Satellite data reveal resilience loss across global ecoregions

A global analysis of 22 years of satellite vegetation records has identified statistical early-warning signals in 27% of Earth’s terrestrial ecoregions, covering 35.3% of the global terrestrial area analyzed. The study combined spatial clustering, directional vegetation trends and temporal autocorrelation in monthly kernel normalized difference vegetation index data, identifying ecoregions where vegetation dynamics increasingly differ between coherent subregions while recovery becomes slower. These are statistical warning signals, not evidence that ecosystems have crossed collapse thresholds, and the researchers did not directly observe or predict a collapse event.

The work used 266 monthly observations from July 2002 through August 2024, derived from the MODIS/Aqua Vegetation Indices Monthly L3 Global 0.05° product. For every valid pixel, the investigators fitted separate linear trends to vegetation observations for each calendar month, producing a 12-value vector that described the direction and strength of change through the year. These vectors captured seasonal patterns that would be obscured if all months were pooled. The researchers then used K-means clustering, with the number of candidate groups selected by the Calinski-Harabasz criterion, to divide each ecoregion into spatially coherent subregions with distinct temporal behavior. Ecoregions whose clusters changed too little between the first and last years were classified as spatially uniform and excluded from the resilience analysis.

The study’s central indicator was lag-1 temporal autocorrelation, or TAC, defined as the correlation between consecutive monthly values in a detrended vegetation time series. In the framework of critical slowing down, a system approaching a bifurcation may recover more slowly from perturbations, retain a longer memory of earlier deviations and therefore show increasing autocorrelation. The researchers calculated TAC separately for the first and last 133 months and defined the change as ΔTAC, with positive values interpreted as declining resilience. They removed the mean and long-term trend before this calculation so that directional changes or abrupt disturbances would not automatically inflate the metric. A 1° by 1° Gaussian filter was then applied to the resulting ΔTAC maps to reduce pixel-scale noise, and cluster-level values were calculated as arithmetic means across valid pixels.

Directional vegetation trends and resilience were treated as complementary rather than interchangeable signals. A sustained increase or decrease in vegetation index describes directional change in the average ecosystem state, while increasing autocorrelation describes slower recovery around the local state. The authors also required spatially coherent clusters, sufficient differentiation between them and at least one cluster with statistically significant monthly trends in at least three calendar months. A monthly trend was considered significant when the absolute value of its t statistic exceeded 2.086, the two-sided 95% critical value for 20 degrees of freedom. An ecoregion was retained only if at least one cluster had a positive mean ΔTAC. This combination allowed the analysis to distinguish directional reorganization and resilience loss from either signal alone. A repeated analysis using conventional NDVI rather than kernel NDVI produced practically the same numbers of statistically significant ecoregions, suggesting that the main results were not dependent on the transformed vegetation index.

Two detailed case studies illustrated how the method works. In the Angolan wet miombo woodlands, clustering separated the ecoregion into two regions covering 79% and 21% of its area. The smaller red cluster showed declining kNDVI in every calendar month, with each monthly trend significant at the 95% confidence level, whereas the larger cluster showed a small increase from April through June. TAC rose by 3.3% for the ecoregion as a whole, 11.1% for the declining red cluster and 0.3% for the larger blue cluster. In the Madeira-Tapajos moist forests of the Amazon basin, a browning cluster covering 14% of the ecoregion had significant trends in eight months. TAC increased by 4.1% for the whole ecoregion and by 4.9% and 3.8% for the two clusters. In both examples, the spatial separation suggested that an ecoregion can contain subregions following different ecological trajectories even when its overall monthly signal appears comparatively uniform.

Four additional ecoregions illustrated other combinations of greening, browning and resilience change. In Madagascar’s subhumid forests, 36% of the ecoregion area showed significant browning trends in 11 months, while TAC rose by 5.9% overall and by 4.7% and 7.7% in the two clusters. Significant greening appeared in 43% of the east Deccan moist deciduous forests, 47% of the Deccan thorn scrub forests and 35% of the Western European broadleaf forests, with at least seven significant months in each case. Rising TAC accompanied the greening in these examples. The authors cautioned that vegetation gains can reflect biome shifts, woody encroachment or other structural reorganization rather than ecological stabilization. Both increasing and decreasing vegetation trends were therefore treated as potentially relevant when they occurred alongside a resilience signal and spatially coherent differentiation.

Across the global analysis, the 27% of ecoregions meeting the criteria were distributed across tropical South America, Africa and Asia, Central and South Asia, northwestern North America, central Europe and high-latitude regions. Most affected ecoregions in Asia and Europe were expected to experience greening, whereas those in South America, tropical Africa and northwestern North America were mostly expected to experience browning. The researchers estimated 20-year changes by multiplying statistically significant monthly trends by 20 years. They noted that an average kNDVI difference of roughly 0.2 could represent a transition from rainforest to savanna or from Mediterranean forest to desert, illustrating how persistent but individually small trends can accumulate into major ecological differences. This interpretation concerns the study’s trend-derived estimates rather than observed transitions in every flagged ecoregion.

Comparisons with climate, biome and human-modification classifications showed the warning signals across multiple environmental settings. Approximately 45% of vegetated area in cold and polar climates met the criteria, compared with about 40% in tropical areas and roughly 20% in arid and temperate climates. The polar percentages referred only to vegetated pixels because the analysis could not detect signals where vegetation was absent or permanently ice-covered. In several biome groups—including tropical and subtropical moist broadleaf forests, temperate conifer forests, boreal forests, tropical and subtropical grasslands, savannas and shrublands, flooded and montane grasslands, and mangroves—40% to 65% of the area met the criteria. Human-modification classes divided affected ecoregions into locally and globally affected categories. Twenty-six percent of the area in locally affected ecoregions and 41% in globally affected ecoregions showed warning signals, and 74% of the total flagged area belonged to globally affected ecoregions.

Despite this broad pattern, the study cannot assign the observed changes to particular causes. Vegetation indices are indirect proxies for ecosystem state and may be influenced by canopy structure, background reflectance and phenology. Temporal autocorrelation can also be altered by external forcing and stochastic variability, so increasing TAC does not uniquely demonstrate proximity to a tipping point. The authors note that approximately three-quarters of the flagged area lay in ecoregions with relatively low direct human modification, which is consistent with a contribution from large-scale environmental drivers such as climate change, but they do not claim that climate change caused the signals. The global, ecoregion-scale analysis may also obscure fine-scale ecological processes. Regional ecological, climatic and land-use investigations would be required to determine the mechanisms behind each pattern and to establish whether particular systems ultimately cross critical transitions.

Subject of Research: Global satellite detection of early-warning signals for ecological transitions in terrestrial ecoregions

Article Title: Early-warning signals of critical transitions in global ecoregions

Article References: Hirsch, E., & Koren, I. (2026). Early-warning signals of critical transitions in global ecoregions. iScience, 29(11), Article 117964. https://doi.org/10.1016/j.isci.2026.117964

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.117964

Keywords: ecoregions, ecosystem resilience, critical transitions, remote sensing, vegetation indices, temporal autocorrelation, spatial clustering, climate change, land degradation, MODIS, Early-warning, signals