Landscape patterns of shrubification in the Siberian Low Arctic: A machine learning perspective

Abstract

Tundra shrub expansion is a central form of change in warming Arctic ecosystems, but the pace of shrubification varies across spatial scales, complicating efforts to understand its drivers and consequences. Here, we apply convolutional neural networks (CNNs) to very‐high resolution satellite image pairs acquired 10–15 years apart (circa 2005–2019) to identify spatio‐temporal patterns of tall shrub (> ~1.5 m height) expansion and their relationships to environmental covariates and antecedent shrub cover in three upland‐dominated Siberian Arctic landscapes. We developed human‐interpreted training datasets for CNN modelling of tall shrub occurrence and change for nearly 1 million 12 × 12 m image tiles using four canopy cover classes: Tundra lacking tall shrubs; Colonization with isolated shrubs; Open Shrub with discontinuous cover of mature shrubs; and Closed Shrub with dense cover of mature shrubs. F1 scores for the canopy cover maps ranged 0.83–0.92, and classification confidence was high (>0.8) in both time periods for 62% of image tiles. We evaluated canopy class occurrence with respect to landscape‐scale environmental covariates related to topography, insolation, wetness and proximity to established shrubs. We detected increases in tall shrub cover in all three landscapes, but the rate of increase varied substantially (+2.4 to 26.1% decade−1). Locally, the distribution of canopy cover classes was strongly influenced by topographically derived metrics of wetness and potential insolation. Shrub colonization was further conditioned by proximity to pre‐existing shrubs. Synthesis. We found that mature, long‐established shrubs (i.e. the Open Shrub and Closed Shrub classes) were closely linked to well‐drained landscape positions; however, closed shrub stands were more likely to occur on warmer, south‐facing slopes, while Open Shrub predominated on cooler slope aspects. Contemporary shrub colonization has occurred on flatter landscape positions and across wider gradients of insolation and moisture. In our study region, northward shrubline advance appears most closely tied to microsites with favourable drainage and edaphic conditions; subsequent seed production and ameliorating effects support canopy infilling, particularly on warmer hillslope aspects. The techniques employed here provide insights into the susceptibility of landscapes to future shrub expansion and help address geographic bias regarding Arctic shrubification and its environmental drivers.