Abstract: This study investigates the impact of the dimensionality of model input parameters on the prediction accuracy of neural network models using overtopping data of rubble mound breakwaters from the European CLASH project database. For three types of breakwaters with varying structural complexity—simple double-slope breakwaters, double-slope breakwaters with a smooth berm, and complex double-slope breakwaters with a crown wall—three neural network overtopping models (ANN1, ANN2, ANN3) were developed, and their prediction accuracies were compared and analyzed. Based on these models, the random forest method was employed to rank the importance of input parameters (design wave factors, breakwater geometry and position, and armor layer characteristics) on the prediction results. The results indicate that the roughness coefficient γ_f is a primary influencing parameter shared by models ANN1, ANN2, and ANN3. The relative crest freeboard significantly affects model ANN1, while the relative berm width has a notable impact on the results of model ANN2. By comparing the contribution of the roughness coefficient γ_f to the predicted values of different structural types, it was found that different structural types exhibit varying sensitivities to γ_f. A single roughness coefficient value is insufficient to reasonably reflect the influence of changes in the geometric shape of armor blocks, the number of block layers, and other structural characteristics on overtopping.