Data-driven methods with deep learning as their core have been gradually applied in Earth science; however, challenges remain regarding the interpretability of models and physical consistency. With the background of remote sensing big data, combining deep learning and data assimilation methods to develop new techniques for the simulation and prediction of terrestrial water cycle processes has become an important research direction in Earth science. Τhe progress in deep learning in recent years combines improving the quality of observation data of terrestrial water cycle components and reducing the uncertainty of physical models. Furthermore, the key scientific issues regarding data assimilation in terrestrial hydrology based on deep learning fusing remote sensing big data are classified according to the observations, physical models, and system integration: \mathrm{①} How can the temporal and spatial representativeness of samples be enhanced when deep learning inverts remote sensing products? \mathrm{②} How can a new physics-guided deep learning method be developed within the framework of data assimilation? \mathrm{③} How can the predictability of the terrestrial water cycle be improved through the “data-model” dual drive? Relevant research and exploration should help promote the in-depth application of the “data-model” hybrid modeling method in the field of hydrology and improve the simulation and prediction capacity of the terrestrial water cycle process.