A Fusion of Deep Learning and Time Series Regression for Flood Forecasting: An Application to the Ratnapura Area Based on the Kalu River Basin in Sri Lanka

Flooding is the most frequent natural hazard that accompanies hardships for millions of civilians and substantial economic losses. In Sri Lanka, fluvial floods cause the highest damage to lives and properties. Ratnapura, which is in the Kalu River Basin, is the area most vulnerable to frequent flood events in Sri Lanka due to inherent weather patterns and its geographical location. However, flood-related studies conducted based on the Kalu River Basin and its most vulnerable cities are given minimal attention by researchers. Therefore, it is crucial to develop a robust and reliable dynamic flood forecasting system to issue accurate and timely early flood warnings to vulnerable victims. Modeling the water level at the initial stage and then classifying the results of this into pre-defined flood risk levels facilitates more accurate forecasts for upcoming susceptibilities, since direct flood classification often produces less accurate predictions due to the heavily imbalanced nature of the data. Thus, this study introduces a novel hybrid model that combines a deep leaning technique with a traditional Linear Regression model to first forecast water levels and then detect rare but destructive flood events (i.e., major and critical floods) with high accuracy, from 1 to 3 days ahead. Initially, the water level of the Kalu River at Ratnapura was forecasted 1 to 3 days ahead by employing a Vanilla Bi-LSTM model. Similarly to water level modeling, rainfall at the same location was forecasted 1 to 3 days ahead by applying another Bi-LSTM model. To further improve the forecasting accuracy of the water level, the forecasted water level at day t was combined with the forecasted rainfall for the same day by applying a Time Series Regression model, thereby resulting in a hybrid model. This improvement is imperative mainly because the water level forecasts obtained for a longer lead time may change with the real-time appearance of heavy rainfall. Nevertheless, this important phenomenon has often been neglected in past studies related to modeling water levels. The performances of the models were compared by examining their ability to accurately forecast flood risks, especially at critical levels. The combined model with Bi-LSTM and Time Series Regression outperformed the single Vanilla Bi-LSTM model by forecasting actionable flood events (minor and critical) occurring in the testing period with accuracies of 80%, 80%, and 100% for 1- to 3-day-ahead forecasting, respectively. Moreover, overall, the results evidenced lower RMSE and MAE values (<0.4 m MSL) for three-days-ahead water level forecasts. Therefore, this enhanced approach enables more trustworthy, impact-based flood forecasting for the Rathnapura area in the Kalu River Basin. The same modeling approach could be applied to obtain flood risk levels caused by rivers across the globe.