Mapping storm damage across the world’s forests

Storms are natural weather events, varying greatly in frequency and intensity across the world. They are characterised by strong winds that can disturb forests via tree defoliation, breakage and uprooting. Despite close monitoring of forest disturbance occurrences in recent years, we still lack information on the contribution of storms in driving forest dynamics worldwide. Here, we build a machine learning classification model to identify wind-related forest disturbances at the global scale. Forest disturbance patches detected between 2002 and 2014 were associated with multiple covariates as potential indicators of wind damage. These covariates include structural metrics inherent to the disturbances, such as patch size, elongation and spatiotemporal clustering, as well as environmental variables describing topography, weather, and soil conditions. We used these data for 20,000 reference patches (10,000 wind and 10,000 non-wind), widely distributed across forest biomes, to train a random forest classifier. Cross-validation with 20,000 other reference patches over 10 runs showed that the model achieved satisfactory performance scores. It yielded omission errors of ca. 20% and commission errors of ca. 5%, mostly associated with harvest, selective logging and biotic outbreaks.  The most important variable was maximum wind speed, followed by patch temporal clustering. Model deployment at the global scale will provide quantitative insights into storm damage biogeography, regime characteristics and relative contributions to global carbon fluxes and forest dynamics.