Use of neural networks for the retrieval of water parameters from hyperspectral data

Some indicators of water quality parameters such as chlorophyll-a, suspended sediments and dissolved organic matter, alter, depending on their concentration, the optical properties of water, influencing, in this way, what is the signal spectrum measured by a remote sensor. Multi-spectral sensors like MERIS and MODIS are often used to estimate the 'bio-optical' parameters in oceanic and coastal waters. Nevertheless, it is difficult, using multi-spectral sensors, to estimate accurately the concentrations of water constituents, because of the wide bandwidth and the few number of bands of these sensors. With the availability of hyper-spectral data able to distinguish significant features of the spectral signature of water mass, it became possible to estimate much more accurately the concentrations of chlorophyll-a, suspended sediment and organic matter dissolved in the water. However, two types of problems have to be managed for the design of an effective retrieval procedure. On the one hand it has to be taken into account that the inverse problem maybe rather complex due to the high nonlinearities characterizing it. On the other hand, the high dimensionality of the hyperspectral measurements suggests that appropriate feature extraction techniques have to be considered in a pre-processing stage. Neural networks can be a possible solution for both the aforementioned issues. Indeed neural algorithms, besides having been proven a suitable approach for discovering subtle nonlinearities in multi-dimensional data, can also be considered for feature extraction problems. In this latter case a particular architecture called Autoassociative Neural Networks can be used. The AANN topology includes an internal “bottleneck” layer and is trained in order that the input layer is approximated at the output layer. If the training phase is satisfactorily completed, the bottleneck nodes must represent or encode the information in the inputs for the subsequent layers. In this work, the potential of neural networks algorithms for the retrieval of water quality parameters from hyperspectral data has been investigated. A simulated dataset based on 6S – (Second Simulation of a Satellite Signal in the Solar Spectrum) radiative transfer model has been generated. The data were used for the design of a double stage neural architecture where in the first stage a features reduction algorithm has been performed by means of AANN, in the second stage the reduced measurement vector has been used as input to an inversion NN scheme for the quantitative retrieval of water quality parameters. The data set has been split into a training and a testing set. The first one has been exploited for determining the networks adaptive parameters while the second one for evaluating their performance. A final validation based on experimental Hyperion data has been also considered.