Author(s): Nobuaki Kimura; Daichi Baba
Linked Author(s):
Keywords: CNN; Transfer learning; Flood forecast; Time-series data; Insufficient dataset
Abstract: A deep neural network (DNN) model has recently been employed in a riverine flood forecast system. The accurate DNN model requires lots of data for machine learning. However, large data for flood events are hardly obtained in field measurements because abundant manpower and expensive sensor installation and maintenance are required in current research. Additionally, only a few flood events are normally recorded per year even if long-term measurement is successfully conducted. To improve data limitation, we introduced transfer learning, which pre-trains a model in a source domain and then reuses it in a different domain. This method can be beneficial when implementing the DNN model to the flood forecast at the location where insufficient data is available. The transfer learning with time-series DNN model potentially has not been successful. Instead, we focused on a convolutional neural network (CNN) model, a deep learning architecture and an image analysis tool, because it has numerous examples of image analyses with a transfer learning approach. Our previous study implemented the CNN model-based transfer learning to time-series flood prediction. The model pre-trained in large-size flood events in a source watershed was retrained with a small size data in a target watershed. This study performed the extension and improvement of the model with transfer learning in the previous study. An accuracy improvement of the model was achieved with a change of training datasets in a validation method. The model prediction was able to be extended till the 3h-lead time with reduced or equivalent errors when compared with those of no-pre-trained model.
Year: 2020