The retrieval of ocean constituents from satellite ocean color measurements is very sensitive to the atmospheric correction. Improved atmospheric correction algorithms, which simultaneously estimate ocean and aerosol optical properties, have recently been developed. These new methods perform well in the presence of absorbing aerosols. However their cost in computer-time prevents to use them within operational ocean color data processing. Advanced programming techniques such as Neural Network (NN) can help us in implementing improved atmospheric correction algorithms in operational satellite data processing. The method we propose is decomposed in two parts. First, the information in the near infrared (NIR) wavelengths is used to provide a first estimation of the aerosol optical parameters. A direct inversion using NN estimates these parameters with a quite good accuracy. These values are used as a first guess in the neuro-variationnal inversion dealing with the visible wavelengths. To perform this inversion, we have modeled the oceanic and atmospheric radiative transfer equations with respect to the aerosol and oceanic parameters (the direct model) by using Neural Networks. The accuracy is better than 5% in term of root-mean-square error. The variationnal method consists in inverting the direct model to retrieve the aerosol and oceanic parameters. We minimize the distance between the observed reflectance at the top of the atmosphere and this calculated by the NN model, these parameters being the control variables. The computation of the gradient of the cost function is straightforward due to the proprieties of the NN. We invert the global image in one time which allows us to take into account the spatial coherence of the parameters . Results for the Mediterranean Sea and for the east coast of the USA are presented.