2.5. PROSAIL Model

PROSAIL RTM is a coupled model of PROSPECT leaf optical properties model [24] and SAIL canopy reflectance model [47]. Currently, various versions of the PROSAIL model have been developed, PROSAIL5B, a combination of PROSPECT5 and 4SAIL, was used to simulate vegetation canopy reflectance spectra under different conditions [48,49]. Using the randomly generated input parameters by running the PROSAIL model in forward mode, it is possible to generate vegetation canopy reflectance simulation datasets with spectral intervals of 1 nm in the spectral range of 400–2500 nm, and the simulation datasets generated by the PROSAIL model can contain different vegetation types under different conditions. To simulated sensor noise, 2% random Gaussian noise was added to the canopy reflectance of the simulated dataset. This study mainly analyzed LAI estimation of ZH-1 hyperspectral data. Therefore, according to the spectral response function and the range of each band of ZH-1 hyperspectral data, the simulated datasets were generated by the PROSAIL model. The key input parameters of PROSAIL are summarized by many literatures in Table 2 [22,50,51,52,53,54], where N: structure index, Cab: chlorophyll, Car: carotenoid, Cw: equivalent water thickness, Cm: dry matter per area, LAI: leaf area index, ALIA: average leaf inclination angle, hspot: hot-spot parameter, and psoil: soil brightness factor. The hspot parameter was introduced to correct for the hotspot effect problem of the PROSAIL model, which was generally expressed using the ratio of leaf size to canopy height [21]; the psoil parameter characterized the degree of soil dryness (psoil = 1) and wetness (psoil = 0), and N actually described the internal structure of the leaf.

Ranges and distributions of PROSAIL input parameters for the simulated datasets generation.

Hence, the input parameters of PROSAIL and the spectral response function of the ZH-1 satellite were used to generate the simulated dataset. The reflectance information of the simulated dataset was used as input to the GPR model, and the corresponding LAI parameter was used as output to train the GPR model. A squared exponential kernel function was used as the kernel function of the GPR model and the other parameters were default values. Then, LAI estimation was made by the trained GPR model using correlation bands from ZH-1 satellite data through sensitivity analysis. The simulated datasets, which had size of 5000, were used for training and verification of the GPR, among which 2500 datasets were used as the training dataset of GPR for LAI estimation, and the rest were used as verification datasets for accuracy verification of the trained model.