Towards Entity-Aware Amortized Variational Inference: An application in Streamflow modeling
Abstract
Machine Learning is beginning to provide state-of-the-art performance in a range of environmental applications such as streamflow prediction in hydrologic basins. However, building accurate broad-scale models for streamflow remains challenging in practice due to the variability in the dominant hydrologic processes. Those different processes are best captured by sets of process-related basin characteristics. At best, direct measurements of basin characteristics suffer from uncertainty due to measurement noise and spatial variability, which adversely impact model performance. Some basin characteristics are not observable or not well-understood. For basins with an incomplete set of known characteristics, providing uncertainty estimates alongside predictions is essential for responsible decision making. To tackle the above challenges, we propose a novel framework for amortized variational inference that can extract system characteristics from driver (input) and response (output) data and provide a suite of predictions which allow uncertainty estimation. This first-of-its-kind framework can produce predictions even when characteristics are corrupted or missing. We evaluate the proposed framework in the context of streamflow modeling using CAMELS (Catchment Attributes and MEteorology for Large-sample Studies) which is a widely used hydrology benchmark dataset. We also compare learned representations from the data to known basin characteristics in order to determine whether the representations hold physical meaning, and whether the framework extracts basin characteristics from drivers and responses.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H35C..06G