This paper presents a methodology to build representative railway track geometries thanks to a stochastic modelling. This modelling, which has to integrate the statistical and spatial variabilities and dependencies, is a key issue when using simulation for conception, maintenance or certification purposes, as the dynamic behaviour of the trains is mainly induced by the track geometry. The stochastic process theory is used, combining Karhunen-Lo'eve and polynomial expansions. Through a practical example, this paper finally shows to what extent this methodology gives rise to new promising opportunities for the track geometry maintenance.