Geographical Discrimination Of Yunnan Wild Boletes Based On Near Infrared Spectroscopy Combined With Machine Learing

Wild porcini is a precious natural resource with rich nutrition and high medicinal value.It has attracted the attention of countries all over the world.China is a major producer and exporter of edible fungi in the world,and there is broad consensus on the rational and efficient utilization of wild edible fungus resources.In recent years,with the improvement of people’s living standard and the changing of diet structure,wild edible fungus has become an indispensable part of people’s daily diet.Yunnan Province wild edible fungus variety,excellent quality,high yield,resources are extremely rich.Wild porcini with unique flavor,high nutritional value,the power of anti-cancer,antioxidant and immune-enhancing functions are popular with consumers.The growth and development of wild bovine liver bacteria is influenced by many factors.Yunnan has different climatic conditions and different climatic factors in different production areas,which results in different nutritional value of bovine liver bacteria.There are many sources of wild bovine liver bacteria,which makes its effective utilization difficult.At present,infrared spectroscopy,as a nondestructive,efficient,rapid and reliable analytical technique,has been widely used in the identification and resource evaluation of wild bacteria.In this study,FT-NIR was used to collect spectra of Lanmaoa asiatica(L.asiatica)and Butyriboletus roseoflavus(B.roseoflavus),two representative wild edible porcini bacteria in Yunnan Province.Spectral data were combined with a variety of pattern recognition methods to identify porcini from different regions.In addition,1,383 synchronous,asynchronous,and synthetic 2DCOS were generated using two-dimensional correlation infrared spectroscopy techniques to establish a deep learning model based on 2DCOS images.By comparing the accuracy and performance of different identification models,the best identification method for wild edible porcini was found.It can provide a theoretical basis for the establishment of a reasonable and efficient evaluation system for the resources of porcini fungi,and a reference for the quality evaluation of other edible fungi.The main findings from the study were as follows:(1)The results of unsupervised pattern recognition showed that principal component analysis(PCA)could not effectively isolate samples from different regions.In contrast,samples from hierarchical clustering analysis(HCA)showed clustering at very low distances suggesting that chemical information was more similar in samples from different regions.(2)Partial least squares discriminat analysis(PLS-DA)model showed 99.5%accuracy and 99.02%accuracy in both the FD and SD pretreated training and test sets of the B.roseoflavus identification model.Higher R~2(0.877 and 0.87)indicated good fit of the model.The best pretreatment method for L.asiatica was SD,with 100%accuracy in both training and test sets,0.815 for Q~2and 0.935.200 for R~2replacement test results confirming no fit risk in established PLS-DA models.(3)Extreme Learning Machine(ELM)model of B.roseoflavus results showed that the SD pretreated model test set was 99.21%accurate and the training set was 100%accurate when the number of neurons was 50.In the ELM model based on L.asiatica,the number of implicit layer neurons was set at 30 to obtain the optimal model.After SD pretreatment,the model had the highest accuracy,98.51%accuracy of test set and 100%accuracy of training set.(4)Randomized forest(RF)model results showed that the raw data and FD,SD,SNV,MSC pretreated data of B.roseoflavus were 73.53%,91.18%,96.08%,83.33%,and83.33%accurate,.The model accuracy of L.asiatica dataset was 71.70%,96.23%,98.11%,86.79%and 92.45%.Compared with the original data,the accuracy and fitting effect of SD pretreatment model were significantly improved.However,RF models performed poorly compared to other models.(5)SVM results showed that SVM models based on raw data was significantly different from the SVM model based on pretreated data for both B.roseoflavus and L.asiatica.,and the best pre-treatment methods were SD with 97.5%,100%and 99.04%,100%,respectively.However,the g value was low and the model was too slow to perform well compared to other provenance identification models.(6)The results of the residual convolution neural network(Res Net)model showed that at iteration 18,the training and test sets of the B.roseoflavus identification model based on synchronous 2DCOS were 100%accurate and the model stabilized,at which point the loss value was 0.13.Synchronous 2D spectra were used to identify the origin model of L.asiatica with 100%accuracy and 0.09%loss.In addition,the external validation accuracy of both models is 100%,with good generalizability.The asynchronous,integrated 2DCOS model is not effective.(7)A brief analysis of local climate factors showed that the average annual temperature in the 14 sampled sites ranged between 15.1℃and 17.6℃Rainfall and average temperatures varied significantly from month to month.Precipitation will be concentrated in June-September with a range of 89.94-230.52 mm.From the temperature and precipitation of 14 sites in this study,all sites had favorable conditions for the growth of porcini.The results of the comparison between different models indicate that the combined synchronous two-dimensional spectroscopy and Res Net model performs better overall than traditional chemometric models.This provides a new approach for rapid identification of the origin of boletes in the market.