River water-level prediction is crucial for mitigating flood damage caused by torrential rainfall. In this paper, we attempt to predict river water levels using a deep learning model based on radar rainfall data instead of data from upstream hydrological stations. A prediction model incorporating a two-dimensional convolutional neural network (2D-CNN) and long short-term memory (LSTM) is constructed to exploit geographical and temporal features of radar rainfall data, and a transfer learning method using a newly defined flow–distance matrix is presented. The results of our evaluation of the Oyodo River basin in Japan show that the presented transfer learning model using radar rainfall instead of upstream measurements has a good prediction accuracy in the case of torrential rain, with a Nash–Sutcliffe efficiency (NSE) value of 0.86 and a Kling–Gupta efficiency (KGE) of 0.83 for 6-h-ahead forecast for the top-four peak water-level height cases, which is comparable to the conventional model using upstream measurements (NSE = 0.84 and KGE = 0.83). It is also confirmed that the transfer learning model maintains its performance even when the amount of training data for the prediction site is reduced; values of NSE = 0.82 and KGE = 0.82 were achieved when reducing the training torrential-rain-period data from 12 to 3 periods (with 105 periods of data from other rivers for transfer learning). The results demonstrate that radar rainfall data and a few torrential rain measurements at the prediction location potentially enable us to predict river water levels even if hydrological stations have not been installed at the prediction location.

In our previous work, we proposed a river water-level prediction method using deep learning, incorporating radar rainfall data in place of water-level and rainfall stations upstream of the prediction point. By introducing a newly defined flow distance matrix, transfer learning becomes available, i.e., even when data at the prediction point is scarce, accurate water-level predictions are made using inundation data from other rivers. However, this approach requires pre-selecting rivers that behave similarly to the prediction point for training, making it laborious to build prediction models for multiple rivers. Furthermore, the previous study only performed predictions for a single river, raising uncertainty about whether the method is applicable to water-level prediction for other rivers with different conditions. In this paper, we construct a generalized river water-level prediction model commonly applicable to multiple Japanese rivers by using inundation data from all Japanese Class-A rivers (the major river systems managed by the government) for pre-training, rather than only the rivers similar to the prediction site. Through evaluation, we showed that pre-training using all Class-A rivers yields higher prediction accuracy than pre-training using similar rivers across multiple rivers with varying conditions. This demonstrates that using all Class-A rivers for pre-training enables the construction of a generalized river water-level prediction model applicable to a wide range of rivers.