The estimation of tropospheric gradients in GNSS data processing is a well-known technique to improve positioning precision. To study the correlation between the tropospheric gradients and the topography, we computed Zenith Wet Delay (ZWD) and tropospheric gradients using the GIPSY-OASIS II SOFTWARE from 2 years of GPS observations recorded at 52 worldwide permanent stations, focusing on regions with significant relief. We observe that gradient directions are stable over time and point towards the relief for most of the considered stations. Based on these results, we discuss the physical meaning of the tropospheric horizontal gradients and we investigate why gradients have this particular direction for stations located nearby high mountains. The GPS stations were selected and classified into four main categories: stations close to a mountain range or an isolated mountain (class 1 and 2), stations surrounded by isolated mountains in several directions or in all directions (class 3 and 4). The correlation between the gradient direction and their magnitude with respect to mountain slopes was analysed. A very clear correlation appears for stations of classes 1 and 2 whereas no correlation is obvious for stations of classes 3 and 4. For 89% of stations in classes 1 and 2, a relevant correlation appears, varying between 0.4 and 1. For 64% of stations in classes 1 and 2, a relevant correlation appears, varying between 0.6 and 1. Horizontal gradients estimation show very significant amplitude and a stable direction all along the year, this main direction is most of the time pointing towards the direction of mountains. This behaviour can be explained by a vertical shift of the tropospheric layer due to the presence of mountains, close to the station and up to the maximum distance of 60 km from the station. This orientation does not seem to depend on seasons because no annual or bimonthly means variations appear for all stations. Moreover diurnal variations do not appear on the spatial distribution of the gradients and results are similar for neighbouring stations, separated by few km, which show that local effects such as multipath propagation have influence.