Developing a Machine Learning Model to Predict Dry Thunderstorms Associated with Haboobs

Convectively driven dust storms, or haboobs, are complex meteorological phenomena that are difficult to predict. They require strong, organized outflows that sufficiently outrun the precipitation core to loft dust while sustaining repeated convection in their wake. In this work, we aimed to develop a machine learning model capable of predicting dry thunderstorms associated with haboobs. An inventory of haboobs and windstorms (storms where no dust was reported) from 2013-2021 was gathered for the study location of Arizona, USA. 2021 was set aside to be used as an independent test set. Various atmospheric variables obtained from a numerical weather prediction model (NWP) were tied to each event and used as features. In previous research, we investigated potential variables that could be utilized as additional features for the model. These features included NASA Land Information System (LIS) data parameters, active dust source areas from the NRL Dust Source Database for North America, atmospheric sounding variables, and a collection of engineered features. In this work, we take those features and feed them to different machine learning architectures to gauge whether a model can differentiate between a haboob and a windstorm (no dust). The four machine learning architectures that we applied were Random Forest, Weighted Random Forest, Multilayer Perceptron (MLP), and XGBoost. Different experiments were run using the aforementioned machine learning architectures. These experiments are defined by the subset of features that were used as inputs to the model such that we could investigate how different input features affected the model performance. Cross-validation was used to determine whether the model was generalizable or not. We found that some experiments led to an increase in performance, however no experiment necessarily performed well. The same goes for the choice of model architecture. Surprisingly, while the LIS parameters showed a noticeable difference in distribution between haboobs and windstorms, the addition of these parameters as features did not improve model performance as much as we expected. Based on the cross-validation results, for some years a certain model architecture or input feature subset would shine, while in other years they would perform poorly. This was especially true for imbalanced years, such as 2015 and 2017, where the number of windstorms greatly outweighed the number of haboobs. This led to low Critical Success Index (CSI) scores and high False Alarm Rate (FAR) scores. We believe that finding and adding more events, specifically haboob events, would improve model performance. A more thorough feature selection process may also be beneficial. More research may need to be done towards understanding the difference between haboobs and windstorms and discovering what features would better discriminate between the events.