Observed and simulated local climate responses to tropical deforestation

Tropical deforestation has local and regional effects on climate, but the sign and magnitude of these effects are still poorly constrained. Here we used satellite observations to evaluate the local land surface temperature and precipitation response to tropical deforestation in historical simulations from 24 CMIP6 models. We found tropical forest loss leads to an observed local dry season warming and reduced wet and dry season precipitation across the range of scales (0.25°-2°) analysed. At the largest scale analysed (2°), we observed a warming of 0.018 ± 0.001 °C per percentage point of forest loss (°C %−1), broadly captured in the multi-model mean response of 0.017 ± 0.005 °C %−1. The multi-model mean correctly simulates reduced precipitation due to forest loss in the dry season but simulates increased precipitation due to forest loss in the wet season, opposite to the observed response. We found that the simulated dry season surface temperature and precipitation changes due to forest loss depend on the simulated surface albedo change, with less warming and less drying in models with greater increases in surface albedo due to forest loss. Increased recognition of the local and regional climate benefits of tropical forests is needed to support sustainable land use policy.


Introduction
Land cover change alters energy and water fluxes between the surface and atmosphere affecting the local and regional climate (Bonan 2008, Pongratz et al 2021).Tropical regions are experiencing rapid changes to land cover, particularly from deforestation (Hansen et al 2013) and forest degradation (Vancutsem et al 2021).Tropical deforestation has been shown to cause local surface warming of greater than 2 • C (Alkama and Cescatti 2016, Bright et al 2017, Duveiller et al 2018, Baker andSpracklen 2019).The effect on precipitation is more complex and scale-dependent (Lawrence and Vandecar 2015), with increases in precipitation over or near smallscale deforestation (Garcia-Carreras and Parker 2011, Khanna et al 2017, Taylor et al 2022) and reductions over and downwind of large-scale deforestation (Spracklen and Garcia-Carreras 2015).Analysis of remotely sensed precipitation suggests tropical forest loss causes reductions in local precipitation, particularly at scales larger than 50 km (Smith et al 2023).
Climate models have different representations of the land surface and the biophysical responses to land cover change, leading to different simulations of the climate response to land cover change (Boisier et al 2015, Boysen et al 2020, Baker et al 2021a, Luo et al 2022, De Hertog et al 2023).Most models agree that deforestation in the tropics causes local surface warming but disagree on the magnitude of the temperature response (Winckler et al 2019b, Boysen et al 2020).In contrast, some models simulate local cooling over tropical deforestation due to strong increases in simulated surface albedo (Robertson 2019).The simulated response of local precipitation to land cover change is even more varied.Luo et al (2022) simulated the impacts of idealised deforestation scenarios and found a multi-model mean reduction in precipitation over regions of forest loss of −2.2%, with a range of −5.5% to +0.1% across 11 models.Spracklen and Garcia-Carreras (2015) synthesised simulated impacts of deforestation in the Amazon basin, finding an average of 12 ± 11% reduction in annual precipitation due to basin-wide deforestation.
Previous assessments of climate model responses to land cover change have analysed both idealised (e.g.Davin (Duveiller et al 2018) have largely focused on temperature from satellite (Li et al 2015, Alkama and Cescatti 2016, Bright et al 2017, Duveiller et al 2018) or in-situ measurements (Lee et al 2011).Simulations of the impacts of land cover change on precipitation (Luo et al 2022) have not yet fully been evaluated.We build on this previous work by evaluating the impacts of tropical deforestation in the historical CMIP6 simulations on both local land surface temperature (T) and precipitation (P) in a consistent manner.We focus on tropical deforestation because of the urgent need for clear evidence to support conservation of remaining tropical forests for climate change adaptation and mitigation (Windisch et al 2021).We explore how the climate sensitivity to the extent of forest loss depends on simulated changes to surface albedo, evapotranspiration (ET), and leaf area index (LAI).We evaluate the simulated response using satellite observations, applying a before-aftercontrol-impact approach, where the change in local climate over regions of forest loss is compared against the change in climate over control areas with no forest loss.This allows us to analyse the simulated and observed responses to deforestation identically.

Data and methods
We analysed data from 24 CMIP6 models (CMIP6 Tier 1: historical; dataset information listed in table 1), with spatial resolution varying from 0.56 to 2.79 degrees latitudinally.We downloaded and processed monthly mean surface albedo, ET, LAI, land surface temperature and precipitation for 1850-2014.
To evaluate the CMIP6 models, we used satellite data from the period 2003-2019.We calculated forest loss from the Global Forest Change (GFC) version 1.9 (Hansen et al 2013), using forest canopy cover in 2000 and subsequent annual forest loss from 2003 to 2019 at 30 metre (m) resolution.We used MODIS albedo (MCD43A3), ET (MOD16A2GF) and LAI (MOD15A2) available at 500 m resolution and land surface temperature day-night mean (MOD11C3) available at 1 km resolution.We used precipitation data from nine datasets, spanning a range of native resolutions from ∼4 to 25 km (approx.at equator, table 1 lists the details).
We analysed the observed impacts of forest loss across four spatial scales (0.25 spanning the spatial resolution of the CMIP6 models.We performed spatial regridding using the Python package Iris (Met Office 2023) with the area-weighted regridding scheme.Datasets were regridded to coarser resolutions using the highest available resolution as listed in table 1.Two alternative regridding methods (xESMF (Zhuang 2022): 'conservative-normalised' and 'bilinear') were tested and had little impact on our results.We calculated forest loss at each spatial resolution as the sum of all 30 m pixels within each larger pixel.
We constrained our analysis to the tropics (30 • S-30 • N).We additionally constrained satellite datasets by the tropical evergreen broadleaf biome, defined by the MODIS land cover dataset (MCD12Q1), and CMIP6 models by areas where their forest cover was greater than 70% at the start of the discrete analysis periods.This accounted for the fact that simulated forests may be in different geographical areas within each model.We tested both constraining CMIP6 models by MODIS evergreen broadleaf and by areas of forest cover greater than 70%, finding similar results with both methods.We analysed separately over the Amazon and Congo Basins and southeast Asia using shapefiles to geographically constrain the analysis as outlined in supplementary figure 1.
Detecting a robust local climate response to deforestation requires long simulations (Winckler et al 2017).For this reason, we analysed data over 16 year periods.For the satellite datasets, this period was 2003-2019, as this was the longest common period of precipitation data.For the CMIP6 models, we analysed ten 16 year periods starting in 1854 and ending in 2014.We selected 16 year periods to match the length of the satellite record and report model values as the median across the ten periods.In addition to this, we analysed the CMIP6 models over five 32 year periods, finding similar results over this longer time period (supplementary figures 2-5).To reduce the impact of interannual variability in temperature and precipitation, we compared 5 year means at the start and end of each analysis period.
Land cover change causes both local and nonlocal climate impacts (Pongratz et al 2021).The local climate impacts of land cover change can be assessed  Comparing the change over a pixel with forest loss with its immediate neighbour with little or no forest loss removes the impacts of climate change and variability.We adopted this approach and analysed the local climate response to forest loss using a moving window nearest neighbour approach as used by previous studies (Baker and Spracklen 2019, Smith et al 2023), here employing a 3 × 3 grid size.We calculated the forest loss of each deforested pixel relative to neighbouring control pixels as the forest loss of the deforested pixel minus the forest loss of the control.To be included in the analysis, deforested pixels must have experienced more than 0.1 percentage points of forest loss compared to their neighbouring control pixels.We calculated the change in each variable over the deforested pixel relative to the change of the control pixel.We report changes as a function of forest loss by dividing by the difference in forest loss between deforested and control pixels.
We focused our temperature analysis on the dry season, where there was better availability of satellite data for albedo, ET, LAI, and T, as the wet season has more clouds which obstruct satellite retrievals.Dry season temperature is also more sensitive to tropical deforestation (Baker and Spracklen 2019).For precipitation, we analysed dry, wet and transition seasons as the driest 3 months, wettest 3 months and remaining 6 months, respectively, of each year for each pixel.The satellite remotely sensed precipitation value is based on the median of the nine satellite precipitation datasets, whilst for the CMIP6 models, we derived each model's season from its own precipitation data.
To test the relationships between climate variables, we fitted linear regressions using Pearson's correlation coefficient, (calculated using SciPy (Virtanen et al 2020)) to identify whether the computed correlation coefficients were found to be statistically significant and different from zero at the 5% level (p<0.05).We report errors throughout as the standard error of the mean.

Results and discussion
Figure 1 shows the observed impacts of forest loss on local land surface temperature and precipitation.We observed dry season warming due to forest loss across all spatial scales analysed (figure 1(a)).This demonstrates that tropical forest loss caused local warming at spatial scales simulated in regional (0.25 The local land surface warming we report here is similar to previous studies such as Alkama and Cescatti (2016) who reported that tropical forest deforestation caused a warming of 0.015 ± 0.001 • C % −1 .Duveiller et al (2020) used a space-for-time approach and reported a warming of 0.018 ± 0.001 • C % −1 for wet tropical forests using 1.0 • × 1.0 • resolution data.In the Amazon, Baker and Spracklen (2019) reported deforestation caused dry season land surface warming of 0.014 • C % −1 using 0.05 • resolution data.
We observed reductions in precipitation over regions of tropical forest loss at both an annual scale and in the dry, wet and transition seasons (figure 1(b)).Forest loss causes a decrease in precipitation across all analysed resolutions, with larger reductions as the scale of forest loss increases.At 2 • resolution, the annual reduction was −0.18 ± 0.07 mm month −1 % −1 .This sensitivity is slightly lower than reported by Smith et al (2023) (−0.25 ± 0.10 mm month −1 % −1 at 2 • ) due to small methodological differences, including a longer analysis period (2003-2019 compared to 2003-2017).Reductions in precipitation were observed throughout the year, with the largest absolute reductions in precipitation over regions of forest loss in the wet season (−1.12 ± 0.32 mm month −1 % −1 ) compared to −0.06 ± 0.05 mm month −1 % −1 in the dry season and −0.33 ± 0.24 mm month −1 % −1 in the transition season.
Figure 2 compares the simulated and observed impact of forest loss on local dry season land surface temperature.Most models (22 out of 24) simulate a warming response consistent with the satellite observations.The simulated surface temperature response to forest loss varies from −0.038 ± 0.008 In idealised deforestation simulations, Boysen et al (2020) found that the near-surface air temperature response simulated by CMIP6 models varied between −0.02 The local surface warming due to forest loss is relatively insensitive to spatial scale, both in the models and observations.The multi-model mean warming due to forest loss is +0.017 ± 0.005 • C % −1 (0.016 ± 0.002 • C % −1 for models <1 • resolution and 0.017 ± 0.006 • C % −1 for models >1 • resolution), which compares well to the observed warming of 0.018 ± 0.001 Whilst the multi-model mean is close to the observed value, figure 2 highlights the large variability across models.We also analysed the simulated temperature change due to forest loss for each model separately over the ten 16 year model periods (supplementary figure 6).Only 7 models show consistent warming across all periods.Most models (17 out of 24) show warming and cooling in different periods, five of which (CanESM5, CMCC-ESM2, GISS-E2-1-G, MPI-ESM1-2-HR and UKESM1-0-LL) show a cooling response in four or more of the ten 16 year periods contrary to the observed temperature response.This further confirms the need for long simulations to robustly diagnose a climate response to land use change from climate models.When analysing the simulated temperature change over the longer 32 year periods (supplementary figure 2), we find a very similar multi-model mean warming response of 0.019 ± 0.004 • C % −1 .Most models show a consistent warming in both the 16 year and 32 year analysis (supplementary figures 3 and 7), with only GISS-E2-1-G and MPI-ESM1-2-HR showing an overall cooling in the 16 year analysis and UKESM1-0-LL and MPI-ESM1-2-HR showing a cooling in the 32 year analysis.
Figure 3 compares the simulated and observed changes in dry and wet season precipitation due to tropical forest loss.The simulated precipitation response to forest loss is less consistent than for temperature.In the dry season, 6 of the 24 models simulate increases in precipitation due to forest loss, whilst the remaining 18 models simulate reductions.In the wet season, 10 of the 24 models simulate an increase, whilst the remaining 14 simulate a decrease.Across all models, the multi-model mean response of dry season precipitation to forest loss is −0.06 ± 0.08 mm month −1 % −1 , comparable to the observed change of −0.06 ± 0.05 mm month −1 % −1 (at 2 • ).The multi-model mean response in the wet season is 0.11 ± 0.65 mm month −1 % −1 , opposite to the observed response of −1.12 ± 0.32 mm month −1 % −1 .The individual CMIP6 models tend to be oversensitive to forest loss (either large increases or decreases) compared to observed changes.Analysing over 32 years, we find a very similar multi-model dry season precipitation response (−0.063 ± 0.073 mm month −1 % −1 , supplementary figures 4(a) and 5).In the wet season, where there is large intermodel variability in both analysis periods, the multimodel mean response changes to a slight drying (−0.010 ± 0.430 mm month −1 % −1 ).
At the annual scale, the multi-model mean precipitation sensitivity to forest loss is +0.06 ± 0.23% per percentage point of forest loss (% % −1 ) (supplementary figure 8), opposite in sign to the observed sensitivity of −0.12 ± 0.11% % −1 (at 2 • ).Previous studies have also reported a wide range in the simulated precipitation response to tropical deforestation.Luo et al (2022) reported Amazon deforestation resulted in a regional annual mean precipitation response of −11% to +2% for a 50% reduction in forest cover, equivalent to −0.18% to +0.04% per percentage forest loss, with eight out of the 11 models simulating decreased precipitation over regions of forest loss in the western and southern Amazon basin.Spracklen and Garcia-Carreras (2015) reported multi-model mean annual mean sensitivity of −0.16 ± 0.13% per percentage point forest loss in the Amazon.
To explore the different regional impacts of forest loss on climate, we analysed the changes in temperature and precipitation (supplementary figures 9 and 10 respectively) across the Amazon, Congo Basin and Southeast Asia.Local dry season warming due to forest loss is seen across all regions in both satellite data and the multi-model mean.The models simulate the greatest sensitivity of temperature to forest loss in the Amazon and Congo, however this is not seen in the satellite data.The observed and simulated response of dry season precipitation  1) are the median over ten 16 year periods from 1854 to 2014.Observed results from satellite (orange bars) (table 1) are for 2003-2019, regridded to three resolutions (0.5 to match the range of resolutions from the models.Satellite observed forest loss data were from GFC v1.9 (Hansen et al 2013).Model results are for areas where initial forest cover exceeds 70%, whilst the satellite analysis is constrained by MODIS evergreen broadleaf land cover.Error bars show the standard error of the mean calculated across the 10 time periods for each model.The dashed line separately shows the multi-model mean value for <1 and >1 • spatial resolution models.
to forest loss is consistent across the tropics as a whole, however regionally the results are divergent, with opposite responses in the Amazon and Congo (supplementary figure 10(a)).In the wet season, there is consistent observed drying due to forest loss across all regions, however in the simulated response, there are increases in precipitation in the Congo and decreases in Southeast Asia (supplementary figure 10(b)).The simulated precipitation responses in the Congo and Southeast Asia have especially large variability, reflecting the divergence of model responses.
Figure 4 compares the median dry season sensitivity of temperature and precipitation to forest cover loss against the equivalent sensitivity of different land surface variables (albedo, ET, LAI) to forest loss.There is substantial variability in the simulated sensitivity of albedo, ET and LAI to forest loss.We find large variability in the simulated sensitivity of surface albedo to forest loss varying from ∼0 to 5.1 × 10 −4 % −1 , with 23 of the 24 models simulating an increase in surface albedo in regions of forest loss (INM-CM4-8 simulates a decrease).A previous assessment of the CMIP5 models also found large variability in the simulated albedo response to land use change (Lejeune et al 2020).For ET, we find simulated sensitivity ranges from −1 to +0.5 mm month −1 % −1 .Luo et al (2022) reported that forest loss caused annual mean changes of +50 to −150 mm year −1 .For LAI, we find a simulated sensitivity of −0.05 to 0.03 m 2 m −2 % −1 .In the Amazon, Luo et al (2022) also reported a wide range in the sensitivity of LAI to forest loss ranging from −2 to +1 m 2 m −2 , equivalent to −0.02 to +0.01 m 2 m −2 % −1 .
The local warming due to forest loss is caused by reduced surface roughness, which reduces turbulent heat fluxes and ET (Bright et al 2017, Duveiller et al 2018).For dry season temperature, we find statistically significant relationships (P < 0.05) with albedo (r 2 = 0.299) and ET (r 2 = 0.292).As would be expected (Bright et al 2017, Duveiller et al 2018, Winckler et al 2019a), models with a stronger sensitivity of surface albedo (greater surface brightening) and weaker sensitivity of ET (smaller ET decreases) to forest loss tend to show less warming from forest loss.
Albedo measurements from satellite also suggest increased albedo due to forest loss with a sensitivity of 8.0 × 10 −5 -1.31 × 10 −4 % −1 , equivalent to an increase in albedo of 0.008-0.013for complete forest loss (figure 4).This albedo sensitivity to forest loss is relatively well captured by some models (3 simulating albedo within the satellite range), whereas 9 models underestimate (<8.0 × 10 −5 % −1 ) and 12 overestimate (>1.31 × 10 −4 % −1 ) the sensitivity.Models that overestimate the albedo sensitivity to forest loss underestimate the warming due to forest loss.  1) are the median over ten 16 year periods from 1854 to 2014.Observed results from satellite (orange bars) (table 1) are for 2003-2019, regridded to three resolutions (0.5 to match the range of resolutions from the models.Model results are for areas where initial forest cover exceeds 70%, whilst the satellite results are constrained by MODIS evergreen broadleaf land cover.We calculate the standard error of the mean from the ten time periods for each model but from one time period for the satellite data.The dashed line separately shows the multi-model mean value for <1 and >1 • spatial resolution models. Most models simulate a reduction in ET over forest loss (multi-model mean −0.19 ± 0.06 mm month −1 % −1 ), although there is large variability across models with a range of −1.17 to +0.62 mm month −1 % −1 .The sensitivity of ET to forest loss is related to the change in LAI, as has been shown previously (Luo et al 2022), with models that simulate larger decreases in LAI tending to simulate larger decreases in ET following forest loss (supplementary figure 11).Forest loss causes a reduction in simulated ET due to the replacement of forests by grasses with lower ET rates, matching the response in idealised deforestation simulations (Boysen et al 2020).Increased ET over regions of forest loss in some models (e.g.CESM) may be due to tropical forests being replaced by C4 grasses that are over productive in the moist tropics (Boysen et al 2020), whilst in other models (e.g.GISS-E2-1-G) it may be due to increased simulated precipitation over regions of forest loss.
We found significant positive relationships for dry season precipitation with ET (r 2 = 0.564) and albedo (r 2 = 0.176).Luo et al (2022) also reported positive relationships between changes in precipitation and ET due to deforestation.They also found that the inter-model spread in precipitation response to forest loss primarily results from divergent responses of ET.Previous work has also suggested albedo as an important parameter controlling precipitation changes (Dirmeyer and Shukla 1994, Berbet and Costa 2003, Costa et al 2007).Dirmeyer and Shukla (1994) found that the local precipitation  d)-(f) precipitation (P) to surface albedo (Alb), evapotranspiration (ET), leaf area index (LAI), per percentage point of forest loss.Simulated (blue) values show the median change for each model's ten 16 year periods.We report the linear Pearson correlation coefficient squared (r 2 ) and the p-value (p) and plot the linear fit where p < 0.05.Results are for areas where initial forest cover exceeds 70%.Satellite values are plotted as orange circles, regridded to four resolutions (0.25 response to forest loss showed a strong sensitivity to the assumed increase in albedo with forest loss over a range of 0-0.09 (9.0 × 10 −4 % −1 ).However, they found forest loss reduced precipitation when the albedo sensitivity was greater than 3.0 × 10 −4 % −1 , opposite to our results of increased precipitation in models with greater brightening.
We note that the satellite-based sensitivity of albedo, ET and LAI to forest loss is less than would be expected based on in-situ measurements.In the Amazon, (Culf et al 1995) observed annual mean albedo of 0.13 for tropical forest and 0.18 for pasture, suggesting deforestation causes increased albedo of 0.05 or 4.6 × 10 −4 % −1 (plotted as a red star in figures 4(a) and (d)), about a factor 4 greater than in the satellite measurements.In-situ data represents a complete conversion from forest to pasture with correspondingly large changes in albedo.In comparison, satellite data observes forest loss at larger scales where remaining tree cover and vegetation regrowth may reduce the change in albedo caused by forest loss.In situ observations of dry season ET in the Amazon are around 110 mm month −1 for tropical forests and 70 mm month −1 for pasture (Restrepo-Coupe et al 2013), suggesting deforestation causes a reduction of 40 mm month −1 or 0.4 mm month −1 % −1 (plotted as a red star in figures 4(b) and (e)), around 3.5 times greater than seen in the satellite measurements.Challenges with remote-sensed ET data which combine remote sensed and model data (Baker et al 2021b) may explain the discrepancy with in-situ data.The simulated temperature response to forest loss is strongly related to albedo and ET in the dry season but less so in the wet season (Baker et al 2021b).
Our analysis focused on assessing the simulated local climate response to tropical deforestation and understanding how this depends on the modelled treatment of the land surface change.Tropical deforestation drives changes to the local energy balance that are dominated by changes in the turbulent energy flux (Boysen et al 2020, De Hertog et al 2023).Our analysis shows a large disagreement in the simulated response of ET flux to deforestation.However, uncertainty in measurements of the ET flux (Baker et al 2021b) are a challenge to constraining the simulated response of ET to forest loss.In contrast, satellitederived datasets of surface albedo are more reliable (He et al 2014) and may provide a stronger constraint on the large spread of model simulated albedo responses to forest loss.We suggest that evaluating and improving the surface albedo response to forest loss may be a logical and practical initial step to model improvement.Constraining the albedo sensitivity to deforestation is also important for accurate simulation of the radiative forcing due to historical land-use change (Lejeune et al 2020).
Previous work has shown that the climate response to deforestation depends on the background climate (Pitman et al 2011).To explore whether changes in background climate have changed the climate response to deforestation, we calculated how the simulated sensitivity of temperature to forest loss varied over the 1854-2014 period.We found that there was no significant trend in the simulated response of temperature to forest loss over this period (supplementary figure 12).The climate response to deforestation also depends on simulated atmospheric feedbacks through altering mesoscale circulations (Khanna et al 2017).Boysen et al (2020) found that increased shortwave radiation due to reduced cloud cover over regions of tropical deforestation was more important than changes in surface albedo in some models.Luo et al (2022) found mean reductions in ET over deforested areas (16.9 mm yr −1 ) were about 4 times greater than reductions in mean flow convergence (−4.3 mm yr −1 ), suggesting local reductions in ET dominate reduced rainfall rather than changes in circulation.
In addition to impacts on temperature and precipitation, deforestation can also impact other important climate variables such as causing reductions in low level cloud cover (Duveiller et al 2021).We focused on the local land surface warming due to forest loss, though we note that air temperature's response to deforestation may differ (Winckler et al 2019b).Deforestation can also cause important changes in the timing and intensity of precipitation.In Amazonia, deforestation has extended dry season and delayed the onset of the rainy season (Leite-Filho et al 2021, Commar et al 2023).In West Africa, deforestation has enhanced storm frequency (Taylor et al 2022).In addition to local impacts, deforestation may also change regional climate (Leite-Filho et al 2020).Tropical deforestation can cause reductions in downwind precipitation through reductions in moisture recycling (Spracklen et al 2012, Zemp et al 2017, Staal et al 2018) and can alter regional temperatures up to 50 km away from the location of land-use change (Cohn et al 2019).Deforestation may even alter precipitation in regions far removed from the land use change through teleconnections (Werth and Avissar 2005, Pitman et al 2009, De Noblet-Ducoudré et al 2012, Luo et al 2022).

Conclusions and implications
Our analysis provides further evidence of the local surface warming and drying (reduced precipitation) due to tropical deforestation.The multi-model mean captures the observed surface warming due to tropical forest loss, with 22 out of 24 CMIP6 models analysed simulating warming in response to tropical forest loss.The multi-model mean suggests increased annual mean precipitation over regions of tropical forest loss, opposite in sign to the observed response.There is large variability in the magnitude of the modelled temperature and precipitation responses to deforestation, some of which we attribute to different implementations of land use change within CMIP6 models and the subsequent changes to albedo and ET.We find the simulated local land surface warming due to forest loss is sensitive to the simulated surface albedo change.
The local warming and drying due to tropical deforestation will have negative impacts on human health (Wolff et from a single simulation through comparing the climate change over regions of land cover change compared to neighbouring regions with little or no land cover change(Kumar et al 2013, Lejeune et al 2017).

Figure 1 .
Figure 1.Observed local response of temperature and precipitation to tropical forest loss from 2003 to 2019.(a) Median dry-season land surface temperature change (∆T) per percentage point of forest loss and (b) median precipitation change (∆P) per percentage point of forest loss for annual mean, dry, wet and transition seasons.Temperature data are from the MODIS MOD11C3 product and precipitation data from nine products (table S1).Results are shown at four spatial resolutions (0.25 • , 0.5 • , 1.0 • and 2.0 • ).Error bars indicate the standard error of the mean.

Figure 2 .
Figure 2. Change in simulated and observed dry season local surface temperature per percentage point of forest loss (∆T, • C % −1 ), ordered by latitudinal resolution.Simulated changes from the CMIP6 models (blue bars; datasets listed in table1) are the median over ten 16 year periods from 1854 to 2014.Observed results from satellite (orange bars) (table 1) are for 2003-2019, regridded to three resolutions (0.5• × 0.5 • , 1.0 • × 1.0 • , 2.0 • × 2.0 • )to match the range of resolutions from the models.Satellite observed forest loss data were from GFC v1.9(Hansen et al 2013).Model results are for areas where initial forest cover exceeds 70%, whilst the satellite analysis is constrained by MODIS evergreen broadleaf land cover.Error bars show the standard error of the mean calculated across the 10 time periods for each model.The dashed line separately shows the multi-model mean value for <1 and >1 • spatial resolution models.

Figure 3 .
Figure 3. Change in simulated and observed (a) dry and (b) wet season precipitation per percentage point of forest loss (∆P, mm month −1 % −1 ), ordered by latitudinal resolution.Simulated changes from the CMIP6 models (blue bars) (datasets listed in table1) are the median over ten 16 year periods from 1854 to 2014.Observed results from satellite (orange bars) (table1) are for 2003-2019, regridded to three resolutions (0.5• × 0.5 • , 1.0 • × 1.0 • , 2.0 • × 2.0 • )to match the range of resolutions from the models.Model results are for areas where initial forest cover exceeds 70%, whilst the satellite results are constrained by MODIS evergreen broadleaf land cover.We calculate the standard error of the mean from the ten time periods for each model but from one time period for the satellite data.The dashed line separately shows the multi-model mean value for <1 and >1 • spatial resolution models.

Table 1 .
CMIP6 model and satellite datasets used in this analysis.Models are grouped by spatial resolution (<1 • , and ⩾1 • resolution in latitude).
(Lovejoy and Nobre 2019)et al 2020, Li et al 2022)e (Lawrence and Vandecar 2015, Leite-Filho et al 2021), surrounding forests(Zemp et al 2017, Staal et al 2020, Li et al 2022)and biodiversity(Pardini et al 2017).A warmer and drier climate will also exacerbate the risk of forest fires causing additional forest loss and the potential for positive climate feedbacks(Cochrane et al 1999).Some work has suggested the Amazon is close to a tipping point where additional deforestation would drive sufficient drying to induce forest dieback(Lovejoy and Nobre 2019).Future work is needed to assess the resilience of remaining tropical forests to a warmer and drier climate.Overall, our analysis provides additional impetus for policymakers to account for the local climate impacts of tropical deforestation(Duveiller etal 2020, Pongratz et al 2021).European Union's Horizon 2020 research and innovation programme (DECAF project, Grant agreement No. 771492), and the Newton Fund, through the Met Office Climate Science for Service Partnership Brazil (CSSP Brazil).We also acknowledge that DIK was supported by the NC-International programme [NE/X006247/1] delivering National Capability.No competing interests.We acknowledge the World Climate Research Programme, which, through its Working Group on Coupled Modelling, coordinated and promoted CMIP6.We thank the climate modelling groups for producing and making available their model output, the Earth System Grid Federation (ESGF) for archiving the data and providing access, and the multiple funding agencies that support CMIP6 and ESGF.Bull.Am.Meteorol.Soc.96 69-83 Baker J C A, De Souza D C, Kubota P Y, Buermann W, Coelho C A S, Andrews M B, Gloor M, Garcia-Carreras L, Figueroa S N and Spracklen D V 2021a An assessment of land-atmosphere interactions over south America using satellites, reanalysis, and two global climate models J. Hydrometeorol.22 905-22 Baker J C A, Garcia-Carreras L, Gloor M, Marsham J H, Buermann W, Da Rocha H R, Nobre A D, De Carioca Araujo A and Spracklen D V 2021b Evapotranspiration in the Amazon: spatial patterns, seasonality, and recent trends in observations, reanalysis, and climate models Hydrol.Earth Syst.Sci. 25 2279-300 Baker J C A and Spracklen D V 2019 Climate benefits of intact Amazon forests and the biophysical consequences of disturbance Front.For.Glob.Change 2 1-13 Berbet M L C and Costa M H 2003 Climate change after tropical deforestation: seasonal variability of surface albedo and its effects on precipitation change J. Clim.16 2099-104 Boisier J P, Ciais P, Ducharne A and Guimberteau M 2015 Projected strengthening of Amazonian dry season by constrained climate model simulations Nat.Clim.Change 5 656-60 Bonan G B 2008 Forests and climate change: forcings, feedbacks, and the climate benefits of forests Science 320 1444-9 Boucher O et al 2018 IPSL IPSL-CM6A-LR model output prepared for CMIP6 CMIP historical Earth System Grid Federation (https://doi.org/10.22033/ESGF/CMIP6.5195)Boysen L R et al 2020 Global climate response to idealized deforestation in CMIP6 models Biogeosciences 17 5615-38 Bright R M, Davin E, O'Halloran T, Pongratz J, Zhao K and Cescatti A 2017 Local temperature response to land cover and management change driven by non-radiative processes Nat.Clim.Change 7 296-302 Cochrane M A, Alencar A, Schulze M D, Souza C M, Nepstad D C, Lefebvre P and Davidson E A 1999 Positive feedbacks in the fire dynamic of closed canopy tropical forests Science 284 1832-5 Cohn A S, Bhattarai N, Campolo J, Crompton O, Dralle D, Duncan J and Thompson S 2019 Forest loss in Brazil increases maximum temperatures within 50 km Environ.Res.Lett.14 084047 Commar L F S, Abrahão G M and Costa M H 2023 A possible deforestation-induced synoptic-scale circulation that delays the rainy season onset in Amazonia Environ.Res.Lett.18 044041 Costa M H, Yanagi S N M, Souza P J O P, Ribeiro A and Rocha E J P 2007 Climate change in Amazonia caused by soybean cropland expansion, as compared to caused by pastureland expansion Geophys.Res.Lett.34 2-5 Culf A D, Fisch G and Hodnett M G 1995 The albedo of Amazonian forest and ranch land J. Clim.9 1544-54 Danabasoglu G 2019a NCAR CESM2 model output prepared for CMIP6 CMIP historical Earth System Grid Federation (https://doi.org/10.22033/ESGF/CMIP6.10023)Danabasoglu G 2019b NCAR CESM2-FV2 model output prepared for CMIP6 CMIP historical Earth System Grid Federation (https://doi.org/10.22033/ESGF/CMIP6.11281)Danabasoglu G 2019c NCAR CESM2-WACCM model output prepared for CMIP6 CMIP historical Earth System Grid Federation (https://doi.org/10.22033/ESGF/CMIP6.10023)Danabasoglu G 2019d NCAR CESM2-WACCM-FV2 model output prepared for CMIP6 CMIP historical Earth System Grid Federation (https://doi.org/10.22033/ESGF/CMIP6.11282)Danek C, Shi X, Stepanek C, Yang H, Barbi D, Hegewald J and Lohmann G 2020 AWI AWI-ESM1.1LRmodel output prepared for CMIP6 CMIP historical Earth System Grid Cecil L D, Nelson B R and Prat O P 2015 PERSIANN-CDR: daily precipitation climate data record from multisatellite observations for hydrological and climate studies Federation (https://doi.org/10.22033/ESGF/CMIP6.9301)