The past decade has witnessed astonishing growth in the volume and diversity of data for water resource research, owing to the increasing number of installed observation sensor systems (McCabe et al., 2017) and the latest crowdsourcing and opportunistic sensing technologies (Jiang et al., 2019). A new challenge arising in this context is extracting information and knowledge from the data deluge (Reichstein et al., 2019). The concept of data mining, which aims to elicit implicit relationships and structures hidden within variables through modern data-driven science (Mount et al., 2016), has been used for over two decades in the field of hydrology (Abbott, 1999). Machine learning (ML), a major branch of artificial intelligence (AI), offers practical tools for data mining in hydrologic analysis (Babovic, 2005). Numerous studies have examined the roles of various ML models, such as genetic programming (Babovic & Keijzer, 2000), classification and regression trees (Fang & Shen, 2017), and self-organizing maps (Ley et al., 2011), in providing insights into hydrological processes. However, a significant tradeoff between model accuracy and interpretability persists for early generation ML models (Murdoch et al., 2019).Deep learning (DL) is a cutting-edge AI paradigm whose superiority in automatically capturing abstract information hidden in data has been acknowledged across different fields (LeCun et al., 2015;Reichstein et al., 2019). To date, one of the most successful DL architectures for modeling dynamic hydrological variables is long short-term memory (LSTM) networks (Feng et al., 2020;Kratzert et al., 2018), whose recurrent structure and unique gating mechanism can accurately capture nonlinear and temporal dependencies between variables (Hochreiter & Schmidhuber, 1997). While substantial efforts have been made to demonstrate and enhance the predictive power of LSTM networks in hydrology, few studies have illustrated how to gain scientific insights from the model (Shen, 2018). This is largely due to the accuracy-interpretability dilemma (Florez-Lopez & Ramon-Jeronimo, 2015), which is