Regional detection, characterization, and attribution of annual forest change from 1984 to 2012 using Landsat-derived time-series metrics

The examination of annual, gap-free, surface reflectance, image composites over large areas, made possible by free and open access to Landsat imagery, allows for the capture of both stand replacing and non-stand-replacing forest change. Furthermore, the spatial and temporal information extracted from the time-series data enables the attribution of various forest change types over large areas. In this paper we apply spectral trend analysis of Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM +) data from 1984 to 2012 to detect, characterize, and attribute forest changes in the province of Saskatchewan, Canada. Change detection is performed using breakpoint analysis of the spectral trends and change events are characterized using a set of metrics derived from an image time series that relate the temporal, spectral, and geometrical properties on an object basis. Change objects are attributed to a change type (i.e., fire, harvesting, road, and non-stand-replacing changes) using a Random Forest classifier. Non-stand-replacing changes are generally low magnitude, punctual, trend anomalies that relate year-on-year ephemeral changes that do not lead to a change in land cover class (i.e., phenology, insects, water stress). The results confirm that land cover changes are detected with high overall accuracy (92.2%), with the majority of changes labeled to the correct occurrence year (91.1%) or within ± 1 year (98.7%). Characterization of changes enables accurate attribution both at the object (91.6%) and area levels (98%), with fire and harvesting events the most successfully attributed (commission error < 10%), and roads, the most challenging to attribute correctly (commission error > 13%). Our approach, prototyped over the forested area of Saskatchewan, has enabled a highly automated and systematic depiction of a 30-year history of forest change, providing otherwise unavailable insights on disturbance trends including spatial, temporal, and categorical characteristics. The generation and application of metrics that relate a range of change characteristics allow for depiction of a broad range of change events, types, and conditions.