During the last three decades, electronic devices have conquered the railway domain, taking the place previously held by electromechanical devices, thanks to higher performances and lower costs. The price of this "revolution" is the fact that, in order to work properly and reliably, electronic systems must be fairly immune to the effect of external interferers, while, at the same time, they are not to electromagnetically pollute the environment they work in. These issues are dealt with by electromagnetic Compatibility (EMC) whereas several international EMC standardization committees work on the definition of tests and rules the manufacturers must comply with. In the European Union, the reference for EMC issues in the railway domain is set by the CENELEC standard EN 50121, which deals with several aspects of a generic railway system, from the power-supply infrastructure to rolling stocks and signalling circuits. The introduction of this standard in 1996 has had a strong impact on rolling stock manufacturers, who are now required to test their products for EMC compliancy. As opposed to the automotive domain, the testing of trains cannot be performed in standard facilities, such as anechoic chambers, so that they have to be tested on actual railway lines, typically on the customer's. Industrial experience has shown that results obtained in this way are usually site-dependent, something that is against the very idea of a standard. The aim of this work is to prove the importance of the infrastructure in radiated emission tests, showing that the test results are site-dependent, thus subject to misunderstandings and misinterpretations. To this end, the features of a generic railway system are briefly described, pointing out the great variability in actual configurations, together with the absence of standard solutions. Subsequently, the electromagnetic modelling of a railway system is introduced, dealing with both propagation and radiation phenomena; in particular, the main topic here addressed is the modelling of supply-lines, through a quasi-TEM approach. The finite conductivity of the soil is taken into account by means of a closed-form formulation, thus avoiding numerical methods, and overcoming the limitations of Carson's model. Moreover, special attention is paid to discontinuities that would increase the model complexity, proposing approximated descriptions supported by numerical results. Results obtained with this model are then validated through several measurement campaigns carried out on actual railway lines, proving the effectiveness of the approach here pursued. The model is then employed in order to prove that some criteria in the standard EN 50121, specifically introduced in order to avoid site-dependency, are not realistic, thus leaving this issue unresolved. To this end, numerical examples are considered, assessing the impact of the infrastructure by comparing results obtained with realistic site configurations and with the ideal one envisaged by the standard. These comparisons are at the base of a tentative procedure that would allow to avoid the misinterpretations that triggered this work. Unfortunately, this approach requires an accurate description of the test-site. Since this is hardly the case, an alternative experimental characterization of the site is proposed, based on magnetic field measurements. This approach, involving the solution of an inverse problem, is shown to be feasible through a numerical validation, though its practical utilization requires efficient optimization techniques.