| dc.description.abstract | Flash floods, tornadoes, damaging winds, and large hail are costly and difficult to predict, even for state-of-the-art, high-resolution numerical weather prediction (NWP) systems. Current operational NWP ensembles have a variety of shortcomings: they are under-dispersive for precipitation, contain biases in precipitation magnitude and convection placement, have suboptimal forecast reliability, and use horizontal grid-spacing too coarse to explicitly depict some high-impact hazards (e.g., severe hail and tornadoes). Thus, post-processing techniques are required to obtain skillful probabilistic hazard forecasts from raw NWP ensemble guidance. Common post-processing methods include the use of proxies (e.g., climatologically large values of 2-5 km updraft helicity; UH2-5km) to represent simulated high-impact weather events and/or the use of spatially smoothed raw ensemble probabilities to improve forecast reliability. However, these methods use limited data and thus tend to be suboptimal. In this dissertation, I develop and analyze a random forest- (RF-) based procedure for obtaining more skillful precipitation and severe weather probabilistic forecasts for next-day lead times (i.e., 12-36 h forecasts valid from 1200 UTC – 1200 UTC). While past studies have used RFs to better predict high-impact weather, my RF procedure is unique because it uses temporally-aggregated, spatially-upscaled, point-based ensemble forecast predictors over the full contiguous United States (CONUS). This method of generating predictors is relatively simple but skillfully accounts for uncertainties in simulated convection timing and placement.
For precipitation and severe weather hazard prediction, I show that my RF procedure improves forecast reliability and resolution relative to top-performing (human and non-human) baseline forecasts. I find that RF post-processing is most beneficial for convection-parameterizing ensembles (which have more initial biases than convection-allowing ensembles) and more-common events (e.g., lighter precipitation thresholds and severe wind and hail compared to tornadoes). For precipitation, I find that RF-based post-processing reduces spatial biases and note that a season of training data is sufficient to produce skillful probabilistic precipitation forecasts for thresholds up to 3-inches. For severe weather, I show that RF-based forecasts have verification metrics similar to or better than corresponding Storm Prediction Center (SPC) day-1 human forecasts for most hazards, seasons, and regions. By discretizing RF forecast probabilities and making SPC probabilities continuous, I show that this result is only partly due to the ability of RFs to generate continuous forecast probabilities.
Through RF sensitivity tests, I find that ensemble mean (EM) predictors are more skillful than individual member (IM) predictors for severe weather forecasting, since EM predictors contain less noise. By conducting additional sensitivity tests and using the Tree Interpreter (TI) Python module, I find that storm predictors are most important for severe weather prediction, followed by index-based predictors, although I note that RFs using both storm and index-based predictors are most skillful. With TI analysis, I show that RFs emphasize different and physically-relevant predictors for each hazard. Further, I demonstrate that RFs learn to implicitly “weigh” multiple appropriate storm and index variables at and near the point of prediction, suggesting that RFs learn to account for model error. Importantly, my work shows that RFs are not constrained by the exceedance (or non-exceedance) of a simple UH2-5km threshold at one point and suggests that RFs can discern between similar ensemble forecast UH2-5km values as well as the same UH2-5km value in different environments. | en_US |
