Temporal gaps in the Gravity Recovery and Climate Experiment (GRACE) and GRACE–FO missions pose difficulties in analyzing spatiotemporal variations of terrestrial water storage (TWS) anomalies over Indian river basins. In this study, we developed a deep learning-based CNN–LSTM model to address these temporal gaps by integrating GRACE–TWS with other meteorological variables. The model achieved a strong Pearson’s correlation coefficient (median of 0.96 in training, 0.92 in testing), high Nash–Sutcliffe efficiency (0.91 in training, 0.85 in testing) and low normalized root-mean-squared error (0.064 in training, 0.098 in testing) when compared to the original GRACE–TWS. Moreover, a comparison with two publicly available global datasets suggests the superior performance of CNN–LSTM in predicting TWSA (terrestrial water storage anomalies). The study also highlighted potential biases up to 2 cm/yr in long-term TWSA trends due to temporal gaps in GRACE/GRACE–FO. Additionally, the estimated TWSA matched well with in situ wells across majority of the Indian river basins. The findings revealed significant groundwater depletion in the northern and northwestern river basins but positive trends in the central and southern basins in India. Overall, the estimated TWSA product developed in this study provides continuous data records and valuable insights into long-term trends in groundwater storage, making it useful for groundwater monitoring.

重力恢复和气候实验（GRACE）和GRACE-FO任务中的时间差距给分析印度河流域陆地水储存（TWS）异常的时空变化带来了困难。在本研究中，我们开发了一种基于深度学习的 CNN-LSTM 模型，通过将 GRACE-TWS 与其他气象变量集成来解决这些时间间隙。该模型实现了强大的 Pearson 相关系数（训练中的中位数为 0.96，测试中的中位数为 0.92）、高 Nash-Sutcliffe 效率（训练中为 0.91，测试中为 0.85）和低归一化均方根误差（训练中为 0.064，测试中为 0.098）。测试）与原始 GRACE-TWS 相比。此外，与两个公开的全球数据集的比较表明，CNN-LSTM 在预测 TWSA（陆地水储存异常）方面具有优越的性能。该研究还强调了由于 GRACE/GRACE-FO 的时间差距，长期 TWSA 趋势可能存在高达 2 厘米/年的偏差。此外，估计的 TWSA 与印度大部分河流流域的原位井吻合良好。研究结果显示，印度北部和西北部河流域的地下水严重枯竭，但印度中部和南部盆地的地下水呈积极趋势。总体而言，本研究开发的估计 TWSA 产品提供了连续的数据记录和对地下水储存长期趋势的宝贵见解，使其可用于地下水监测。