Support vector machines regression for retrieval of leaf area index from multiangle imaging spectroradiometer

The retrieval of biophysical variables using canopy reflectance models is hindered by the fact that the inverse problem is ill posed. This is due to the measurement, model errors and the inadequacy between the model and reality, which produces similar reflectances for the different combination of the input parameters into the radiative transfer model. This leads to unstable and often inaccurate inversion results. The ill-posed nature of the inverse problem requires some regularization. Regularization means that one tries to consider only those solutions that are in the proximity of the true value. In order to regularize the model inversion, we propose kernel-based regularization by support vector machines regression (SVR) method. rmulation of the SVR contains meta-parameters C (regularization) and ε-insensitive loss. The SVR generalization performance (estimation accuracy) depends on these two parameters and the kernel parameters. Often the meta-parameters are selected using prior knowledge and/or user expertise. In this paper we adopt methods for the estimation of the meta-parameters from the input data itself instead of relying on any prior information. This paper is focused on the retrieval of leaf area index (LAI) from multiangle imaging spectroradiometer (MISR) data. The proposed methodology was implemented by inverting a 1D canopy reflectance model (PROSAIL) using SVR over MISR data. The results were validated against the LAI retrievals at the Alpilles EOS validation core site. An RMSE of 0.64 was obtained using both near-infrared (NIR) in conjunction with the red band and an RMSE of 0.50 using only the NIR band.