Author(s): Ryosuke Arai; Yasushi Toyoda; So Kazama
Linked Author(s): So Kazama
Keywords: Absent of discharge data; Observed streamflow; Cross validation
Abstract: Japan has a diversity of climatic division and one of the heaviest snowfalls in the world. To estimate streamflow characteristics throughout Japan is one of the greatest challenges for hydrologists. Recently, artificial neural networks (ANNs) gained an attention as an approach to estimate streamflow (Q) characteristics in absent of discharge data. We developed ANNs to estimate the Q characteristics inputting enormous basin characteristics throughout Japan. The Q characteristics were obtained from observed discharge data in 448 target basins. We employed the 14 Q characteristics including mean annual runoff height and flow percentiles in flow duration curves. The 175 basin characteristics including climate, land use, geology, soil, and topography were used as input data of the network. The network performed the best in mean annual runoff height (R2 = 0.72) and the worst in 99 percentile of flow duration curves (R2 = 0.18). We also evaluated the relationship between Q characteristics and basin characteristics. The results showed that the Q characteristics were explained dominantly by precipitation and aridity index. We found a limited geological effect on low flow regime, whose effect may have been weakened by extreme snowmelt contributions. We validated a relationship between the number of training data and the performance of ANNs. The results showed that increasing training data had a possibility for improvement of the accuracy, especially in low flow regimes. This will be a hope to improve the performance in low flow regimes.