Assessment Of Species Diversity In The Subtropical Forest Using Multi-source High Resolution Remote Sensing Data

Biodiversity underpins the health of ecosystems and the services they provide to society.Monitoring is an essential part of biodiversity conservation,allowing governments and civil society to identify problems,develop solutions,and assess effectiveness of actions and progress toward meeting the Aichi targets set by the Convention on Biological Diversity.Hyperspatial data with their increased pixel resolution are possibly best suited at estimating the distribution of forest species diversity.But the high pixel resolution may be the limitation for forest species diversity estimation,particularly when spatial resolution becomes too fine and pixels are smaller than the size of the object(e.g.,tree canopy)being identified.Airborne lidar is an active remote-sensing tool of increasing importance in ecological and conservation research due to its ability to characterize three-dimensional vegetation structure.If forest species diversity and composition can be related to vegetation structure,the assessments of forest species diversity by LiDAR data may be possible.The objective of this paper was to investigate the ability of remote sensing data to predict tree alpha diversity by using high spatial resolution satellite GF-2 multi-spectral imagery,airborne LiDAR and hyperspectral data.This paper explored the relationships between alpha species diversity indices(Richness,Shannon-Wiener and Simpson diversity indices,Pielou evenness indices)and remote sensing data derived parameters such as the variance of spectral indices and texture derived from GF-2 multi-spectral imagery and airborne hyperspectral imagery and forest vertical structure parameters derived from airborne LiDAR data.Random forest was used to reducing the input remote sensing features and find most meaningful inputs.Four different machine learning methods(Random Forest,RF;Support Vector Regression,SVR;K-nearest neighbor,KNN,Cubist)were tested using 10-folder cross-validation and the SVR model was selected to predict forest species diversity in this study.The results and conclusions of this paper as follow:(1)The machine learning method can be used to find most meaning features and explore the distribution of features in specific study area.Then they improved the ability of remote sensing data to estimate the forest speices diverisity.(2)The spectral and texture features derived from high resolution remote sensing data show a satisfactory power for estimating Shannon-Wiener,Simpson diversity index and Pielou's evenness index(R~2 > 0.54).The heterogeneity of spectral or texture of forest canopy are the most meaningful features for forest species estimation.(3)With the help of forest vertical structure parameters derived from LiDAR data,the SVR model for species diversity estimation show better result than using GF-2 multi-spectral or hyperspectral data only.The combination of LiDAR and hyperspectral data show the best estimation result specially for species richness estimation(R~2 = 0.74,RMSE = 3.36).(4)Despite of lower spectral and special resolution,the GF-2 multi-spectral data show the similar result with hyperspectral data for Shannon-Wiener,Simpson diversity index and Pielou's evenness index estimation.The accuracy of these diversity indices estimation using remote sensing data didn‘t increase significantly with the increse in spatial and spectral resolution.(5)The merging derived waveform parameters from synthesize waveform and the variation of spectral indices and texture that derived from GF-2 imagery showed a satisfied result for four different species diversity estimation(Species richness: R~2 = 0.58,RMSE = 4.2;Shannon-Wiener diversity: R~2 = 0.7,RMSE = 0.47;Simpson diversity: R~2 = 0.73,RMSE = 0.164;Pielou evenness: R~2 = 0.68,RMSE = 0.15).Synergizing passive and active remote sensing offers tremendous potential for forest species diversity estimation in regional area.