Data-driven Hydrological Models Using Gaussian Processes
Abstract
The NASA Land Information Systems model (LIS) is a driver for high-performance terrestrial hydrology modeling. It enables modeling of several state-of-the-art hydrology models including NOAH, NOAH-MP, CLM, VIC, CABLE, Sacramento/SNOW17, and SUMMA. The land surface models can be validated using ground-based observations collected at hundreds of locations around the globe. The data products from LIS are used in many important applications including flood prediction, weather forecasting, and agriculture to name a few.
With a large amount of observational data being collected, data-driven modeling offers a potential to improve the accuracy of current land-surface models. In particular, we show the potential to add a machine-learning model to the time-step integration of land surface models in LIS. A data-driven model is trained on the deviation between observations and a terrestrial hydrology model (e.g., NOAH), and then used at every timestep of the NOAH run to reduce simulation errors. In effect, this is a form of real-time data assimilation, and can also be combined with more common types of (e.g., Bayesian, variational, ensemble) data assimilation. By using the terrestrial hydrology model as the mean field or mean function for the data-driven model, simulations revert to the hydrology model in cases that are dissimilar to past observation data used for training. In this work, Gaussian Processes (GPs) are used to construct the data-driven model; these are trained and added to the LIS simulation code at the timestep level. The model is trained using data from 40 AmeriFlux tower sites with a two-year split-record time period for calibration and validation. In addition a leave-one-out strategy is used to test spatial extrapolation - here, one tower site is reserved for validation and all others are used for training. The GP model is run within LIS and compared with the NOAH model, and the results are presented and discussed.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFM.H14C..03K
- Keywords:
-
- 1869 Stochastic hydrology;
- HYDROLOGYDE: 1895 Instruments and techniques: monitoring;
- HYDROLOGYDE: 1914 Data mining;
- INFORMATICSDE: 1942 Machine learning;
- INFORMATICS