Radial Basis Function and Multilayer Perceptron neural networks for sea water optically active parameter estimation in case II waters: a comparison

This paper deals with the problem of retrieving optically active parameters of the water from multispectral remotely sensed data. We analyse the neural networks approach applied to the estimation of chlorophyll concentration in coastal waters (Case II Waters) and discuss the use of two types of networks: the Radial Basis Function neural network and Multilayer Perceptron. We present a brief summary concerning their architectures and training methods. For proving the concept we analyse the procedures and the performances on a simulated data set reproducing the data acquired from the MERIS (Medium Resolution Imaging Spectrometer), the multispectral sensor on board the ENVISAT satellite. The multispectral subsurface reflectance data have been generated by means of a three component ocean colour direct model and statistically reproduce the case II waters. The neural networks performances have been analysed in terms of MSE (Mean Square Error), correlation coefficient and relative error. We provide a detailed discussion and comparison of the two types of networks and the obtained results confirm the effectiveness of the neural approach in such an application.