Abstract: Most popular unsaturated-saturated water flow models were developed based on solving different forms of Richards' equation. Because these models all have their unique applicability and computational accuracy, it is important to select the most reasonable and efficient model for specified problems. This research used a Bayesian Model Averaging (BMA) method to select the optimal one from six types of unsaturated water flow models (Picard-h model, Picard-θ model, Picard-mix model, Ross model, Kinematic wave model, and water balance model), using real data from field infiltration experiments. The BMA method was first implemented to estimate the model accuracy. Then a modified version of BMA for model selection was proposed to consider the model accuracy and computational cost simultaneously. The results demonstrated that Ross is the optimal model with the best model accuracy and computational efficiency. The modified BMA method will increase the probability of an efficient model and can provide a more comprehensive and reasonable model selection.