Study On Hyperspectral Identification Model Of Typical Tree Species In Changting County

Forest resource investigation is an important field of remote sensing application.Tree species identification based on hyperspectral technology is a challenging and practical development direction of forest remote sensing.Tree species identification model based on measured spectral data is an important basis for tree species identification based on space-borne or airborne imaging spectral data.Changting County used to be the worst-hit area of soil and water loss in China.After long-term ecological management and restoration,Changting County has become a banner of soil and water loss control in the red soil area of South China.The forest area has been greatly increased and the forest coverage rate has been greatly improved.Investigating the vegetation restoration status in this area,especially the type and distribution of tree species,is of great significance for forest resource management and biodiversity assessment.Therefore,this study selected 7 tree species in Changting County(Pinus massoniana,Cunninghamia lanceolata,Castanopsis fissa,Cinnamomum camphora,Citrus reticulata Blanco,Eriobotrya japonica and Myrica rubra)as the research objects,and used the ASD Field Spec4 field spectrometer to measure the hyperspectral data of different tree species canopy.By analyzing the spectral characteristics of different tree species under a variety of spectral transformations,the spectral characteristic bands were screened,and three classification algorithms of Random Forest(RF),e Xtreme Gradient Boosting(XGBoost))and Gradient Boosting Decision Tree(GBDT)were used to identify tree species,so as to construct the hyperspectral recognition model of typical tree species.The applicability of XGBoost and GBDT tree species classification based on hyperspectral data was evaluated,and the importance of spectral characteristic bands was analyzed using XGBoost and RF classification algorithms.The main conclusions of this study are as follows:(1)Through the analysis of the spectral characteristics of different data types,it can be seen that the characteristic bands with large differences mainly concentrate on the green and red region in the visible band,the two water absorption bands near 960 nm and 1100 nm in the near-infrared band,and the reflection peak near 1650 nm in the short-wave infrared band.(2)Comparing RF,XGBoost and GBDT classification algorithms for tree species identification and classification of different types of spectral data,it can be seen that the three classification algorithms have the best classification effect on continuum removal spectral data,and the classification effect of XGBoost algorithm is better than that of RF and GBDT.The overall classification accuracy of XGBoost for the original spectral data,the first-order differential spectral data,the second-order differential spectral data and the continuum removal spectral data were 95.83%,91.67%,94.64% and 95.83%,respectively,and the average F1-score was 0.96,0.92,0.95 and 0.96,respectively.In the study of tree species recognition with different data types and classifier combinations,the "original spectral data-XGBoost" model and the "continuum removal spectral data-XGBoost" model have the best tree species recognition effect.(3)Through the comparative analysis of the classification effect of XGBoost and GBDT with the RF classification effect,it can be seen that the overall classification accuracy and average F1-score of the original spectrum data,second-order spectrum data and continuum removal spectrum data are shown as: XGBoost> GBDT> RF,The overall accuracy of the first-order differential spectrum data is: GBDT> RF>XGBoost,and the average F1-score is GBDT> RF = XGBoost.Therefore,XGBoost and GBDT are suitable for tree species identification research in hyperspectral data,and the classification effect is good.(4)Through the XGBoost and RF algorithm to analyze the band importance of the selected spectral characteristic bands,it can be seen that compared with the near-infrared band and the short-wave infrared band,the visible light band is the most important band for tree species identification.