Attribution of forest disturbance agents in the European Alps: a multidecadal analysis based on Landsat time series

Alpine forests have been shaped by human and natural disturbances for millennia, resulting in highly heterogeneous and fragmented landscapes. Interactions between land use change and natural disturbances have been observed in recent decades. In addition, climate change is altering the disturbance regime in many forest ecosystems, including those in the Alps, by increasing the severity, extent and frequency of natural disturbance events. In this study, we aimed to attribute the disturbance agent to forest patches that experienced stand-replacing and non-stand-replacing events during the last four decades in the European Alps. In particular, we considered both natural disturbance agents, i.e., wind, fire, snow, insects, ice, and drought, and human activities. The latter included both primary and secondary disturbances, i.e., salvage logging following a natural disturbance. We trained an eXtreme Gradient Boosting (XGBoost) machine learning model using disturbed forest patches detected annually by an automated algorithm based on Landsat time series from 1984 to 2022. We obtained information on the disturbance agent using both historical field data from several European countries and visual interpretation of remote sensing data, e.g., Landsat imagery, aerial orthophotos, and high-resolution satellite imagery. We built the final classification model after selecting predictor variables from several disturbance, topography, patch, and climate-related metrics. Preliminary results showed that the model had good predictive performance, as highlighted by the accuracy metrics obtained from a fivefold cross-validation approach, i.e., Cohen’s Kappa equal to 0.81 and balanced accuracy of 0.82. The elevation range, pre-disturbance spectral values, the climate moisture index, and the range of the spectral change magnitude of each disturbed forest patch were among the most important variables. In particular, the elevation range emerged as a key predictor for discriminating between natural and anthropogenic disturbances. Similarly, pre-disturbance spectral values were important for distinguishing between certain natural disturbances, such as windthrows and snow avalanches. Spatially explicit results from this study are expected to allow a thorough characterisation of the changes in disturbance regimes in the European Alps that have occurred over the last four decades, and to provide useful information on the main drivers that have determined these recent shifts.