Collaborative Inversion Of Grassland Leaf Area Index In Eastern Inner Mongolia Using Ground Measurement And Multi-source Remote Sensing Data

Grassland ecosystems play a critical role in maintaining biodiversity,protecting water and soil,and mitigating climate change in ecological,socioeconomic,and agricultural activities.Satellite remote sensing has the advantage of large-area and long-term observation of the earth’s surface,providing a low-cost and efficient means of monitoring grassland growth at macroscopic scales.Leaf area index(LAI)is the most basic characteristic parameter of many ecological models,such as grassland productivity models and carbon cycle models,and is related to gas-vegetation exchange processes such as photosynthesis,evaporation and transpiration,and rainfall interception.Obtaining the optimal LAI estimation model and inverting the LAI at the pixel scale of eastern Inner Mongolia grassland are of great significance to the management and ecological research of eastern Inner Mongolia grassland.This paper takes the eastern Inner Mongolia grassland as the research area,and establishes LAI inversion models from both empirical and mechanistic model perspectives using two data sources:ground cover spectral and remote sensing images.The ability of these models to estimate LAI is analyzed comprehensively.Specifically,the LAI parameters of the eastern Inner Mongolia grassland canopy in July-August 2022 are discussed based on measured spectral data.The characteristics of the original and nine transformed spectra of the grassland canopy are analyzed,and single-band functions and stepwise regression multi-band function models are built using feature bands and vegetation indices.Machine learning models are established using random forest and radial basis functions and compared.LAI estimation ability is analyzed using vegetation indices,random forests,radial basis functions,and Ada Boost Regressor algorithms based on ground reflectance obtained from Landsat 9 and Sentinel-2 satellite remote sensing images corresponding to sampling points.The input parameters of the PROSAIL model are determined by sensitivity analysis,and the simulated spectra are obtained using a radiative transfer model.The most similar spectra are obtained using a cost function,and the corresponding LAI is obtained as an estimation value.The model is optimized by replacing the cost function and adding ecological rules.The results show that:(1)The characteristics of grassland canopy spectra are that they absorb less in the visible part of the spectrum,forming a reflection peak with a central wavelength of 550 nm,while LAI exhibits strong reflectivity in the near-infrared part of the spectrum,resulting in the"red edge"phenomenon.Spectral transformation can more easily select bands that are more sensitive to LAI inversion.The diversity of spectral transformation methods provides more possibilities for the inversion of physical and chemical parameters of surface features.(2)For grassland canopy spectra,after performing Pearson correlation coefficient tests,the wavelengths with the highest correlation were concentrated in the red edge band,with correlation coefficients ranging from 0.503 to 0.742.The best model for estimating LAI was a random forest model that used SG-smoothed spectra as the independent variable,with a validation set R~2 of 0.494,RMSE of 0.524,and MAE of 0.404.Overall,machine learning models produced better results than empirical models and provided better LAI estimation capabilities.In addition,vegetation indices also provided a simple and practical method for LAI estimation.In summary,combining machine learning with spectral transformation can effectively improve the LAI estimation ability of grassland ecosystems.(3)Among the two types of satellite image processing results,Sentinel-2outperformed Landsat 9,and machine learning yielded better results than vegetation indices for inversion.Regarding LAI estimation,Sentinel-2 can be estimated using either RF parameter tuning or RBF neural network models,and the results indicate that the difference between the two methods is insignificant,with R~2 values of 0.39 and 0.407,respectively.However,the RF model exhibited a smaller MAE of 0.473.The optimal estimation model,based on the Ada Boost Regressor algorithm,outperformed other models in terms of R~2,RMSE,and MAE on the validation set.The respective values were 0.55,0.543,and 0.427.(4)When using the PROSAIL model for LAI inversion,different cost functions with different minimum distances can be used to estimate LAI using a look-up table method to reduce the influence of parameter instability.The ecological rule of hspot,which can be approximated to 0.5/LAI,and multiple minimum values are introduced to further improve the inversion accuracy of the model.Using a look-up table with a simulated spectrum capacity of 300,000,the spectral angle function is used as the cost function,and the method of using the average of the two minimum values as the estimated value has the highest precision for estimating LAI,with MAE,MSE,RMSE,and R~2 of 0.388,0.268,0.518,and0.525,respectively.Overall,when establishing LAI inversion models using ground cover spectra and remote sensing images from both empirical and mechanistic model perspectives,the use of mechanistic models such as the PROSAIL model resulted in the smallest errors in estimated LAI,with an MAE of 0.388.The use of the Ada Boost Regressor algorithm to establish models had the strongest fitting ability,with an R~2 of 0.55.By utilizing both ground cover spectra and remote sensing images from both empirical and mechanistic model perspectives and using various methods,this study established a model for the inversion of the Leaf Area Index of the grassland ecosystem,providing methodological references and data support for monitoring the growth of grasslands and serving better management and ecological research of the grassland ecosystem.