Machine-Learning Downscaling of Extreme Heat Events in New York City

Motivated by the increased probability of extreme heat events in the future, we developed a method to downscale a regional forecast model to a fine scale to produce risk assessments for a range of stakeholders. On a local scale, heat is strongly influenced by land-surface and land-cover properties. New York City is a particularly complex environment where patterns of extreme heat are complicated by the pronounced urban heat island and sea-breeze dynamics. To capture local variability in extreme heat, we use the Weather Research and Forecasting (WRF) model at 1-km grid-spacing with the NoahMP land-surface model to simulate the effects of land-surface characteristics on the near-surface air temperature. To extend this analysis to variability on the city-block scale, we then use several predictors from the National Land Cover Database (NLCD) as well as latitude and longitude to downscale the WRF output to 30-meter resolution using random forests. The random forest technique (Breiman 2001) is a flexible and powerful machine learning tool well-suited to this purpose. We model and downscale two historical extreme heat events in New York City and compare these results to satellite-derived surface temperatures and in situ meteorological station measurements. The observed fields validate the downscaled fields, and verifications confirm the utility of this approach for hyperlocal projections of historical and future heat waves.