Study On Wood Identification Methods For Dalbergia And Pterocarpus Species In Combination With Machine Learning

Tropical wood species have been the foci of Convention on International Trade in Endangered Species of Wild Fauna and Flora(CITES),with the sharply decrease of global forest resource and aggravation of illegal logging and associated trade,herein,it is a severe challenge to strengthen the surveillance of timber import and export for the implementation of CITES in China.The scientific and accurate identification of wood species provides an important channel to protect the global forest resource and promote the healthy development of timber import and export in China.Aiming at revealing the variation rules of wood inherent features on the species level,and breaking the barrier of conventional wood anatomy methods on species-level identification,this study emphasized on wood identifications including quantitative wood anatomy,DNA barcodes and Computer Vision in combination with machine learning approaches for tropical Dalbergia and Pterocarpus species.This study illuminated the variation rules of wood inherent features on the species level,providing scientific evidence for the species-level identification of timber and technical support for the implementation of CITES regulations.The results of this study were listed as following.Firstly,the wood macroscopic features,i.e.color and texture,and microscopic anatomy features,i.e.vessel,parenchyma and ray,exhibited relatively high similarity across eight Dalbergia species.D.melanoxylon showed divergence on wood color and ray width with other seven Dalbergia species.For the six Pterocapus species,P.angolensis,P.indicus and P.macrocarpus showed high similarity on wood macroscopic and microscopic features,furthermore,P.santalinus and P.tinctorius exhibited extreme similarity.Nevertheless,P.soyauxii showed divergence on parenchyma with other Pterocarpus species.In conclusion,it is impossible to discriminate eight Dalbergia and six Ptercarpus species at species level accurately with wood macroscopic and microscopic features.Secondly,D.melanoxylon and P.soyauxii could be discriminated from eight Dalbergia and six Pterocarpus species with Naive Bayes algorithm,respectively.The decision-tree classifier could recognize D.latifolia,D.melanoxylon and D.odorifera from eight Dalbergia species,and sepertate P.soyauxii from six Pterocarpus species.SVM classifier could identify D.cochinchinensis,D.hupeana,D.latifolia,D.melanoxylon,D.odorifera and D.oliveri,as did P.angolensis,P.indicus,P.macrocarpus and P.soyauxii.The rule-based algorithm,Jrip,could separate D.melanoxylon,D.odorifera and D.oliveri successfully,as well as P.macrocarpus and P.soyauxii.Among the four machine learning classifiers,SMO exhibited the highest success rates,95%(eight Dalbergia species)and 93.3%(six Pterocarpus species),respectively.Thirdly,for the DNA barcoding of eight Dalbergia species,the combination of ITS2+trn Hpsb A was proposed as the best-performance barcode combination,by mean of analyzing the characters of candidate barcodes,inter-and intra-specific distance,and success rates of Taxon DNA and NJ tree.The results of machine learning approaches(BLOG and WEKA)showed that SMO classifier exhibited the highest identification success rate,which is higher than Taxon DNA and NJ tree.Moreover,the classifiers BLOG,J48 and Jrip also performed relatively higher identification success rates,and generating logic rules and decisions for species recognition,which is simple and readable.Additionally,the non-vouchered specimens were identified accurately using machine learning approaches.On the other hand,for the discrimination of six Pterocarpus species,Taxon DNA and NJ tree could not discriminate six Pterocarpus species.While the two-locus combination,ITS2+mat K could discriminate six Pterocarpus species successfully when using machine learning approaches,which is much costeffective than three-locus combination mat K+ITS2+ndh F-rpl32 proposed by previous study.P.santalinus and P.tinctorius were identified successfully using machine learning approaches with mini-barcode,ndh F-rpl32.The results demonstrated that machine learning approaches outpermed Taxon DNA and NJ tree methods on accuracy and efficiency for DNA barcoding of Dalbergia and Pterocarpus.Lastly,digital images of transverse section were collected from wood specimens with Xylotron,building up the wood images database with 10,237 images for 15 Dalbergia and 11 Pterocarpus species.Deep learning models types of VGG16 and Alex Net were constructed with convolutional neural networks,and trained with training set,thereafter tested with testing set.The accuracy of VGG16 model on species-level was 85.44% when identifying 14 Dalbergia species,and 65.44% for 11 Pterocarpus species,which could discriminate D.frutescens var.tomentosa,D.oliveri,D.hupeana,D.sissoo,D.melanoxylon,D.nigra and D.stevensonii,as well as P.soyauxii.The accuracy of Alex Net model on species-level was 99.34% for identification 15 Dalbergia and 11 Pterocarpus species,among those 12 Dalbergia and 7 Pterocarpus species achieved the 100% of success rate,which performed much better than VGG16 model.The reuslts of the model configuration demonstrated that the Alex Net model could exhibit relatively high accuracy with more than 10 specimens,100 high-quality images,and patch size of 1000 × 1000 × 3 for per species,when runing the model.The deep learning model could reach the goal of species-level identification of endangered Dalbergia and Ptercarpus species rapidly and accurately.Consequently,we concluded that the deployment of machine learning approaches with quantitative wood anatomy,DNA barcoding and Computer Vision illustrated the variation rules of wood inherent features on the species level,and reach the goal of species-level of wood species.The results demonstrated in this study poivide scientific and technical supports for combating illegal logging and implementation of CITES regulations.