Binata Roy

While Machine Learning (ML) surrogate models achieve accuracy and transferability in street-scale flood forecasting, they are still unconstrained by the underlying physical processes governing flood routing. To address this limitation, this study explores a Physics-Informed Long Short-Term Memory (PI-LSTM) surrogate model that integrates physics-based prior information into the loss function of the neural network by embedding the mass balance equation, without relying on the calculation of continuous derivatives. The approach was implemented across 22 flood-prone streets in Norfolk, Virginia, where each street segment was treated as a control volume and mass balance was applied in both exact and inequality forms. The study shows that the PI-LSTM exact-based mass balance model outperformed the PI-LSTM inequality-based model. The PI-LSTM Exact model improved overall flood depth forecasting accuracy by 13.9% compared to the LSTM model. The benefits of mass balance enforcement varied across streets, with the majority of streets showing higher accuracy for the PI-LSTM model. Rain-dominated streets with high flood conditions generally show greater PI-LSTM improvement (mean ∆ RMSE 0.006 m vs 0.0 m) than tide-dominated streets, further suggesting the mass balance formulation better captured the flood dynamics of rain-driven flooding than those of tide-driven flooding. Overall, this study aims to improve street-scale flood prediction by providing reliable and physically consistent street inundation forecasts through the integration of physics equations with ML surrogate models for real-time transportation decision-making.

An important challenge with Machine Learning (ML) is its transferability; that is, whether an ML model trained on one set of data can be applied to a second set of data without requiring full retraining 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 an 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 retraining proved most effective and minimal retraining 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 the target data increased beyond 18 flood events, the benefit of TL diminished relative to training an 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.

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.

Climate change is expected to raise river discharge and sea level in the future, and these near-term changes could alter the river flow regime and sedimentation pattern of future floods. Present hazard assessment studies have limitations in considering such morpho-dynamic responses in evaluating flood hazards or risks. Here, we present a multi-model-based approach to quantify such potential hazard parameters influenced by climate change for the most vulnerable communities living on river bars and islands of the Brahmaputra–Jamuna River. River flood-flow and flood wave propagation characteristics are predicted to be affected by changing temporal distribution patterns of precipitation as a result of enhanced global warming. Increased incidences of large multi-peak floods or uncommon floods resulting in long-duration floods driven by sea-level rise may happen as a result of this. To assess it, we have set up a hydromorphic model, Delft3D, for the Brahmaputra–Jamuna River forced by upstream flow, generated from a hydrological model SWAT, over the Brahmaputra basin. The simulations cover moderate, wettest, and driest conditions of the RCP8.5 scenario, and the results reflect the flooding consequences of the near-future, mid-century, and end-century. Floods in the Brahmaputra–Jamuna River are becoming more severe, frequent, and long-lasting, as a result of climate change, and are expected to last until the end of November rather than the current September timeline. While assessing the hazard, we found that the pattern and timing of the flood are as equally important as the peak of the flood, as the river continuously adjusts its cross-sectional area with the flow. The study also demonstrates that, depending on their location/position, climate-induced hazards can affect sand bars/islands disproportionally. The high flood depth, duration, and sedimentation have a significant impact on the sand bars downstream of the river, making them more vulnerable.

Bangladesh is situated at the confluence of GBM basins, with 90% of the basin area locating outside the country. Future climate change will lead to intense, prolonged, and frequent floods in Bangladesh. An integrated flood risk assessment that transforms flood risks from transboundary river basins to the local administrative level is necessary. A 1D/2D hydrodynamic model is developed for flood vulnerable Arial Khan River feed by basin-scale hydrologic model for low (RCP2.6) and high (RCP8.5) climate scenarios. An increasing trend in flood depth, duration, and area is observed from the early (2020s) to the end (2080s) of the century for both scenarios. The difference between both RCPs is minimal from the 2020s to 2050s but becomes very pronounced in the 2080s. The depth-duration-area with equal weightage provides better hazard results for the area. Flood risk is assessed using the IPCC AR5 framework incorporating vulnerability and exposure. Some medium-hazard zones fall into high-risk zones due to high exposure and vulnerability to flooding. The areas along the left reach are found more hazard-prone, while the areas on the right side are more risk-prone in the 2080s of RCP8.5. The hazard/risk maps will help policymakers identify priority areas for planning a sustainable flood management strategy.

Bangladesh is one of the largest flood-prone deltas of the GBM (Ganges–Brahmaputra–Meghna) basins, and recently, it is categorized as the 7th worst climate-affected country in the world. Future climate change along with economic development, urbanization, and increase in population may worsen this situation manifolds. To cope with future flood situations and lessen probable flood losses, it is essential to develop flood maps of the major flood-prone rivers of Bangladesh considering climate change scenarios. In this study, the flood inundation of the Arial Khan River and its floodplain has been assessed for the predicted climate change scenario of RCP 8.5 (Representative Concentration Pathway 8.5) using open-source mathematical models. A calibrated and validated hydrologic model of GBM basins in SWAT (Soil and Water Assessment Tool) model has been used to estimate the future flow magnitudes at Bahadurabad Transit (Brahmaputra River) and Hardinge Bridge (Ganges River) using extreme emission scenario RCP 8.5. Using the flow magnitude of these two stations as the upstream boundaries, an HEC-RAS 1D model has been set up for the Brahmaputra, Ganges, and Padma rivers for generating future flow magnitude at the offtake of the Arial Khan River. Later, an HEC-RAS 1D-2D coupled model is set up for the Arial Khan River floodplain and flood maps are prepared considering flood depth, duration, and inundation extent. The flood assessment for different projections of RCP 8.5 shows that there is an increasing trend of flood in terms of depth, duration, and inundation from the 2020s to the 2080s. Hence, the floodplain becomes more hazardous by the end of this century. The climate change impact on the projected population for the RCP 8.5 scenario is assessed under SSP5 (Shared Socioeconomic Pathways 5) which indicates that the total flood-affected population will be nearly twice in the 2080s compared to the 2020s. So, future climate change is going to have a dreadful effect on the flood situation of the Arial Khan River floodplain.