Explore Spatio‐Temporal Learning of Large Sample Hydrology Using Graph Neural Networks

Abstract

Streamflow is the spatial integral of basin input, storage, and runoff processes. It provides a direct measure of water availability and the impact of climate change and human activities at the basin level (Milly et al., 2005;Hannah et al., 2011). Quantifying streamflow regimes and their spatiotemporal similarity can reveal valuable insights on the controlling hydroclimatic, geomorphic, and ecological processes, improving water resources planning and management (Coopersmith et al., 2012;Ghotbi et al., 2020). Current practices in hydrological modeling are either depth driven or breadth driven (Gupta et al., 2014). The former category seeks to improve process understanding at the catchment scale, aiming to provide detailed and physically realistic catchment models for a specific basin through in situ sampling and long-term monitoring (Blöschl et al., 2019); while the latter focuses on the development and application of hydrological models at the continental and global scales, seeking to provide a global assessment of water scarcity (

Differentiable, Learnable, Regionalized Process‐Based Models With Multiphysical Outputs can Approach State‐Of‐The‐Art Hydrologic Prediction Accuracy

“…A particularly popular architecture in hydrology is long short‐term memory (LSTM) (Hochreiter & Schmidhuber, 1997). LSTM's accuracy has been demonstrated for many hydrologic variables on large data sets including soil moisture (Fang & Shen, 2020; Fang et al., 2017, 2019; Liu et al., 2022), streamflow (Feng et al., 2020; Konapala et al., 2020; Kratzert, Klotz, Shalev, et al., 2019; Sun et al., 2021; Xiang & Demir, 2020; Xiang et al., 2020), dissolved oxygen (Zhi et al., 2021), groundwater (Solgi et al., 2021; Wunsch et al., 2021), and water temperature (Rahmani, Lawson, et al., 2021; Rahmani, Shen, et al., 2021), covering every part of the hydrologic cycle (Shen et al., 2021). It is widely publicized that LSTM represented a “step‐change” in performance which also suggests our traditional models were far from optimality (Nearing et al., 2021).…”

Stream Temperature Prediction in a Shifting Environment: Explaining the Influence of Deep Learning Architecture

“…Although the importance of spatial awareness is well recognized in fields such as traffic prediction (Li et al., 2018; Wu et al., 2019), it is much less commonly implemented in deep learning hydrology with a few notable exceptions (Jia et al., 2021; A. Y. Sun et al., 2021, 2022). This lack of spatial focus contrasts a substantial effort to capture temporal dependencies.…”

Machine Learning for Bayesian Experimental Design in the Subsurface

“…As noted by Sun et al (2021), ML has been used in hydrology for decades (e.g., Coulibaly et al 2000;Dawson and Wilby 2001;lin Hsu et al 1995;Maier et al 2010;Sun 2013;Zealand et al 1999). Furthermore, Sun et al (2021) emphasize that the current wave of hydrological ML applications has greatly benefited from more accessible cyber-infrastructures and a new breed of deep learning algorithms, bolstered by the exponential growth of Earth observation data (Peters-Lidard et al, 2017;Shen, 2018;Sun and Scanlon, 2019).…”

“…As noted by Sun et al (2021), ML has been used in hydrology for decades (e.g., Coulibaly et al 2000;Dawson and Wilby 2001;lin Hsu et al 1995;Maier et al 2010;Sun 2013;Zealand et al 1999). Furthermore, Sun et al (2021) emphasize that the current wave of hydrological ML applications has greatly benefited from more accessible cyber-infrastructures and a new breed of deep learning algorithms, bolstered by the exponential growth of Earth observation data (Peters-Lidard et al, 2017;Shen, 2018;Sun and Scanlon, 2019). These developments allow for the importation of priors or domain knowledge into ML models, the exportation of knowledge from learned models back to the scientific domain, the generation of a vast amount of synthetic data, the quantification and analysis of uncertainty in models and data, and the inference of causal relationships from the data (Lavin et al, 2021).…”