Research On Inversion Of Grassland Leaf Area Index Based On PROSAIL Model And Machine Learning Algorithm

The grassland ecosystem is an important part of the terrestrial ecosystem andplays an important role in regulating climate,soil and water conservation,and the ecosystem carbon cycle.Leaf Area Index(LAI)is an important index to monitor the growth of plants at various stages.Studying the leaf area index of grassland is of greater significance to the health evaluation of grassland ecosystems and the development and utilization of grassland resources.Statistical regression method based on vegetation index is a widely used vegetation inversion method today,but it is very dependent on measured data and cannot carry out long-term,dynamic and continuous efficient monitoring work.The PROSAIL model is based on the principle of radiative transfer and generates a large amount of canopy simulated data according to the physiological characteristics of the grassland.The machine learning regression algorithm has the characteristics of flexibility and high computational efficiency.Combining the PROSAIL model with the machine learning algorithm can obtain inversion results comparable to the accuracy of the statistical model,overcome the dependence of model construction on measured data,and achieve efficient,fast,and accurate LAI estimation.This paper takes the Bashang grassland in Zhangbei County,Hebei Province as the research area.Based on Landsat-7 ETM+remote sensing reflectance data and measured data,a statistical model of vegetation index is established,the PROSAIL model is used to build a simulated data set,and a machine learning regression algorithm is used to establish an inversion.The accuracy of each inversion model was compared,and finally,the LAI inversion model was used to carry out the spatial mapping of the Landsat 30-meter resolution leaf area index,and the temporal and spatial variation of the LAI in the study area was analyzed.The main work and conclusions of this paper are as follows:(1)Construction of the inversion model based on vegetation index.Correlation analysis was carried out on the vegetation index and the measured LAI data,and seven vegetation indices with correlation coefficients such as NDVI greater than 0.85were selected.The vegetation index inversion model was constructed by using 6regression methods such as linear,logarithmic,quadratic,power,sigmoid and exponential,and the measured LAI is used to evaluate the model accuracy.The results show that the logarithmic regression model(y=2.81 ln x+3.24)of GNDVI has the highest accuracy:R~2is 0.9402,RMSE is 0.1237.(2)Sensitivity analysis of PROSAIL model and construction of simulated dataset.Local and global sensitivity analysis was performed on the input parameters of the PROSAIL model.The LAI,Cab,and Cw were highly sensitive.The ALA,N,and Psoil were moderately sensitive.The rsoil,Cm,and Hotspot were lowly sensitive.So the change input parameters of the PROSAIL model are set as:LAI,Cab,Cw,ALA,N,Psoil,and other parameters were selected to be fixed empirical values.The spectral reflectance of the PROSAIL simulated canopy was converted into multispectral data according to the Landsat-7 ETM+spectral response functions.(3)Construction of the machine learning inversion model based on PROSAIL model.The Kernel Ridge Regression algorithm is compared with the current mainstream Multilayer Perceptron and Random Forest Regression algorithm,and the parameters of the three machine learning algorithms are optimized by using the learning curve and the grid search method respectively,and the LAI inversion under the default parameters and the parameter optimization are constructed models.The comparison results of the accuracy of each model show that the KRR model under parameter optimization has the highest accuracy,with an R~2of 0.8404 and an RMSE of 0.2162.The research shows that the PROSAIL model based on the KRR algorithm is a robust and fast inversion of the grassland LAI method.(4)Analysis of temporal and spatial variation of LAI in the study area.Based on the PROSAIL-KRR inversion model,the LAI distribution map with 30m spatial resolution was drawn.Through time series analysis,it can be seen that from January to December 2002,the LAI showed a changing trend of the Chinese character"几",and it reached the maximum value in August,which was in line with the characteristics of vegetation growth and change;From 2002 to 2011,LAI showed a fluctuating upward trend as a whole,with a change range of 0.35,an increase of about33%.According to Moran’s I index and Slope trend analysis,the LAI distribution was positively correlated and showed a significant clustering situation,and the LAI showed an increasing trend as a whole,indicating a slight improvement in the growth of grassland vegetation in the study area.