This paper presents a new rainfall estimation method, EPSAT-SG which is a frame for method design. The first implementation has been carried out to meet the requirement of the AMMA database on a West African domain. The rainfall estimation relies on two intermediate products: a rainfall probability and a rainfall potential intensity. The first one is computed from MSG/SEVIRI by a feed forward neural network. First evaluation results show better properties than direct precipitation intensity assessment by geostationary satellite infra-red sensors. The second product can be interpreted as a conditional rainfall intensity and, in the described implementation, it is extracted from GPCP-1dd. Various implementation options are discussed and comparison of this embedded product with 3B42 estimates demonstrates the importance of properly managing the temporal discontinuity. The resulting accumulated rainfall field can be presented as a GPCP downscaling. A validation based on ground data supplied by AGRHYMET (Niamey) indicates that the estimation error has been reduced in this process. The described method could be easily adapted to other geographical area and operational environment.