Application of an artificial neural network in canopy scattering inversion

Abstract

Owing to their recent success in other inversion tasks, application of an artificial neural network to the development of an inversion algorithm for radar scattering from vegetation canopies is considered. Because canopy scattering models are complicated functions of the desired biophysical parameters (vegetation biomass, leaf area index, soil moisture content, etc.), the development of an effective inversion algorithm is not a straightforward task. The Michigan Microwave Canopy Scattering (MIMICS) model, which has shown remarkable success in predicting the radar response to vegetation canopies, was used, as were measured polarimetric backscatter values. Hence, the radiative transfer simulation code, MIMICS, was used to produce some of the training data. The inputs to the neural network were the expected polarimetric backscatter values from specific canopies, while the outputs were the desired parameters, such as tree heights, crown thickness, leaf density, etc.

Two special cases were examined: (1) inversion of MIMICS given modelled aspen stands of different ages; (2) inversion of measured data from the Duke forest loblolly pine stands. The MIMICS inversion shows that neural networks are capable of accurately inverting some of the parameters of such a complicated model. The implication is that once MIMICS is made to model the radar data for a specific application, then inversion of the radar data may be accomplished. The measured data inversion shows that, even without a model such as MIMICS, one may train a neural network to invert several parameters of interest. However, this depends on accurate and complete surveys of the ground truth data to be useful.