Preliminary Investigation Of Wood Identification By Information Fusion Between Visible-near Infrared Spectroscopy And Digital Image Features

The wood identification based on visible-near infrared spectroscopy and digital image processing could make full use of the advantages of both the spectroscopy method and the image method,and effectively overcome the limitations of the two methods when used alone.the technology plays an significant role in wood identification.In the study,visible-near infrared spectroscopy and digital image processing were respectively used to preliminarily identify 51 wood species.Simultaneously,these wood species were identified by the fusion of spectra features and digital image features.In addition,the influencing factors including different spectral bands,different wood surfaces and different pattern recognition methods on the prediction effect of wood identification models were investigated for exploring more stable models and good prediction results.The results are as follows:(1)Raw spectral data from different spectral bands(400?780 nm,780?1100 nm,1100?2500 nm,780?2500 nm,and 400?2500 nm)were collected for the establishment of wood identification models,which of recognition accuracies were 67.22%,70.59%,78.29%,86.14%,and 86.96%,respectively.results indicated that it was feasible for visible-near infrared spectroscopy to rapidly discriminate 51 wood species.Simultaneously,the effects of different wood surfaces(cross,radial and tangential surface)on identification models and prediction ability were investigated.the classification accuracies of cross,radial and tangential surface were 86.96%,75.07%,and 75.63%,respectively.results indicated that it was better to build models with visible-near infrared spectra collecting from cross surface in wood.(2)Firstly,gray-level co-occurrence matrix(GLCM)was utilized to extract 14 texture feature parameters of wood image.Then these data of feature parameters were processed by principal component analysis and correlation analysis between parameters and established predicted models,which of recognition accuracies were 82.49%and 80.11%,respectively.Results indicated that the texture identification models based on principal component analysis had a better result.In addition,the effects of different wood surfaces(cross,radial and tangential surface)on texture identification models and prediction ability were investigated.the classification accuracies of cross,radial and tangential surface were 82.49%,80.53%,and 81.51%,respectively.Results indicated that it was better to build back-propagation artificial neural network(BP-ANN)models with texture feature parameters collecting from cross surface in wood.(3)Respctively,combing the features of near infrared spectra and visible-near infrared spectra with digital image features,the prediction results of wood identification models were 91.04%and 92.85%.the results showed that the prediction effect for the addion of visible spectra into near infrared spectra had some improvement,but the effect was not obvious.meanwhile,the wood idntification models were established with the fusion features of cross,radial and tangential surface in wood,and the prediction accuracies of the models were 92.85%,88.79%,and 90.19%,repectively.the results indicated that it was better to build BP-ANN models with fusion features collecting from cross surface in wood.Comparing with wood identification using independent visible-near infrared spectroscopy or digital image processing,the combination of the two methods could build more robust identification models.(4)Pretreated methods including Savitzky-Golay convolution smoothing,first derivative,second derivative,and standard normal variate transformation(SNV),were used to treat raw spectra,the classification accuracies of predicted models based on these pretreated spectra and raw spectra were 96.43%,93.57%,90%,95.71%,and 97.14%,respectively.the results indicated that prediction effects of wood identification models from pretreated spectra had no obvious improvement.In addition,different pattern recognition methods including soft independent modeling of class analogy(SIMCA)method,partial least squares discriminant analysis(PLS-DA)method,and BP-ANN method,were used to identify wood species,and identification accuracies of the three methods were 93.57%,98.57%,97.14%,respectively.results indicated that the predicted effect of PLS-DA and BP-ANN were better than of SIMCA.