Although rainfall estimates are a key environmental parameter, they are not always available at an appropriate scale and with a sufficient accuracy. Satellite based rainfall estimation methods supply valuable information and their enhancement is a key issue. In this paper we describe a rainfall estimation method based on MSG, the new European meteorological geostationary satellite. Computations have been carried out on West-Africa during 2004 rainy seasons. Our algorithm is composed of two parts: assessment of rainfall probability and actual rainfall estimation. Rainfall probability is computed by comparing MSG derived indicators with TRMM/PR rain detection. After the selection of input nodes, a feed forward neural network is trained. The network coefficients being assessed, rainfall durations can be computed. The second part of the algorithm merge rain probability data with actual rainfall measurement. This part is highly dependent of available rainfall data. If precipitations are supplied as gridded data, a potential intensity is first computed and then precipitations are downscaled at intitial satellite resolution. On 2000 data, resulting rainfall estimates have been tested against GPCP on a dense raingauges network and have demonstrated better performances.