Toward Global Terrestrial Hydrology with Theory Guided Machine Learning
Abstract
Machine learning (ML) is quickly becoming an indispensable part of hydrologic prediction, and the community is developing strategies to use ML to advance basic knowledge of hydrologic processes in general. ML is well known to outperform physically-based models in most side-by-side comparisons, but this comes with a caveat that ML's predictive accuracy breaks down when conditions fall outside the range of what was included in the training data. This is a serious problem because these scenarios are often the more critical periods for forecasting ( e.g. , droughts and floods). Models that are based on an understanding of physical processes are still necessary for these scenarios. We have developed a hybrid (physics-based + ML) approach to leverage the predictive power of ML with the robustness of process-based models. I will present results from this hybrid approach.
A long-term goal will be to use ML to improve model predictions anywhere, and on a large scale, using gridded satellite data. This requires ML to make accurate, site independent, predictions, because the hydrologic response at any particular site does not usually represent an average response for a broader area. In our example, Gaussian Process Regression (GPR) is used for the ML component, which predicts dynamic corrections to model structural error in top layer soil moisture state of the Noah-MP land surface model. We test this method over annual cycles at FluxNet and SCAN sites with high quality observations. The performance of the GPR was tested and verified out-of-sample both in time and in space. In addition to the results of individual predictive performance, I will present results on the potential for cross-site information transfer and the ability for ML to learn and predict with site independence. This is an early, but critical, step for a global model using theory guided machine learning.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.H43I2149F
- Keywords:
-
- 1816 Estimation and forecasting;
- HYDROLOGY;
- 1839 Hydrologic scaling;
- HYDROLOGY;
- 1843 Land/atmosphere interactions;
- HYDROLOGY;
- 1847 Modeling;
- HYDROLOGY