Data Driven Approaches for Water Level Prediction



Journal Title

Journal ISSN

Volume Title



Floods are a hazardous disaster which threaten the United States and its territories as they affect millions of people every year. In this context, having suitable approaches for forecasting flooding will benefit people by reducing property damage and saving lives, by warning citizens of dangerous flooding events in advance. Consequently, many cities have developed sensor-based early warning systems which report water levels in real-time. The goal of this thesis is to take advantage of the data collected by such systems to develop data-driven water level prediction techniques. In past research, physics-based hydrology models have been developed, such as the National Water Model, which predict water levels by simulating the rise and fall of water. The purpose of this research is to generate alternative, complementary, data-driven water-level forecasting models using existing statistical models and recurrent neural networks which extrapolate the past into the future. We investigate various time series forecasting approaches, in particular: Vector Autoregressive (VAR) and Long Short-Term Memory (LSTM) Networks. The investigated forecasting techniques are applied and evaluated using USGS datasets. Moreover, we analyze the role of soil moisture — a less explored parameter — in flood incidence and conduct some experiments that explore the relationship between rainfall and soil moisture. Finally, we develop a web application called FloodNet, a real time water level prediction system, with a forecast horizon of 2 hours for a location along Buffalo Bayou in Houston, Texas.



Flood forecasting, Deep learning, LSTM, Buffalo Bayou