Main

Over the past few decades, rapid forest loss2,3 in the Amazon region has markedly affected a series of crucial services4,5,6,7,8 that support the global carbon cycle, climate change mitigation, societal well-being and biodiversity. The interplay between deforestation and the water cycle is particularly noteworthy, as it is an important determinant of ecosystem resilience and stability1,9. Deforestation not only reduces evapotranspiration (ET), thereby affecting precipitation in the deforested and neighbouring areas1,5,9,10, but also triggers changes through atmospheric and oceanic feedback11, second-order effects12 and large-scale teleconnection effects13. A recent satellite-based study demonstrated that tropical deforestation leads to substantial reductions in precipitation, showing a scale-dependent deforestation–precipitation relationship across various spatial resolutions1. They use a ‘space-for-time’ moving window approach, quantifying the impacts of deforestation by comparing the precipitation changes over pixels experiencing forest loss with neighbouring pixels that had experienced less forest loss, under the assumption that all pixels share similar climatic backgrounds11.

However, the moving window approach primarily emphasizes local effects11,12, which are driven by substantial decreases in water vapour resulting from ET reductions over deforested pixels. Moreover, nonlocal effects related to mesoscale atmospheric circulation and regional precipitation recycling due to deforestation are underestimated or blended with the local effects13,14,15. The crucial role of nonlocal effects in the deforestation–precipitation relationships over the Amazon is important, as 50–60% of the water vapour contribution to precipitation comes from mesoscale atmospheric circulation involving both the atmosphere and the ocean16,17. The important roles of precipitation recycling18 on forest–precipitation relationships over the Amazon have been widely reported17,19,20. Ignoring these mechanisms, for example, in the moving window approach, leads to an incomplete and potentially inaccurate understanding of the deforestation–precipitation feedback14,15,16. Until now, there is limited understanding of how deforestation affects precipitation across different spatial scales in terms of distinguishing local and nonlocal effects, particularly regarding the variations in precipitation recycling between wet and dry seasons.

Climate models are useful tools for accessing both local effects induced by ET reductions and nonlocal effects from mesoscale atmospheric circulation, demonstrating their ability in diagnostic attributions14,15,21. Here we use a regional climate model integrated with a high-resolution satellite-based forest cover dataset22,23 spanning from 2000 to 2020 (for dataset details, refer to Extended Data Table 1) to access both the local and nonlocal effects induced by deforestation. Our regional climate model, a recent version of the Weather Research and Forecasting (WRF) model24, coupled with the Water Vapor Tracers module (WVT)25, enables tracking of the moisture evaporated from the Amazon basin region, covering the whole Amazon rainforest eco-region since 2000 (area outlined in Fig. 1c). Furthermore, the CLM-tiling land surface scheme26 is activated, adopting an activated mosaic approach with homogeneous subgrid tiles (model configurations are detailed in Extended Data Table 2), exhibiting more reasonable climate sensitivity in response to forest cover change compared with other land surface schemes of WRF in the tropics26.

Fig. 1: Simulated contrasting precipitation responses to deforestation at different scales (0–1,000 km) over the Amazon region.
figure 1

a,b, Bars represent the mean ΔP of pixels in each buffer with different distances from deforested grids during the wet (a, DJF) and dry (b, JJA) seasons. The 0-km buffer corresponds to deforested grids. Insets represent the relationships between ΔP and subgrid-scale forest cover fraction in each deforested grid. Error bars show the standard error of the mean. Asterisks indicate distributions with averages that are statistically different from zero (two-sided Student’s t-test; **P < 0.01, *P < 0.05) for ΔP (S2020 minus S2000). Linear regression is used to fit the relationship between the forest loss fraction and ΔP in the insets, with the P value from the Wald test, testing the null hypothesis that the slope is zero. c,d, Spatial pattern of ΔP (S2020 minus S2000) during the wet (c, DJF) and dry (d, JJA) seasons. Stippling indicates pixels locally statistically different from zero (P < 0.05, two-sided t-test), all passing a more rigorous field significance test corrected for the FDR (α = 0.05).

Full size image

In the analysis, we compare two simulation experiments, S2000 and S2020, representing forest cover changes from 2000 to 2020 (Extended Data Fig. 1). In our methodology, the precipitation effects of Amazon deforestation are quantified by S2020 minus S2000. The total precipitation effect (ΔP) is divided into local effects (ΔPt, effects from tagged moisture) and nonlocal effects (ΔPnt, effects from untagged moisture). Local effects represent changes in precipitation from tagged moisture, where ‘tagged’ refers to tracking and identifying the source of atmospheric moisture throughout the simulation. This moisture originates from the land surface of the Amazon (area outlined in Fig. 1c) and reflects the contribution of ET reductions. Nonlocal effects (ΔPnt) are derived by subtracting ΔP from ΔPt, capturing the precipitation changes associated with mesoscale atmospheric circulation. We also apply buffers (for buffer settings, refer to Supplementary Fig. 1) to investigate the response of ΔP, ΔPt and ΔPnt at different scales ranging from deforested grids to continents. Moreover, four high-resolution satellite-observed rainfall datasets (for details, refer to Extended Data Table 1) are used to quantify the Amazon deforestation–precipitation response based on the moving window approach, serving as a validation and constraint for our modelling results (for more details, refer to the Methods).

Precipitation response to deforestation

From 2000 to 2020, forest loss occurred in 60.19% of grid cells in the Amazon region (Extended Data Fig. 1c). Partial forest gain occurred during this period but was limited to a low level (Extended Data Fig. 1c), the impacts of this on our presented results of forest loss are negligible (Supplementary Fig. 2). During the wet season (December, January and February (DJF)), the climate simulations indicate that deforestation causes a significant increase in precipitation (ΔP = 5.75 ± 0.35 mm month−1) over deforested grid cells (two-sided Student’s t-test; P < 0.01; Fig. 1a, buffers = 0 km). Grid-scale positive precipitation effects and forest loss fraction are positively correlated (Fig. 1a, inset), with precipitation increasing at a rate of 0.96 mm month−1 per percentage point loss in forest cover (P < 0.01). However, the large increase in ΔP weakens greatly within the 20-km distance buffers (1.22 ± 0.22 mm month−1) and approaches zero in the 40 km buffer (0.18 ± 0.17 mm month−1). Small decreases in precipitation are found in buffers of 60 km (−0.28 ± 0.14 mm month−1) and 80 km (−0.46 ± 0.13 mm month−1). Continuous reductions in precipitation are detected in wider buffers, becoming gradually negligible beyond 600 km (Fig. 1a).

During the dry season (June, July and August (JJA)), a nearly negligible relationship (0.09 mm month−1 per percentage point loss in forest cover; R2 = 0.02) between subgrid-scale deforestation and precipitation is found over deforested grid cells (Fig. 1b, inset). The most pronounced reductions in dry season precipitation appear within buffers of 40 km (−1.55 ± 0.06 mm month−1) and 60 km (−1.49 ± 0.05 mm month−1), and precipitation continues to decrease to 1,000 km (−0.52 ± 0.02 mm month−1; Fig. 1b). The spatial heterogeneity of precipitation changes is much lower (Fig. 1d), indicating a continuous signal of reductions in precipitation over each buffer (Fig. 1b). Precipitation changes clustered on the northern edge of southern America (Fig. 1d) owing to high background rainfall (Supplementary Fig. 3). At the continental scale (all land surface grids within the model domain), the simulations show that Amazon deforestation causes reductions in precipitation during both wet seasons (DJF, −0.25 ± 0.002 mm month−1) and dry seasons (−0.67 ± 0.001 mm month−1) (Fig. 1c,d).

Local and nonlocal precipitation impacts

The water vapour tracer module enables decomposing the modelled precipitation effects (ΔP) into two components: (1) a local component induced by ET-related water vapour changes from the tagged Amazon region and (2) a nonlocal component resulting from changes in mesoscale atmospheric circulation within the model domain (Fig. 2). During the wet season, local ΔPt exhibits a positive precipitation effect of 1.08 ± 0.13 mm month−1 in the deforested grid cells but is negative in all buffers, with the greatest reductions occurring at distances of 60–100 km (−0.55 ± 0.04 mm month−1 to −0.56 ± 0.05 mm month−1) (Fig. 2a). The negative local precipitation effects then weaken gradually from 100 km to 1,000 km buffers. The wet season simulations show a greater increase in nonlocal precipitation in deforested grid cells (ΔPnt = 4.67 ± 0.24 mm month−1) resulting from changes in mesoscale atmospheric circulation. Positive nonlocal precipitation effects weaken sharply in buffers ranging from 20 km (1.45 ± 0.15 mm month−1) to 60 km (0.27 ± 0.10 mm month−1), diminishing for buffers larger than 80 km (0.11 ± 0.09 mm month−1) (Fig. 2a). During the dry season, two sources of precipitation change are both negative over all the buffers ranging from scales of 0 km to 1,000 km (Fig. 2b). At the continental scale, local effects from ET reductions dominate negative precipitation effects over all buffers, with the largest ΔPt occurring at buffers of 20–100 km (largest intensity of ΔPt = −0.90 ± 0.02 mm month−1).

Fig. 2: Comparisons between the contributions of local and nonlocal effects to the simulated ΔP during wet and dry seasons.
figure 2

a,b, Mean ΔPt (green bars, local effects from tagged moisture) and ΔPnt (blue bars, nonlocal effects from untagged moisture) of pixels within each buffer at different distances from deforested grids during the wet (a) and dry (b) seasons. Error bars show the standard error of the mean. c, Spatial distribution of ΔPt, showing local effects induced by ET reductions from a tagged Amazon region during the wet season (DJF). d, Spatial distribution of ΔPnt, showing nonlocal effects induced by changes in mesoscale atmospheric circulation during the wet season (DJF). e,f, Spatial distribution of local (e) and nonlocal (f) precipitation effects during the dry season (JJA). The stippling (c–f) indicates pixels with values that are locally statistically different from zero (P < 0.05, two-tailed Student’s t-test), all passing a more rigorous field significance test corrected for the FDR (α = 0.05).

Full size image

The consistent patterns of local effects on precipitation within surrounding buffers during both seasons (Fig. 2c,e) indicate the dominant role of ET from rainforests in regulating the regional hydroclimatic cycle17,19,20. During the wet season, ET reductions govern regional precipitation effects, whereas substantial positive nonlocal precipitation effects are detected mainly over the deforested grid cells (Fig. 2d). This result aligns with potential increases in precipitation induced by deforestation reported in previous studies27,28,29. The precipitation recycling ratio ρ (ρ < 50% indicates that more precipitation comes from water vapour from mesoscale atmospheric circulation; ρ > 50% indicates that more precipitation comes from local ET; for more details, refer to the Methods) is low during the wet season over the Amazon region. Less than 1% of the precipitation recycling ratio of the region exceeded 50% (Extended Data Fig. 2a), indicating the dominance of mesoscale atmospheric circulation for regional precipitation. Conversely, during the dry season, forest ET is important in regulating drought events over the Amazon17. This regulating effect is facilitated by a high precipitation recycling ratio over the Amazon, with approximately 30% of regions maintaining a precipitation recycling ratio exceeding 50% (Extended Data Fig. 2b). Overall, continental precipitation patterns are primarily driven by mesoscale atmospheric circulation, which redistributes rainfall across the Amazon region without notably affecting the overall amount of precipitation at the continental scale (Fig. 2d,f). These results agree with those from previous studies based on different methodological approaches16,30.

Mechanisms of precipitation response

To further explore the potential mechanisms governing the different Amazon deforestation–precipitation responses during the wet and dry seasons, we compare the ET changes (ΔET) on the continental land surface (Fig. 3a,b) and water vapour changes (ΔWV) in the vertical cross-sections above the land area between 10° S and 15° S (Fig. 3c,d). This area surrounds the southern part of the Amazon region and overlaps with grid cells that experienced substantial deforestation during the study period. During the wet season, considerable increases in water vapour (ΔWV > 0.10 g kg−1) occur at pressures of 600–800 hPa over the deforested grids, resulting in notable wetting. By contrast, the land surface ET during the dry season was largely negative, resulting in drying. In the vertical cross-section, ET reductions cause substantial decreases in moisture near the land surface during both seasons (Fig. 3c,d). Different water vapour responses in the 600–800 hPa pressure levels above deforested grids account for the differing impacts of deforestation on precipitation in the Amazon during the wet and dry seasons. This occurrence of water vapour compensation over deforestation regions has been reported in previous research on South Asia27 based on simulations using RegCM31 (International Centre for Theoretical Physics Reginal Climate Model) and CESM32 (Community Earth System Model).

Fig. 3: Simulated water vapour changes over land surface and vertical levels induced by deforestation.
figure 3

a,b, ET changes at the land surface level during the wet (a) and dry (b) seasons. c,d, Water vapour (shaded; g kg−1) and wind (arrow; m s−1) changes at vertical levels (mean between 10° S and 15° S) during the wet (c) and dry (d) seasons. The stippling in a,b indicates pixels with locally statistically different from zero (P < 0.05, two-tailed Student’s t-test), all passing a more rigorous field significance test corrected for the FDR (α = 0.05).

Full size image

Our results indicate that the mechanisms governing the Amazonian deforestation–precipitation response are related to the balance between direct drying (related to local effects) and indirect warming (related to nonlocal effects) processes during the wet and dry seasons (Fig. 4), in which the indirect warming-up process is the reason for the reverse of precipitation effects between seasons. Local warming detected over deforested grids (Supplementary Fig. 4) during both seasons, which is consistent with previous research based on satellite data12,13 and numerical simulations5, causes substantial vertical air movement and surface air convergence, reducing land surface pressure and triggering changes in mesoscale atmospheric circulation. However, water vapour compensation from mesoscale atmospheric circulation is more pronounced during the wet season (Fig. 4a) than during the dry season (Fig. 4b), leading to reversed precipitation responses between seasons.

Fig. 4: Mechanisms of deforestation–precipitation feedback in the Amazon during the wet and dry seasons.
figure 4

The results are derived from simulations conducted between 2001 and 2022 for the wet (a) and dry (b) seasons. Variables T, ET, q and P (labels in this figure) represent surface air temperature, evapotranspiration, water vapour and precipitation, respectively. Repartition represents the water vapour interplay across different spatial scales regulated by local and nonlocal effects of deforestation. Variations in q are regulated by different mesoscale atmospheric circulation patterns and precipitation recycling ratios drive the different regimes during the wet and dry seasons. The visual elements used in this figure were adapted from the Integration and Application Network, University of Maryland Center for Environmental Science (https://ian.umces.edu/media-library), under a Creative Commons licence CC BY-SA 4.0, and the figure was developed using Microsoft PowerPoint software.

Full size image

This type of response induced by warming effects is reported in research findings that deforestation affects regional climate primarily through changes in precipitation patterns and atmospheric stability27,28,29. For example, deforestation on the Maritime Continent was found to increase both surface temperature and local precipitation because of enhanced atmospheric instability and dynamic moisture advection27. The enhanced precipitation induced by water moisture compensation leads to an increase in the annual mean precipitation27. Moreover, deforestation in sub-Saharan Africa generally reduced rainfall, particularly in areas north of the equator, with the severity of the impact varying on the basis of the extent of tree cover loss28. Elsewhere, deforestation in southern West Africa increased the frequency of afternoon storms, especially over larger deforested patches, because of mesoscale circulations and enhanced land–sea thermal contrast29.

During the Amazon dry season, the effect of local warming is similar to that in the wet season (Fig. 4b) but without sufficient water vapour compensation16, indicated by a higher precipitation recycling ratio (Extended Data Fig. 2). Over the deforested grid cells, a net decrease in precipitation is driven by strong local effects from ET reductions, rather than weaker nonlocal effects from mesoscale atmospheric circulation. At the continental scale, ET reductions during the dry season are pronounced, dominating decreases in water vapour17 and suppressing the weak wetting effects from mesoscale atmospheric circulation, thereby causing net negative precipitation effects (Fig. 4b). Consequently, deforestation is associated with exacerbating drought conditions during the dry season or in drought year. Moreover, the two-way feedback between drought and deforestation represents a pronounced environmental pressure on the Amazon rainforest33. Although our study addresses the feedback of Amazon deforestation to precipitation, it does not examine the potential feedback of drought on vegetation degradation34, which is reported as a key factor in accelerating forest loss over the Amazon35,36.

Robustness of the analyses

We assessed the robustness of our analyses concerning the influences of background climate (for example, La Niña and El Niño years), variations in forest cover datasets and the influence of unrestricted water convergence into the model domain within the prescribed boundary conditions. Regarding the influences of background climate, we compared precipitation effects for wet and dry seasons from 2001 to 2022 (Supplementary Fig. 5). Our results indicate that precipitation effects within buffers remain consistent under different climate forcing, regardless of variations in the intensity of precipitation changes (for example, during moderate and strong La Niña and El Niño years, refer to Supplementary Figs. 6a–e and 7a–e). Furthermore, spatial patterns of precipitation response to Amazon deforestation show strong alignment across different climate-forcing scenarios (Supplementary Figs. 6f–j and 7f–j).

To evaluate how different forest cover datasets might affect our results, we compared the GLAD (Global Land Analysis and Discovery)23 dataset with the TMF (tropical moist forest)2 and ESA-CCI (European Space Agency Climate Change Initiative)37 datasets, as shown in Extended Data Fig. 3. By repeating the simulation with an expanded model boundary (600 × 480 grids versus 350 × 285 grids in the main results; 20 km resolution) and allowing for unrestricted water convergence, we observed results consistent with the original simulation (Extended Data Fig. 4). Future studies using alternative methodologies to assess water compensation outside the Amazon region are necessary to quantify the extent of nonlocal effects. Moreover, the separation of local and nonlocal temperature effects exhibits uncertainties across different Earth System Model experiments14,15,38. Therefore, further research using various climate models and methodologies is needed to explore potential model dependencies in distinguishing local and nonlocal precipitation effects caused by deforestation.

Comparison of approaches

When comparing our discovered patterns from different methodologies, we first applied the space-for-time moving window approach based on satellite data, following the framework in ref. 1 (for more details, refer to Methods and Supplementary Fig. 8). We also detected reversed seasonal precipitation effects of Amazon deforestation, with values of DJF (0.05 ± 0.06 mm month−1) and JJA (−0.04 ± 0.04 mm month−1) as shown in Supplementary Fig. 9. Next, when comparing the two statistical approaches in the moving window framework, namely, the binary method used in ref. 1 and a regression method using all samples within a window12,13, we observed variations in the intensity of precipitation impacts over deforested pixels in our simulations (Supplementary Fig. 10c–f). This variation may be attributed to the inherent characteristics of the moving window approach (Supplementary Fig. 8; see comparisons of climate modelling and the moving window approach in the Methods, which tends to emphasize local effects12,13 while overlooking nonlocal impacts). Importantly, ref. 1 was the first to use the space-for-time moving window approach, to our knowledge, to quantify precipitation responses to deforestation. Given the variation among different statistical approaches, further research is needed to evaluate the effectiveness of this framework in capturing precipitation responses and to investigate potential discrepancies between methodologies1,13 (Methods), especially because these statistical approaches are more commonly used in studies focused on temperature responses11,12,13.

The Amazon region is projected to experience intensive deforestation under a regional rivalry scenario (SSP3-RCP4.5 in the Global Change Analysis Model), with an estimated mean deforestation rate reaching −8.79% over the region39 by 2100, compared with its baseline in 2015 (Supplementary Fig. 11). In even more pessimistic scenarios, such as the previous A2 scenario described in the Special Report on Emission Scenarios, projections suggest that the Amazon rainforest could nearly disappear altogether40,41. However, determining whether the nonlocal effects of projected deforestation will mirror those in the current deforestation scenario is challenging, especially considering the anticipated climate changes in future projections42. Therefore, further research should provide more robust evidence on the nonlocal effects of deforestation and clarify the climate feedback mechanisms associated with projected deforestation in the Amazon region. It is crucial to avoid simplistic extrapolations of future impacts based solely on the linear hypothesis1 using current precipitation sensitivity to deforestation (Supplementary Fig. 12), as the biophysical impacts of deforestation on precipitation involve both linear6 and nonlinear36 processes.

Climate mitigation implications

The contrasting seasonal precipitation responses to deforestation at different scales in the Amazon necessitate different strategies for addressing environmental risks related to biodiversity protection43, forest management44 and agricultural production6,10. At the scale of deforested pixels, given the crucial part that rainforests play in regulating the regional climate17, more rainfall during the wet season over deforested regions may exacerbate local floods, especially for agricultural land use45. Our analyses indicate that deforestation could lead to increased rainfall through mesoscale atmospheric circulation. If substantial enough to affect extreme events or elevate catchment wetness, these increases could exacerbate the wet season flooding in certain deforested regions, harming regional agriculture and the social economy45,46. Relevant, for example, are locations proximal to the Amazon in which agriculture conversions have resulted in elevations of local groundwater tables45. Although our results show that high-intensity deforestation may slightly increase rainfall in some deforested regions during the dry season, its magnitude is negligible in mitigating the large-scale regional drought induced by forest cover loss.

At larger scales ranging from neighbouring regions to continents, continued deforestation in the Amazon could lead to declines in total rainfall, threatening endemic species in conjunction with additional habitat loss and a warming climate43. Reduced regional precipitation might result in substantial economic losses in agriculture, with crop yields declining by 0.5% for each percentage point reduction in rainfall47. Intensified regional droughts could increase the risks associated with wildfire frequency48, carbon sequestration49 and human and livestock health50. These potential impacts at different scales highlight the need for targeted strategies in environmental management and land use planning to mitigate the adverse effects of deforestation on ecosystem services in the Amazon region, especially considering climate variability and ongoing land pressure.

In conclusion, our results, when considered alongside those in a previous assessment1, collectively advance the understanding of the local and nonlocal precipitation to deforestation at different scales of observation, as well as for different analysis methods. Owing to their pivotal roles in regulating regional and global climate, sustained efforts are needed to protect the remaining forest in the Amazon, as well as rehabilitate degraded lands. The understanding of the deforestation–precipitation relationship in the Amazon region reported here calls for dedicated forest management in consideration of afforestation and expanding agricultural activities in hotspot areas of forest conversion to croplands, as well as the frontiers to which deforestation may expand in the future.

Methods

Coupled land–atmosphere regional weather model

The WRF model, developed by the National Center for Atmospheric Research24, is a fully coupled mesoscale model widely used for regional climate studies4,26,51. The Advanced Research version of the WRF model (WRF-ARW, v.4.3.3) was used in this study24. The water vapour tracer tool25 was coupled into the WRF model (WRF-WVT) to track moisture sources for precipitation by replicating the prognostic equation for total moisture and creating new variables for tracers of different moisture species. The method is classified as an online Eulerian approach, as the predictive equations for the tracer moisture are solved simultaneously with the governing equations of the model16,25.

Using the WRF-WVT tool, we can obtain new variables, such as raint (precipitation), that represent the amount of precipitation originating from evapotranspiration (ET) within the Amazon basin region. We consider the border of extension of the Amazon rainforest biome in 2000, excluding surrounding tropical savanna biomes52, covering the whole Amazon rainforest owing to deforestation, which has occurred mainly in the marginal southern Amazon rainforest region since 2000 (Extended Data Fig. 1c). In equation (1), total precipitation (raintotal) is the original variable from the WRF output; raint represents precipitation originating from local evaporative moisture (tagged); rainnt represents rain that formed from water vapour (not tagged) outside the Amazon region and is controlled by regional mesoscale atmospheric circulation. The precipitation recycling ratio (ρ) is calculated as the ratio of local precipitation (raint) to total precipitation (raintotal) (equation (2)), referring to the proportion of precipitation originating from local evaporative moisture to the total precipitation in each model grid18. The coupling of WRF-WVT requires specific settings of parameterization schemes (more details are provided in Extended Data Table 2).

$${{\rm{rain}}}_{{\rm{t}}{\rm{o}}{\rm{t}}{\rm{a}}{\rm{l}}}={{\rm{r}}{\rm{a}}{\rm{i}}{\rm{n}}}_{{\rm{t}}}+{{\rm{r}}{\rm{a}}{\rm{i}}{\rm{n}}}_{{\rm{n}}{\rm{t}}}$$
(1)
$$\rho ={{\rm{rain}}}_{{\rm{t}}}/{{\rm{r}}{\rm{a}}{\rm{i}}{\rm{n}}}_{{\rm{t}}{\rm{o}}{\rm{t}}{\rm{a}}{\rm{l}}}$$
(2)

Moreover, to enable the subgrid vegetation cover fraction in WRF, we coupled the Community Land Model (CLM)53 with the activated subgrid vegetation cover representation approach (CLM-tiling)25 into the WRF model to accurately represent the land and atmosphere interactions. CLM is a state-of-the-art land surface model originally coupled with the Community Earth System Model32,53. The subgrid vegetation cover representation function is deactivated when coupled with the WRF model51. CLM-tiling fixed the deactivation of the subgrid vegetation cover approach in WRF-CLM, considering surface fluxes over four vegetation cover types in the subgrid scale, with weightings based on fractional coverage26. The subgrid vegetation cover representation approach is important in researching the climate feedback of deforestation as it provides more accurate land properties for each model grid51.

Experimental design

We designed a single domain simulation with a horizontal spatial resolution of 20 km and 350 × 285 horizontal model grid cells. The model domain covers most of the southern American region (Supplementary Fig. 1) to account for the effects of large-scale monsoons originating over the oceans. The wet and dry seasons are set as DJF (wet season) and JJA (dry season) according to the climatology of southern Amazon regions in which deforestation mainly occurs (Supplementary Fig. 13). Two paired WRF simulations were conducted during the wet (DJF, from 15 November to 28 February) and dry (JJA, from 15 May to 31 August) seasons. The simulations were running from 2001 to 2022, covering the period of deforestation happening over the Amazon and accounting for the potential influence of interannual variability in climate conditions. The first 16 days of each simulation are referred to as the spin-up period, which is not included in the validations and analyses. For details of evolution and equilibrium of soil moisture (10 soil layers), ET and precipitation, refer to Supplementary Figs. 14–17. All the simulations were performed with 30 vertical levels extending from the surface to 100 hPa. Initial and lateral boundary conditions were derived from the European Centre for Medium-Range Weather Forecasts fifth-generation reanalysis (ERA5) with a spatial resolution of 0.25° × 0.25° and an hourly temporal resolution. Sea surface temperatures in each simulation were dynamically updated every 6 h from the ERA5 data. Specifically, leaf area index (LAI) parameters for different land cover types in the ecosystems of the Amazon region were updated using the MODIS LAI product at 0.05° resolution for the year 2020 (data detailed in Extended Data Table 1) according to the settings used in a previous study51.

We used high-resolution (30 m) forest cover data for 2000 and 2020 from the Global Land Analysis and Discovery (GLAD) forest extent and height change dataset22,23 to resample forest cover change information at a 20-km resolution (Extended Data Fig. 1c). GLAD forest cover data from 2000 and 2020 were mapped by attributing pixels with forest height ≥5 m as the ‘forest’ land cover class, which agrees with the definition of ‘forest’ by the Food and Agriculture Organization of the United Nations. We derived forest cover changes by comparing the years 2000 and 2020, defining ‘deforestation’ as grid cells in which forest cover decreased from 2000 to 2020 at a 20-km resolution.

Two forest cover scenarios were devised for the domain: (1) a control scenario (S2000) representing land cover and land use fractions derived from MODIS land cover in 2000 (default in WRF); and (2) a deforestation scenario (S2020) created by overlaying GLAD forest cover changes from 2000 to 2020 onto the MODIS land cover and land use fractions in 2000. The percentages of non-forest categories (for example, bare ground, grassland, shrubland and cropland) increased proportionally for grid cells experiencing net forest losses, and the percentages of forest categories decreased proportionally from the original percentages in 2000. Conversely, in grid cells with net forest gain, the percentages of forest categories increased proportionally, whereas the percentages of non-forest categories decreased. As a result, the S2020 scenario maintained the same relative proportion among the five MODIS forest cover types (for example, evergreen needleleaf forest, evergreen broadleaf forest, deciduous needleleaf forest, deciduous broadleaf forest and mixed forests) as that in 2000, varying only in total forest cover percentage. After we derived the deforestation scenario (S2020) over the whole domain, we substituted the land cover and land use fraction of the Amazon region with MODIS land cover in 2000 and obtained the S2000 scenario. Finally, the biophysical climate effects of Amazon deforestation were quantified as the results of S2020 minus S2000, as shown in the main text. Another paired simulation for wet and dry seasons under a larger model domain (600 × 480 grids, 20-km resolution) was repeated to assess the influence of the model boundary (Extended Data Fig. 4).

Model performance

The simulations of the S2020 experiment under climate forcing in 2020 are used for model validation. The first 16 days of simulation constitute the spin-up time, and the remaining periods (JJA, 1 June 2020 to 31 August 2020; and DJF, 1 December 2020 to 28 February 2021) are used for validation analyses. Model performance in terms of temporal patterns was evaluated by comparing the S2020 experiments with daily surface air temperature and precipitation data from the Global Surface Summary of the Day54 (GSOD) in situ weather stations in South America. The GSOD dataset is a comprehensive, real-time, updated collection of daily weather summaries from global weather stations produced by the National Centers for Environmental Information, which is part of the National Oceanic and Atmospheric Administration. The GSOD dataset provides historical observations on temperature, precipitation, wind and other meteorological variables from more than 9,000 global in situ stations. In South America, 196 GSOD stations recorded available surface air temperature during the wet season in 2020, whereas 195 stations did so during the dry season in 2020. Moreover, a total of 62 and 119 stations recorded available daily precipitation in the wet and dry seasons, respectively. We also used satellite-based precipitation datasets (GPM55, CHIRPS56, PERSIANN-CCSCDR57 and PERSIANN-NOW58) (Extended Data Table 1) to validate spatial patterns of the simulated precipitation.

First, the WRF simulations were validated with GSOD by comparing the GSOD records of surface air temperature and daily precipitation with the simulated results of corresponding grids in the model domain. The results indicate that our model accurately represented the temporal variability of daily surface air temperature and precipitation. The model represented the historical variability in the near-surface air temperature compared with the GSOD weather station data during both wet (R = 0.62, P < 0.01) and dry (R = 0.95, P < 0.01) seasons (Extended Data Fig. 5). Our model successfully captures the temporal variability compared with GSOD weather station data, showing good consistency in daily precipitation (DJF: R = 0.59, P < 0.01; JJA: R = 0.59, P < 0.01).

To validate the spatial variability of surface air temperature and precipitation, the monthly accumulated precipitation from the WRF simulations was compared with satellite records at a 20-km resolution. We resampled four rainfall datasets, for which the original resolutions are finer than 20 km, at a resolution of 20 km for validation. The results indicated that our model generally captured the spatial patterns of monthly precipitation during the wet and dry seasons (Extended Data Fig. 6). During the wet season, precipitation is complex in the Amazon region; our simulated precipitation captured the spatial patterns compared with the satellite data, aligning with the diagonal line and showing a correlation coefficient of 0.48 to 0.56 (P < 0.01) (Extended Data Fig. 6a–d). During the dry season, the results indicate strong spatial relationships between the simulated and satellite-recorded precipitation (R values ranging from 0.80 to 0.87, P < 0.01) (Extended Data Fig. 6e–h).

Comparisons of climate modelling and moving window approach

In this study, climate modelling and the moving window approach were used to quantify the precipitation impacts of Amazon deforestation. However, these methods differ in their assumptions and strengths and weaknesses (Supplementary Fig. 8). For climate modelling, we can obtain the comprehensive impacts of deforestation and further explore the local and nonlocal effects on buffers at different scales. For the moving window approach using satellite-observed datasets, the realistic impact of deforestation on precipitation can be quantified. However, this quantification is limited to deforested pixels, and the effects on neighbouring and remote pixels cannot be quantified. Moreover, the intrinsic effects are ignored over deforested pixels when deforested pixels are compared with neighbouring pixels. In this section, we provided a comprehensive comparison of these two methods using model simulations and satellite observations.

In climate modelling, the biophysical climate effects of Amazon deforestation can be quantified as the results of S2020 minus S2000 from factorial experiments (for example, ΔP, ΔPt and ΔPnt). To investigate the impacts of Amazon deforestation at different scales, the precipitation response (ΔP) of different buffers (distances ranging from 0 km to 1,000 km from deforested grids) was analysed (Supplementary Fig. 1). The 0-km buffer represents the deforested grids in the model simulations, indicating the precipitation response in the deforested grids. The 20–1,000 km buffers conclude all the grids within the distance of 20–1,000 km from the deforested grids (including inner buffers and deforested grids), analysing precipitation response ranging from smaller (20 km) to larger (1,000 km) scale. Local (ΔPt) and nonlocal (ΔPnt) precipitation effects at different buffers are shown as with ΔP (Fig. 2). We focus on the Amazon deforestation and water vapour evaporated from the tagged Amazon basin region. Moisture evaporated from the tagged Amazon region could contribute to precipitation (ΔPt) on nearby or remote model grids owing to wind transport. Nonlocal effects from atmospheric circulation can also influence nonlocal precipitation (ΔPnt) in each buffer easily.

A two-sided Student’s t-test was used to statistically assess whether the distributions of mean precipitation effects (for example, ΔP, ΔPt and ΔPnt) are statistically different from zero. Moreover, we conducted a more rigorous field significance test using the Benjamini–Hochberg method to control the false discovery rate (FDR) at α = 0.05. Only model grids (ΔP) with locally significant P values (P < 0.05) that remain significant after FDR correction are highlighted in the results (Fig. 1c,d). Linear regression is used to fit the relationship between forest loss fraction and ΔP over deforested grids. The P value is obtained from the Wald test, testing the null hypothesis that the slope is zero (Fig. 1a,b).

We then apply the moving window approach to investigate if the contrasting seasonal precipitation impacts of Amazon deforestation can be detected by satellite observation. We applied the space-for-time moving window approach following the framework in ref. 1. The widely used moving window approach1,59,60 quantifies the observed precipitation effects of deforestation on the basis of four high-resolution satellite rainfall datasets (Extended Data Table 1) at a resampled spatial resolution of 20 km (the results are shown in Supplementary Fig. 9). The four precipitation datasets (GPM, CHIRPS, PERSIANN-CCSCDR and PERSIANN-NOW) have the finest resolution among all the satellite-based rainfall datasets used in ref. 1, whose original spatial resolutions are all finer than 20 km (detailed in Extended Data Table 1) and of monthly temporal resolution from 2000 to 2020. The mean monthly precipitation from 2001 to 2003 represents the precipitation climatology before deforestation over the Amazon region. The mean monthly precipitation from 2018 to 2020 represents the precipitation climatology after deforestation. The moving window approach compares the forest loss fraction and precipitation change of each pixel with its immediate neighbours (3 × 3 moving windows are used in our analysis). The results from mean monthly precipitation changes of 5 years and 5 × 5 moving windows are handled as sensitivity tests, showing the same results among different parameters used in the method1.

We applied the moving window approach based on the threshold parameters in ref. 1: (1) deforested pixels must have experienced 0.1% more forest loss over time than their neighbouring control pixels, ensuring pixels experiencing high-intensity deforestation to influence climate; and (2) the difference between the precipitation of deforested pixels and neighbouring control pixels must be lower than 10% to keep a similar climate background between pixels. We updated the data on forest cover changes and satellite-recorded rainfall during the period 2000–2020 instead of the previous period 2003–2017 (ref. 1). These thresholds were also used to constrain our results when deriving precipitation sensitivity to deforestation (Supplementary Fig. 12).

Moreover, there are two distinct statistical approaches within the moving window framework, namely, the binary method used in refs. 1,60 and a regression method using all samples within a window61,62. To directly compare the differences between these two methods, we applied them to the output of WRF simulations. The mean monthly precipitation from 2001 to 2003 in the S2000 experiment was used to represent the precipitation climatology before deforestation, whereas the mean monthly precipitation from 2018 to 2020 in S2020 experiment served as the precipitation climatology after deforestation. Reversed seasonal precipitation can also be detected by both methods (binary-based: 1.40 ± 0.31 mm month−1 for DJF and −0.18 ± 0.14 mm month−1 for JJA; regression-based: 2.60 ± 0.24 mm month−1 for DJF and −0.05 ± 0.10 mm month−1 for JJA), as shown in Supplementary Fig. 10.

Future deforestation

We used forest cover change projections derived from the Global Change Analysis Model (GCAM) at a 0.05° resolution for 2020–2100 under the Shared Socioeconomic Pathway 3-Representative Concentration Pathway 4.5 scenario39 (Supplementary Fig. 11). The GCAM model incorporated climate and land use effects on future forest cover, and the deforestation rate was computed annually relative to the 2015 baseline, with the data resampled to a resolution of 20 km.