NNs emulating inverse models a practitioner's view

Helmut Schiller

Institute for Coastal Research, Germany

Abstract:

The Medium Resolution Imaging Spectrometer (MERIS) is a sensoronboard the Envisat satellite to be launched in October 2001. The primary objective of MERIS is support for research in biological oceanography. For the MERIS ground segment, GKSS has developed an algorithm for the retrieval of water constituent concentrations in coastal waters from MERIS spectra. The algorithm is based on neural networks (NNs) emulating an inversion of the radiation transfer calculations.

Because NN's are good interpolators but bad extrapolators, an important step was inclusion of an error model of the sensor signals in the NN design. For the inversion to be used operationally a scope check also had to be included. This is done using a NN emulating the forward model. The importance of error model inclusion as well as of the domain check is demonstrated using a low dimensional toy problem.