This study proposes a method for predicting the long-term temporal two-dimensional range and depth of flooding in all grid points by using a convolutional neural network (CNN). The deep learning model was trained using a large rainfall dataset obtained from actual flooding events, and the corresponding raster flood data computed using a physical model. Various rainfall distributions (at different times or over different accumulation periods), the mesh of the simulated area, and the topography of the simulated area were considered when evaluating the performance of two CNNs: a simple CNN and Inception CNN. Neither CNN architecture could converge when the coordinate information was not included in the input data. Adding terrain elevation information to the rainfall data already containing coordinates increased the accuracy of flood prediction. Our findings indicated that in the proposed method, real-time flooding observation data are not required for corrections, and we concluded that the method can be used for long-term flood forecasting. Our model can accurately pinpoint when the water level changes from rising to falling. Once meteorological forecasted rainfall data are obtained, a corresponding long-term forecast of the two-dimensional flooding range and depth can be obtained within seconds.

In this study, the prediction interval method was used in simple regression models to filter continuous river surface velocity microwave radar data. To evaluate the model performance, two data sets from monitoring stations with mild and steep channel slopes were used. A human–machine interface software program developed in LabVIEW was used to sample data from big continuous data for establishing the relationships between surface velocity and water level, two surface velocities, and their prediction intervals. Filtering by coupled relationships detected the most noise in the surface velocity and the original data, and the results for different cases were compared. The results were also compared with widely used modern smoothing methods. It was found that raw data cannot always be post–processed using these smoothing methods. Moreover, peaks become distorted. This study provides a method for filtering noise signals in continuous river surface velocity data without data contamination, which makes the surface velocity data more reliable and applicable for advanced studies, such as machine learning applications, and can be applied for the quality control of surface velocity data in the future.