SIParCS 2019 - Katrina Wheelan

Testing Machine Learning for Regional Climate Applications in the Pacific Northwest

Recorded Talk

There is increasing demand from decision makers for fine scale climate information that is relevant for regional and local adaptation planning. Water resource managers in particular need detailed estimates for future precipitation. While dynamical downscaling is an important tool, output from regional climate models (RCMs) is often still too coarse to be useful. In this project, we compare three methods of statistically downscaling coarse RCM output to the finer resolution end users need.  For each method we downscale precipitation from the WRF ERA-Interim driven RCM simulations (50km) to the Maurer observed precipitation data (12km) from 1980 to 2010.  We use even years for training the models and odd years to evaluate the models’ performance. 

The first model uses a logistic method to predict the probability of non-zero precipitation and then linearly predicts precipitation intensity. The second model uses one random forest to predict the probability of precipitation for each cell and a second random forest to predict precipitation intensity. The third model uses a U-Net convolutional neural network. For each method, we consider computational efficiency, ease of implementation, and prediction accuracy for precipitation frequency and intensity. The linear predictions often cluster around the mean, the U-Net systematically under-predicts the magnitude of precipitation but captures interannual variations, and the random forest performs the best overall. However, all three models often fail to predict extreme values. We apply the random forest model to future climate simulations performed with the WRF RCM to produce detailed information about climate-change-induced trends in precipitation.

Mentors: Ethan Gutmann, Rachel McCrary

Slides