Ensemble Kalman Inversion with Limited Observation for Parameter Estimation and Upstream Streamflow Prediction
Abstract
In modeling physical processes, parameter estimation is a crucial step in making quality predictions. These models often rely on simplifications of the complex underlying processes. Estimating these parameters is difficult, and direct measurements are often sparse, if available. Over the past several decades, data assimilation (DA) frameworks have been developed to exploit the increasingly available data to improve model predictions. Often these methods focus on the state estimation problem, attempting to correct the model output directly. The assumptions necessary to apply these methods result in states which no longer satisfy desirable physical properties built into the underlying models (e.g. conservation of mass and flux). One standard DA approach applied in many fields to solve this state estimation problem is the Ensemble Kalman Filter (EnKF). The Ensemble Kalman Inversion (EKI) framework applies the EnKF to solve Bayesian inverse problems, such as parameter estimation. Unlike in the state estimation problem, the EKI approach ensures posterior state estimates will necessarily satisfy the model properties. This method also gives uncertainty quantification that can give insight into the certainty of the prediction at each location. We test this technique to learn spatially distributed routing parameters in a distributed rainfall-runoff hydrologic model using Stage IV radar-rainfall and MODIS seasonal evapotranspiration for forcings. Given only measurements at the outlet, this method can learn a set of distributed parameters that show substantial improvement in prediction skill across several metrics at locations upstream from the outlet over open loop simulations. In the virtual catchment case, these results well approximate the parameters used to generate synthetic measurement simulations along the links of higher Horton order. In addition, the uncertainty in the parameters mirrors the sites with higher approximation error. This work shows promise for learning model parameters that improve predictions upstream and quantifying the uncertainty for larger river networks. Additionally, this work aims to show the potential for future operational approaches.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMNG35B0467P