Runoff Prediction in a Data Scarce Region Based on Few-Shot Learning

In this study, we propose a new method (metric-LSTM fusion model) for runoff prediction in data scarce regions by combining few-shot learning and deep learning. To validate the effectiveness of the proposed model, taking the upper Yellow River basin as a case study, we compare its performance with that of four state-of-the-art data driven models (LSTM, SVR, RF, ANN) on monthly runoff prediction during 1970 to 1995. Results indicate that the proposed model performs well with the best NSE (Nash-Sutcliffe model efficiency coefficient) of 0.83, outperforming all the comparative models. Furthermore, the less the data used for model training, the more obvious the advance of the proposed model. Findings imply that the proposed model based on few-shot learning can provide an effective tool for runoff prediction in data-scarce regions.