To understand the Earth’s global water cycle and its susceptibility to climate change, as well as to make appropriate risk assessments of ecosystems, wildfires, agriculture and water management, a continuous monitoring of the distribution and movement of continental water masses is necessary. To gain comprehensive insights into continental hydrology, further developments of numerical models and the combination with innovative Deep-Learning-Methods will take up an important role in future.
Introducing this approach, the authors of the paper “Self‐Validating Deep Learning for Recovering Terrestrial Water Storage from Gravity and Altimetry Measurements,” published in Geophysical Research Letters use satellite data from GRACE, the Gravity Recovery and Climate Experiment. These satellite observations depict the Earth’s gravity field and its anomalies, which are, over the continents, dominated by spatiotemporal changes in terrestrial water storage (TWS).
Thus, the gravity variations studied by GRACE are in principle very useful to determine water mass distribution on land. For example, by comparing current data to the average over time, scientists can generate an anomaly map to see where ground water has depleted or increased.[i]
The problem of the previous studies was to detect multiscale terrestrial water storage anomalies (TWSA) from GRACE that range from large-scale total water storage to locally resolved river structures. Especially the latter proved to be a strong problem. Earlier terrestrial water storage anomalies extracted from GRACE measurements are blurred. Because the spatial resolution of satellite gravimetry is limited to approximately 300 km, the underlying distribution of water masses in rivers, lakes, aquifers, and ground water basins remain elusive.
The authors of this recent study combine deep learning, forward hydrological modeling and space-borne observation systems to explore whether a trained artificial neural network can derive high resolution TWSA from the smooth and blob-like GRACE observations, too.
“For the so called down-scaling, we are using a convolutional neural network, in short CNN, in connection with a newly developed training method”, primary author Christopher Irrgang says. “CNNs are particularly well suited for processing spatial Earth observations, because they can reliably extract recurrent patterns such as lines, edges or more complex shapes and characteristics.”
In order to learn the connection between continental water storage and the respective satellite observations, the CNN was trained with simulation data of a numerical hydrological model over the period from 2003 until 2018. Data from satellite altimetry in the Amazon region was then used for validation. This training approach allows the CNN to dynamically adapt and validate its learning process based on independent altimetry observations of surface water storage, and ultimately outperforms the numerical model used in the training process. In this way, the authors show how CNN training can be improved by including additional constraints. With the use of satellite altimetry observations, the novel constrained machine learning approach provides further improvement in the downscaling of GRACE-observations.
“Subsequent applications of our approach can be beneficial for several related topics, including assessments of climate change impacts on continental hydrology, identification, and analyses of ecosystem stresses, or aiding water management in agricultural and metropolitan areas,” the paper concludes.
To find more details read the paper: 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, 17