A Russian-American team of researchers used convolutional neural networks to study high-resolution satellite images of three Siberian landscapes taken at intervals of 10-15 years
The neural network analysed nearly a million 12×12 m plots and found that the rate of shrub spread varies from 2.4% to 26.1% per decade, depending on local conditions. The study showed that mature shrubs prefer well-drained slopes, while new colonisation occurs on flatter areas, which opens up opportunities for predicting further shrubification of the Arctic under climate change conditions
In August 2025, the Journal of Ecology published an article entitled ‘Landscape patterns of shrubification in the Siberian Low Arctic,’ which analysed the shrubification of the tundra in the southern part of the Siberian Arctic using high-resolution satellite images taken at 10–15-year intervals. A Russian-American team involving researchers from the MLLE FGT and the IPEE RAS trained a convolutional neural network to recognise the stages of development of shrubby alder (Alnus alnobetula ssp. fruticosa) – from the absence of shrubs to a closed canopy – and analysed nearly a million 12×12 m plots, which made it possible to speed up mapping hundreds of times compared to traditional methods without any loss of quality.
The Arctic is a region with some of the most extensive evidence of the impact of climate change, particularly rising temperatures. These changes are easy to observe as ecosystems and landscapes are restructured. The rapid spread of tall shrubs into areas previously occupied by low-growing tundra plants is an example of this and one of the key components of the observed ‘tundra greening’. The spread of shrubs includes both the closure of existing shrub canopies (i.e. their filling) and the colonisation of new areas. Shrub encroachment has far-reaching consequences for the biophysical properties and functioning of ecosystems. Tall shrubs significantly alter the surface energy balance through their impact on albedo, snow properties and soil shading. Increased snow retention in shrub thickets can significantly raise winter soil temperatures. Ecosystems with high shrub cover can become sources of carbon at lower temperatures under high light conditions. This poses a potential threat of enhancing the greenhouse effect as the Arctic continues to warm and become more shrubby.
Researchers have shown that shrub encroachment is occurring in all three study areas, but its rate varies greatly, ranging from 2.4% to 26.1% of the area per decade, depending on local conditions. The main factors in the spread of shrubs were heat supply and drainage: mature shrub communities form on well-drained, warmed slopes and near existing thickets, while the initial stages of colonisation are more common on flatter and slightly wetter surfaces. The use of a four-class system (tundra, colonisation, open and mature shrub communities) instead of a binary division allowed for a more detailed description of the ecological mechanisms and trajectories of shrub encroachment.
The proposed methodological approach demonstrates how machine learning technologies and archived satellite data can be used for retrospective and operational monitoring of vegetation over large areas with high spatial resolution. The procedure combines AI-based satellite image processing, extraction of ecological factors from a digital elevation model, and their statistical evaluation, and can be extended to larger regions of the Arctic. The data obtained are important for climatic, biogeochemical, and permafrost models, as the transition from grass cover to tall, closed shrubs radically changes environmental conditions and affects many components of the ecosystem.
The interim results of this study were presented at the Laboratory Seminar in May 2024. The work was carried out as part of the HSE University's fundamental research programme (International Laboratory of Landscape Ecology).
Derkacheva A., Frost G.V., Epstein H. E., Ermokhina K. (2025). Landscape patterns of shrubification in the Siberian Low Arctic: A machine learning perspective. Journal of Ecology, 113(10), 2813–2831. https://doi.org/10.1111/1365-2745.70129