Abstract:

Accurate estimation of leaf area index （LAI） plays an important role in ecological， environmental and climate change research. Large-scale LAI estimates can be obtained from satellite remote sensing technology， but they rely on a large amount of ground-measured data with and they have low spatial resolution， which often does not meet the needs of high-precision and large-scale research. In this study， using surface reflectance data with a spatial resolution of 30 m， we tested an inversion method combining the Kernel Ridge Regression （KRR） algorithm and the PROSAIL physical model to invert LAI without a large number of ground measured data. First， the sensitivity analysis was performed on the input parameters of the PROSAIL model to determine the input parameters and generate the simulated data sets. Then， the KRR model inversion between the simulated reflectance and LAI was established. For comparison， we linked two other models， the Multilayer Perceptron （MLP） algorithm and the Random Forest Regression （RFR） algorithm， with the PROSAIL model， to perform high spatial resolution LAI inversion. Finally， we used ground measured data to compare the outputs and performance of the three inversion models. We found that the LAI inversion accuracy of the KRR-PROSAIL model was the highest with an R^{\mathrm{2}} of 0.8089 and root-mean-square error （RMSE） of 0.2492. The inversion accuracies of the PROSAIL model linked with MLP and RFR were inferior with R² values of 0.7726 and 0.7118， respectively and RMSE values of 0.2781 and 0.2432， respectively. Based on this study we recommend the combination of the Kernel Ridge Regression algorithm and PROSAIL models to invert satellite data to LAI for improved accuracy and high spatial resolution of the inverted LAI data. This methodology provides a method for rapid and accurate inversion of regional high-precision LAI information.

Key words: leaf area index, Kernel Ridge Regression algorithm, PROSAIL model, inversion