ABSTRACT Floods are a significant problem in Sylhet Sadar, Bangladesh, where regular flash flooding, along with other flood types such as riverine flooding, pluvial flooding, and coastal flooding, incurs substantial environmental and socioeconomic costs. This research examines how machine learning and deep learning techniques can enhance the accuracy of flood forecasting. To develop two different model types, a Long Short-Term Memory (LSTM) network was used to improve short-term forecasting accuracy, and the Prophet model was employed to represent long-term seasonal trends. A 30-year dataset (1994–2023) of hydrological and meteorological variables was utilized for a single monitoring site to build these models. The LSTM model achieved strong performance, with R ² = 0.96 and RMSE = 0.65 m for a 1-day lead time, and R ² = 0.85 with RMSE = 1.28 m for a 7-day lead time. The Prophet model also performed well in modelling simultaneously multi-year flood patterns ( R ² = 0.83, RMSE = 1.39 m). Together, these models provide a complementary system of short- and long-term flood prediction. Although all analysis is based on a single gauging station, they demonstrate the potential for integrated forecasting methods to be extended to a multi-station system, thereby developing comprehensive flood forecasting and early warning frameworks in vulnerable regions.
Rana et al. (Mon,) studied this question.