The introduction of deep learning (DL) (LeCun et al., 2015) into hydrology around 2016–2018 (Tao et al., 2016; Laloy et al., 2017, 2018; Shen, 2018; Shen et al., 2018), especially the use of long short-term memory (LSTM) as a dynamical modeling tool for soil moisture and streamflow (Fang et al., 2017; Kratzert et al., 2019), has ignited a surge in machine learning applications across all domains of hydrology. At the core, machine learning is a set of tools that allow us to build and train models that extract and reproduce the spatial and temporal patterns in the datasets they encounter. In particular, the central philosophy of DL has been to minimize the intervention of the human experts in feature design and to facilitate maximal extraction of information from data (Goodfellow et al., 2016). Improved prediction quality in hydrologic machine learning (ML) models has been achieved not by infusing process-based assumptions into the models, but by conducting extensive training of the models with large quantities of a priori data. It has been argued by Nearing et al. (2020) that there could be significantly more information in large-scale hydrological data sets than hydrologists have been able to translate into theory or process-based models. The hydrology community is poised to fully explore the power in the vast amount of data using machine learning in various subdomains of hydrology.
|Number of pages||4|
|Journal||Frontiers in Water|
|State||Published - 11 May 2021|