Datasets

We used 18 precipitation datasets, listed in Extended Data Table 1. All datasets were downloaded at the highest available spatial resolution, which for some datasets was 0.04°, or approximately 4 km at the Equator. Data were obtained as monthly means or converted to monthly mean using the Python package xarray52. We categorized precipitation datasets as satellite (n = 10), station (n = 4) and reanalysis (n = 4). Satellite datasets are those based primarily on data from satellite sensors and include datasets that have both satellite and station-based data (that is, merged datasets). Station datasets include only ground-based information from weather stations and rain gauges. Reanalysis products are models constrained by surface and satellite data. Precipitation datasets have been compared previously over the Amazon20 highlighting the limited station data over tropical forest regions. Time series of precipitation (Supplementary Fig. 1) reveal variability across the different datasets highlighting the need to analyse impacts of deforestation across multiple datasets.

To analyse the changes in forest canopy cover, we used data from the Global Forest Change (GFC) version 1.9 (ref. 10). GFC v1.9 provides forest canopy cover in the year 2000 and subsequent annual forest loss from 2001–2020 at 30-m resolution. We analysed forest cover and precipitation changes over the period 2003 to 2017, which was the period common to all datasets.

Analysis across multiple spatial scales

We analysed the impacts of forest loss across a range of scales (0.05°, 0.1°, 0.25°, 0.5°, 1.0° and 2.0°). Each precipitation dataset was analysed at its native resolution and at all lower resolutions across this range of scales. Spatial regridding was carried out using the Python package xESMF53 with a bilinear regridding scheme. Two alternative regridding methods (xESMF: conservative-normalized; and iris: area weighted) were tested and had little impact on our results. For GFC data, we calculated forest loss using the original 30-m data and converted the resulting values to each of the six spatial resolutions analysed by taking the sum of all 30-m pixels within each larger pixel. Change in canopy cover from 2003 to 2017 at each resolution is shown in Fig. 1.

Assessing impact of historical deforestation on precipitation

We used a moving-window nearest-neighbour approach54 to compare the forest loss and precipitation change of each pixel with that of its immediate neighbours. We tested the sensitivity of the analysis to the size of the moving window and found similar results for 3 × 3 and 5 × 5 (Extended Data Fig. 2) moving windows. Results from the 3 × 3 moving-window approach can been seen in the main paper. We calculated the forest loss of each deforested pixel relative to neighbouring control pixels as the forest loss of the deforested pixel minus forest loss of the control. We constrained our analysis to the tropical evergreen broadleaf biome using the Moderate Resolution Imaging Spectroradiometer land cover dataset55. To be included in the analysis, deforested pixels must have experienced 0.1% more forest loss over time than their neighbouring control pixels. The number of deforested pixels analysed varied between analysis resolutions as follows: 0.05°, n = 243,254; 0.1°, n = 58,660; 0.25°, n = 9,604; 0.5°, n = 2,303; 1.0°, n = 586; 2.0°, n = 123. We observed similar distributions of canopy change for all spatial resolutions analysed (Supplementary Fig. 6).

We calculated the precipitation change of the deforested pixel relative to the precipitation change of the control pixel (ΔP) as the precipitation change of the deforested pixel over the analysis period (for example, 2003–2017) minus the precipitation change over the control pixel. To reduce the impact of interannual variability in precipitation on our results, we calculated 5-yr means for periods at the start (2003–2007) and end (2013–2017; Extended Data Fig. 5) of the analysis period. We then calculated the change in precipitation as the difference between the start and end of these multi-year means. We report precipitation changes (ΔP) as a function of forest loss by dividing by the difference in forest loss between deforestation and control pixels (units of millimetres per month per percentage point). We also report precipitation change as the percentage change in precipitation (ΔP/P, in units of per cent) as a function of forest loss (in units of per cent per percentage point).

To ensure that control pixels and deforested pixels experience a similar background climate, we conducted a sensitivity test in which we restricted our analysis to pixels for which the pre-deforestation precipitation across the control and deforested pixels differed by less than 10%. Restricting our analysis in this way had little impact on our results (Supplementary Fig. 7) showing that our nearest-neighbour approach is effective even at the largest scales analysed here.

To explore the role of the analysis period on our results, we compared the results for 5-yr means to those for shorter 3-yr means (2003–2005 versus 2015–2017) and found consistent results (Extended Data Fig. 3). Our analysis period includes the strong 2015/2016 El Niño that resulted in reductions in precipitation over most tropical land regions, particularly in 2015 (Supplementary Fig. 1). To explore the potential impacts of the 2015/2016 El Niño on our analysis, we estimated the impact of forest loss on precipitation using 3-yr (2003–2005 versus 2018–2020) and 5-yr (2003–2007 versus 2016–2020) multi-annual means spanning an extended time period. The 3-yr analysis completely excludes the 2015/2016 ENSO, and the 5-yr analysis excludes 2015, which was the driest year (Extended Data Fig. 3). Two datasets (TRMM and UDEL) were not available after 2017, so they were removed from this sensitivity analysis.

Statistical analysis

For each category of precipitation data (satellite, station and reanalysis), precipitation change values were grouped together for all deforestation pixels and all control pixels. We found that precipitation changes for deforested pixels and control pixels, and the difference in precipitation change between deforested and control pixels (Extended Data Fig. 4), were normally distributed. Error bars (Figs. 2 and 3) show ±1 standard error from the mean calculated and displayed using the Python package Seaborn56. To test whether mean precipitation changes over regions of deforestation were statistically different from changes over the control areas, we used a Student’s t-test. We also used the Mann–Whitney test to test for significant differences in median precipitation change between control and deforested pixels and found similar results.

Seasonal analysis

For the satellite datasets alone, in addition to calculating precipitation changes at the annual timescale, we calculated changes for the dry season (driest 3 months of each year), wet season (wettest 3 months of each year) and transition season (remaining 6 months). The driest, wettest and transition months were identified for each pixel using each individual precipitation dataset. For each season and dataset, we calculated the median change in precipitation across all of the pixels within the region of interest (Supplementary Figs. 8–10).

Predicting future precipitation change due to forest loss

We used projections of forest cover change available at 0.05° from the Global Change Analysis Model (GCAM) for 2015–2100 based on the Shared Socioeconomic Pathway 3–Representative Concentration Pathway 4.5 scenario, which represents a high-deforestation future57. GCAM includes the impacts of climate and land use on future forest cover. We summed forest cover from all forest categories and calculated forest cover loss in each year compared to a 2015 baseline. Forest cover loss data were regridded to 2°. We estimated the impact of forest loss on future precipitation at the 2° scale through multiplying the projected percentage point forest loss for each pixel by the observed median change in precipitation (millimetres per month) per percentage point forest cover loss across the satellite datasets. To estimate the uncertainty in our predictions, we applied an upper and lower limit on the sensitivity of precipitation to forest loss based on the median value ±1 standard error from the mean (see error bars in Fig. 2) and rescaled by forest loss. This provides a range of estimated precipitation impacts of future forest loss. We also tested the impact on our results of capping future forest loss in each pixel at 30%, which is the upper range of forest loss that is well sampled in the observations (Supplementary Fig. 3). For each region, we applied the tropical satellite precipitation response to forest loss (Fig. 2f), meaning that our projected regional precipitation changes are a product of the regional canopy cover change and the median tropical precipitation response. Our approach assumes a linear precipitation response to forest loss, which recent work suggests could provide a conservative estimate of deforestation impacts31. We tested the sensitivity of assuming a linear response of precipitation to canopy cover loss. We fitted a nonlinear function to the data presented in Extended Data Fig. 1 through applying the median sensitivity of precipitation to forest cover loss (millimetres per month per percentage point) within each forest cover loss bin. We then scaled by the projected forest cover loss. This approach reduces the projected reduction in precipitation to 2.4 mm per month in SEA and 1.5 mm per month in the Congo (Supplementary Fig. 5).