Lightning Prediction using Machine Learning and Explainable Artificial Intelligence

Machine learning (ML) has become increasingly popular in geosciences due to its ability to skillfully model complex processes. One commonly used ML method is eXtreme Gradient Boost (XGBoost) which uses a decision tree-based method to process tabular data. An XGBoost classification model creates decision trees such that each new tree attempts to improve on the errors of the previous tree, creating a more accurate model. The outputs of these trees are then combined to create the model prediction. Another commonly used ML method is U-Nets which use convolutional neural networks with a contracting path and an expanding path to analyze spatial information at multiple scales. This allows the models to understand more complex patterns that occur on multiple spatial scales. Previous work within the NWS Meteorological Development Laboratory has explored the use of ML methods such as Random Forests (RF) and Convolutional Neural Networks (CNNs) to predict lightning probability across the Contiguous United States using observed total lightning from Earth Networks and Vaisala lightning detection networks which is advected in time, lightning climatology, and High Resolution Rapid Refresh (HRRR) output. These ML predictions would then be used to improve the Localized Aviation Model Output Statistics (MOS) Program (LAMP) lightning probability guidance product. XGBoost was first used to attempt to improve on these previous ML methods. When using XGBoost with inputs from a single location, the model produced results similar to that of the RF model. When using inputs from the area around a point (“spatial context”), the model produced significant improvements over the single-point model with results similar to the spatially-aware CNN. In addition to model development, explainable artificial intelligence (xAI) methods which break down predictions into the contributions from each input were applied to the XGBoost model. These methods showed that the observed lightning input was the largest contributor for short-term forecasts while thermodynamic variables were the largest contributors for longer lead forecasts. The U-Net model was used to expand on the positive results from the spatial context model. The U-Net produced results with higher skill than the XGBoost spatial context model. xAI analysis on the U-Nets showed similar patterns to the XGBoost models and some additional understanding of larger-scale patterns such as atmospheric boundaries. In this presentation, the techniques used will be described and the results summarized.