SIParCS 2022 - Eliot Kim
XAI and Active Learning for Predicting Winter Weather Precipitation Type
Winter weather precipitation events threaten economic activity, encumber day-to-day life and at worst, endanger human lives. However, quantifying winter precipitation is difficult due to biases in precipitation observations and the inability of numerical models to fully capture complex atmospheric phenomena. This project harnesses the predictive power of deep learning methods by training multi-layer perceptron (MLP) neural networks to predict precipitation type. We also employ an Evidential MLP to output the uncertainty associated with each prediction. Atmospheric profiles were obtained from the RAP (Rapid Refresh) model as inputs to the neural networks. ASOS (fixed ground monitors) and mPING (crowd-sourced) rain, snow, sleet, and freezing rain observations comprise the model targets. We utilized Explainable AI (XAI) methods to interpret the “black-box” neural networks and verify the physical consistency of the information encoded by the models. XAI results show that complex MLP and simple Evidential MLP models are able to encode information consistent with physical properties of precipitation formation through the atmospheric column. Active Learning is implemented to improve model accuracy using fewer training examples, by training on the most uncertain data determined by the Evidential MLP. Using an initial Active Learning pipeline, we achieved a nearly three-fold increase in sleet prediction accuracy using just 20% of the full dataset. Further research aims to conduct extensive active learning experiments and benchmarking, improve the spatiotemporal consistency of precipitation type predictions, and ultimately facilitate more efficient and effective deployment of deep learning models for winter precipitation prediction.
Mentors: John Schreck, David John Gagne
Slides and poster