Artificial neural network inversion of tree canopy parameters in the presence of diversity

Artificial neural networks have shown themselves to be able to invert many parameters of interest from multi-frequency and multi-polarization radar data when the tree canopies exhibit low variability in their parameters for a given age (Pierce, et. al., 1992). However, when this same model is applied to a forest stand whose parameters are slightly off the modeled age curve, then the inversion fails. This paper describes the authors´ efforts to develop an inversion technique for Loblolly pine stands with the natural diversity of their canopy parameters accounted for. Given a set of Loblolly pine stands from the Duke forest a set of parameters that sufficiently model the stands using MIMICS (Ulaby, et al., 1988) were developed. Next, several important parameters were chosen to be varied over a specified interval with a Gaussian distribution to allow for natural variation. This generated a large data set of 1000 different Loblolly pine canopies. MIMICS was then employed to generate the expected radar backscatter from this set at P, L, and C bands. This data, was then used to train two artificial neural networks: one to invert for the average trunk diameter and subsequently another to invert for the large branch density. This cascaded approach worked well giving errors of only a few percent for trunk diameters and 5% for branch density, when using data with trunk diameters varying between 1 cm and 10 cm. Beyond 10 cm the radar signal begins to saturate due to the high biomass, and was not invertible