Rapid increase of climate extremes reveals new areas of concern in Amazonia

The frequency and intensity of climate extremes have increased in many regions of the world¹ and are predicted to increase further under future climate change scenarios². These extremes are responsible for some of the greatest climate-linked impacts on nature and people, driving increases in human morbidity and mortality^3,4, losses of forest species⁵ and ecosystem degradation⁶. Climate extremes are also responsible for deleterious changes in some of the world’s most important ecosystems. In Amazonia, recent exceptionally hot or dry periods have led to extensive and unprecedented forest fires^7,8,9, large-scale tree mortality¹⁰, and localised animal mortality¹¹, extreme warming of waters¹², and negative outcomes for human access to services¹³ and health^14,15,16. An increase in the incidence and severity of climate extremes could therefore help push Amazonia past critical thresholds, leading to much larger-scale deterioration of social and ecological conditions^17,18,19.

Given the importance of climate extremes in Amazonia, it is crucial to understand the temporal and spatial distribution of change. Although the central trends of temperature change in Amazonia are tracking global averages^20,21,22, these are often poor predictors of the trends of climate extremes¹. This is particularly the case in South America, where the divergences between changes in mean and extreme temperatures are widespread, but variable¹. Inferences from global, pantropical or regional studies on changes in climate extremes are currently insufficient. Many report only annual averages, which are known to mask ecologically important seasonal changes²³. Others simplify the substantial spatial and temporal heterogeneity in dry season length and onset that exists across Amazonia²⁴ (Supplementary Fig. 1) by defining dry season timing according to the drier period in the southern Amazon^20,25,26 or across large sub-regions^21,22,27. In addition, most focus on changes in either temperature, precipitation or water deficit, and therefore fall short of the integrated approach required to understand drivers of key ecological responses such as vegetation productivity and mortality²⁸ and forest flammability²⁹.

Here, we provide a comprehensive assessment of the changing central (50th percentile) and extreme trends (5th and 95th percentiles) of multiple climate variables across the 7 million km² region of Amazonia (see Methods), based on 11 km spatial resolution (57,900 grid cells) from 1981 to 2023 (43 years). Our analysis of extreme trends is not intended to identify particular extreme events as others have done^8,10,30, but instead, focuses on the trends in the tails of conditional distributions (climate variables over time). This approach provides an integrated way to understand the changes in both the frequency and the intensity of extreme events across the whole time series.

To do this, we first divide the year into 73 pentads to explore changes across hydrological years, following approaches used to identify changes in dry season timing across extensive subregions^22,31. However, we extend this so that each grid square has its own hydrological periodicity, with annual or seasonal periods starting with the wettest pentad or month instead of calendar years or fixed months (Supplementary Fig. 2). This allowed us to assess daily averages of minimum, mean and maximum temperature and Vapour Pressure Deficit (VPD) during the driest period as well as across the hydrological year. For rainfall, we evaluated the driest and wettest periods and the intensification of the water cycle through changes in rainfall amplitude. We also integrate changes in temperature with changes in rainfall or humidity, analysing variation in VPD and Maximum Cumulative Water Deficit (MCWD) that are well-established predictors of both plant physiology and fire activity^30,32. MCWD was calculated in two ways: using a fixed 100 mm of evapotranspiration per month commonly used in the literature^30,33 (MCWD₁₀₀) and by allowing evapotranspiration to increase with temperature (MCWD_Temp) following modelled estimates of temperature and atmospheric CO₂ responses³⁴ that are broadly supported by recent observations³⁵. Finally, we identified the metrics and regions showing the greatest rates of change and also assessed whether higher rates of change are associated with historically warmer or drier conditions.

We show that Amazonia’s climate has changed rapidly, but not uniformly. Extremes are changing at a much higher rate than central trends, and Amazonia-wide shifts in temperature—rather than precipitation—are driving most of the changes in water deficit and VPD. Crucially, we identified a hitherto under-emphasised and spatially distinct region of high climate change in the central-north Amazon. We discuss these results considering their implications for climate adaptation, conservation and sustainability in Amazonia.

Climate change across Amazonia

On average, the rate of change of extreme temperatures from 1981 to 2023 was faster than the central trend (Fig. 1, Supplementary Fig. 3). For central trends, annual mean temperatures increased by a mean rate of 0.21 °C per decade (dec^-1), (Interquartile Range IQR: 0.15 °C–0.27 °C dec^-1) and 0.90 °C between 1981 and 2023, which is in line with the global average and previous estimates for Amazonia²⁰ (Fig. 1a). However, the 95^th percentile of mean temperatures increased at a mean rate of at least 0.30 °C dec^-1 (IQR:0.24 °C–0.36 °C dec^-1) and 1.29 °C over 43 years (Fig. 1a). As expected, the rate of increase was greater in the driest period than across the whole hydrological year²¹, with maximum temperature at the 95th percentile increasing by 0.49 °C dec^-1 (IQR: 0.35 °C–0.63 °C dec^-1) (Fig. 1a). This corresponds to an increase of 2.11 °C over 43 years, 2.34 times higher than the annual mean central trend (Fig. 1a).

Temperature-linked measures of maximum cumulative water deficit (MCWD_Temp) and the vapour pressure deficit (VPD) also showed large changes over time, with higher values for extreme percentiles. For VPD, increases in the 95th percentiles were about one-third higher than for the central trend, for both annual and driest periods (Fig. 1b). VPD also had the greatest rate of change during the driest period, increasing by 0.06 kPa dec^-1 at the central trend (IQR: 0.02-0.09 kPa dec^-1) and 0.26 kPa over 43 years. This was even higher at the 95th percentile, increasing by 0.08 kPa dec^-1 (IQR :0.04–0.12 kPa dec^-1; 0.34 kPa over 43 years) (Fig. 1b). For MCWD, which is scaled towards the negative, the 5th percentile represents the drier extreme. The mean rate of change at the 5th percentile of MCWD_Temp was −12.75 mm dec^-1 (IQR: −22.88 to −0.96 mm dec^-1) and −54.82 mm over 43 years. This was twice that of the central trend of MCWD_Temp of −6.61 mm dec^-1 and −28.42 mm over 43 years (Fig. 1c). The temperature-invariant measure of MCWD₁₀₀ showed no directional trend, with the 50th percentile changing by just 0.03 mm dec^-1 (IQR: −5.13 to 4.01 mm dec^-1) and 0.13 mm over 43 years ( Fig. 1c).

Precipitation increased across the central, 5th and 95th percentiles for both annual and wettest periods of the year, but the trends were close to zero when assessing only the driest period (Supplementary Fig. 3d). For the central trend, annual precipitation increased by a mean rate of 31.29 mm dec^-1 (IQR: −8.14 to 65.28 mm dec^-1) and 134.55 mm over 43 years; Supplementary Fig. 3d). Precipitation amplitude showed a slight increase at all percentiles, with the greatest at the 95th percentile, where intensification of the water cycle—the difference between the wettest and driest periods—rose by a mean rate of 14.95 mm dec^-1 (IQR: −19.29 to 51.42 mm dec^-1) or 64.28 mm over 43 years (Supplementary Fig. 3e).

Spatial variation and hotspots of climate change

Amazonia-wide averages encompass considerable regional variation, with pronounced spatial divergence among the regions most affected by the highest rates of change at the 50th and extreme percentiles (Fig. 2, Supplementary Figs. 4–8). For the central trend of change across the hydrological year, our results are broadly consistent with previous work, and the southern Amazon stands out as the region most affected by increases in temperature and temperature-linked variables such as VPD (Fig. 2 and Supplementary Figs. 4, 5)^21,36. However, at the 95th percentile, the greatest rates of change for temperature and VPD are concentrated in a central-northern region (Fig. 2a, b) that has rarely been identified as being imperilled by climate change. Here, changes in average maximum temperature and mean VPD in the driest period were particularly alarming, with 10% of the biome surpassing rates of increase of 0.75 °C dec^-1 ( > 3.22 °C over 43 years) and 0.14 kPa dec^-1 ( > 0.60 kPa over 43 years). The region was also a locus of change in MCWD_temp, with extreme years seeing rates of change of −35.02 mm dec^-1 (< −150.59 mm over 43 years) (Fig. 2).

There was a strong spatial signal to the difference between central and extreme rates of climate change (Fig. 3). In much of the southern and southeastern Amazon, central trends are growing at a similar rate to—or even slightly faster than—the 95th percentile of temperature and VPD or 5th percentile of MCWD_temp. In contrast, in the central-north region the rate of change of climate extremes (95th percentiles for temperature and VPD, 5th percentiles for MCWD_temp) diverged greatly from both the central trends and the opposite extremes (Fig. 3a). Overlaying the top deciles of rates of change at the central and most extreme trends (95th percentile for mean temperature in the hydrological year, maximum temperature and VPD in the driest period of the hydrological year with the 5th percentile for MCWD_Temp) highlights critical regions where the fastest rates of change are co-occuring (Fig. 4). For central trends, there was strong spatial agreement between temperature and VPD, while MCWD_Temp was also prevalent on the southern borders of Amazonia. For the extreme trends, temperature-VPD overlaps were also extensive, but critical areas for MCWD_Temp included additional regions at the periphery of Amazonia (Fig. 4). These regions of greatest risk from rising central and extreme trends are predominantly located within Brazil and overlap with many important protected areas and indigenous lands (Supplementary Fig. 9).

Linking climate change to historical averages

Given the strong spatial signal of the observed climate trends, we investigated whether central or extreme trends were spatially associated with the historical averages from the start of the time series (1981–1990; Supplementary Fig. 10). The data were variable across metrics, and most correlation coefficients were generally non-significant (Supplementary Fig. 10). However, central trends were more positively associated with historical averages than the extreme trends, and there was a significant positive internal association between rates of change at the 50th percentile and the historical climate for MCWD_Temp (r = 0.38) and for VPD measured across the hydrological year (r = 0.27) and in the driest period (r = 0.34). These associations indicate that regions with historically higher MCWD and VPD are experiencing the highest rates of change for central trends.

Sensitivity of results to the 2023–24 extreme event

As our time series ends in an extreme year (2023–24), we assessed whether this event was driving the findings from the full time series (Supplementary Methods, Supplementary Fig. 11). For central trends, there was very strong agreement between the rates of changes calculated for both periods across all climate metrics ( r = 0.98–0.996). These correlations were weaker for extreme percentiles (r = 0.78–0.91) and, as expected, removing the most recent extreme event lowers the magnitude of change (Supplementary Fig. 11). However, the main spatial trends were not sensitive to removing 2023–24 from the time series, and the central-northern region retained the highest rates of change in 95th percentiles of maximum dry-season temperatures and VPD as well as highest rates of change in the 5th percentile of MCWD_Temp (Supplementary Fig. 11).

A new region of climate concern

While the southern Amazon is confirmed as the fastest-warming region on average, it is the central-northern region that is experiencing the greatest rates of increase in extreme temperature and VPD. This central-northern region is a critical part of Amazonia’s cultural and biological wealth, encompassing extensive areas of high forest cover, native savannas, and key indigenous territories, including the Coata‑Laranjal, Waimiri‑Atroari, Yanomami and Trombetas/Mapuera (Supplementary Fig. 9). The historical record of maximum temperatures (Supplementary Fig. 4) and water deficit (Supplementary Fig. 6) suggests this region has experienced less extremely hot or dry conditions than forests in south and eastern Amazonia. Multiple lines of evidence demonstrate that climate extremes are already affecting socio-ecological systems in this central-northern region. Examples include the first observed mass mortality of forest mammals in Amazonia¹¹ and the loss of understorey bird populations and changes in bird lifespans, behaviour and morphology^37,38,39. Aquatic systems have undergone extreme reductions in surface water⁴⁰ and abrupt shifts in fish assemblages⁴¹. Extreme heat and water deficit have contributed to extensive forest fires in the Brazilian state of Roraima⁴² and in flooded forests along the Rio Negro⁴³. These and other fire events have likely contributed to record levels of air pollution and hospitalisations in the largest Amazonian urban centre of Manaus^44,45.

The drivers of the increase in extremes in the central-northern region require further investigation. The record drought of 2023/24 provides insights, as the low rainfall and high temperatures in the northern Amazon were linked to sea surface temperature (SST) anomalies in the eastern tropical and central equatorial Pacific relating to the transition from La Niña to El Niño conditions²⁶. The region has also been identified as being at risk from major global climatic events, such as a slowing down of the Atlantic Meridional Overturning Circulation (AMOC)⁴⁶ or changes in cloud cover linked to a narrowing of the Intertropical Convergence Zone⁴⁷. However, the extent of these changes remains uncertain^48,49. Changes in soil moisture have been shown to exert a strong influence on extreme temperatures in regional hotspots of warming⁵⁰, but much of the central-northern region has experienced increases in annual and dry-period precipitation (Supplementary Fig. 7), and soils are variable and not consistently different from those in regions where extremes have changed less⁵¹.

Regardless of the underlying cause, the high rate of increase of extreme temperatures, VPD and water deficits demonstrates the critical importance of preventing further frontier advance in this region. Additional deforestation would amplify temperature changes and expose the remaining forests to heightened fire risks exacerbated by edge creation, within-forest disturbances, such as logging, and the spread of ignition sources^17,52. Sustained investment in adaptation will be required⁵³, particularly where preexisting infrastructure and development deficiencies exacerbate people’s vulnerability and the risks to health⁵⁴, wellbeing in urban environments^15,45, and the viability of sustainable socio-bioeconomies^55,56. Support will also be required to maintain the viability of some of the key climate change and conservation interventions, as the increase in extreme climate conditions jeopardises both protected areas through megafires⁵⁷ and the regrowth of secondary forests^58,59. All of this will require improvements in monitoring to evaluate risks and identify the most effective adaptation approaches.

Precipitation changes and uncertainty

There was no clear signal of strong reductions in precipitation across Amazonia or major differences between central and extreme trends. The only region where rainfall declined significantly in our analysis (Supplementary Fig. 7) and in other studies^18,60 was small and bisected the Brazilian states of Rondônia and Amazonas. In fact, overall, there was a slight increase in biome-wide averages. This could reflect increased moisture arriving from the warming tropical Atlantic Ocean^23,61,62 acting to offset the local and regional losses of precipitation resulting from deforestation^27,63.

The lack of a wider decline of rainfall in southern Amazonia contrasts with research based on rainfall gauges and other rainfall reanalysis products that identify a strong and consistent drying signal across a similar period^31,36. Although satellite products can underestimate periods of high rainfall and dry days⁶⁴, CHIRPS is considered one of is the best products for tropical precipitation during extreme climate events⁶⁵ and is even more accurate in drier regions of the Amazon⁶⁴, where the greatest drying trend has been identified in other studies³⁶. Furthermore, the overall increase in rainfall detected by CHIRPS is consistent with increased river flows²³ and groundwater storage⁶⁶. In contrast, the signal from rainfall gauges used in previous work³¹ is difficult to interpret due to their sparse coverage⁶⁷, inconsistent directional trends across stations^68,69, and their biased distribution towards the most deforested areas⁷⁰ and along rivers where the breeze influences rainfall⁷¹. As such, the source of the relatively low correlation between CHIRPS and Gauges (54%, see Supplementary Fig. 12) is unclear, as gauges are also beset by uncertainties, highlighting the importance of improving the observational network across tropical South America⁶⁹.

Temperature drives changes in indices of water stress

Overall, changes in temperature rather than precipitation are driving the most significant increases in VPD and reductions in MCWD_Temp, with the greatest recorded changes in the central-north region occurring where there is a good coverage of weather stations⁶⁹ to inform and constrain climate reanalysis data (Supplementary Figs. 13–15). However, the large difference between our temperature-invariant and temperature-linked measures of water deficit highlights the urgent need to reduce uncertainty around temperature and CO₂ effects on evapotranspiration in tropical forests. Elevated CO₂ induces stomatal closure, causing plants to reduce transpiration under elevated CO₂^72,73. Malhi et al.³⁴ predict this reduction would be smaller than the temperature-induced increases in evaporation from the soil or transpiration from the leaf. Although other models suggest a stronger influence of CO₂ than temperature⁷⁴, an overall increase in evapotranspiration is consistent with the relationship between temperature and potential evapotranspiration in the Penman-Monteith model⁷⁵ and from models and remote sensing observations conducted across the arc of deforestation³⁵. Furthermore, our use of 8.9 mm for every degree increase in temperature³⁴ is potentially conservative compared to the observed 11% increase in dry season ET over the arc of deforestation over 23 years, when mean temperatures increased by c. 0.40 °C (Supplementary Fig. 4) and a higher increase in ET was offset by forest loss³⁵.

While our results demonstrate the importance of incorporating ET in the assessment of MCWD, many uncertainties remain in moving beyond the fixed ET threshold of 100 mm month⁻¹ that has been the established estimate of evapotranspiration in tropical forests for the past two decades^10,30. Changes in evapotranspiration are likely to be spatially variable⁶⁶, will vary across years and seasons³⁴, and may be sensitive to forest disturbances⁷⁶ as well as the drought indicator and data sources that are used⁷⁷. While ET estimates based on remote sensing can reduce uncertainty, the marked divergence between products⁷⁸ will only be resolved with further validation and ground-based measures. In addition, remote sensing estimates actual ET – which often summarises multiple land use land cover classes – while MCWD is based on potential ET.

We show that climate change in Amazonia is neither gradual nor homogeneous, and that public policies must address the enormous challenge of rapidly increasing extreme temperatures and water deficits. Crucially, these are occurring in the central-northern region, far from the arc of deforestation. This spatial dissociation with the main loci of forest loss—the main local driver of climate change²⁷—demonstrates that the world’s highest greenhouse gas-emitting countries bear a strong responsibility for Amazonia’s rapidly changing socio-ecological condition. This underscores once again the urgent need for rapid reductions in greenhouse gas emissions², which are projected to amplify the region’s hydrological instability⁷⁹, and strengthens arguments that high-emitting countries should contribute to adaptation and conservation interventions in tropical forest regions through mechanisms such as the Fund for responding to Loss and Damage or novel initiatives such as the Tropical Forest Forever Facility. It is vital that this support reaches the most vulnerable regions and peoples and prevents further frontier advance that could amplify the consequences of climate extremes.

Datasets of climate variables

We used the following climate variables and resampling procedures to undertake our assessment of climate change in the Amazon region delimited by ref. ⁸⁰.

Temperature and vapour pressure deficit (VPD)

We used the ERA5-Land post-processed daily statistics dataset with 1-hour sub-daily frequency sampling over a 44-year period (1981-2024). We use 1981 as the starting point to standardise with precipitation data availability, and because 1981 is after the widely observed change in the rate of climate change that occurred in the 1970s^65,81. ERA5-Land is a global climate reanalysis product produced by the European Centre for Medium-Range Weather Forecasts (ECMWF) available at a spatial resolution of 0.1° – approximately 11 × 11 km⁸². For temperature metrics, we used the 2-m temperature minimum, mean and maximum daily statistics. We calculated the mean VPD using the ‘plantecophys’ R package^83,84, combining the ERA5-Land daily mean statistics from 2 m temperature, 2 m dewpoint and surface pressure.

Precipitation

We used the daily Rainfall Estimates from Rain Gauge and Satellite Observations (CHIRPS v2) dataset over a 44-year period (1981 to 2024). The CHIRPS algorithm integrates satellite data, ground-based observations and precipitation estimates to produce rainfall time series at a spatial resolution of 0.05°⁸⁵. CHIRPS has been described as one of the most accurate datasets to represent the rainfall regime across Amazonia, demonstrating superior performance compared to other data sources when validated against station-based observations^86,87. Validations across the basin have shown accuracy above 70%^68,88,89. We resampled the CHIRPS dataset using bilinear interpolation to match the 0.1° spatial resolution of the ERA5-Land dataset.

Maximum cumulative water deficit (MCWD)

We used CHIRPS to calculate the annual maximum cumulative water deficit (MCWD) from 1981 to 2024, an indicator of water stress in forests (Aragão et al., 2007). A water deficit occurs when evapotranspiration (ET) exceeds precipitation, with a 100 mm per month (month⁻¹) ET threshold commonly applied for tropical forests³⁰. However, increases in temperature have affected Amazonia’s ET by increasing atmospheric evaporative demand³⁴. To incorporate this warming effect, we calculated MCWD using fixed 100 mm month⁻¹ ET (MCWD₁₀₀) and temperature-linked ET (MCWD_temp). In this case, we considered that for each 1 °C above the mean temperature of a reference period (1971–1980), the 100 mm month⁻¹ ET threshold is increased by 8.9 mm month⁻¹. This coefficient is based on simple linear extrapolations of simulations of climatic scenarios that predict an increase of evapotranspiration in the Amazon forest to 140 mm month⁻¹ under conditions of increased temperature (4.7 °C) and CO₂ (850 ppm)³⁴. Rescaling this adds 8.9 mm month⁻¹ for every 1 °C increase in temperature, assuming atmospheric CO₂ concentrations and temperature increase linearly as per the model. We calculated the MCWD for both the fixed-ET (MCWD₁₀₀) and temperature-linked ET (MCWD_temp) using the R routine developed by ref. ⁹⁰.

To assess trends in the climate variables, we defined a cell-wise hydrological year that follows Amazonia’s highly variable precipitation cycle²⁴ rather than a fixed (January-December) calendar year or predetermined months. This approach considers the pronounced spatiotemporal variation in the timing and length of wet/dry periods across Amazonia. First, we aggregated the daily CHIRPS precipitation into 5-day periods (pentads), resulting in 73 pentads for each calendar year from 1981 to 2024. Leap days were removed from the analyses. Then, for each cell, we calculated the annual mean precipitation of each pentad over the 44-year period and identified the pentad with the highest mean precipitation. For each cell, the pentads of the calendar year were reordered according to the position of the wettest peak pentad, becoming the first of the 73 pentads of each hydrological year. For example, if the wettest pentad occurs in position 32 in the calendar year, the hydrological years for that cell span from pentad 32 of the current year (t) to pentad 31 of the next year (t + 1), (Supplementary Fig. 2).

We aggregated the daily statistics for the minimum, maximum, and mean temperature and mean VPD into mean 5-day periods (pentads) and reordered these variables according to the hydrological year. For MCWD, we applied the same methodology, but as this metric is calculated based on monthly thresholds for evapotranspiration, the hydrological years were defined based on the month with the highest precipitation. Because the wettest pentad/month generally did not occur in the first pentad/month of the calendar year, the hydrological year covered periods that extended across two calendar years in 99.84% of cells in the pentad approach and in 84.31% in the monthly approach, demonstrating the unsuitability of calendar years for these assessments. As a result, even though the CHIRPS and ERA5-Land datasets were available for 44 years (1981–2024), only 43 complete hydrological years were available for trend analysis (1981–2023).

To consider the spatiotemporal variation of precipitation across Amazonia²⁴, we used the CHIRPS hydrological year time series to identify the driest period for each cell. To identify the onset and end of the driest period, we applied the methodology proposed by ref. ³¹, comparing the mean precipitation of each pentad with the annual mean precipitation over the 43-year period. The onset corresponds to the first pentad in a sequence of at least six out of eight consecutive pentads with precipitation below the baseline. Conversely, the end of the driest period corresponds to the first pentad in a sequence of at least six out of eight with precipitation above the baseline (Supplementary Fig. 1). The wettest period was defined as all pentads outside the driest period.

Set of variables analysed

We calculated mean values for the entire hydrological year and for the driest period for temperature variables (minimum, mean and maximum) and for mean VPD. For MCWD, we calculated the values for the entire hydrological year. For precipitation, we obtained the annual sum and the sum for both driest and wettest periods. The precipitation in the wettest period was calculated as the difference between the precipitation in the hydrological year and the driest period. Finally, we calculated the precipitation amplitude by taking the difference between precipitation in the wettest and driest periods within the hydrological year.

Central and extreme trend analyses

To analyse central and extreme trends in climatic variables over time, we used non-parametric Quantile Generalised Additive Models (qGAM). We chose this model because it allowed us to test and obtain coefficients for central and extreme trends through percentile (i.e., quantile) models, while still accounting for the temporal dependence of the data (analysis over years). To keep the trend analysis linear, we did not include a smoothing parameter for the term “hydrological year” [Response ~ hydrological year (1981-2023)]. For each cell and each variable, we ran a qGAM for the 50th (median – central) and 5th and 95th (extremes) percentiles and obtained the coefficients, which can be interpreted as the rates of change of a given variable (change in the variable per year). We did that using the function ‘qgam’ from R package ‘qgam’⁹¹ version 2.0.0 in the R software⁸⁴ version 4.5.1.

To pinpoint the areas with greatest divergence between extreme and central trends, we subtracted the values of the decadal rates of change in the 50th percentile from the rates of change in the extreme (95^th percentile for temperature and VPD and 5th percentile for MCWD_temp) for mean temperature in the hydrological year, maximum temperature and VPD in the driest period of the hydrological year and MCWD_Temp of the hydrological year. We selected the top and bottom 10% values of the difference between central and extreme rates of change (i.e., 10% values for where extreme trends were higher than central trends and vice versa) to display the average curves for the regions with greatest divergence between extreme and central trends. For this, we built the trend curves (50th and 5th / 95th percentiles) for all the 5790 cells (10% of cells – 5,735 cells for MCWD) based on their intercepts and coefficients and calculated the median curve and IQR (25-75 – 50% of the data distribution). To identify overlaps between regions with the highest rates of change of different climate variables, we selected the top 10% values of decadal rates of change for mean and maximum temperature, VPD and MCWD_temp of the extreme (95th percentile for temperature and VPD and 5th percentile for MCWD_temp) and central (50th percentile) trends.

Associations of rates of change with climatic initial conditions

To investigate if climate change is spatially associated with the average climatic conditions at the start of the time series, we ran 1000 bootstrap analyses of Spearman correlation, using 100 samples (cells) each time. We used the bootstrap approach because i) we wanted to have a confidence interval for the coefficients of correlation, ii) there is high spatial autocorrelation in the data, and iii) using > 50,000 points would increase the chances of type I error. We first calculated the mean values for each climatic variable in the period between 1981 and 1990. Then, for each combination of a variable, a percentile (5th, 50th and 95th), a period (year, Driest Period – DP and Wettest Period – WP) and a method (100 mm threshold and Temp-linked threshold for MCWD) we used the ‘cor.test’ function in R software to bootstrap the Spearman correlation test and obtained the coefficient of correlation (r) between the rate of change and the mean value (1981–1990) of a given variable. We defined the confidence intervals as the distribution of 95% of the r coefficients (2.5th–97.5th percentiles) obtained from the 1000 bootstraps. We defined non-significant correlations as those whose confidence intervals crossed zero.