Deep learning-bases downscaling of satellite gravimetry for the estimation of high-resolution terrestrial water storage
The GRACE satellites provide us with valuable global observations of the Earth's gravity field. These observations are analyzed to investigate and to untangle spatial and temporal variations of moving masses on Earth. Terrestrial water storage is one collective mass component that is used to study water masses on land and various related processes, such as cycles, trends, or extreme events like floods and draughts.
Due to the limited spatial resolution of GRACE observations of approximately 300 km, only large-scale mass anomalies can be detected, while small-scale structures, like water masses in individual river branches, remain elusive. In this study, we have developed a deep learning approach that is able to recover such small-scale mass features from GRACE satellite observations. This so-called downscaling is possible due to a novel training method that integrates high-resolution data from both numerical hydrology modelling and satellite altimetry. In particular, the developed neural network is able to validate its downscaling process and can produce more accurrate estimates of terrestrial water storage than its trainer model.
Irrgang, C., Saynisch‐Wagner, J., Dill, R., Boergens, E., & Thomas, M. (2020). Self‐validating deep learning for recovering terrestrial water storage from gravity and altimetry measurements. Geophysical Research Letters, 47, e2020GL089258. https://doi.org/10.1029/2020GL089258