Tree Species Classification Based On Canopy Hyperspectral Images

Quick and precise identification of species has always been an essential requirement of any ecological research and application.However,this requirement is not entirely fulfilled in identifying and monitoring canopy tree species of high diversity forests.In this thesis,I studied spectral separability across plant species,the roles of spectral and textural information,as well as the potential of species classification in unmanned aerial vehicle(UAV)based hyperspectral canopy images of two green lands in school and nursery of Chenshan Botanical Garden by deep learning.Results showed that: 1)spectral reflectance difference exists among texted species.Simple spectral statistics can capture the difference,but they are not robust enough for precise species identification.2)A model with combined spectral and textural information performed better than any model with only one kind of information,indicating that both types of information should be used to classify tree species.3)Deep learning classification model,resulting in over 80% overall accuracy,performed better than SVM in the task of classification of 162 tree species.These results suggest that UAV-based hyperspectral canopy images,combined with the power of deep learning,have great potential for precise and efficient identification of canopy tree species in biodiversity monitoring and forest application.More samples of hyperspectral images and sophisticated structures of deep learning models will improve the accuracy of tree species classification.