Recent acceleration of shrub expansion in Siberian tundra detected at high resolution using convolutional neural networks

Shrubification is among the most conspicuous and pervasive vegetation changes being observed in Arctic ecosystems and is expected to continue in the future. We applied convolutional neural networks (CNN) to map changes in tall shrub cover since the early 2000s in very-high resolution (VHR; < 1 m) commercial satellite imagery for Low Arctic ecotones of northwestern Siberia. The study landscapes include extensive areas of patterned ground (sorted circles) with a known history of shrub increase dating to the mid-1960s. The VHR satellite constellation has grown and acquired data with ever-increasing frequency and detail, enabling comparative analysis of images from sensors with similar spatial and spectral resolutions across time periods long enough to detect meaningful changes in shrub abundance and permafrost geomorphology. In addition, the development of high resolution, wall-to-wall spatial data products for landscape properties including topography, soils, disturbance, permafrost characteristics, and bedrock geology permit analysis of landscape covariates of shrub expansion. Our ultimate vision is to develop workflows to quantitatively monitor shrub distribution and related successional changes in the circumpolar Low Arctic and adjacent forest-tundra ecotone. Here we seek to answer the following research questions: Are CNNs effective at discriminating tall shrubs and changes in their cover in VHR imagery over decadal timescales? Has the pace of shrubification and related successional changes accelerated in recent decades at the study landscapes? What is the relative importance of landscape covariates, such as topography, periglacial geomorphology, snow properties, and disturbance regime, in influencing observed patterns of change (or stability) in shrub cover?