Study On Tree Species Identification Based On Leaf Hyperspectral Images

It is of great significance for rational planning,utilization and protection of forest resources to accurately identify tree species.In recent years,it has become a hotspot in forestry remote sensing research to identify forest tree species based on hyperspectral remote sensing technology.In many research results at home and abroad,hyperspectral identification models for tree species were established in only single season and the same region.However,the changes of season and region can affect the identification results of hyperspectral models for tree species.To solve this problem,this paper carried out the research on tree species identification in different seasons and regions based on leaf hyperspectral images.This paper aims to establish a hyperspectral identification model for tree species suitable for different seasons and different regions,and to mine more effective hyperspectral dimension reduction and identification algorithms.Based on the tree species on the campus of Beijing Forestry University,this paper collected leaf samples of 50 tree species,and carried out identification research on 20 tree species in different seasons(spring,summer and Autumn),and on 5 tree species in different regions(Beijing Forestry University and Xiling Lake Park,Wu’an city,Hebei Province).Furthermore,the hyperspectral identification model was established for the above 50 tree species,which can identify the tree species in different seasons and different regions,and ensure a high identification accuracy for each tree species.The research contents and results of this paper are as follows:(1)The results of leaf hyperspectral response of tree species and modeling analysis in different seasons and regions show that the changes of season and region have significant effects on the reflectance spectra of tree leaves,especially in the near infrared region of 760-1000 nm.Moreover,the identification results of tree species identification models established in a single season and the same region are significantly affected by the changes of season and region.When the spectral information of different seasons and different regions was added into the models,the identification accuracy of the models is improved significantly.(2)Different preprocessing methods were used to preprocess the original hyperspectral data.Through comparison,it is found that the models established based on spectral data preprocessed by the logarithmic transformation combined with the first derivative method achieve the best classification performance.(3)To solve the problem that the band subsets selected by the ranking-based clustering methods are easy to fall into the local optimal solution,this paper proposes a band selection method based on shared nearest neighbor and correlation analysis(SNNCA),and the SNNCA method is compared with multiple band selection methods.The results show that the feature bands selected by the SNNCA method achieve the best classification performance.(4)To solve the problem that Tikhonov regularization can effectively improve the classification performance of collaborative presentation classification models,but reduce the operation efficiency of algorithms,this paper proposes an Euclidean distance-based adaptive cooperative representation classifier with Tikhonov regularization(EDACRT),and the classification performance and running time of EDACRT method are compared with those of multiple traditional algorithms under all bands and feature bands.The results show that the EDACRT method achieves the best classification performance both on all bands and feature bands,in which the overall accuracy(OA),average accuracy(AA)and Kappa coefficient(Kappa)based on all bands are 94.46%,95.97% and 94.29%,respectively,and OA,AA and Kappa based on feature bands are 85.02%,87.80% and 84.56%,respectively.Moreover,the operating efficiency of EDACRT method is 22.5% and 27.45% higher than that of CRT method under all bands and feature bands,respectively.(5)To solve the problem that the bit 0 in random sparse coding matrix of error-correcting output codes(ECOC)algorithm can affect the classification performance of ECOC algorithm,this paper proposes supervision mechanism-based ECOC(SM-ECOC)algorithm,namely SM-ECOC-V1 and SM-ECOC-V2 algorithms.And the classification performance of SM-ECOC algorithm is compared with those of multiple traditional algorithms under all bands and feature bands.The results show that the SM-ECOC-V2 algorithm achieves the best classification performance both on all bands and feature bands,in which OA,AA and Kappa based on all bands are 96.98%、97.68% and 96.88%,respectively,and OA,AA and Kappa based on feature bands are 93.89%、95.24% and 93.70%,respectively.The conclusions and methods presented in this paper provide theoretical basis and technical reference for the fine identification of forest tree species based on hyperspectral remote sensing technology and the identification of other ground objects based on hyperspectral images.