Convolutional Neural Network Coupled with a Transfer-Learning Approach for Time-Series Flood Predictions

• Published in: Water (Basel) Vol. 12; no. 1; p. 96
• Main Authors: Kimura, Nobuaki, Yoshinaga, Ikuo, Sekijima, Kenji, Azechi, Issaku, Baba, Daichi
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.01.2020
• Subjects: Analysis; Classification; cnn; Computer applications; Datasets; Domains; Evaluation; Flood control; Flood forecasting; flood prediction; Flood predictions; Floods; Image classification; Japan; Neural networks; Rivers; Time series; time-series datasets; Transfer learning; Trends; Two dimensional models; Water levels; Watersheds; Japan
• ISSN: 2073-4441; 2073-4441
• DOI: 10.3390/w12010096

East Asian regions in the North Pacific have recently experienced severe riverine flood disasters. State-of-the-art neural networks are currently utilized as a quick-response flood model. Neural networks typically require ample time in the training process because of the use of numerous datasets. To reduce the computational costs, we introduced a transfer-learning approach to a neural-network-based flood model. For a concept of transfer leaning, once the model is pretrained in a source domain with large datasets, it can be reused in other target domains. After retraining parts of the model with the target domain datasets, the training time can be reduced due to reuse. A convolutional neural network (CNN) was employed because the CNN with transfer learning has numerous successful applications in two-dimensional image classification. However, our flood model predicts time-series variables (e.g., water level). The CNN with transfer learning requires a conversion tool from time-series datasets to image datasets in preprocessing. First, the CNN time-series classification was verified in the source domain with less than 10% errors for the variation in water level. Second, the CNN with transfer learning in the target domain efficiently reduced the training time by 1/5 of and a mean error difference by 15% of those obtained by the CNN without transfer learning, respectively. Our method can provide another novel flood model in addition to physical-based models.