| Publication Abstract: | One of the important challenges with Machine Learning (ML) is its transferability; that is, models trained for one region and their ability (or inability) to predict for another region without being completely trained for that region. Transfer Learning (TL) addresses this challenge by transferring the learned knowledge from one domain (Source) to another domain (Target). First, the weights of the pre-trained model act as learned knowledge and are used as the initial status of the transfer model. Then, the remaining weights in the transfer model are re-trained using the data available from the target domain. TL is commonly used when developing an ML model from scratch is complicated due to having less data in the target domain. TL has already been highly recognized in the artificial intelligence community but less explored in the hydrology domain, especially in an urban setting for street-scale flood forecasting, which has challenges due to the complexity of the built urban landscape and the highly variable, rapid and localized nature of urban hydrological processes. Hence, the research question of the study is - Can an ML-based surrogate model trained in one region effectively predict street-scale flooding in another region, where it was not originally trained, through applying different TL strategies? For this research, an ML-based surrogate model has been developed for the flood-prone Norfolk City of Virginia, using the environmental features (rainfall, tide), topographic features (Digital Elevation Model (DEM), Topographic Wetness Index (TWI), Depth To Water (DTW) index) and flood depth from a high-fidelity physics-based model. A set of flood-prone street segments was used to pre-train the base model in the source domain, while another set of flood-prone street segments was used to re-train that model (termed as the transfer model) in the target domain. The TL technique has been applied in two senses: weight initialization (i.e. pre-training the base model at the source domain) and weight freezing (i.e., preventing some weights from further changes during re-training the transfer model at the target domain). The preliminary results demonstrate that TL improves forecasting accuracy for street-scale flooding in urban areas, particularly when training data is limited at the target domain, by leveraging a well-trained model from the source domain. |
