High-resolution groundwater storage anomalies in the Middle and Lower Yangtze River Basin of China using machine learning fusion of in-situ wells, satellite gravity and hydrological model

Groundwater plays a key role in the water cycle and is used to meet industrial, agricultural, and domestic water demands. High-resolution modeling of groundwater storage is often challenging due to the limitations of observation techniques and mathematical methods. In this study, two machine learning (ML) algorithms, namely random forest (RF) and artificial neural networks (ANNs), were employed to estimate groundwater level anomaly (GWLA) and groundwater storage anomaly (GWSA) with a 0.25° resolution using multi-source datasets, including in-situ wells, the Gravity Recovery and Climate Experiment (GRACE), land surface models and hydrogeological parameters. The results indicated the ANN algorithm outperformed the RF algorithm in predicting the spatiotemporal variations of the shallow and deep GWLA in the Middle and Lower Yangtze River Basin (MLYRB). Hence, the ANN algorithm was used to construct a model for predicting the GWSA over the 2005–2017 period. The GWSA exhibited an increasing linear rate of about 0.77 ± 0.30 mm/yr in almost the entire area, except in the Han River Basin (HRB), where GWSA decreased by −1.18 ± 0.38 mm/yr due to decreased precipitation amounts. The occurrence of seasonal variations in the deep GWSA showed lead time effects compared with those in the shallow GWSA, ranging from 0 to 1 month and 1 to 2 months in the humid and relatively dry areas, respectively. It was found that the ANN-based model results showed pronounced responses of the GWSA to the drought events. The standard groundwater drought index (SGDI) was further calculated to assess the spatiotemporal characteristics of the groundwater drought events. The results revealed the occurrence of a severe drought event in 2011, as well as pronounced impacts of the El Niño-Southern Oscillation (ENSO) events on the GWSA. This study demonstrated the effectiveness of in-situ groundwater measurements in enhancing the reliability of ML-based GWSA prediction. Furthermore, the constructed high-resolution groundwater variation features in this study can provide water resource managers with enhanced information and valuable insights into climate-induced groundwater changes. •A 0.25°-resolution GWSA model was built using machine learning algorithms.•The model was trained with 415 groundwater well data to achieve accurate results.•Spatiotemporal differences in the shallow and deep aquifers were studied.•The SGDI of the constructed GWSA model was in line with the El Niño/La Nina events.