Research On Near-infrared Detection Technique And Transfer Learning Method Of Microplastics In Soil

Microplastics are extremely harmful to soil ecosystem and human health,and its toxicity is closely related to concentration and residual time of microplastics in soil.Detection for soil microplastics usually requires complicated pretreatment steps,laboratory optical observations,chemical analysis and other methods.There exist many problems such as long detection cycles,many experimental steps,and high reagent loss.Near-infrared(NIR)spectroscopy technology has the advantages of simple sample pretreatment,high detection efficiency,and stable equipment,which can realize rapid detection of soil pollution.This research took soil as the object,and conducted in-depth research on detection method of polyvinyl chloride(PVC),low-density polyethylene(LDPE)and polystyrene(PS)microplastics in soil based on NIR technology.The main research contents were as follows:(1)NIR-based detection models for microplastic concentration in soil were studied.For soil samples contaminated with PVC-,LDPE-and PS-type microplastics,spectral analysis were performed based on the following two NIR instruments:near-infrared-hyperspectral imaging(NIR-HSI)instrument and NIR sensor.Sensitive wavelengths of same-type microplastics were similar in NIR spectra of soil samples from different regions,while different-type microplastics had different sensitive wavelengths in NIR spectrum of soil samples from the same area.For the same sample set,sensitive wavelengths of NIR-HSI system and NIR sensor were different.Performances of sensitive wavelengths based model and the whole wavelengths based model showed little discrepancies.The detection model based on NIR-HSI system performed slightly better than that of NIR sensor.The best detection accuracy for PVC,LDPE and PS in soil samples were 0.91-0.94,0.97-0.98 and 0.94-0.97 respectively,with the Lowest detection error of 0.79%,0.35%and 0.46%,which improved the detection accuracy of PVC-,LDPE-and PS-type microplastic concentration in the soil.(2)Constructed microplastic detection models which could transfer among different soil regions.Transfer learning algorithms and machine learning algorithms were used to establish models for detecting microplastics pollution degree in different soil regions,and 18 transfer tasks were set to evaluate model performances.Typical machine learning models were also investigated and served as the comparison baseline.The overall results showed that the Manifold Embedded Distribution Adaptive(MEDA)transfer learning model had an average detection accuracy of 0.98 and 0.80 in the source domain and target domain respectively,higher than that of Transfer Component Analysis-Support Vector Machine(TCA-SVM)model and SVM model.And averaged training time of MEDA transfer learning model(0.70 s)was much lower than TCA-SVM model(21.90 s)and SVM model(41.38 s).(3)Proposed an efficient transfer strategy for the transference from high-throughput NIR-HSI system to portable NIR sensor,and realized detection for six-class microplastics in soil.The feasibility of Maximum Mean Difference(MMD)algorithm serving as indicator of the optimal calibration transfer method was verified,and Easy Transfer Learning(EasyTL),a parameter-free transfer learning modeling algorithm was used to construct transferable model.The results showed that after Direct Standardization(DS)and Repetitive File(Repfile)transformation,the MMD values between spectral datasets of the two NIR instruments significantly decreased and the averaged MMD value of Repfile transformation was the smallest one,which proved that MMD could be used as a diagnostic tool to select the optimal calibration transfer method.Averaged accuracy and robustness of Repfile-EasyTL model was the highest than others,with the accuracy value of 0.7.In addition,compared with SVM model,Repfile-EasyTL model saved 839 times of model training time,which greatly improved efficiency of model construction.In summary,this study developed detection method based on NIR spectroscopy,which improved detection accuracy of microplastics in soil.Using transfer learning algorithms,high-efficient methods for model transference were proposed,which could provide theoretical basis and technical support for field diagnosis of microplastics in large-scale soil environment.