This paper proposes a new interpolation method based on Kohonen self-organizing networks. This method performs very well, combining an accuracy comparable with usual optimal methods (kriging) with a shorter computing time, and is especially efficient when a great amount of data is available. Under some hypothesis similar to those used for kriging, unbiasness and optimality of neural interpolation can be demonstrated. A real world problem is finally considered: building a map of surface-temperature climatology in the Mediterranean Sea. This example emphasizes the abilities of the method.