Hydrological processes that occur in catchments usually require large space resolution over long periods of time. The advance on numerical methods as well as the increasing power of computation are making possible the physically based simulation of these phenomena. In particular, the 2D shallow water equations can be used to provide distributions of water depth and velocity fields. The necessity of spatial resolution involves the use of a large number of elements hence increasing the computational time when simulating realistic scenarios for a long time period. This work deals with an efficient GPU implementation of the 2D shallow water equations on unstructured meshes analysing the influence of the mesh resolution both on the computational performance and the quality of the results to simulate a rainfall/runoff event. The numerical method to solve them has been developed and compared following three programming approaches: the sequential implementation and its adaptation to the multi-thread and many-core architectures. The particular detail of the influence of the mesh ordering when using unstructured triangular meshes is paid attention in this work to find the best strategy to further reduce the computational time in the context of GPU simulation. The resulting approach is efficient and can become very useful in environmental simulation of hydrological processes.