Summer warming explains widespread but not uniform greening in the Arctic tundra biome

The Arctic tundra is rapidly warming, affecting climate feedbacks, wildlife, and human communities, yet pan-Arctic ecological change is difficult to quantify due to sparse long-term field measurements and dependence on coarse satellite data. Prior field studies show mixed vegetation responses to warming, from increases in plant cover and shrub dominance to little change or even declines. Satellite-based NDVI has been used to infer greenness trends, but coarse AVHRR data have cross-sensor calibration issues and spatial resolution too coarse to resolve heterogeneous tundra landscapes. This study aims to use higher-resolution Landsat data to: (1) quantify the extent of tundra greenness change in recent decades; (2) assess how interannual greenness tracks summer temperatures; (3) identify links between greenness trends and climate, permafrost, topography, land cover, and fire; and (4) evaluate correspondence between satellite greenness and field-based productivity. The overarching hypothesis is that summer warming has stimulated tundra plant productivity across much of the biome, but with spatial variability.

Since the 1980s, NDVI from AVHRR has indicated widespread Arctic greening with regional browning, but discrepancies among AVHRR products arise from cross-calibration challenges across many sensors and the coarse (~8 km) spatial resolution, which hampers attribution and comparison with field data. MODIS provides improvements but still coarser than many ecological processes. Field studies document increases in plant cover, height, biomass, and shrub dominance in some tundra regions, while others show stability or warming-induced growth declines. Previous Landsat-based studies demonstrated regional greenness changes, links to shrub expansion, and relationships with aboveground biomass, but a pan-Arctic assessment with Landsat integrating climate, permafrost, fire, and topography had not been performed. The literature underscores the need for higher-resolution, cross-calibrated satellite analyses aligned with field observations to clarify drivers of greening and browning.

• Study domain and data: 50,000 random sampling sites across the Arctic tundra biome. Vegetation greenness characterized by annual maximum summer NDVI (NDVI_max) from Landsat 5, 7, and 8 surface reflectance (30 m), extracted via Google Earth Engine.
• Time periods: Primary analyses for 1985–2016 (coverage ~64% of domain) and 2000–2016 (coverage ~96%, improved eastern Eurasia coverage).
• Preprocessing and uncertainty: Developed novel cross-sensor NDVI calibration and phenology-based modeling to estimate annual NDVI_max even with sparse summer observations. Propagated uncertainties from radiometric and cross-sensor calibration and NDVI_max estimation, as well as climate and field data uncertainties, using Monte Carlo simulations (n=10^3).
• Greenness trends and temperature covariation: For each site, computed annual NDVI_max and assessed trends with rank-based Mann–Kendall tests and Theil–Sen slopes; significance levels α=0.05/0.10. Quantified summer air temperature via Summer Warmth Index (SWI; sum of monthly means >0°C) using an ensemble of five temperature datasets. Assessed Spearman correlations between annual NDVI_max and SWI (current year and 2-year averages), including detrended series, within Monte Carlo framework to derive median r and 95% CIs.
• Spatial drivers (2000–2016): Built Random Forest classifiers to predict site-level NDVI_max trend class (browning, no trend, greening) using 20 predictors spanning climate (SWI level and trend), soil moisture (minimum summer, trend), permafrost (extent, ground temperature at 1 m, active layer thickness; 2003–2016 where available), thermokarst vulnerability, land cover (ESA CCI), fire (MODIS burned area 2001–2016), and topography (elevation, slope, aspect, roughness, position; TanDEM-X 90 m). Balanced classes by downsampling majority classes, removed highly correlated predictors (|r|>0.75), tuned models for out-of-bag accuracy, evaluated on held-out data, computed variable importance (mean decrease in accuracy) and partial dependence.
• Field validations: Compared Landsat NDVI_max with (1) graminoid ANPP (1990–2017, Bylot Island, clip harvests), (2) 22 shrub ring-width index (RWI) chronologies (Alnus, Salix, Betula) from six countries, using detrended NDVI_max and RWI series, and (3) ecosystem GPP from 11 eddy covariance flux towers (AON, FLUXNET). Assessed temporal and spatial correlations with Monte Carlo uncertainty (n=10^3).

• Biome-scale greening and warming: Mean Arctic NDVI_max increased by 7.3% [7.0, 7.7] from 1985–2016 and 3.6% [3.4, 3.7] from 2000–2016. SWI increased by 5.0°C [4.9, 5.1] (1985–2016) and 2.5°C [2.3, 2.7] (2000–2016). Annual mean Arctic NDVI_max and SWI anomalies were positively correlated (r=0.68 [0.66, 0.70] for 1985–2016; r=0.76 [0.73, 0.78] for 2000–2016), strengthening when using 2-year average SWI (r≈0.86–0.89). Lowest mean NDVI_max occurred in 1992 (Mount Pinatubo cooling), highest in 2012 and 2016.
• Site-level trends: From 1985–2016, greening occurred at 37.3% [36.3, 38.4] of sites, browning at 4.7% [4.4, 5.2], no trend at 58.0% [57.1, 58.7]; greening:browning ratio 7.9:1 [7.1, 8.7]. From 2000–2016, greening at 21.3% [20.8, 21.7], browning at 6.0% [5.8, 6.3] (ratio 3.6:1 [3.4, 3.8]). Greening more common in Low and Oro Arctic; browning less common but widely distributed.
• NDVI–temperature covariation at sites: 1985–2016 showed positive NDVI_max–SWI correlations at 28.2% [27.3, 29.1] of sites (mean r=0.41±0.06) and negative at 1.0% [0.8, 1.1] (mean r=−0.40±0.06). About 41% of greening sites had positive NDVI–SWI correlations; ~6.5% of browning sites had negative correlations.
• Spatial patterns: Extensive greening in western Eurasia (e.g., Gydan, southern Yamal) and North America (e.g., Ungava Peninsula, Northwest Territories, NW Nunavut) over 1985–2016; additional greening in eastern Eurasia (Chukotka, Yakutian mountains) over 2000–2016. Recent browning noted along southwestern Greenland coast.
• Drivers (Random Forests, 2000–2016): Cross-validated overall accuracy 55% [53, 58]; class accuracies: greening 70% [68, 73], browning 73% [70, 75] (chance=33.3%). Top predictors: change in SWI (2000–2016), annual mean soil temperature (1 m), early-2000s SWI, elevation, change in minimum summer soil moisture, change in annual mean soil temperature (2003–2016). Greening more probable at warm, higher-elevation sites with increases in summer air temperature, soil temperature, and soil moisture. Browning more probable at colder, lower-elevation sites experiencing cooling and drying. Greening probability declined and browning increased where early-2000s soil temperatures exceeded 0°C. Recent fires were rare (~1.1% of sites burned) and not an important pan-Arctic predictor.
• Field validation: NDVI_max correlated with graminoid ANPP on Bylot Island (rs=0.43 [0.24, 0.58]; stronger using 2-year mean NDVI_max, rs=0.68 [0.55, 0.78]). Across 22 shrub RWI chronologies, median rs=0.42 [0.34, 0.50] (range −0.12 to 0.84). Spatially, median annual NDVI_max correlated with GPP across 11 flux towers (rs=0.72 [0.54, 0.88]).

Findings support the hypothesis that summer warming has broadly stimulated tundra plant productivity, reflected in widespread greening and strong NDVI–temperature covariation, while acknowledging substantial spatial heterogeneity with many stable areas and some browning. Landsat’s higher spatial resolution captures heterogeneous change patterns that may be obscured in coarser AVHRR/MODIS data; however, inter-sensor differences lead to varying greening:browning ratios across datasets. Lack of greenness trends at many sites despite warming suggests additional constraints (e.g., low soil temperatures, nutrient or moisture limitations, genetic adaptation), or offsetting influences such as herbivory or mixed pixels with water/snow/bare ground. Browning, though uncommon, is linked in some regions to warming-induced drought stress, insect defoliation, increased browsing, local cooling/drying, and permafrost-related hydrologic change, extreme events, or development. Fires were rare during the study period and had limited pan-Arctic influence but may grow in importance with continued warming. The ecological changes have mixed climate feedbacks: greening may increase plant carbon uptake (negative feedback) but can reduce albedo and interact with permafrost carbon release (positive and mixed feedbacks). Vegetation shifts alter wildlife habitats, benefiting some species (e.g., moose, beaver) while potentially disadvantaging others (e.g., caribou/reindeer via lichen declines), with implications for northern communities. Results emphasize the need for integrated Earth system modeling and continued high-resolution monitoring to resolve net feedbacks and ecological impacts.

Using three decades of 30 m Landsat observations, the study provides a pan-Arctic assessment showing extensive but non-uniform tundra greening strongly associated with summer warming, increases in soil temperature, and soil moisture, and validated against graminoid, shrub, and ecosystem productivity metrics. While greening predominates, many areas show no trend and some show browning, reflecting diverse local drivers and constraints. The work highlights Landsat’s value in resolving heterogeneous tundra dynamics and underscores regional discrepancies with coarser satellite products. Future research should: (1) integrate newer higher-resolution sensors (e.g., Sentinel-2, Planet, WorldView) and UAV data with the Landsat record; (2) improve mapping of drivers (permafrost processes, hydrology, disturbance) at higher spatial/thematic resolution; (3) expand long-term field measurements across under-observed regions; and (4) use coupled Earth system models to assess net climate feedbacks of concurrent vegetation and permafrost changes.

• Spatial data gaps: Limited Landsat coverage prior to ~2000 in parts of eastern Eurasia restricted 1985–2016 analyses to ~64% of the Arctic domain.
• Observation frequency: Fewer clear-sky acquisitions at 30 m impede phenology characterization; addressed via modeling but residual biases may remain.
• Sensor calibration and NDVI limits: Despite cross-sensor calibration, uncertainties in radiometry and NDVI saturation or mixed-pixel effects persist, especially in patchy tundra with water/snow/bare ground.
• Attribution constraints: Random Forest analysis identifies associations, not causation; key predictors like permafrost variables are only available since 2003, limiting longer-term driver analyses.
• Temperature proxies: SWI represents summer air temperature; plant responses may lag or depend on other seasons and subsurface conditions.
• Trend detection power: Many sites show no significant trend or weak NDVI–SWI correlations at annual scales; local processes (herbivory, hydrology, extreme events) and data uncertainties can obscure relationships.
• Fire influence: Low fire frequency during 2001–2016 limits inference on fire impacts at pan-Arctic scale.