Machine learning forecasts of storm damages, forest fires, pedestrian slippings, and vehicle accidents for operational early warnings of weather impacts

Accurately predicting the impacts of severe weather events is crucial for improving meteorological warnings and aiding weather-vulnerable organizations in decision making. However, achieving a high-quality estimate of impacts is difficult because of complex and detail-sensitive interactions between the weather and the impacted quantity. For instance, forest damages caused by a storm are influenced not only by the strength of gusts but also by factors such as ground frost, tree leaf maturity, and wind direction. 

 

To address this challenge, the SILVA project developed a nation-wide weather-impact database with over 10 individual impact datasets, which were then utilized as targets for machine learning modeling. Storm damage clearance tasks, wildfire fighting tasks, traffic accidents, and pedestrian slipping accidents stored in the database were aggregated to the counties of Finland, and a separate gradient boosting model was fitted between the historical ECMWF HRES weather forecasts and the impact data in each county.

 

Subsequently, a reliable production system was set up to produce forecasts automatically twice per day. Four risk classes were determined based on the extremeness of the forecasted incidents, taking into account the seasonal differences, and the Meteoalarm-like four step color scheme was used to visualize the forecasts as traditional warning maps and as time series. Two of the products primarily describe weather risks in summer (storm damage clearance tasks and wildfire fighting tasks), while the other two are more useful for warning the risks in winter (traffic accidents and pedestrian slipping accidents). 

 

The impact forecasts were tested by numerous end-users during a seven-month pilot phase. Both the feedback from the pilot participants and the numerical validation results clearly indicate the value of the products. The selected modeling method, gradient boosting, was found to be effective in taking into account the nonlinear and complex interactions when explaining the variability of the impact data. 

 

The forecasting system is currently undergoing further development in the Europe Horizon CREXDATA project, with efforts focused on exploring possibilities for expanding the geographical areas of application and incorporating new impact datasets.