This study proposes a novel hybrid method that substantially accelerates and improves deep learning (DL) model development for streamflow prediction. The method leverages a combination of a long short-term memory (LSTM) network and random forests. A hybrid encoder-decoder model is designed using a pre-trained LSTM to extract temporal features, followed by a random forest decoder for streamflow prediction. Our method was tested on 421 catchments in the continental United States and 324 in Germany. The hybrid method has several benefits. First, it shows competitive performance compared to regional LSTM models (e.g., 6.66% improvement in the median in Nash-Sutcliffe Efficiency (NSE)) when trained with less data. Second, it is more efficient and robust than training LSTMs on each catchment individually (∼14x faster) and achieves superior accuracy, and third, it is much less computationally expensive than regular fine-tuning (i.e., feasible on a CPU-based workstation). • This study presents a hybrid long short-term memory–random forest model. • The hybrid model can be deployed as an efficient and accurate fine-tuning method. • The hybrid model is effective in improving the performance of regional deep learning models.
Jahangir et al. (Wed,) studied this question.