Reliable Flood Inundation Mapping (FIM) is a key component in the flood forecasting and analysis framework. The selection of FIM prediction solvers is a trade-off between computational and data costs and accuracy. Physically based (hydraulic) models are considered more accurate but computationally and data expensive, while low complexity terrain-based solvers offer efficiency and scalability at the expense of accuracy. The NOAA Office of Water Predictions (OWP) operational FIM framework is based on the Height Above Nearest Drainage (HAND) approach. In this study, a deep Convolutional neural network (CNN) technique is explored to enhance the OWP HAND-FIM using a surrogate modeling approach. Among various architectures, an encoder-decoder U-Net architecture performs well to classify the flooded area precisely. U-Net's skip connections help to maintain high-resolution spatial information by combining characteristics from multiple Convolutional layers, which improves prediction accuracy, whereas skip connections help reduce vanishing gradient concerns, which improves the model's generalization capabilities. The model is trained using physics-based HEC-RAS simulations and a suite of other input features (digital elevation model, topographical wetness index, land use, land cover, flow accumulations, etc.) over three HUC8-scale watersheds. The results show a considerable improvement in FIM accuracy when using a surrogate CNN model compared to the original OWP HAND-FIM. The transferability of the model is tested in a fourth HUC8 region, showing considerable enhancement in the OWP HAND-FIM after the trained surrogate model is applied. The enhanced FIM was also evaluated using building footprints, resulting in a significant improvement in the surrogate model impact detection accuracy. This demonstrates that CNN holds a generalizing capability by capturing the spatial correlation between pixels. Introducing physical laws to attain behaviors during modeling and increasing the size and diversity of its training data is expected to further enhance the model's transferability and predicting capability.