Comparison of neural networks with other ocean colour algorithms and effects of noise on reflectance band weighting

Artificial radiance sets are used as inputs to Multi-layer Perceptron and k NearestNeighbour algorithms to study their retrieval capabilities for optically active constituents in the sea water. The radiative transfer model Hydrolight has been used to produce 18,000 artificial reflectance spectra representing various Case 1 and Case 2 water conditions. The remote sensing reflectances were generated at the MERIS wavebands 412, 442, 490, 510, 560, 620, 665 and 682 nm from randomly generated triplet combinations of chlorophyll a, non-chlorophyllous particles and CDOM concentrations. These data are used to assess the performance of the K-Nearest Neighbour and the Multilayer Perceptron algorithms, which are compared to some more traditional band ratio regression algorithms that have been a popular choice for CZCS and SeaWiFS imagery. Using a novel and computationally quick method involving unit impulse and single perturbation matrices, the spectral channels most important for an accurate computation by the neural network of all optical constituents are singled out. Furthermore a detailed analysis shows at which concentration ranges the two most accurate algorithms from the initial comparison, the neural network and multilinear regression perform with the highest confidence.