Predicting river flooding using deep learning models and historical data on temperature, precipitation, and river depth.

نوع: Type: Thesis

مقطع: Segment: masters

عنوان: Title: Predicting river flooding using deep learning models and historical data on temperature, precipitation, and river depth.

ارائه دهنده: Provider: Keyvan Doostkam

اساتید راهنما: Supervisors: Dr. Jalal Sadeghian

اساتید مشاور: Advisory Professors:

اساتید ممتحن یا داور: Examining professors or referees: Dr. Jalal Akbari and Dr. Moharram Mansourizadeh

زمان و تاریخ ارائه: Time and date of presentation: 2026

مکان ارائه: Place of presentation: دانشکده مهندسی

چکیده: Abstract: Accurate and timely flood forecasting, as one of the most destructive hydrological hazards, plays a crucial role in crisis management and damage mitigation. In this study, two independent models based on the CNN–LSTM deep learning architecture were developed in parallel: first, a classification model for estimating the probability of flood occurrence; and second, a multi-output regression model for the simultaneous prediction of water level (stage) and streamflow discharge. The dataset comprised 87,650 hourly records spanning the period from 2013 to 2023. The model input variables included water level, streamflow discharge, and land surface temperature (ERA5-Land) at the San Luis hydrometric station, together with satellite-based precipitation data (GPM IMERG) spatially disaggregated across 21 upstream subregions of the station. To prevent data leakage, the data were split chronologically before scaling, and class weighting was applied during the training of the classification model to address class imbalance. To evaluate performance, 20 temporal scenarios based on different combinations of historical input sequence length (time steps) and forecast horizon (lead time) were examined. The overall performance of the two models was ranked using a Final Score, calculated by combining the models’ mean regression performance—based on the R² values for water level and discharge—with their classification performance, measured by the F1-score. The ranking results showed that the 12–48 scenario achieved the highest overall performance, with a Final Score of 0.9512. However, considering engineering requirements and the need for sufficient warning time, the 36–48 scenario was selected as the preferred scenario. Under this scenario, the R² values for water level and streamflow discharge were 0.9907 and 0.982, respectively, while the classification model achieved an F1-score of 0.9037, resulting in a Final Score of 0.945. Finally, to facilitate the practical application of the developed models, a graphical user interface was designed and implemented for the forecasting system