A new few-shot learning model for runoff prediction: Demonstration in two data scarce regions

Most existing hydrologic models and machine learning models failed to perform well on runoff prediction in data scarce regions. As an alternative to this, the Long Short-Term Memory (LSTM)-prototypical network fusion model based on few-shot learning is proposed, where the strong learning ability of LSTM and the low data dependence of prototypical network are combined. The proposed model was calibrated and implemented on monthly runoff prediction in the Lancang River basin (LRB) and the source region of the Yellow River basin (SRYRB). Compared with eight state-of-the-art data driven models (LSTM, SVR, ANN, ARMA, Random Forest, SimpleRNN, GRU, and BiLSTM), the proposed model outperformed especially when less training data were used. Results in the LRB indicate NSE of the proposed model achieved 0.802 and 0.832 when the proportion of training data (K) was 20% and 45%, improved by 0.527 and 0.222 relative to the mean NSE of other models, respectively. In the SRYRB, NSE reached 0.830 and improved by 0.354 when K was 40%. The findings imply that the new few-shot learning model provides a promising tool for runoff prediction in the two investigated basins and possibly other data -scarce basins where precipitation dominates runoff change, which will benefit regional water resources man-agement and water security.