Groundwater (GW) plays a crucial role in coastal aquifers and arid regions, serving as a lifeline for communities by providing a reliable and resilient water source, making its monitoring essential for sustainable water management. This study aimed at modeling GW via regionalization of the Gravity Recovery and Climate Experiment (GRACE) data based on two methods. The first method directly regionalized the GRACE data for modeling GW via in situ measurements, including the lake level, precipitation, temperature, observed GW and Penman-Monteith-Leuning (PML) evapotranspiration data. The second method included two stages, in the first stage, the GRACE data were downscaled via the Famine Early Warning Systems Network Land Data Assimilation System (FLDAS) data which contains satellite based precipitation, temperature, soil moisture, and snow water equivalent data. In the second stage, the downscaled GRACE was bias corrected to provide regionalized data. Artificial intelligence models consist of shallow networks (Feed Forward Neural Network (FFNN), Adaptive neuro fuzzy (ANFIS), Support Vector Machine (SVR)), the ensemble of shallow networks and Long-Short Term Memory (LSTM) deep learning method were employed in the modeling process and the observed GW level data were targeted for the regionalization. The Link CluE clustering ensemble method was implemented to cluster the piezometers of the aquifer to separate different GW patterns in the area. The proposed methodology was examined over the Miandoab plain, one of the sub-basins of the Lake Urmia, located in Northwest Iran. The modeling results demonstrated that the first method could exhibit superior performance with the Nash-Sutcliffe Efficiency (NSE) of up to 17% higher than the second method. Thus, using in situ observed data for downscaling proved to be more accurate than relying on the data based on the satellite imagery. The results indicated that the ensemble of shallow networks could lead to more precise results than using the deep and shallow learning models, individually, where the NSE for the ensemble of shallow networks was up to 50% higher compared to the LSTM model.

中文翻译：

---

通过人工智能方法集合对海岸线 GRACE 数据进行区域化

地下水 (GW) 在沿海含水层和干旱地区发挥着至关重要的作用，通过提供可靠且有弹性的水源成为社区的生命线，使其监测对于可持续水管理至关重要。本研究旨在通过基于两种方法的重力恢复和气候实验（GRACE）数据的区域化来模拟重力波。第一种方法直接对 GRACE 数据进行区域化，通过原位测量来模拟 GW，包括湖水位、降水、温度、观测到的 GW 和 Penman-Monteith-Leuning (PML) 蒸散发数据。第二种方法包括两个阶段，在第一阶段，GRACE数据通过饥荒预警系统网络陆地数据同化系统（FLDAS）数据进行降尺度，其中包含基于卫星的降水、温度、土壤湿度和雪水当量数据。在第二阶段，对缩小规模的 GRACE 进行偏差校正以提供区域化数据。人工智能模型由浅层网络（前馈神经网络（FFNN）、自适应神经模糊（ANFIS）、支持向量机（SVR））组成，采用浅层网络集成和长短期记忆（LSTM）深度学习方法在建模过程中，观测到的GW级数据是区域化的目标。采用 Link CluE 聚类集成方法对含水层的压力计进行聚类，以分离该区域中不同的 GW 模式。所提出的方法在位于伊朗西北部的乌尔米亚湖的子流域之一的米安多阿布平原进行了检验。建模结果表明，第一种方法可以表现出优越的性能，纳什-萨特克利夫效率 (NSE) 比第二种方法高出 17%。因此，使用现场观测数据进行降尺度被证明比依赖基于卫星图像的数据更准确。结果表明，浅层网络集成可以比单独使用深度和浅层学习模型获得更精确的结果，其中浅层网络集成的 NSE 比 LSTM 模型高出 50%。