Coastal urban flooding presents significant challenges due to the complex interactions between surface runoff, storm tides, and drainage systems. In highly impervious urban areas, infiltration may seem negligible, particularly during extreme events when soil saturation is rapidly exceeded. However, for nuisance flooding-becoming more frequent due to sea level rise and shifting rainfall patterns-the role of infiltration warrants closer examination. This study aims to assess the impact of incorporating infiltration processes into an urban flood model by evaluating flood extent under 1-, 10-, 50-, and 100-years return periods. Specifically, we performed our test case over an urban coastal watershed in Norfolk, Virginia using a detailed 2D/1D hydrodynamic model. For the evaluation of the model, we incorporated community-sourced inputs from residents and drivers collected through data portal application and a mobile navigation platform. The results show that infiltration losses reduced the flooded areas by 33 % during the 1-year return period and by 12 %, 9 %, and 5 % during 10-, 50-, and 100-years return periods respectively. In terms of flood volume, infiltration led to a 29 % reduction for the 1-year event, while reductions were smaller for extreme events due to rapid soil saturation. We also found significant variability in water depth, with for example a nearly 20 cm difference between full and zero saturation conditions during the nuisance flooding event on a selected street in the Park Place neighborhood. While less intense nuisance storms did not achieve full soil saturation, thereby reducing flood impacts, extreme events rapidly overwhelmed infiltration capacity, leading to increased surface runoff and flood inundation. Finally, the study showed how crowdsourced flood reports can help to validate urban flood models. In our case, the model accurately predicted flood locations with 88 % of community reports and 89 % of driver reports of flooding aligning with simulated inundation maps.

This study explores the use of Deep Convolutional Neural Network (DCNN) for semantic segmentation of flood images. Imagery datasets of urban flooding were used to train two DCNN-based models, and camera images were used to test the application of the models with real-world data. Validation results show that both models extracted flood extent with a mean F1-score over 0.9. The factors that affected the performance included still water surface with specular reflection, wet road surface, and low illumination. In testing, reduced visibility during a storm and raindrops on surveillance cameras were major problems that affected the segmentation of flood extent. High-definition web cameras can be an alternative tool with the models trained on the data it collected. In conclusion, DCNN-based models can extract flood extent from camera images of urban flooding. The challenges with using these models on real-world data identified through this research present opportunities for future research.

Deep learning surrogate models trained with mean-squared-error loss produce statistically accurate but physically unconstrained flood predictions: water may flow uphill, appear spontaneously, or smooth over street-level corridors. In this work, a physics-informed training framework is developed for CNN-LSTM models that predict urban flood depths at 15 min intervals over a $128\times 128$ spatial grid. Three differentiable penalty terms are embedded directly into the loss function: (i) a gravity loss that penalizes depth increases against the water-surface-elevation gradient, (ii) a continuity loss enforcing local mass conservation with rainfall-adaptive thresholds, and (iii) a topography-aware false-alarm penalty modulated by the topographic wetness index (TWI). The framework is evaluated on the Norfolk, Virginia, flood dataset spanning two major storm events (August 2017 and September 2022) comprising 300 samples, with all variants trained on identical splits and robustness assessed over repeated random splits and leave-one-storm-out tests. A road-proximal evaluation restricted to a TWI-derived street mask quantifies street-level skill. The physics-constrained model achieves near-zero gravity violations (∼10⁻⁶) and the highest street-channel recall (0.77 ± 0.09 versus 0.44 ± 0.10 for the unconstrained baseline), the capability most relevant to downstream traffic routing, and its recall advantage more than doubles on a held-out storm, while a uniform false-alarm variant attains 16% lower mean absolute error but suppresses street recall to 0.25. The proposed TWI-modulated penalty reconciles this trade-off: it improves upon the uniform variant on every metric measured, recovering 60% higher street recall at the lowest MAE among all constrained variants and the best street-level F1 score. These results expose a fundamental tension between aggregate pixel-level error metrics and application-specific physical plausibility, and demonstrate that terrain-aware loss modulation offers a principled resolution.

An important challenge with Machine Learning (ML) is its transferability; i.e., whether a ML model trained on one set of data can be applied to a second set of data without requiring full re-training of the model. Transfer Learning (TL) addresses this challenge by transferring knowledge learned in the source domain (the data it was trained on) to the target domain (a second set of data that is statistically different but related, which the model was not trained on). This study investigates the use of TL for street-scale nuisance flood forecasting by exploring whether a ML model trained for one set of streets can effectively forecast flooding for another set of streets in the same city. TL is explored using a Long Short-Term Memory (LSTM) model trained on data for the flood-prone streets of Norfolk City, Virginia. The results show that full-weight re-training proved most effective and minimal re-training of only the output layer was insufficient. The advantage of TL was most pronounced when target data was limited, meaning data collected at the new water depth sensor location included generally less than 18 flood events. As target data increased beyond 18 flood events, the benefit of TL diminished relative to training a ML model directly on the local flood events. These findings can assist cities as they implement street-scale flood sensing systems to create accurate forecasts for new sensing locations that do not yet have sufficient data records to train a local ML model.