Research On Non-destructive Detection Of Lead Content In Lettuce Leaves Based On Hyperspectral Image Technology And DBN

With the intensification of industrial pollution,lead gradually invades China's arable land,water and air,posing a great chal enge to the safety of agricult ur a l production.High concentration of lead not only inhibits the growth of vegetables,causing the reduction of production,but also gradually enriches in the organism through the food chain,which will do harm to human health.Therefore,it is of great significa nce to explore a method to detect lead content in vegetables for the development of vegetable industry and human health.At present,the detection of heavy metal lead is mainly based on the chemical methods,which are destructive,consumable and cannot meet the requirement of real-time detection.As a rising non-destructive testing technology,hyperspectral image technology is widely used in the quality testing of agricultural and sideline products.However,there exits strong correlation and high redundancy between the bands of hyperspectral data.Thus,how to effectively process the nonlinear hyperspectral data is the focus of this research.Conventional processing methods have low modeling accuracy and can not satisfy the requirement of accurate detection.However,deep belief network(DBN)can extract more essential deep features from the original data,and thus it has stronger prediction ability.In this paper,DBN is applied to the processing of hyperspectral data.The normalized spectral data was directly used as the input of the network to realize the prediction of lead content in lettuce leaves,and the traditional analysis models were established for comparison.Specific research contents are as follows:(1)The hyperspectral data of lettuce leaves under different lead stress levels was obtained and divided.Hyperspectral images of 360 lettuce samples were collected,and the average spectral values in the region of interest(ROI)were calculated as the spectral information of the samples.Then,SG algorithm was used to preprocess the spectral data after removing the jitter bands.Besides,SPXY algorithm was used to divide the calibration set and the prediction set with the ratio of 3:1.Finally,there were 270samples in calibration set and 90 samples in prediction set.(2)The conventional models were established to predict lead content in lettuce leaves.First,SPA and PCA algorithm were used to obtain the feature spectral data.Afterwards,BP neural network prediction model was established for analysis.SPA-BP model had better performance than PCA-BP,which indicated that SPA was better than PCA in reduction of spectrum dimension.Moreover,in order to improve the prediction performance of BP,PSO algorithm and GWO algorithm were used to optimize BP model.The results showed that SPA-GWO-BP model has the best prediction abilit y with R_c~2 of 0.932,R_p~2 of 0.910,RMSEC of 0.342 mg/kg,RMSEP of 0.383 mg/kg,RPD of 3.886,and operation time of 20.427s.(3)DBN model was established to predict lead content in lettuce leaves.The preprocessed spectral data was taken as the input of DBN,and the influence of network structure,learning rate,batch size on DBN training results was analyzed respectively.When the structure of DBN was 399-150-100-50-1,the learning rate of DBN was 0.1,and batch size was 90,the prediction accuracy was the highest,with R_c~2 of 0.950,R_p~2of 0.935,RMSEC of 0.275 mg/kg,RMSEP of 0.332 mg/kg,RPD of 4.492,and the running time of 10.457s.(4)DBN was improved to establish prediction model of the lead content in lettuce leaves.In order to avoid local optimization of DBN model in the process of training,PSO and GWO algorithm were used to optimize the initial weight and bias of DBN respectively.Although new parameters were introduced into the optimized DBN,leading to more training time,the prediction performance and stability of the DBN model were significantly improved,with R_c~2 and R_p~2 reaching more than 0.95,RMSEC and RMSEP less than 0.3 mg/kg,and RPD greater than 5.Compared with PSO-DBN model,GWO-DBN model had better preformance,stronger stability and shorter training time,R_c~2 and R_p~2 were 0.980 and 0.973 respectively,RMSEC and RMSEP were 0.204 mg/kg and 0.226 mg/kg respectively,the RPD reached 6.582,and the running time was 25.382s.Compared with BP models,it is obvious that the feature modeling effect of DBN model is better,which verifies the efficiency of deep learning model.What's more,the performance of DBN model was further improved after optimization,which showed that the combination of hyperspectral image technology and GWO-DBN model could efficiently realize rapid,non-destructive and accurate detection of lead content in lettuce leaves.This study can providetechnical support for real-time monitoring of heavy metal content in vegetables.