Estimation of crop LAI using hyperspectral vegetation indices and a hybrid inversion method

Leaf area index (LAI) is an important indicator of crop growth. In this paper, a hybrid inversion method was developed to estimate the LAI values of crops. Based on PROSAIL simulation datasets, 43 hyperspectral vegetation indices (VIs), including the optimized soil-adjusted vegetation index (OSVAI) and modified triangular vegetation index (MTVI2), were analyzed to identify optimal VIs for estimating LAI values. Hybrid inversion models were then built to determine the relationships between optimal simulated VIs and the LAI values (generated by the PROSAIL model) using modeling methods, including curve fitting, artificial neural network (ANN) and random forest regression (RFR) algorithms. Finally, remote sensing mapping of a Compact High Resolution Imaging Spectrometer (CHRIS) image was completed using the inversion model to verify the LAI estimation accuracy. The remote sensing mapping of the CHRIS image yielded an accuracy of R2 = 0.928 and RMSE = 0.485 for OSAVI and R2 = 0.910 and RMSE = 0.554 for MTVI2, demonstrating the feasibility of high-accuracy estimation of crop LAI using hyperspectral VIs and a hybrid inversion method. The estimation results of various VIs suggested that the identification of the appropriate VIs is critical to improve the inversion accuracy. In addition, to obtain the appropriate VIs, the factors must be evaluated with respect to two aspects, i.e., the sensitivity to target parameters and the insensitivity to interference. In this study, OSVAI and MTVI2 were sensitive to LAI and relatively insensitive to the effects of interference factors, such as chlorophyll, soil background, sky scattered light and observed geometry. Therefore, these indices could be primarily used as VIs for an LAI estimation. The inversion results of different datasets demonstrated that prior information is critical for improving the inversion accuracy and identifying the optimal VIs. Additionally, based on the comparison of the curve fitting, ANN, and RFR algorithms, RFR was an optimal method for modeling in this study, as indicated by the higher R2 and lower RMSE values for different datasets and various VIs.