CNN-Based Flood Forecasting Model: Case Study in Thung Song District, Thailand

Flooding is a significant natural disaster that poses serious risks to lives and property. Accurate flood forecasting is essential to provide advance notice to residents in flood-prone areas, enabling them to prepare for evacuation. Most traditional methods for flood forecasting are based on local rainfall accumulation and/or water levels in nearby rivers. However, these methods provide only a few hours up to at most a day of lead time or their required measurements may not be available in practice. These limitations highlight the need for advanced forecasting techniques that can offer longer lead times and require input that is always available. In this study, we developed a Convolutional Neural Network (CNN) model that classifies flood events into binary categories: flood and no flood, at least two days in advance, leveraging forecasted total precipitation data from the European Centre for Medium-Range Weather Forecasts (ECMWF) found in The International Grand Global Ensemble (TIGGE) dataset. The model was trained and tested using historical data of Thung Song District from 2009 to 2022. The model’s performance was evaluated at two levels: event (or pulse) level and sample (or day) level. At the event level, it achieved 100% Pulse Recall (PR) and 100% Pulse Precision (PP) on the testing dataset. At the sample level, it achieved 99.86% Accuracy and 92.86% F1-Score on the same dataset. These results indicate that the proposed model is promising, with the potential to enhance early warning capabilities and support the decision-making processes of relevant organizations.