The tsunami impact assessment should be done rapidly and accurately to ensure adequate emergency risk responses and management so the city can recover swiftly. Significant progress has been made in numerical analysis techniques to yield more precise assessments and predictions of tsunami forces. However, this achievement comes with an expensive computational cost, which might hinder their applicability for uncertainty quantification and probabilistic risk analysis. Reduced-order modeling is a numerical approach that attempts to find a model in lower-dimension space but still manages to represent the physical features of the full-order model with acceptable accuracy. Proper orthogonal decomposition is one of the reduced-order modeling methods that has been used in many engineering fields. This study proposed a framework to estimate tsunami propagation and land inundation using a surrogate-based reduced-order modeling method, neural network method is used as the surrogate model. Two error measures are used to analyze its performance: the root mean square error and the average relative error of projection. We found that, the root mean square error are decreasing when we use more bases. However, the average relative error of projection in the online phase does not significantly reduce when the number of bases is increased. Suggesting that moderate number of bases, between five to 10 bases, seem sufficient for the proposed framework.