Complementary to the reduction of fertilizers and pesticides applications, a promising measure against non-point source pollution from agriculture is the implementation of constructed wetlands intercepting polluted surface water. They are widely used to protect sensitive water resources (e.g. aquifers, lakes or rivers), especially when grass strips are bypassed by a subsurface drainage network. Constructed wetlands, intermittently fed during autumn and spring in a temperate climate, are composed of three interacting compartments: sediment layer, vegetation and water column. Pollutant mitigation increases with the time spent by water in contact with the sediment and with vegetation. Its removal potential is mainly driven by the inflow and its distribution over the wetland area. When sparse or in patches, vegetation makes the flow path more complex and can cause shortcuts and dead zones. The challenge is to link the yearly patterns of discharge, hydraulic performance and removal rate, in order to provide guidelines for the design of constructed wetlands. In this study, the three-dimensional hydrodynamic model Delft3D-Flow was used to simulate flow within a constructed wetland, forced by observed meteorological conditions. Contrary to commonly used bidimensional models, Delft3D-Flow can differentiate flow conditions through immerged and above submerged vegetation. The study site is the constructed wetland (0.5 ha built in 2010) of Rampillon, France, a buffer zone between drained agricultural lands (355 ha subcatchment) and the direct recharge into a karstic aquifer trough sinkholes. The model was validated on continuous outflow concentration measurements and on areal images during two tracing experiments in different flow conditions, using the conservative tracer Rhodamine WT. The hydraulic performance assessment at the yearly scale is based on shortcut and mixing indicators, which are extracted from the distribution of the tracer residence time. The calculation of these indicators at the yearly scale showed that the mixing level does not change significantly and that the shortcuts and dead zones decrease slightly as the flow rate increases.