In order to understand the Earth's global water cycle and its vulnerability to climate change, and to be able to carry out appropriate risk assessments for ecosystems, forest fires, agriculture, and water management, continuous observation of the distribution and movement of continental water masses is urgently needed. So-called hybrid models, i.e., numerical models combined with AI methods, now play a central role in gaining deeper insights into the continental hydrology.
Jan Saynisch-Wagner, GFZ Helmholtz Centre for Geosciences
For 24 years, satellite data from the GRACE (Gravity Recovery and ClimateIn contrast to weather, which refers to daily or very short-term events, climate refers to an average condition in the atmosphere over a longer period of 30 to 40 years. All processes such as average temperature, precipitation, wind direction, wind s... Experiment) missions have provided important insights into the mass distribution of the Earth, e.g., the mass balance of ice sheets, but also continental hydrology. For example, by comparing current data with long-term averages, anomaly maps can be created that show where groundwater has decreased or increased, and the associated risks of droughts and forest fires can be better determined.
A key problem with earlier GRACE analyses was resolving small-scale anomalies in terrestrial water storage from GRACE data, as the spatial resolution of satellite gravimetry is limited to about 300 km. This meant that the local distribution of water in rivers, lakes, and groundwater reservoirs remained hidden in detail.
AI can now help here, and low-resolution satellite data can be downscaled to a higher working resolution. Successful deep learning AI models use methods such as convolutional neural networks (CNN) or generative adversarial networks (GAN), which are trained with state-of-the-art numerical hydrology models. To do this, satellite observations of the hydrology model outputs are simulated and the respective relation is learned by an AI (see Fig. 1).
Both methods, GAN and CNN, are particularly well suited for processing spatial Earth observation data, as they can reliably recognize recurring patterns such as lines, edges, or more complex structures (Fig. 2). Further possibilities are based on the Transformer architecture, which is also used in chatGPT, for example.
The current GRACE missions and those already in the planning stage guarantee essential, consistent long-term observations of climate-relevant parameters. However, the spatio-temporal resolution of the measurements will only improve to a very limited extent in the future. In conjunction with modern AI downscaling, numerous relevant questions can nevertheless be answered - including the assessment of climate change impacts on continental hydrology, the identification and analysis of pressures on ecosystems, and the support for water management in agricultural and urban regions with regard to future challenges.
Further Reading
- Hassanibesheli, F., Jensen, L., Dobslaw, H., Thomas, M., Saynisch-Wagner, J. (2024). Downscaling Satellite Gravimetry Observation for Terrestrial HydrologyThe science of the water cycle. Hydrology analyses the temporal and spatial distribution of water, its properties and its interactions with the environment. The following sciences are related disciplines or sub-disciplines of hydrology, depending on ... Using a Conditional Generative Adversarial Network. ESS Open Archive. https://doi.org/10.22541/essoar.172493940.07693855/v1
- 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 https://doi.org/10.1029/2020GL089258