Long-term streamflow forecasting in data-scarce regions: Insightful investigation for leveraging satellite-derived data, Informer architecture, and concurrent fine-tuning transfer learning

Abstract

Accurate multistep forecasting of the long-term streamflow (Qflow) in poorly gauged basins is pivotal for sustainable water resource management and decision-making. The purpose of this study is to improve the performance of Qflow predictions in data-scarce regions. To achieve this, we employed geo-spatiotemporal satellite-derived datasets, advanced Informer architecture, and transfer learning methodologies. The deep-transfer-learning-based Informer model enables the perception of time dependencies and is specifically tailored to facilitate multistep-ahead Qflow prediction. Initial investigations identified correlations between historical Qflow records and geo-spatiotemporal satellite metrics, and this helped select grids that encapsulated the highest correlation between Qflow and climatic parameters. Subsequent assessments contrasted three main dimensionality reduction techniques—principal component analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), and multi-dimensional scaling (MDS)—to craft deep-learning-amenable endogenous inputs. The introduction of the current fine-tuning transfer learning (CFTL) methodology, augmented with incremental learning, which was evaluated against conventional transfer learning (CTL) across two distinct target domains with varying similarities to the source domain, was central to this study. A holistic comparison involving a range of univariate multistep-ahead prediction models including DNN, CNN, LSTM, AR-LSTM, and Transformer, was also undertaken. In addition, multivariate models, notably CFTL and CTL, were set against the two former models under different data availability scenarios for each target domain. The findings revealed that the MDS-Informer model significantly outperformed the t-SNE-Informer and PCA-Informer models, showing an improvement of approximately 23 % and 40 % in terms of MAE, respectively. Moreover, the results confirmed that the CFTL approach effectively enhanced long-term Qflow prediction accuracy compared to conventional transfer learning (CTL) and the limited data scenario (LD) without TL by approximately 30 and 47 % improvement in terms of MAE, respectively. This study demonstrates that a target domain possessing a higher similarity index with the source domain is more likely to facilitate the transfer of positive noise, thereby enhancing the predictability of the model. This phenomenon underscores the importance of domain similarity in the efficacy of TL approaches.