Abstract: The influences of input parameter dimension on prediction accuracy of the neural network model are investigated based on the data of wave overtopping at sloping breakwaters of block revetment from the European CLASH Project Database. For the selected three types of breakwaters with varying structural complexity, which are simple double-slope breakwaters, double-slope breakwaters with smooth berm and complex double-slope breakwaters with crown wall, three neural network（ANN） wave overtopping models (ANN1, ANN2, ANN3) are established and their prediction accuracies are compared and analyzed. For these models, the random forest method is used to rank the importance of the input parameters such as design wave factors, breakwater geometry and position and revetment characteristics. The results indicate that the roughness coefficient γ_f is a primary influencing parameter shared by the three models ANN1, ANN2 and ANN3. The relative crest freeboard significantly affects model ANN1, while the relative berm width has a notable impact on model ANN2. By comparing the contribution of roughness coefficient γ_f to predicted values of different structural types, it is found that different structural types exhibit various sensitivities to γ_f. A single roughness coefficient value is not enough to reflect reasonably the influences of the changes in geometric shape of revetment blocks, number of block layers and other structural characteristics on the wave overtopping.