Testing Machine Learning for Regional Climate Applications in the Pacific Northwest

Water managers and end users need detailed estimates for future precipitation, but regional climate models (RCMs) produce predictions on a grid that is often too coarse to be useful. In this project, we compare three methods for statistical downscaling from the coarse RCM output in the American Pacific Northwest to the finer resolution end users need. We downscale the ERA-Interim WRF output, and we train all models using the Maurer observed precipitation for even years between 1980 and 2010. We then evaluate the models’ performance on odd years.

The first model uses a logistic method to predict the probability of non-zero precipitation for each cell, and then it linearly predicts precipitation intensity where the logistic model expects non-zero precipitation. The second model uses one random forest to predict the probability of precipitation for each cell and a second random forest to conditionally predict the precipitation intensity. All random forests use early-stopping based on the out-of-bag error rate. The final model is a U-net, a type of convolutional neural network. For each method, we consider computational efficiency, ease of implementation, and prediction accuracy for precipitation frequency and intensity. We find that the linear predictions often cluster around the mean, smoothing temporal variation and losing essential information. The U-net picks up much of the variation, but it systematically underpredicts precipitation. All three models struggle to predict extreme values. Although we can optimize further, analyzing these models is an important step toward understanding detailed information about climate-change-induced trends in future precipitation.