold on the market timely. 　　In recent years， many scholars at home and abroad have studied pesticide residues in fruits and vegetables by liquid chromatography[4]， gas chromatography[5]， gas chromatography??tandem mass spectrometry[6-7]， enzymatic inhibition method[8]， supercritical fluid extraction （SFE）[9] and biosensor[10]， and some achievements have been made. However， these detection methods， especially chemical methods， generally have the disadvantages of long time， high cost， serious damage to samples and serious waste of samples and reagents， and could not realize rapid economic nondestructive examination， though they have high detection precision[11]. Therefore， it is urgent to find a real??time nondestructive online technique for rapid screening of pesticide residues， which could satisfy the rapid healthy development of corps feature forestry and fruit growing， and has higher economic value and broad application prospect for the improvement of brand effect.
　　Near??infrared spectroscopy （NIR） could rapidly quantitatively or qualitatively analyze an analyte using full spectrum or partial??band spectrum in the near infrared region[12-13]. This technology has been applied to the quality detection， species identification and pesticide residue analysis of agricultural products[14-17]. However， near infrared spectrum generally contains thousands of wavelength variables， while during the modeling using full??wave band data， not every wavelength could provide useful information， and a lot of redundant data would increase modeling workload. The selection of wavelength variables not only could get rid of irrelevant variables， but also could improve the prediction precision and robustness of the model. Meanwhile， the selected characteristic wavelengths could be used for the construction of online rapid spectral measurement system， thereby reducing production cost[18]. Successive projections algorithm （SPA） performs analysis using the projections of vectors to find the variable combination of least redundant information， so as to reduce the colinearity between variables. This method is widely applied in the selection of characteristic spectral wavelengths[19-20]. Partial least squares （PLS） method is a common multivariate statistic method， which is widely applied to the building of spectral quantitative models using near??infrared， hyperspectral， raman， nuclear magnetism and mass spectrum， and has nearly become a universal method for the building the quantitative calibration models in spectral analysis. 　　Based on the above considerations， with fresh winter jujube from the southern region of Xinjiang as the object of study， the absorbance information of winter jujube sample surface was obtained through NIR technology， for the building of a full??wave band PLS model of fresh winter jujube sample sprayed with different concentrations of pesticide （taking chlorpyrifos as an example） as well as an SPA??PLS model which was built with the characteristic wavelengths extracted with SPA as the input variables for PLS， and the prediction precision of the two kinds of models was compared. This study aimed at providing a new thinking and method for the rapid detection of pesticide residues on fresh winter jujube， and reference for the spectral detection of pesticide residues on fruits and vegetables.
　　Materials and Methods
　　Experimental materials
　　Fresh winter jujube purchased from some farm in southern region of Xinjiang was selected as the object of study. Specifically， 180 winter jujube samples with good maturity and uniform color were selected randomly. The used pesticide was chlorpyrifos， which was purchased from State Center for Standard Matter， in the type of emulsifiable concentrate （EC）， with a weight percentage content of 40%. The pesticide was prepared into different concentrations of solutions （1?? 10， 1?? 50， 1?? 100 and 1?? 1 000）， and distilled water was used as control （CK）.
　　The 180 winter jujube samples were naturally air??dried， and randomly divided according to experimental requirements into five groups， each of which included 36 samples. The prepared four concentrations of pesticide solutions and blank control were sprayed onto one group of winter jujube， respectively. From each concentration， 30 winter jujube samples were selected to form a modeling set， and the remaining six winter jujube samples were used for the prediction set. The spraying process was completed with an sprayer to ensure uniform spraying.
　　After spraying， the winter jujube samples were placed at a well??ventilated place for 12 h， for the acquisition of near??infrared spectra. The whole process was performed in laboratory at 20-25 ?? with a relative humidity of 30%-40%.
　　Spectrum acquisition instrument and methods
　　The experimental equipment was an Antaris II fouriertransform near??infrared spectrometer （FT??NIR） （Thermo Scientific， USA）. The instrument uses internal air as the background without the need for other sampling background. The determination wavelength was 4 000-10 000 cm-1， and the resolution was 4 cm-1. There were 1 557 sampling numbers， and each sample was scanned for 32 times for one spectrum. During the determination， the instrument was preheated for 30 min， and each winter jujube sample was scanned at three parts for 32 times totally， obtaining the average values as the original spectrum. Normalized spectral data were collected with OMNIC software and matched standard white plate and converted to other format. Data analysis and processing were performed with MATLAB2010b （Mathworks， America）. Fig. 1 is the schematic diagram of near infrared spectrum acquisition device. 　　Successive projections algorithm （SPA）
　　During spectral analysis and modeling， SPA[18] could extract the characteristic wavelengths of spectral data， which could effectively avoid information overlapping when reducing data dimension. Using this method， explain information could be extracted to the largest extent， so that prediction precision and speed could be improved， and the model could effectively realize the practical use purpose.
　　Supposing X（n*p） is the absorbance matrix of different concentrations of pesticide residue on fresh winter jujube surface （wherein n is sample size， and p is the number of full??spectrum waves）， and xk is the primary iteration quantity （wherein M is the range of the number of extracted wavelength variables）， the specific SPA is：
　　The spectral matrix X（n??m） of the prediction set gives the number of selected wavelengths， the SPA is：
　　（1） before iteration （m=1）， assign the random column， the kth column （k=1...p， p is the total number of wavelength） of the calibration set??s spectral matrix to xk；
　　For each pair of var（m） and M， after one time of cycle， multiple linear regression （MLR） is performed， obtaining the root mean standard error prediction （RMSEP）， and the var（m） and M corresponding to the lowest RMSEP value are the optimal values.
　　Modeling method and model validation indices
　　PLS is to establish a multi??statistical regression model between spectral data and pesticide concentration， so as to perform analysis. Except linear??regression analysis， PLS integrates the characteristics of principle component analysis （PCA） and canonical correlation analysis in modeling process. Therefore， it could establish optimized regression model， and is thus widely applied in spectral analysis[19].
　　In the experiment， the samples sprayed with the 1?? 10 chlorpyrifos solution was assigned to 1； the samples sprayed with the 1?? 50 chlorpyrifos solution was assigned to 2； the samples sprayed with the 1?? 100 chlorpyrifos solution was assigned to 3； the samples sprayed with the 1?? 1 000 chlorpyrifos solution was assigned to 4； and the samples sprayed with distilled water was assigned to 5. The difference between PLS values was 1. The judgment was performed with the various values as reference， and the values 0.5 unit higher or lower than the corresponding value represents the pesticide concentration， as shown in Table 1.
　　In order to effectively evaluate the precision of models， correlation coefficient （R）， standard error cross??validation （SECV）， standard error prediction （SEP） and precision were selected for model analysis and check， and the computational formulas are shown in Table 2. Among them， the nearer the R value approximates to 1， the better the regression （or prediction） result is； the smaller the SECV， the higher of the model??s predictive ability； the smaller the SEP value， the higher the model??s predictive ability to external samples； for the same batch of samples， the smaller the SECV and SEP values， the higher the model??s precision， and the closer the two values， the better the model??s stability； and precision reflects the correctness of the model and is used to verify the correctness of the model. 　　Results and Analysis
　　Original spectrum curve and characteristic analysis
　　Fig. 2 shows the original spectra of sample groups （sprayed with the four different concentrations of chlorpyrifos） and the blank control group. It could be seen from Fig. 1 that the samples prayed with the four different concentrations of chlorpyrifos and the blank control samples had basically the same trend in the near infrared range of 4 000-10 000 cm-1， without remarkable differences. There were obvious absorption peaks around 6 800 and 5 200 cm-1， while the samples sprayed with different concentrations of chlorpyrifos differed in the absorbance at the two wave peaks. In the measurement range of 4 000-10 000 cm-1， the absorbance was in negative correlation with the concentration of chlorpyrifos， i.e.， in the range， the lower the chlorpyrifos concentration is， the lower the peak value its spectrum had， which accords with results of previous studies.
　　Preprocessing results
　　Savitzky golay first derivative （SGFD）， savitzky golay second derivative （SGSD） and multiplicative scatter correction （MSC） were used to preprocess the spectra， respectively， and the preprocessed spectra are shown in Fig. 3. Then， modeling was performed to different concentrations of the pesticide combining with PLS. The PLS models built with the three preprocessing methods and that built using the original spectra were used to classify pesticide residue in winter jujube， and their modeling precision and predictive ability are shown in Table 3. Through comparison， the model built using the original spectra was optimal （as shown in Fig. 4）， so the SPA method was performed on the basis of the original spectra.
　　Selection of specific wavelengths by SPA
　　The division of samples in the calibration set was performed on the original spectra by SPA in the full spectrum range using SPXY （Sample set partitioning based on joint x-y distance）， and finally， compression of spectral variable was performed by SPA method. The range for the number of extracted wavelength variables （M） was assigned to 2-20， and the final variable number was determined by root??mean??square error （RMSE）. It could be seen from Fig. 4 that when the RMSE had the minimum value of 0.689 34， five characteristic wavelengths were selected from the original spectra： 3 999.64， 4 578.18， 5 179.861， 5 283.999 and 7 162.325 cm-1， the significance of which weakened sequentially. It could be known from the selected wavelengths that the chlorpyrifos on winter jujube surface had spectral characteristics with higher correlation in the wave band of 4 000-5 300 cm-1， which could be used for the rapid detection of chlorpyrifos residue on winter jujube surface. In the wave band， 5 179.861 and 5 283.999 cm-1 had higher contribution values， which is accordant with the existing result that there is a sensitive wave band around 5 200 cm-1 for the classification of pesticide concentration. 　　PLS models based on SPA characteristic wavelengths
　　Five characteristic wavelengths were selected from the original spectra by SPA， and SPA??PLS models were built still with the three preprocessing methods and original spectra. Their modeling precision and predictive ability in the classification of pesticide residue on winter jujube are shown in Table 4. It could be seen from the modeling precision and prediction ability of the PLS models in Table 3 that for the original spectra， the SEP increased from 0.653 683 to 0.691 782； RMSECV increased from 0.573 286 to 0.782 928； R decreased from 0.886 780 to 0.872 192； and the precision was improved from 0.737 027 to 0.795 669. As to SGFD， SEP increased from 0.739 075 to 0.771 944； RMSECV increased from 0.005 520 to 0.365 338； R decreased from 0.852 575 to 0.837 885； and precision was improved from 0.694 676 to 0.761 169. For MSC， SEP decreased from 1.232 677 to 0.708 189； RMSECV decreased from 0.689 179 to 0.033 439； R increased from 0.490 157 to 0.865 583； and precision was improved from 0.478 865 to 0.668 338. In the case of MSC， SEP decreased from 1.117 406 to 0.916 856； RMSECV decreased from 1.609 444 to 1.000 208； R increased from 0.612 946 to 0.761 372； and precision was improved from 0.485 502 to 0.558 819. The results showed that SPA??PLS only used five characteristic wavelengths for modeling， and the number of used variables only accounted for 0.32% of the full??wave band， while the accuracy and precision of the built model for pesticide residue on winter jujube surface were better than those of the PLS model built in the full??wave band. This is because that the peaks of near infrared spectra seriously overlap and comprise much redundant information， and the full??wave band also comprises massive information not related to chlorpyrifos， and the modeling using all the information of the spectra reduced model performance to a certain extent[20]. However， the models built on the basis of the effective characteristic variables selected from the full??wave band by SPA reserved useful information to the largest extent， and got rid of massive useless and redundant information， and the precision and stability of the models were improved.
　　Conclusions
　　SPA were used to select variables from the 180 groups of spectra data， and five wavelengths， 3 999.64， 4 578.18， 5 179.861， 5 283.999 and 7 162.325 cm-1， were selected from the original spectra， by which the calculated quantity of PLS model was reduced. And SPA method highly summarized most information in sample spectra， avoided information overlapping when removing redundant information， and thus simplified models which had improved precision and stability. Meanwhile， using the obtained characteristic wavelengths for the development of portable pesticide residue instrument could greatly reduce the requirements to hardware， providing reference and theoretical basis for the development of portable spectrometer for pesticide residue on fruit surface. 　　The results of this study showed that visible and near??infrared reflectance spectroscopy （Vis??NIRS） combined with PLS and SPA is a kind of method capable of effectively differentiate different concentrations of pesticide residue on winter jujube surface. In future study， more pesticides could be introduced， and the gradients of pesticides would be increased， to further validate the effectiveness of the method. However， the precision and accuracy of the built model were relatively lower， and the methods for improving model precision and accuracy should be studied and discussed further.
　　References
　　[1] WANG YP. Research progress of winter jujube storage and preservation technology[J]. Chinese Agricultural Science Bulletin， 2006， 22（2）： 82-85. （in Chinese）
　　[2] QU ZZ， WANG YH. Records of national fruit trees： Jujube volume[M]. Beijing： China Forestry Publishing House， 1993： 229-230. （in Chinese）
　　[3] XUE L， LI J， LIU MH. Detecting pesticide residue on navel orange surface by using hyperspectral imaging[J]. Acta Optica Sinica， 2008， 28（12）： 2277-2280. （in Chinese）
　　[4] CHUANSHAN YU， SUFANG FAN， YANJIE LI， et al. Representative commodity for five brassica vegetables based on the determination and dissipation of six pesticide residues[J]. Intemational Journal of Environmental Analytical Chemistry， 2015， 59（5）： 419-433.
　　[5] WU HY， CHEN JJ. Determination of organophosphorus pesticides in vegetables by gas chromatography[J]. Food & Machinery， 2010， 26（4）： 76-77. （in Chinese）
　　[6] LOZANO A， LUKASZ R， BELMONTE??VALLES N， et al. Pesticide analysis in teas and chamomile by liquid chromatography and gas chromatography tandem mass spectrometry using a modified QuEChERS method： validation and pilot survey in real samples[J]. Journal of Chromatography A， 2012， 1268（14）： 109-122.
　　[7] SU JF， ZHONG MS， CHEN J， et al. Multi??residue determination of 295 pesticides in tea and its products by gas chromatography??mass spectrometry and gas chromatograph[J]. Journal of Instrumental Analysis， 2015， 34（6）： 625-638.??（in Chinese）
　　[8] LIU YJ， ZHANG JZ， CAO MZ， et al. Study and application of enzymatic inhibition method in rapid detection of pesticide residue in agricultural products[J]. Modern Agrochemicals， 2004， 3（2）： 25-27. （in Chinese）
　　[9] SANDRA R， M?BRIO S， F?BTIMA RN， et al. Supercritical fluid extraction for pesticide multiresidue analysis in honey： determination by gas chromatography with electron??capture and mass spectrometry detection[J]. Journal of Chromatography A， 2004， 1048（2）： 153-159. 　　[10] CESARINO I， FERNANDO C， MARCOS RV， et al. Electrochemical detection of carbamate pesticides in fruit and vegetables with a biosensor based on acetylcholinesterase immobilised on a composite of polyaniline??carbon nanotubes[J]. Food Chemistry， 2012， 135（3）： 873-879.
　　[11] YUE XT， GAO GZ. Advances in detections methods of agrochemical residues in vegetables[J]. Mod Agric Sci Technol， 2008， 12： 111-115. （in Chinese）
　　[12] WU XIAOHONG， SUN JUN， WU BIN， et al. Qualitative nnalysis model of near infrared spectra of pork based on synergy interval partial least squares discriminant analysis[J]. Laser & Optoelectronics Progress， 2015， 52（4）： 242-247.
　　[13] GONG ZYU， LI YF， LIU YD， et al. Study on influence of irradiation angle on detection of sugar content of apple by near infrared spectroscopy[J]. Laser & Optoelectronics Progress， 2016， 53（2）： 235-240.
　　[14] ZHONG YP， ZHONG ZS， CHEN LZ， et al. Qualitative identification of floral origin and adulteration of honey by near??infrared spectroscopy[J]. Modern Food Science & Technology， 2010， 26（11）： 1280-1283. （in Chinese）
　　[15] GAO JF， LI XL， ZHU FL， et al. Application of hyperspectral imaging technology to discriminate different geographical origins of Jatropha curcas L??seeds[J] Computers and Electronics in Agriculture， 2013， 99： 186-193.
　　[16] WEN ZC， SUN T， GENG X， et al. Discrimination of pressed and extracted camellia oils by Vis/NIR spectra combined with UVE??PLS??LDA[J]. Spectroscopy and Spectral Analysis， 2013， 33（9）： 2354-2358. （in Chinese）
　　[17] TAN DM， SUN B， LOU HJ. Kind discrimination of egg by near infrared spectroscopy[J]. Chinese Journal of Spectroscopy Laboratory， 2012， 29（5）： 2699-2702. （in Chinese）
　　[18] ZHOU Z， YIN JX， ZHOU SY， et al. Knot defection on coniferous wood surface by near infrared spectroscopy and successive projections algorithm[J]. Laser & Optoelectronics Progress， 2017， 54（02）： 023001. （in Chinese）
　　[19] LIU F， FENG L， CHAI RY， et al. Discrimination of rice canopy leaf blast based on spectroscopic techniques and direct orthogonal signal correction[J]. Acta Optica Sinica， 2010， 30（2）： 585-589. （in Chinese）
　　[20] LUO X， HONG TS， LUO K， et al. Application of hyperspectrum technology in non??destructive measurement of soluble solid content in pitaya[J]. Laser & Optoelectronics Progress， 2015， 52（8）： 309-317. （in Chinese）