Deep Learning Based(Micro)plastic Spectral Reconstruction,Identification And Software Development

With the wide application of plastics,the plastics released to the environment are increasing rapidly.Due to their stable chemistry structure and unknown effect with other pollutants,the plastics in the environment are becoming difficult to handle.Since proposed in 2004,microplastics have become a new type of emerging contaminant.The damages from large-sized plastic particles are obvious and noticeable,while the pollution of microplastics can be more complex due to their small diameters.Before the standards of microplastics monitoring and management are established,it is important to accumulate related researches as much as possible,and the identification and classification were fundamental.The goal of identification and classification for microplastics is to determine their types and abundance,which matches well with many tasks of deep learning,namely classification tasks and,image segmentation and target extraction.In this thesis,deep learning was used for the identification and classification of microplastics.Firstly,based on up to 120 types of blended plastic,this study established the classification models in terms of their FT-IR spectra.Over 170 standard microplastic samples and 70 environment samples were tested and their FT-IR spectra were acquired,then a spectral dataset of common plastics was established.1D and 2D convolutional neural networks(CNN1D and CNN2D)with raw spectral data and images as input were selected to classify microplastics,and decision trees(DT)as well as random forest(RF)were also employed for comparison.The results indicated that CNN1D outperformed other models with an overall accuracy of>96%and>97%on small and large dataset,respectively.With fewer data,the prediction results on the environmental samples revealed that RF and CNN2D had better abilities of generalization,for that the environment samples contained many additives,suffered from ultraviolet rays,and might be disturbed by noises,and also CNN1D itself may be overfitted.For CNN2D,it used images as input and thus not sensitive to interferences.With richer data,the generalization ability of CNN1D was improved,and CNN1D outperformed other models in the prediction of environment samples.More data remitted the overfitting of CNN1D and improved its ability of generalization may be a potential explanation.In the meantime,the accuracy of RF and CNN2D was also improved,the performances of RF and CNN1D were very close.Based on the above results,RF and CNN2D were recommended when identifying uncommon plastics lacking enough data,while CNN1D seems to be the optimal choice for classifying common plastics such as polyethylene.Secondly,to improve the performances in the prediction of environment samples,24types of single-component plastics were used and two spectral reconstruction models,AE based on autoencoders and VCNN based on convolutional neural networks,were established and trained.Like before,four classifiers included DT,RF,linear support vector machines(LSVM),and CNN1D were also employed to classify the dataset before and after reconstruction.The results indicated that VCNN outperformed AE with the signal-to-noise ratio of>16 and R~2 of>0.96,but less stable than AE,both outperformed Savitzky-Golay filter algorithm.On the original dataset,LSVM worked the best with an accuracy of 97.38%,followed by CNN1D with an accuracy of 96.02%.On the AE-constructed dataset,all other models except DT performed worse than before.On the one hand,the characteristics of reconstruction were not included in the training data;on the other hand,AE worked badly and thus brought additional interferences to the dataset.On the VCNN-constructed dataset,the accuracy of all models was improved or similar as before,which should be the award of VCNN’s performance.When adding the characteristics of reconstructors to the training data,all models performed better on all datasets,which proved the above guess,also proved that the reconstructors did produce new features even though the constructed spectra were very close to before.This should be taken into account when combining them to classify microplastics.On the real environment dataset,comparing to blended microplastics,the max top-1 accuracy improved from<20%to>70%,demonstrating that the blended type was an important factor affecting the accuracy of the models on the real samples.Limited by unbalanced dataset,data augmentation,and real sample dataset itself,almost all models did not work as well as before on real samples.Therefore,further studies can focus on exploring a suitable method of augmentation and normalization that fits well with the characteristics of environment samples.Finally,based on the established models,the corresponding graph user interface program,i.e.,microplastic spectral analysis tool(m PSAT),was developed to simplify the process of microplastic identification and classification,and also provide an application example of this research for other researchers.