Estimation of the diffuse attenuation coefficient Kd(lambda) with a neural network inversion

The fine-scale study of the diffuse attenuation coefficient, K_d(λ), of the spectral solar downward irradiance is only feasible by ocean color remote sensing. Several empirical and semi-analytical methods exist. However, most of these models are generally applicable for clear open ocean waters and estimate this coefficient only at 490 nm. They show limitations when applied to coastal waters. A new empirical method based on neural networks has been developed using a relationship between the remote-sensing reflectances between 412 and 670 nm and K_d(λ) between 412 and 490 nm, for the SeaWiFS ocean color remote sensor. The first results concern the estimation of Kd(490). The architecture of the neural network has been defined using synthetical and in situ dataset The comparison with the SeaWiFS empirical algorithms shows similar retrievals accuracies for low values of K_d(490) (i.e. <;0.20 m^-1) and better estimates for greater values of K_d(490). The new model is suitable for open water but also for turbid waters and does not show the limitations of the empirical method.