Prediction of calorific value of Salix psammophila using canopy visible and near infrared spectra
Calorific value is one of the important combustion performance parameters for shrub biomass energy utilization.In order to compensate the shortcomings of the destructive,time-consuming,and laborious detection of a large number of samples for traditional laboratory detection methods,this study aimed at investigating the accuracy difference of visible and near infrared spectroscopy(Vis-NIR)combined with various chemometric methods to predict the calorific values of Salix psammophila.Firstly,the spectral transform technique was used to analyze the performance of two kinds of canopy Vis-NIR spectra of S.psammophila,namely reflectance spectra(R)and logarithm reflectance spectra(lg(1/R)).The optimal canopy Vis-NIR spectra of S.psammophila were preprocessed by various optimization algorithms including standard normalized variate(SNV),normalization,combination of SNV and normalization,and lifting wavelet transform(LWT)methods.Meanwhile,the de-noising performance of LWT based on different de-noi-sing parameters including mother wavelet,wavelet orders and decomposition levels was also analyzed in this study.Then the canopy Vis-NIR spectra processed by the optimal pre-processing method were inputted into convolutional neural net-work(CNN)model.Finally,as for the parameters selection of CNN models,three optimization techniques including whale optimization algorithm(WOA),grey wolf optimizer(GWO),and sparrow search algorithm(SSA)were em-ployed to optimize three main parameters of CNN models including learning rate,batch size,and regularization coeffi-cient.Among these pre-treatment methods,LWT obtained the best de-noising results based on db mother wavelet with wavelet orders and decomposition levels of 4 and 5,respectively.In terms of Vis-NIR models optimization of calorific value prediction of S.psammophila,the results demonstrated that the optimal model was achieved using LWT combined with WOA-CNN method with the coefficient of determination(R2),root mean square error(RMSE)and ratio of per-formance to standard deviation(RPD)of 0.852,0.103 and 2.599,respectively.In addition,the optimal parameters of learning rate,batch size,and regularization coefficient were achieved for CNN models based on WOA.The RPD value increased by 19.11%and 76.80%in comparison to the PLS and CNN models using raw spectra.These results provide data support for the efficient and refined utilization for S.psammophila biomass energy.
Salix psammophilacanopy spectravisible and near infrared spectroscopycalorific valuechemometrics
万方数据
维普
