Mapping forest disturbance and recovery for forest dynamics over large areas using Landsat time-series remote sensing

Sustainable forest management requires consistent and simple approaches for characterizing forest changes through time and space at the landscape scale. Landsat satellite data, with its long archive and comprehensive spatial, temporal and spectral detail, could enable us to achieve this goal. This study develops a consistent approach for mapping both disturbance and recovery for forest dynamic estimation across large areas over a 30 year period (1988 to 2016) using Landsat time series data. We analyzed dynamic Eucalypt/ Sclerophyll public forests in south eastern Australia which have been impacted by a series of disturbances including fire and logging over the last 30 years. We first prepared annual satellite composites and fitted spectral time series trajectories on a per-pixel basis using the LandTrendr algorithm, from which we derived a range of spatial disturbance and recovery metrics. We then simultaneously modeled disturbance and consequent recovery levels using the Random Forest classifier. Using derived change information and a one-off forest cover dataset, we estimated change in forest extent throughout the time series. Disturbance and consequent recovery were simultaneously detected with an overall accuracy of 80.2%, while the model of change levels classification obtained an overall accuracy of 76.5%. Over the 30 year period, approximately 49.5% of the study area was disturbed, 92% of which has fully recovered. Forest extent was found to be quite dynamics throughout the time period and comprised between 80.2% to 88.3% of public forest estate.