Research
Weather and climate models describe the atmosphere on grids that are often tens of kilometers across, but the decisions that depend on them happen at a single river basin, farm, or city. My research is about bridging that gap. I develop and evaluate statistical and machine learning methods that downscale and post-process model output into information that’s tailored to a specific use, with a particular focus on precipitation.
Ensemble precipitation forecasts for streamflow
At NOAA PSL, I work on improving ensemble precipitation forecasts that feed streamflow models. An ensemble gives a range of possible outcomes rather than a single forecast, but raw model ensembles need correcting before hydrologic models can rely on them. I use statistical and machine learning post-processing to make those forecasts more accurate and better calibrated.
Statistical downscaling of climate projections
Global climate models are too coarse to say much about a single weather station, so I’ve built station-level projections for regions where local detail matters. In the Midwest United States, I compared self-organizing maps and random forests for producing daily temperature and precipitation at 279 stations around Illinois from nine CMIP5 models. Random forests worked best for temperature and self-organizing maps for precipitation. Under high emissions, the projections show daily high temperatures rising by about 4.3 °C on average by 2071–2100, 48% more extreme heat days, and a drier south of the region.
In West Africa, where rain-fed agriculture makes precipitation changes especially consequential, I used self-organizing maps to project station precipitation. Under high emissions, the coast gets slightly wetter while the interior Sahel loses about 10% of its precipitation by 2100. Most of that drop comes from fewer rainy days in the wet season, not lighter rain or a shorter season.
Open-source software
To make these methods easier for others to use, I wrote CCdownscaling, a Python package for multivariable statistical downscaling (Environmental Modelling & Software, 2023). It includes several downscaling methods, among them self-organizing maps, along with metrics for evaluating how well a downscaled model performs. Example data is on Zenodo.
How much do precipitation datasets agree?
Any assessment of precipitation change depends on the data underneath it. In the Amazon Basin, I compared precipitation trends from 1980 to 2022 across observational, reanalysis, and climate model datasets. The trends vary a lot by region and season, and the datasets often disagree. For example, reanalyses show significant dry-season drying in the southern and central Amazon that the observations don’t.
See Publications for the full list.