地下水是极其重要的淡水资源，在农业，工业，公共安全和生态环境等各个领域都发挥着举足轻重的作用。掌握中国区域地下水的时空分布及变化特征有助于为国家和区域水资源管理和政策的制定、提高水资源的利用与分配提供可靠的理论支撑。同时，由地下水变化导致的地面沉降也是一个严峻的地质环境问题，评估全国地区由于地下水变化导致的潜在地面沉降危险性，分析其对社会经济各方面可能造成的影响，有助于防范和减少可能对社会和人民生命财产安全造成损失的灾害事件，对社会的稳定和持续发展起着未雨绸缪的重要作用。

       鉴于此，本文基于重力恢复与气候实验（GRACE）卫星和全球陆地数据同化系统（GLDAS）产品，采用垂直方向水量平衡方程反演了中国区域2003?2016年地下水储量变化，并分析了中国区域地下水时空变化规律与降水之间的关系。采用机器学习模型生产了中国区域1km地下水储量（GWS）数据产品，在此基础上，结合自然因素和社会环境因素评估了2004?2016年和2040年中国区域公里网格尺度的潜在沉降危险性，研究结果表明：

（1）在2003?2016年间，中国区域地下水变化在空间上呈现大面积聚集性分布特征，中国新疆西北部天山地区，西藏南部和华北平原一带，地下水呈现显著下降，下降速率最大的地区位于西藏的东南区域，达到–7.98 cm/a，华北平原一带地下水下降速率达到–3.42 cm/a。除华北平原和青藏高原以外，全国大部分地区观测到的地下水变化趋势与降水趋势保持一致，相关系数0.73。从季节波动来看，降水对地下水的补给存在一定的滞后性, 全国尺度上地下水对降水延迟1个月，相关系数0.76。长江流域地下水滞后效应最小，约为1个月，降水与地下水的相关系数达0.63；松辽河流域地下水滞后效应最大，为47个月，降水与地下水相关系数达0.4；海河流域与淮河流域地下水与降水相关性最低，仅为0.27和0.23。地下水季节波动性最大七个省份分别为江西，湖南，福建，广西，浙江，广东和海南。

       （2）在1km空间分辨率下，极限梯度提升树（XGBoost）与随机森林（RF）都能很好的描述自然和社会经济因素与陆地水储量（TWS）及非地下水组分之间的非线性关系。其中XGBoost模型的R²在2004?2016年的变化范围为0.77?0.89（RF是0.74?0.86）,相关系数（CC）为0.88?0.94（RF是0.88?0.93），RMSE为0.37?2.3（RF是0.4?2.53），二者都能很好地实现中国区域1km地下水储量产品的生产。在2004?2016年期间主要影响TWS的变量为降水，DEM和黏土含量，而主要影响非地下水组分的因子为DEM，降水和地表温度。降尺度的数据保留了原始分辨率数据的空间分布、年际趋势和年际时序波动特征。海河流域降尺度前后相关系数最高，达到了0.999，西南流域相对较低，为0.929。降尺度前后与监测站点时序相比，相关系数高于0.5以上站点都有69个，占总站点的27.49%，表明降尺度结果很好保留了原始数据的时序精度特征。本文同时提供XGBoost和RF方法生产的数据产品，为精细的水文模型研究，区域水资源的管理以及政策的制定提供了强有力的数据支持。

       （3）中国区域潜在地面沉降危险性受地下水消耗的影响较大，中国区域沉降高危险性地区集中在华北平原，安徽中部，河南北部，陕西中部，广东南部和江苏?上海一带，高危险性地区面积达到43.52?50.08万km²，占全国总面积的4.53%?5.22%。这些地区是灌溉集中区，社会经济发展较快且地下水消耗严重区域。在未来二氧化碳排放持续增加（RCP8.5）且社会共享经济路径为区域竞争路径（SSP3）情景下，到2040年，中国潜在沉降高危险性区域的空间格局与历史一致，但华北平原，山西西部，江苏—上海一带以及广东省南部广州市周围沉降危险性进一步加剧，潜在沉降高危险区域面积将增加28.72%?43.38%，达到62.40?64.46万km²。暴露于高危险沉降地区的人口和GDP将分别增加32.7%?41.87%和355.91%?358.68%，即约有3.631?4.289亿人口和4.692×10⁵?5.553×10⁵亿元暴露于潜在沉降高危险地区。未来2040年铁路暴露于高危险性的长度将增加15.85%?42.99%，占铁路总长度的22.03%?23.91%；未来公路暴露于高危险性地区的长度将增加19.65%?37.4%，占公路总长度的17.83%?19.57%。

       本文成功实现了中国地下水储量1km降尺度数据产品的生产，为小尺度地下水变化的动态研究、水文模型的参数校正、区域水资源的管理提供了重要的数据基础。并且，本文基于此数据分析了中国历史与未来潜在沉降危险性空间分布，并重点评估了人口、GDP以及公路和铁路在潜在沉降高危险性下的暴露，为后期分析重点沉降区域和沉降地区的治理提供了重要的参考价值。

Groundwater storage (GWS) is an extremely important freshwater resource, which plays a key role in various fields such as agriculture, industry, public safety, ecological environment and so on. Grasping the temporal and spatial distribution characteristics of groundwater in China is helpful to understand in detail the dynamics of GWS change in China, and provides reliable theoretical support for national and regional water resources management, policy formulation and improving the utilization and distribution of water resources. Meanwhile, land subsidence caused by groundwater depletion is also a serious geological environmental problem. To assess the potential land subsidence risk caused by groundwater consumption over China and analyze its possible impact on the society and economy will help prevent and reduce possible impacts on society in advance, which plays an important role in precautions for the stability and sustainable development of society.

       Given those above, based on the Gravity Recovery and Climate Experiment (GRACE) satellite products and Global Land Data Assimilation System (GLDAS), this paper analyzed the spatio-temporal changes of GWS in China from 2004 to 2016 using vertical water balance equation and its relationship with precipitation. Then we used machine learning methods to downscale the GWS to 1 km over China. Furthermore, we evaluated the potential subsidence hazard at the kilometer grid scale from 2004?2016 and in 2040 throughout China by combining natural and social environmental factors. Research indicates that:

(1) The GWS in China presented a large-area agglomeration distribution characteristic in terms of the trend from 2003 to 2016. There showed a decreasing in the Tianshan area of northwestern Xinjiang, southern Tibet and the North China Plain (NCP). The largest decrease was –7.98 cm/a in southeastern Tibet and the decreasing velocity in NCP was –3.42 cm/a. The trend of GWS change observed in most areas in China was consistent with precipitation except NCP and Qinghai-Tibet region and the correlation coefficient was 0.73. There was a certain lag in the replenishment of groundwater by precipitation from the perspective of seasonal fluctuations. There was one-month lag for GWS relative to precipitation on national scale and the correlation coefficient was 0.76. GWS changes in Yangtze River Basin were delayed by precipitation with one month and the correlation coefficient reached 0.63 while Songhua and Liaohe River Basin owned the largest time lags?47 months and the correlation coefficient was 0.4. GWS changes in Haihe and Huaihe River Basins, which covered the NCP, owned the lowest correlation. The seven provinces where the stability of groundwater during the year was most affected by total annual precipitation were Jiangxi, Hunan, Fujian, Guangxi, Zhejiang, Guangdong and Hainan.

(2) Both Extreme gradient boosting tree (XGBoost) and random forest (RF) can well describe the non-linear relationship between natural and socio-economic factors and terrestrial water storage (TWS), as well as non-groundwater components. Three evaluation metrics were employed for the testing dataset for 2004?2016: The R² ranged from 0.77?0.89 for XGBoost (0.74?0.86 for RF), the correlation coefficient (CC) ranged from 0.88?0.94 for XGBoost (0.88?0.93 for RF) and the root-mean-square error (RMSE) ranged from 0.37?2.3 for XGBoost (0.4?2.53 for RF). Both XGBoost and RF models could produce 1km resolution GWS products. The main variables that affect TWS during 2004?2016 are precipitation, DEM and clay content, while the main factors that affect non-groundwater components are DEM, precipitation and surface temperature. The downscaled products retained the spatial distribution, inter-annual trend and inter-annual time series fluctuation characteristics of the original resolution data. The correlation coefficient was the highest in the Haihe River Basin, reaching 0.999 while the lower in the Southwest Basin was 0.929. There were 69 stations whose CCs were greater than 0.5 before downscaling and 69 stations after downscaling, accounting for 27.49% of the total stations. The downscaled products maintained the accuracy of the original data. This study also provided downscaled products based on both XGBoost and RF methods, and provided strong data support for local GWS changes, finer hydrological model construction, water resources management and establishment of policy.

(3) The potential land subsidence in China is badly affected by the groundwater consumption. The main high hazard land subsidence areas in China are concentrated in the NCP, central Anhui, northern Henan, central Shaanxi, southern Guangdong and Jiangsu—Shanghai. The areas reached 435.2—500.8 thousand km², accounting for 4.53%?5.22% of the total areas. These areas were mainly located in irrigation, rapid socio-economic development and serious groundwater consumption. Under the scenario where carbon dioxide emissions continue to increase in the future (RCP8.5) and the shared socioeconomic pathway is the regional competition path (SSP3), China's potential land subsidence areas in 2040 is consistent with history, but the potential land subsidence in NCP, western Shanxi, Jiangsu-Shanghai area and Guangzhou will be further intensified. The high potential land subsidence areas will increase by 28.72%-43.38%, reaching 624.4—644.6 thousand km². By 2040, the population and GDP exposed to high-hazard land subsidence areas will increase by 32.7%—41.87% and 355.91%—358.68%, respectively, which means that there will be approximately 0.3631-0.4289 billion people and 4.692×10⁴-5.553×10⁴ billion Yuan in potential high-hazard land subsidence areas. The length of railways exposed to high hazard will increase by 15.85%-42.99%, accounting for 22.03%-23.91% of the total railway length, and the length of highways exposed to high hazard will increase by 19.65%—37.4%，accounting for 17.83%—19.57% of the total railway length in 2040.

The 1km downscaled GWS products in China provides an important data foundation for the dynamic study of small-scale groundwater changes, parameter correction of hydrological models and the management of regional water resource. Meanwhile, this paper analysized the spatial distribution of potential land subsidence from history to future and focused on assessing the exposure of popultoin, GDP, roads and railway under high potential land subsidence, which provides important references for later analysis and governance of key land subsidence areas.