The Liao River Basin in China is characterized by an arid and semi-arid climate [1
]. Rainfall in the area mainly occurs in July or August [2
]. The Liao River Basin, affected by the alternating effects of warm and moist air in the Southeast Pacific and cold air from the west or north, is prone to rainstorms [5
]. Rainstorms account for a large proportion of annual precipitation, and cause frequent flooding in the basin. Moreover, the river basin suffers from different degrees of drought intensity nearly each year; spring drought is particularly severe [6
]. These hazards harm the national economy, human life, and property. The frequency of hydro-meteorological extreme events has increased substantially due to global warming and high-intensity human activity, such as deforestation and unsustainable use of water resources [1
]. Hence, a need is created to monitor and/or predict occurrences of drought and flood timely and effectively.
Total water storage change (TWSC) integrates the change of water storage in the vertical direction and includes variations in groundwater storage (GWS), soil moisture storage (SMS), snow water equivalent (SWE), and biomass water content. TWSC is derived from the Gravity Recovery and Climate Experiment (GRACE) total water storage anomalies (TWSAs) by default with a monthly temporal resolution. TWSC represents the difference in TWS (i.e., water flux) between two consecutive months, while TWSA is the anomaly with relative to the average during the time period. As a result, the GRACE satellites can be used to monitor floods and droughts by analyzing temporal variations in TWSC [10
]. Droughts and floods in the Liao River Basin are mostly analyzed by statistical data of disasters in Liaoning Province, which may prove difficult to acquire. As an alternative, we used GRACE TWSAs to monitor droughts and floods.
The Earth is a dynamic system that changes over time and space. The changes in the Earth’s gravitational field are mainly caused by the Earth’s mass redistribution. GRACE is established on the basis of the time-varying gravity field of the Earth [17
]. The GRACE satellite adopted a low-low satellite to satellite tracking technology to launch two low-orbit satellites simultaneously that are approximately 220 km apart on the same orbit. The satellite-borne GPS receiver was used to determine the orbit position of the two low-orbit satellites accurately; then, the k-band ranging system was utilized to continuously monitor the distance variation between the two satellites; thus, the change in the Earth’s gravitational field was obtained [19
]. By removing the effects of tidal (solid, oceanic, and polar) and non-tidal (atmospheric and oceanic) influences, in terms of land areas, GRACE time-varying gravity fields mainly reflect the changes in Terrestrial Water Storage (TWS) on a seasonal or short time scale [20
Low-pass filtering is required to reduce high-frequency noise in processing the GRACE spherical harmonic coefficient product. Low-pass filtering (e.g., truncation, destriping and Gaussian smoothing), however, may cause partial signal loss, which can be restored by two signal restoration methods [21
]: The first is based on TWS derived from land surface models (LSMs), e.g., an additive correction approach [22
] and scaling factor [24
]; the other is less dependent on LSMs, such as forward modeling [21
] and multiplicative correction approaches [28
]. The multiplicative correction approach, however, needs to assume that the distribution of TWS changes is uniform, if not, large biases are introduced. The recently released Jet Propulsion Laboratory (JPL) mass concentration block (mascon) solutions is based on a priori constraints in space and time without additional destriping filter, thereby minimizing the effect of measurement and leakage errors [29
]. The mascons can be applied at regional to global scales and do not require prior information provided by LSMs to restore signal loss [30
]. Hence, both the JPL mascon solutions and GRACE spherical harmonic coefficient products are used here.
Drought monitoring systems require (near) real-time inputs of GRACE TWS changes; in addition, developing a drought index requires a long-term time series of data (more than 30 years) [32
]. Monthly GRACE satellite data, however, are only available since April 2002, with a latency of two to six months [13
]. Moreover, the first generation GRACE satellites stopped functioning before the launch of the GRACE follow-on mission in May 2018. Therefore, a method to extend the time series is required.
By assimilating the information obtained from remote sensing into the hydrological model, the integrated soil moisture (SM) can be obtained, and then the monitoring of floodplain inundations can be completed [33
]. However, the vertical accuracy and spatial resolution of remote sensing images constrained the use of this method. Pan et al. [36
] combined the terrestrial water budget estimated from different data sources, including in situ observations, remote sensing retrievals, LSM simulations, and global re-analyses, to enforce the water balance constraint of 32 globally distributed major basins from 1984 to 2016 using data assimilation techniques. In his article, TWS beyond the GRACE period is estimated by the variable infiltration capacity (VIC) model, which lacks GWS change and is arguably ill-suited to simulate the interactions between water storage compartments. Long et al. [13
] reconstructed the GRACE TWSA data for a large karst plateau in Southwest China over the past three decades by developing an artificial neural network (ANN) model and combining this model with the monthly mean temperature, monthly precipitation and SMS to obtain the frequency and severity of droughts and floods. Sun [37
] predicted the groundwater level changes by developing an ANN and combining it with GRACE products. However, the convergence rate of the ANN algorithm is slow and is prone to over-fitting. In comparison with ANN, a general regression neural network (GRNN) is superior to ANN network in terms of learning speed, function approximation, pattern recognition and classification ability [38
GRNN demonstrates a strong nonlinear mapping capability, flexible network structure and high degree of fault tolerance and robustness [39
]. The prediction effect also performs well when the sample data size is restricted, which may be exploited for unstable data processing; this effect can be used to process unstable data [40
]. However, overfitting may occur with the GRNN method reconstructing TWS changes for the training period, resulting in poor performance in other periods, especially over the hindcasting period. To avoid this, cross validation is applied over the training period and early stopping is considered. This paper aims to (1) compare TWS changes of JPL mascon solutions with GRACE spherical harmonic solutions, (2) predict TWSA for the Liao River Basin in China beyond the GRACE period with GRNN models, and (3) monitor droughts and floods across the Liao River Basin with a drought index, where the total storage deficit index (TSDI), which is based on a long-term TWSA time series. The flowchart of this study is shown in Figure 1
2. Study Area
The Liao River originates from Guangtoushan in Qilaotu Mountain in Hebei Province and flows through Hebei Province, Inner Mongolia, Jilin Province and Liaoning Province [2
]. The Liao River Basin (117°00′E–125°30′E, 40°30′N–45°10′N) is situated in the southwest of Northeast China, east of the Di’er Songhua and Yalu Rivers, west of the Inner Mongolia Plateau, south of the Luan River, Daling River Basin, and Bohai, and north of the Songhua River; the total area equals 221,100 km2
]. Figure 2
illustrates the location of the Liao River Basin and the distribution of the meteorological stations.
The main tributaries of the Liao River Basin are the West Liao, East Liao, and Liao and Hun Tai rivers [5
]. Most of the regions in this basin are characterized by a temperate semi-humid and semi-arid monsoon climate. Severity of droughts and floods is related to precipitation [6
]. The frequency of droughts is high in spring due to dry weather conditions when sand winds are common. Precipitation is highly localized in summer, resulting in alternating droughts and floods. Precipitation is virtually absent in autumn and winter, which makes droughts likely to occur during these seasons [44
A reliable TWSC product is important for regional hydrological cycle and other related studies. In this study, the newly released JPL mascon solutions were compared with the GRACE spherical harmonic solutions. The JPL mascon solutions showed higher accuracy than the GRACE spherical harmonic solutions in terms of the uncertainty and the comparison with the Water Resources Bulletin of the Song Liao River Basin.
The TSDI, which was established by the JPL mascon and GRNN-predicted TWSAs, monitored the drought and flood patterns in the semi-arid and semi-humid land of China. The findings indicated that the monitoring results of drought and flood disasters were consistent with those of previous studies and records, and the frequency and severity of drought and floods had intensified in the Liao River Basin in the past 30 years. The comparison between TSDI and previous studies showed that TSDI was more reliable than previous studies in drought and flood monitoring. The reason was that the TSDI drought index was constructed on the basis of TWSA, and the TWSA was recovered by the GRACE satellite and contained all the components of water storage, including SMS, GWS, SWE, and so on. However, the drought index used in previous studies contains only one single variable runoff. Thus, the GRACE satellite and TSDI can provide a solid theoretical basis for drought and flood monitoring in areas where data are lacking or even non-existent.
Extreme climate events are a serious threat to human life and property and reflect the irrational use of water resources in the region to a certain extent. Therefore, strengthening the early warning system for drought and flood disasters in the study area, and establishing a reasonable water resources management mechanism are necessary to reduce the losses caused by meteorological disasters and to develop and utilize water resources rationally. The GRNN method is useful in predicting TWSAs beyond the GRACE period. By combing GRACE TWSA, the extended GRACE TWSAs and TSDI, drought and flood disasters in the Liao River Basin in China and other semi-arid and semi-humid regions can be monitored and predicted.