Heterogeneous long-term changes in larch forest and shrubland cover in the Kolyma lowland are not captured by coarser-scale greening trends

Abstract Changes in shrub and tree cover concurrent with rising air temperatures are a widespread phenomenon in Arctic–Boreal ecosystems. The expansion of tall shrubs and trees can alter ground thermal regimes and soil moisture impacting permafrost and biogeochemical cycling. Changes in shrub and tree cover can be difficult to characterize with limited in-situ observations and moderate/coarse resolution satellite imagery, thereby posing challenges in disentangling changes in vegetation growth from shifts in vegetation composition. We pair high resolution historical (KeyHole9 1971) and current satellite imagery (WorldView-3 2020) with a convolutional neural network approach to predict forest, shrubland, and surface water cover within a region of the Kolyma lowland (171 km 2 ) in eastern Siberia. The overall accuracy of the predictions was 0.90 for 1971 and 0.92 for 2020. We found an overall net increase in shrubland cover of 14 km 2 (8% of study extent) and little net change in forest cover, but changes in both land cover classes were highly heterogenous across the landscape. Increases in shrubland cover were highest in proximity to surface water (<100 m) and in close proximity to areas with stable shrubland cover. We found that changes in shrubland and forest cover did not correspond with trends in vegetation greenness (i.e., NDVI) derived from moderate resolution satellite data time series, which were fairly uniform among the land cover classes. Our findings highlight that ongoing land cover change in Siberian lowlands is highly heterogeneous and the need for a better quantification of the drivers and consequences of landscape change in these carbon- and ice- rich permafrost ecosystems.