Inverting a canopy reflectance model using an artificial neural network

Abstract

An off-nadir canopy reflectance model, the Liang and Strahler algorithm for the coupled atmosphere and canopy (CAC) model, was used to simulate multi-angle reflectances based on various combinations of canopy biophysical parameters. Biophysical parameters such as leaf angle distribution and leaf area index were input to the CAC model along with reflectances of leaf, soil, and aerosol optical depth. The CAC model, however, can only be inverted through numerical iterations and it is extremely difficult to use for retrieval of those biophysical parameters with ordinary inversion methods. In order to retrieve those biophysical parameters, we employed an error back-propagation feed forward neural network program. We constructed a number of neural network models based on the simulated results from the CAC model. Ideally, through network training we would like to have a neural network model that uses the multi-angle reflectances as its inputs and output simultaneously all the biophysical parameters, the component reflectances of leaf and background soil, and the aerosol optical depth of the atmosphere. We have not yet reached this objective due to the requirement of an extremely large amount of calculation. In this paper, we report the results obtained from retrieving any individual parameter from multi-angle reflectances and results obtained from simultaneously retrieving some combinations of two parameters. We tested the use of a different number of multi-angle reflectances as input to the neural networks. This number varies in the range of 1 - 64. The test results show that a relative error between 1-5% or better is achievable for retrieving one parameter at a time or two parameters simultaneously.