Reconstructing forest dynamics in the European Alps through a high-dimensional analysis based on Landsat time series

Abstract

Natural disturbances are increasingly threatening forest ecosystems due to climate change globally. InEurope, disturbance regimes have intensified over the last decades, leading to increased size, frequencyand severity of disturbance events. Satellite remote sensing data acquired over the past decades are crucialfor assessing changes in disturbance regimes as they provide wall-to-wall spatial information from thelandscape to the global scale. In particular, Landsat imagery has been continuously acquired since 1984,and it offers an unprecedented opportunity for mapping land cover changes thanks to its spatial andspectral consistency. Following the opening of the USGS Landsat archive, dense time series have beenexploited through automated algorithms for targeting forest dynamics. Currently, the most widely usedalgorithms aim to detect abrupt and gradual changes by performing a temporal segmentation of Landsattime series at the pixel level. The sensitivity of automated algorithms has been enhanced by includingmultiple spectral and spatial information in the time series though their combined usage is still limited.Here, we present an automated algorithm for detecting forest dynamics named High-dimensional detectionof Land Dynamics (HILANDYN), which exploits the temporal, spatial and spectral dimensions of inter-annualLandsat time series. We tested HILANDYN to map forest disturbances that occurred during the last fourdecades in the European Alps. HILANDYN builds upon a statistical procedure for detecting changepoints inhigh-dimensional time series through a bottom-up segmentation procedure. Our results showed that thealgorithm is sensitive toward a wide range of disturbance severities and can detect stand-replacing events,e.g. windthrows and wildfires, and non-stand-replacing ones, e.g. insect outbreaks and drought-induceddieback. Moreover, we were able to map disturbances occurring in consecutive years, such as windthrowsfollowed by salvage logging. We obtained the best results in terms of accuracy metrics using a combinationof original bands and indices that included the heterogeneous spectral information provided by themultispectral sensors of the Landsat missions. In particular, we achieved an F1 score equal to 83% for thedisturbed class, corresponding to a user’s accuracy of 84,3% and a producer’s accuracy of 82%. Accuratedisturbance maps of the European Alps will enable a thorough analysis of the shifts in the disturbanceregimes over the last four decades, alongside the assessment of forest recovery patterns under differentmanagement practices and environmental conditions.