Neural network for retrieval of concentrations of water constituents with the possibility of detecting exceptional out of scope spectra

For the ground-segment of MERIS a retrieval procedure based on neural network (NN) technology was developed. The NN not only transforms water leaving radiance reflectances pixel by pixel with high efficiency into concentrations of water constituents but also checks if its input is in the domain which was covered during the training of the NN. Since measurements do not cover the data space with sufficient density, the construction of the NN is based on a large table generated by a Monte-Carlo radiative transfer code. Two NNs are trained with this table: (1) invNN to emulate the inverse model (reflectance and geometry → concentrations), and (2) forwNN to emulate the forward model (concentrations and geometry → reflectance). The two NNs together with a comparison network are combined to give a new NN which first uses the invNN part to obtain an estimate of the concentrations. These are fed into the forwNN part and the derived reflectances are compared with the measured reflectances by the NN component cmpNN. Large deviations signal a violation of the necessary condition for a successful inversion; corresponding pixels are then flagged. The present scheme is demonstrated using simulated MERIS data