Abstract Storm damage in forests of Central Europe causes severe economic losses and alters these ecosystems significantly. There are many approaches to analyze past damage and to predict the future damage, often relying on data from field surveys. With the current data abundance there is potential in using existing datasets to develop transferable methodologies for modeling storm damage. We hypothesized that the commonly available datasets, with a large spatial extent and addressing a variety of damage-relevant conditions, could create accurate prediction models, especially with variables generated at different spatial resolutions. We created random forest models using four variable groups: topography, vegetation structure, soil and gust speed, for the study site Tharandt Forest, Germany, a typical Norway spruce dominated even-aged forest in temperate Europe. Our models reached a good level of accuracy in predicting damage for the storms Herwart (2017) and Friederike (2018), with areas under the curve of 95.8% and 77.6%, respectively, showing the merits of our approach. We found that variable groups had varying contribution to model accuracy, with topography variables contributing the most. Additionally, we confirmed that creating topographic variables at different spatial resolutions was key for accurate damage classification. Finally, we used the models to run a simulation in our study site and estimated about 30% of the forest area being damaged over the next 60 years. The results from this approach could be used as an indicator for potential storm damage in other regions of the world, and support decision making and maintenance of forest ecosystems for the present and the future.
Gliksman et al. (Tue,) studied this question.