How to beat your teachers in hydrologic machine learning
Abstract
Deep learning (DL) models trained on hydrologic observations are recently shown to be highly performant. Used directly, however, they inherit certain flaws of their supervising data. In other words, these models are students that cannot exceed their teachers (supervising data). For example, satellite data has global coverage but low resolution/accuracy, while in-situ networks are spatially imbalanced. For another example, we cannot predict a variable at large scales if we do not have extensive observations for it. While some have shown that adding physical constraints could be beneficial, the benefit is limited to minor-to-modest gains in performance. Here we explore several pathways for machine learning models to exceed their teachers. First we explore learning from multiple data sources (soil moisture) at different scales, creating a multiscale forecast model that breaks the confines of the individual supervising dataset. Second we demonstrate how we connect machine learning with physics-based models to predict unobserved variables that help determine future trends of the water cycle. Third we show how network models can be leveraged to learn physics rather than purely making predictions. Overall, there are substantial new paths to take for hydrology to benefit from big data machine learning apart from elevating the prediction accuracy.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H31H..07S