A novel hybrid fine-tuning method for supercharging deep learning model development for hydrological prediction

Abstract. 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, where a pre-trained LSTM is utilized as an encoder to extract temporal features from the input data. Subsequently, the random forest decoder processes the encoded information to make streamflow predictions. Our method was tested on 421 catchments in the continental United States and 324 in Germany, both selected from two CAMELS datasets. The hybrid method has several benefits. First, it is much more efficient and robust than training LSTMs on each catchment individually (~14x faster). Second, it is much less computationally expensive than LSTM fine-tuning (i.e., feasible on a CPU-based workstation). Third, it achieves superior accuracy compared to a catchment-wise training strategy (e.g., 9.2 % improvement in the median in Nash-Sutcliffe Efficiency (NSE)), shows competitive performance compared to regional LSTM models when trained with fewer data, and through fine-tuning, improves regional LSTM performance in out-of-training samples by 13.13 % (median NSE). To our knowledge, this is the first decision-tree model integrated within a DL workflow to enhance fine-tuning efficiency of pre-trained models in new locations. This hybrid approach holds significant promise for future applications in hydrological modeling, particularly considering the imminent rise of geospatial foundation models in hydrology that will rely on transfer learning techniques for effective deployment.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.