Deep Learning Techniques Improve Prediction of Severe Thunderstorm Events

By Lina Sorg

Extreme rainfall events, such as severe thunderstorms, can lead to flash flooding and cause significant damage to both life and infrastructure. Effective forecasting of such events is therefore a high-stakes and complex mathematical modeling problem. “It’s really important to predict severe events in advance,” Sabrina Guastavino of the University of Genoa said. “This problem is very challenging, especially when we consider areas that are characterized by a complex, steep orography close to the coastline, where intense precipitation can be enhanced by specific topographic features.” During a minisymposium presentation at the 2023 SIAM Conference on Computational Science and Engineering, which took place this week in Amsterdam, the Netherlands, Guastavino used deep learning and radar data to predict severe thunderstorms before they occur. She concentrated her efforts on the Italian region of Liguria, which is located on the northwest Mediterranean Sea and characterized by mountains that are only a few kilometers (km) from the coast (see Figure 1).

Guastavino opened her presentation by referencing an unexpected November 2011 flash flood event in Genoa, Italy, that killed seven people and produced substantial economic, monetary, and emotional damages. The use of nowcasting techniques—short-term weather predictions based on a detailed analysis of current conditions—can generate reliable, accurate forecasts that reduce the consequences of these events. Such techniques fall into one of two categories: (i) numerical methods for the solution of dynamic model equations or (ii) data-driven techniques like deep learning. Although the former approach is more prevalent in the existing literature, Guastavino focused on image-based deep learning techniques that account for the presence of lightning. “The extreme event is defined on the basis of a certain level of precipitation and lightning density,” she said.

Deep learning for classification problems has three components:

1. A fixed neural network
2. A historical data set for training and validation
3. A score for the assessment of performance.

“A neural network is a parametric function that approximates the map connecting data to the event probability,” Guastavino said. She trained her deep neural network (DNN) on the training set by solving a minimization problem and minimizing the loss function. Because weight optimization in the training phase is an iterative process, Guastavino selected the best iterates on the validation set and tested the DNN on a test set.

In binary classification, computation of the confusion matrix evaluates the goodness of prediction. However, this method does not quite work in Guastavino’s case because her data set is quite imbalanced (only 1.5 percent of events end up being classified as “extreme”). Classical skill scores do not account for a prediction’s temporal distribution and can hence yield false positives. “In the case of an imbalanced data set, a more suitable skill score is the value-weighted true skill statistic (wTSS),” Guastavino said. This framework better handles false positives and negatives. She then trained her model with the epoch that had the highest skill score on the validation set. “In this way we are proposing an ensemble procedure,” she said.

Next, Guastavino addressed the data. Her input data is a time series of multichannel radar reflectivity images at two, three, and five km above sea level. An event is categorized as “extreme” if the modified conditional merging data for rainfall is more than 50 millimeters per hour in three or more contiguous pixels over the reference area, and there are at least 10 lightnings within 10 minutes in a five-km neighborhood.

After further explaining her DNN, Guastavino discussed the model outcomes. The input data ranges from July 2018 to December 2019: training took place from July 2018 to June 2019, validation occurred from July to October 2019, and testing transpired from October to December 2019. Use of the wTSS ensemble strategy yielded strong results (see Figure 2).

Ultimately, Guastavino used deep learning and radar imaging to recast severe thunderstorm prediction as a binary classification problem. She determined that value-weighted skill scores effectively evaluate predictions over time and demonstrated that an ensemble approach for epoch selection yields better results than classical methods.

Lina Sorg is the managing editor of SIAM News.