Abstract: The aim of current research was to determine the most efficient model for water purification, by studying different combinations of plants and matrixes and related enzyme activities (urease and phosphatase) in small wetland models. Two types of plants: Zizania latifolia and Sagittaria graminea, and two types of matrixes: garnet and magnetite were used to construct six different wetland models. Within a wetland system, the quality of inflow and outflow water and the matrix enzyme activities were measured. The results indicated the wetland models have the highest removal rate of nitrite nitrogen, ranging between 98.87% and 58.06%, and the lowest removal rate of COD_Mn, ranging between 37.91% and 2.51%. Compared with the single plantation model, combination of both plant species yielded a synergy effect in improving water quality, evident by the increased average removal rates of total nitrogen (by 3%), ammonia nitrogen (by 10%), nitrite (10%), total phosphorus (15%), orthophosphate (20%), and COD_Mn (8%). This effect was independent of the use of matrix. The filter material urease activity was highly correlated with the total nitrogen removal rate (r²: 0.903—0.980), whereas phosphatase activity was highly correlated with the COD_Mn removal rate (r²: 0.821—0.992). Hence, urease and phosphatase can be used as predictors of the total content of nitrogen and COD_Mn, respectively, in a wetland system. This research provides a theoretical basis in optimal selection of wetland plants and matrixes and in evaluation of water purification effect by assessing substrate enzyme activities.