ASSESSING URBAN FLOOD DYNAMICS USING DATA- DRIVEN AND PHYSICALLY BASED MODELLING FRAMEWORKS

Author

Dewasis Dahal, Southern Illinois University CarbondaleFollow

Date of Award

12-1-2025

Degree Name

Master of Science

Department

Civil Engineering

First Advisor

Kalra, Ajay

Abstract

Urban floods pose a serious threat to public safety, infrastructural resilience, and sustainable urban development due to the rapid changes in land use and the growing consequences of climate change. In order to address this problem, the current study uses a dual-modeling framework to investigate hydrologic simulation and flood prediction using two approaches: (1) machine learning models applied to climate data, and (2) hydrologic models applied to an urban watershed environment using physical models.Four machine learning models—Support Vector Machine (SVM), Random Forest (RF), Artificial Neural Network (ANN), and Long Short-Term Memory (LSTM)—were employed in the first section of the study to forecast flood events in Sacramento, California, by examining past climate variables, such as temperature, precipitation, and soil moisture. The LSTM model performed far better than the others, with an accuracy of 89.9% in flood prediction. It demonstrated outstanding performance in capturing temporal dependencies and nonlinear interactions in sequential climate data. On the other hand, ANN and RF performed quite well but were limited by their static learning architectures, whilst SVM performed worse in classification.The second section evaluates the rainfall-runoff modeling performance of two widely used hydrologic simulation systems, PCSWMM and HEC-HMS, in the densely populated Briar Creek watershed of Charlotte, North Carolina. Although PCSWMM demonstrated better accuracy in capturing peak discharge and runoff volume, both models successfully simulated flood hydrographs using real storm events for model calibration and validation. During calibration, PCSWMM achieved a Nash-Sutcliffe Efficiency (NSE) of 0.949, a Percent Bias (PBIAS) of 1.4%, and an RSR of 0.2, placing it in the "Very Good" performance category. HEC-HMS also performed well (NSE of 0.937), but having a larger negative PBIAS (–12.53%), which indicates some underestimating of runoff volume.These studies show how effective data-driven and process-based approaches are for evaluating urban flooding. The findings also show that machine learning models, especially LSTM, can produce dependable early warning systems when trained on high-quality climate data, even though hydrologic models like PCSWMM and HEC-HMS are still essential tools for modeling watershed behavior and guiding infrastructure design. The comparative results of the study aid in the selection of modeling approaches based on project goals, watershed conditions, and data accessibility, ultimately enabling more sophisticated and adaptable urban flood management strategies.