Machine Learning Ensembles Can Enhance Hydrologic Predictions and Uncertainty Quantification

Abstract Accurate and robust predictions of hydrologic variables are essential for water resource management. In the past decade, significant progress has been made in using machine learning (ML) models for large‐scale, spatiotemporal predictions of stream flows and other hydrologic quantities. Yet, despite the rich history of ensemble use in hydrology, many ML studies rely on deterministic models using single architectures or simplistic ensembles, which limits their accuracy and uncertainty quantification (UQ). In this study, we demonstrate the utility of ensemble modeling using predictions of stream temperatures in unmonitored basins of the contiguous United States as a use case. We evaluate six different ensemble construction techniques spanning classical and deep learning (DL)‐specific methods to determine optimal modeling strategies. The analysis uses four ML architectures: the long short‐term memory (LSTM), gated recurrent unit (GRU), temporal convolution network (TCN), and extreme gradient boosting (XGB). Ensemble accuracy and spread are quantified using multiple deterministic and probabilistic metrics. Our results clearly demonstrate the universal benefit of using ensembles to improve accuracy compared to deterministic models. Ensemble construction strategies such as varying the input data across models or combining model architectures are most effective in improving predictions for both average and extreme values. Notably, we find XGB has the highest accuracy, but lower spread relative to DL architectures. This study underscores the importance of using ensemble strategies for hydrologic predictions. We conclude that considering diverse ensembles of optimal sizes and probabilistic metrics of performance can enhance the accuracy and reliability of ML models.