Open Data Science Europe Workshop 2021

Windstorms remain one of the most disturbing factors of European forest ecosystems. Nowadays, our deep understanding of all the major drivers standing behind observed changes in forests is essential to improve prediction models. This is important for various scientific disciplines, decision makers on different levels and forest management planners. New statistical learning techniques help with analyses of massive data objects and offer sophisticated explaining tools that help to understand complex models. In the present study, we combined several data sets on tree features, bioclimatic and geomorphic conditions, and the level of forest damage in the Sudety Mountains over the period 2004-2010. We tested four scenarios under five classification model frameworks: logistic regression (binomial GLM), random forest (RF), support vector machines (SVM), neural networks (NN), and gradient boosted modelling (GBM). All models except GLM offer similar level of predictive power (AUC ~ 0.7). GBM and RF feature the best predictive power indicated by AUC = 0.717, while RF model reached AUC = 0.715. Tree volume and age are the most important predictors. Less important are climate and geomorphic variables. The same models performed less accurately for test data from the period 2004-2006. Forest damage probability maps based on forest data from 2020 show overall lower level of damage probability as compared to the end of 20th and the beginning of 21st century. To sum up, using only 11 variables based on the open source datasets, we were able to obtain predictive models of good accuracy.