Monitoring and modeling the dynamics of terrestrial freshwater resources are vital for human beings to cope with the growing global water crisis (Carpenter et al., 2011). Most terrestrial freshwater resources, excluding those in Antarctica and Greenland, are stored as groundwater, which is difficult to sense directly by using conventional satellite-based optical remote sensing (Alsdorf et al., 2007; Wood et al., 2011; Taylor et al., 2012; Gleeson et al., 2015). Although the groundwater table in the wells can be measured, it is difficult to estimate water storage dynamics at the scale of a large basin or continent based on the spatially sparse water table records (Yeh et al., 2006; Voss et al., 2013). Land surface models (LSMs) and global hydrological studies and water resource models (GHWMs) provide alternative ways to understand land water storage and its dynamics on a global basis (Wood et al., 2011; Mo et al., 2016). However, problems can arise, especially over areas in which human activity is intense, as a result of many uncertainties in the experimental parameters, input data and index, and initial conditions of the model (Scanlon et al., 2018). Therefore, the development of alternative methods that can be implemented to monitor and model terrestrial water storage (TWS) is of great significance.

The Gravity Recovery and Climate Experiment (GRACE) satellites, which were launched in 2002, revolutionized the observation and understanding of terrestrial hydro-systems (Rodell, 2004; Yeh et al., 2006; Feng et al., 2013; Save et al., 2016; Scanlon et al., 2016; Wiese et al., 2016; Andrew R. et al., 2017; Scanlon et al., 2018). Compared with global models, GRACE solutions provide a more direct means of measuring variations in global water storage because the satellites measure changes in the Earth’s gravity field by quantifying alterations in the relative distance between the twin GRACE satellites that occur as a result of gravitational changes. After removing noise from the results, such movements can be considered quite relevant to changes in the amount of water stored at a particular place (Save et al., 2016; Sean Swenson, 2006; Watkins et al., 2015). Variations in the TWS that are measured by the GRACE satellites result from the influences of both climatic and human interventions. Therefore, the GRACE solutions provide reliable information about the variation in TWS on a global basis, and the system has been successfully applied in drought mapping (Leblanc et al., 2009), surface water monitoring (Rodell et al., 2009; Huang et al., 2012; Feng et al., 2013), and water balance modeling (Rodell, 2004; Rodell et al., 2011).

Given the higher reliability of this method as compared to LSMs and GHWMs, the results of GRACE total water storage anomalies (TWSA) have the potential for improving models describing TWS dynamics. The assimilation or fusion of the GRACE solutions with land surface models or global reanalysis systems has garnered increasing amounts of attention in recent years (Kumar et al., 2016; Li et al., 2019). The empirical GRACE observations can be integrated using global model outputs by linking TWS-related variables with the TWSA dataset produced by GRACE. Such models can be simplified to linear or polynomial functions that occur between the TWSA and related variables under ideal conditions, and successful experiments have been conducted for the Amazon Basin using this method (Nie et al., 2015; Humphrey et al., 2017). These models are certainly based on a fundamental hypothesis that the mismatches in TWSA and land surface model simulations can be calibrated using typical regression models; however, this assumption is only valid for specific regions and in certain situations. Machine learning is a powerful tool for fitting relations. As reported in previous study (Long et al., 2014), artificial neural networks (ANNs) can produce GRACE TWSA-like predictions for southwestern China that date back to the 1980s by learning the matching patterns between the available GRACE TWSA and independent indicators such as precipitation and soil moisture. Similar experiments were conducted over northwestern China by comparing three machine learning approaches (support vector machine, artificial neural networks, and random forest (RF)) and the general linear regression model (Yang et al., 2018). The empirical results indicated that the RF model outperformed the others and that the linear regression model was not applicable in this area. A recent study has also proven that ensemble learning algorithms perform well when learning the matching patterns between GRACE TWSA and climate forcings (Jing et al., 2020).

However, few experiments have been conducted in areas undergoing intensive irrigation or industrial activity, which are the major human interventions that lead to a decrease in TWS (Feng et al., 2013; Voss et al., 2013). Human withdrawal can be one of the most significant factors that lead to errors in land surface models for heavily irrigated areas (Joodaki et al., 2014; Pokhrel et al., 2017; Tangdamrongsub et al., 2018). In addition, the tools and models that have been initiated for extending the time span covered by GRACE rely heavily on the soil moisture produced by global models or reanalysis systems. This limits the application of the proposed models because they are restricted by the availability of LSMs or the output from reanalysis systems for specific regions and periods. More importantly, it is impossible to separately quantify the contributions of climate and artificial factors with these models due to the dependence on soil moisture variables and the ignorance surrounding human intervention.

This study proposes a machine learning model that can learn the underlying patterns connecting TWS dynamics with variations in the climate and human water withdrawal. The model has been tested in the Huang-Huai-Hai River Basin of China, which is a heavily irrigated area with intensive human activity and is suffering from a water scarcity crisis. A GRACE-consistent TWSA was generated for the study area back to the 1980s and was compared with LSMs/GHWMs and in-situ soil moisture measurements. The significance of the contribution from each factor to the TWSA estimation model is estimated. This study attempts to provide new perspectives for the high-quality modeling of TWS dynamics by using a machine learning framework over areas that suffer from severe human-induced water scarcity. It is highly expected that flexible machine learning tools have great potentials to enrich complex physical models and extend our understanding of TWS modeling for areas such as that studied herein.

The research area covers the Huang-Huai-Hai River Basin in China, which comprises three major river basins: the Yellow River (or Huanghe), Huaihe River, and Haihe River basins. It is located between 95°53′–122°42′E and 30°57′–42°43′N. The areas of the three basins are 752,000 km² (Yellow River Basin), 274,000 km² (Huaihe River Basin), and 318,000 km² (Haihe River Basin). Dominated by low plains and suitable temperatures, the Huang-Huai-Hai River Basin seems abundant as a large block of well cultivated land and is considered one of the most nourished agricultural zones in China. The Huang-Huai-Hai River basin has long been the political and cultural center of China. The cultivated land, population, and gross domestic product (GDP) of this basin account for more than one-third of the national totals, while the amount of water available accounts for approximately 7% of the total (Yan et al., 2013; Yuan et al., 2019). The terrain and climate in the basin changes gradually from west to east; it includes a plateau with a mountain climate, hills with a continental climate, and temperate plains that are affected by monsoon.

This study proposed a machine learning model for estimating historical variations in the TWS in the Huang-Huai-Hai River Basin. Random forest (RF) and Extreme Gradient Boost (XGB) models were used to train a model that linked GRACE TWSA observations with climate data and human withdrawal. The past GRACE-consistent TWSA was reconstructed by applying this model. The reconstructed TWSA results are compared with the global land surface area and hydrological models, as well as the in situ soil moisture measurements. In addition, the contribution of every variable to the machine learning tool/model was quantified.

In general, the presented approach reproduces spatial patterns and temporal variations in the GRACE TWSA, thereby outperforming a set of global models. The inclusion of human intervention improved the performance of the machine learning model. This is rational because groundwater has been overexploited since the 1970s in the plains of North China, and human activities have profound effects on the changes in the TWS of this region (Kendy et al., 2004; Shi et al., 2011; Cao et al., 2013; Huang et al., 2015; Min et al., 2015). The catchment and Noah models, however, cannot consider human intervention and underestimate the decrease in TWS relative to the GRACE TWSA produced for the research area. These findings are in line with previous research (Mo et al., 2016; Scanlon et al., 2018). The catchment model includes groundwater simulation, enriching the terrestrial water storage components compared with the Noah model. However, the groundwater in the catchment is not directly modeled, and the vertical distribution equilibrium of the soil moisture comprises an implicit water table that is located at the depth of the equilibrium saturation (Kumar et al., 2016). Therefore, there are still many differences between the TWS simulations performed using the catchment model and the GRACE solution.

The PCR-GLOBWB model comprises all the terrestrial water storage models, and because human activities are considered, the TWSA from PCR-GLOBWB presents higher correlations with GRACE TWSA at the basin level. However, the declining trend in the TWS of the Haihe River Basin was overestimated by PCR-GLOBWB relative to the GRACE solution. A similar overestimation of the decline in TWS, which was produced by the PCR-GLOBWB model in this region, was also revealed in another study (Feng et al., 2018). The PCR-GLOBWB model without human intervention underestimated the decrease in the TWS of the Haihe River Basin, indicating that the overestimation of the decline in the TWS by PCR-GLOBWB is a result of the overestimation of human water use in the basin. The machine-learning model used the same calculations that were used for describing water withdrawals in the PCR-GLOBWB model and produced consistent results with the GRACE solution. This indicates that the pattern of climatic and human withdrawals produced by the GRACE TWSA has been successfully calibrated and predicted from the machine learning tool for both the present and past records.

The contribution importance estimations provide an intuitive perspective from which to interpret the ensemble learning models used to estimate TWS changes in this study. Precipitation, acknowledged as the major climatic driving force of variations in the TWS in previous studies (Li et al., 2017; Meng et al., 2019; Xie et al., 2019), is the climatic variable with the highest significance value in the present model. The use of water by three sectors (irrigation, domestic, and industry) has equally important contributions in the model, though not livestock. Some individual variables are more important in the models. In theory, the models consider that only the most important variables would produce comparable results with models that include a full set of variables. For this reason, the RF model is used to select important features describing specific issues (Ham et al., 2005; Chen et al., 2014; Ma et al., 2017). Thus, additional experiments are required in the future to investigate the performance of the models with variables of different sizes.

The variable importance rankings derived from the RF and XGB models are generally similar, except that the human factors and geolocation obtained higher importance values in the XGB model than in the RF model. A possible explanation for this is that the geolocation and human factor contribute more to the residuals of the base learners because the simple learners recursively fit the residuals of their predecessors in the boosting-based ensemble learning model (Freund and Schapire, 1996; Friedman, 2002).

Not every tree in the RF and XGB models includes all the characteristics or observations, which guarantees that the trees are decorrelated and therefore less prone to overfitting. However, correlated features will be issued equal or similar importance, which may reduce the significance compared to the same trees built without their correlated counterparts. In addition, it should be mentioned that the significance value is merely a statistical relative score that does not indicate the authentic contribution of a variable to the TWS, although the RF and XGB models are produced by random selection. The reliability of the results in terms of ranking importance relies on the performance of a model and the selection of the variables used. The notable variables used in this study probably do not cover all types of human intervention or the hydrological factors affecting the TWS dynamics within the headwater of the Yellow River, the Loess Plateau, and the surrounding areas, such as the Three Rivers Source Region Reserve, the Grain for Green project, reservoir operation, and the contribution from snowmelt in the surrounding areas (Jin et al., 2017; Yi et al., 2017; Deng et al., 2018; Lv et al., 2019; Meng et al., 2019; Xie et al., 2019). Therefore, the variable importance results should be explained with additional caution.

The era of big data and artificial intelligence is accelerating the development of data-driven Earth system science. Many machine learning techniques have provided tools and exciting new opportunities for accurately predicting the evolution of water cycles and expanding our understanding of the Earth system from multisource data. We can learn from the results of this study that machine learning and artificial intelligence have the potential to elucidate much more from data than traditional land surface and hydrology models can. However, respecting nature’s physical law is important for expanding our knowledge of the Earth by using machine learning. Therefore, the following key challenge is to develop a hybrid model for coupling physical process models and machine learning approaches.

Because the model input data is not exactly the same as the ensemble learning model, in addition, the GLDAS model also contains input data from other sources, such as land use, soil type and texture, which can cause uncertainty in the GLDAS model. The biases caused by these uncertainties cannot be quantitatively analyzed and discussed in this paper, and the comparison results of GLDAS land surface models need to be further studied.

In summary, the models proposed in this study address how both climate and human factors impact the dynamics of TWS while performing as well as or better than a set of global models in terms of describing areas with intense human activities. Instead of modifying the physical models, the machine learning model is a more flexible, and less expensive alternative for directly reconstructing past changes in the total TWS. The findings of this study enrich our conceptions of the changes and driving forces in the TWS over the Huang-Huai-Hai River Basin and have great potential to be assimilated into hydrological models to improve TWS simulations in the future.