Can Machine-Learned Summary Statistics Improve Simulation-Based Inference of Uncertain Parameters? A case study in the Upper Colorado River Basin
Abstract
Simulation-based inference is a technique to estimate uncertain parameters by comparing outputs of mechanistic models to observed data. Due to the curse of dimensionality, 'summary statistics are often needed to 'compress high-dimensional data (e.g. streamflow time series) into a low-dimensional form. Domain experts typically try to define summary statistics (e.g. 95th flow quantile, baseflow index, and many others) that minimize information loss and tailor inference to modeling objectives. However, for a variety of reasons, these summaries are usually insufficient to the task of constraining probable ranges of uncertain parameters. This study confronts the challenges and opportunities of inference with summary statistics using a streamflow model of the Taylor River - a snowmelt-dominated, headwater tributary to the Colorado River. First, a Long Short-Term Memory emulator is trained on an ensemble of ParFlow outputs with systematically varied surface and subsurface parameters. A neural density estimator is then applied using the emulator to explore parameter space and generate posterior distributions in a likelihood-free way. Deep neural network data compression is used to reduce high-dimensional streamflow data down to p-dimensional summary statistics (one per parameter) with the aim of retaining streamflow signatures only relevant to parameter estimation. We evaluate the information content of these machine-learned compression statistics against both traditional hydrologic summary statistics (e.g. flashiness, quantiles, baseflow index) and full streamflow time series. Finally, we compare the performance of parameters inferred via all three approaches in the original ParFlow model. More generally, this investigation provides an example of how machine learning techniques may aid statistical inference with high-fidelity, process-based models.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H35S1253H