Reduction of deforestation by agroforestry in high carbon stock forests of Southeast Asia

Main
Agroforestry, the integration of trees and shrubs into agricultural landscapes, can deliver multiple ecological and socio-economic co-benefits, including enhanced biodiversity, improved soil health, increased agricultural productivity and carbon sequestration1. Given these benefits, agroforestry is now seen as a promising natural climate solution for global climate mitigation2, with increasing efforts to implement agroforestry at landscape scale3. The positive effects of agroforestry have been well studied at local scales, especially carbon and ecological dynamics. However, to fully integrate agroforestry into climate policy will require sufficient understanding of how agroforestry affects land-use systems and thus carbon balances at landscape scale2.
So far, it is unclear whether agroforestry is climate-positive at landscape scale. Various mechanisms have been suggested on how agroforestry could increase or reduce deforestation at that scale, but their cumulative effect is rarely tested. For the local community, agroforestry can provide alternative livelihoods and directly substitute for forest products such as firewood4. This increase in land-use efficiency can reduce incentives to clear forests for agricultural expansion, as per the Borlaug hypothesis5. Apart from economic processes, agroforestry practices that involve tree planting could also transform attitudes by invoking more environmental awareness and discouraging forest clearing6. However, contrary hypotheses have also been suggested: that profitable agroforestry could create further incentives and capacities for forest conversion, whereas increased incomes and the need for transporting products to markets could result in infrastructure expansion and further development7.
These socio-economic mechanisms influencing agroforestry’s impact on deforestation operate within a diverse range of agroforestry systems and practices. In Southeast Asia, agroforestry is present in landscapes ranging from hill slopes in mountainous areas, fire-climax grasslands, peat lands, mangroves and rice paddies and terraces. Major typologies include traditional home gardens, evolution from natural forest or swidden/fallow rotations by promoting desirable tree species among agricultural mosaics, the addition of trees to open-field agricultural systems and the addition of agricultural crops to forestry systems. Agricultural crops commonly integrated with agroforestry include rice, cassava, coffee, cacao, fruits such as pineapple and vegetables; agroforestry tree species range from fast-growing timber (for example, Acacia, Eucalyptus and bamboo), fruit and medicinal trees (for example, coconut, durian and tamarind) and tree crops (for example, small-holder rubber and oil palm)8. Agroforestry systems also frequently supply non-timber forest products such as rattan, mushrooms and herbs such as ginseng. Market demand for these products and market access9, competition with alternative land uses and government incentives and policies10 are known factors driving agroforestry adoption. Also, many of Southeast Asia’s ecosystems contain Indigenous communities, who may maintain or adapt traditional agroforestry practices11.
In addition to this rich and nuanced local context, agroforestry in Southeast Asia has the highest per-hectare aboveground biomass carbon stock of 60–65 tC ha−1 of all agroforestry lands worldwide12 and thus has an importance for climate mitigation beyond its scale. There are 80.5 Mha of agroforestry across Southeast Asia13, comprising 18% of its total land area. Moreover, Southeast Asia contains large tropical forest extents with high carbon stocks and global hotspots for biodiversity and endemism, which are under threat from rapid development14.
In this study, we examined the impact of agroforestry on average annual deforestation rates15 across Southeast Asia between 2015 and 2023. The treatment was defined as the presence of agroforestry within 1-km pixels where more than 5% of the area was covered by agroforestry for the reference year 201513. Using propensity score matching on biophysical, socio-economic and land-cover covariates (Supplementary Table 2), we aligned treated and control groups to ensure comparable characteristics aside from the intervention, effectively simulating the conditions of a randomized controlled trial. Finally, post-hoc analyses in the form of random forest and multiple linear regressions were conducted to explore underlying factors influencing the observed outcomes.
We first performed pixel-level matching at 1-km resolution for each subnational region across Southeast Asia. The presence of agroforestry reduced deforestation compared to the counterfactual in 22 out of 38 subnational regions in Southeast Asia, which had statistically significant results (P < 0.05) (Fig. 1a) by 1.08% points (median; interquartile range (IQR): 0.65–2.01%); this was 318,524 ha yr−1 of avoided deforestation corresponding to 73.6 ± 14.9 MtCO2e yr−1 (hereafter mean ± s.e.m.) of avoided emissions. In the remaining 16 subnational regions, agroforestry increased deforestation by 0.64% points (median; IQR: 0.31–0.99%), which was 68,206 ha yr−1 or 14.8 ± 4.3 MtCO2e yr−1. Across the 38 subnational regions with significant results, agroforestry resulted in a net reduction in deforestation of 250,319 ha yr−1 or 58.8 ± 15.5 MtCO2e yr−1.

a, Treatment effect of agroforestry on deforestation estimated by matching within subnational regions (outlined in grey); a negative treatment effect indicates that agroforestry reduces deforestation, and a positive treatment effect indicates that agroforestry increases deforestation. Southeast Asian countries are outlined in black. Only statistically significant results (P < 0.05) are shown. b, Average annual deforestation rates15 between 2015 and 2023 for subnational regions. c, Agroforestry area13 as a percentage of each 1-km pixel. d, HCS forest area16 with >75 tC ha−1 aboveground as a percentage of each 1-km pixel.
Among areas in Southeast Asia with substantial extents of forests with high carbon stock (HCS) aboveground16 (Fig. 1d), notable areas where agroforestry reduced deforestation across many subnational regions include all of Laos and parts of northern Vietnam, northern and eastern Myanmar, Sumatra, Borneo and peninsular Malaysia, which are already deforestation hotspots14 (Fig. 1b). However, there are some particular subnational regions where agroforestry increased deforestation, such as eastern Cambodia, which is a deforestation hotspot and contains some HCS forests. Some areas where agroforestry reduced deforestation also contained substantial mangrove and peat land extents with high belowground carbon stocks17,18—these include many subnational regions in peninsular Malaysia, Sumatra and Borneo.
Most agroforestry in Southeast Asia occurs in mosaic landscapes, where agroforestry is not the majority land-cover type; for 1-km pixels where agroforestry was present the median agroforestry area was 33.9% (IQR: 15.4–62.8%) and the median forest area was 15.5% (IQR: 0.0–63.4%). As such, the positive results in our study show that agroforestry frequently co-exists with forest fragments in these mosaic landscapes and helps to protect them. Forest fragments have outsized ecological importance, providing habitat heterogeneity to support greater biodiversity and stepping stones, refugia and corridors for species dispersal and re-colonization, thus improving the ecosystem resilience of the landscape19.
To explore underlying factors influencing the observed treatment effects on subnational regions with significant results, we performed post-hoc analyses in the form of random forest variable importance and multiple linear regressions (Supplementary Fig. 2 and Supplementary Table 5). However, these models performed poorly in explaining variance in the treatment effects with R2 ranging from −0.34 to −0.18, resulting from low sample size (n = 38) and reflecting varied local responses to drivers that cannot be well-generalized across Southeast Asia.
Prior studies on agroforestry have highlighted the complex interplay of socio-economic and ecological factors at local scales, such as personal demographic and socio-economic factors affecting agroforestry practitioners’ motivation to expand production20, land tenure21, government policies and donor support10, local market demand for timber and non-timber forest products9 and the role of community governance among local forest user groups22, can influence agroforestry. Whereas we have not directly examined these local factors in our study, it is notable that despite these complex local contexts, we found clear causal evidence that agroforestry leads to a net reduction in deforestation overall across Southeast Asia. Our findings on notable areas where agroforestry reduced or increased deforestation and areas with mixed results may provide useful context for local case studies.
These findings provide important support and additional nuance to guide climate and land-use policies in Southeast Asia. Leakage accounting is an essential component of natural climate solution projects; our findings suggest that agroforestry generally provides a form of beneficial leakage by reducing deforestation. For instance, Indonesia has rapidly expanded land designated as ‘social forestry areas’ from 1.8 Mha in 2018 to 5 Mha in 20223, which includes community-managed agroforestry among other allowed uses; if well-implemented and managed, this may make an important contribution, through both in situ carbon sequestration and avoided deforestation to Indonesia achieving net zero in its forestry and other land-use sector. Further research and practice on factors needed to ensure the success of agroforestry and reducing deforestation, such as securing land tenure and collaborating with Indigenous and local communities, will be crucial for Southeast Asia to achieve its climate goals.
Methods
Overview
We used propensity score matching, a robust causal inference methodology, to attribute changes in deforestation rates (treatment effect) to the presence of agroforestry. To identify factors influencing these treatment effects, we performed random forest and multiple linear regressions as post-hoc analyses. The study focused on ten Southeast Asian countries (shown in black outlines in Fig. 1), with some divided into subnational regions (Supplementary Table 1). Spatial datasets were processed in Google Earth Engine v1.5 and QGIS 3.34.2 using Mollweide equal-area projection, whereas causal inference and statistical modelling were performed in R version 4.3.1, using the packages ‘MatchIt’ v4.5.5, ‘marginaleffects’ v0.18.0 and ‘ranger’ v0.16.0.
Data
Agroforestry data were extracted for year 2015 at 100-m resolution from the Global Forest Management dataset13, which defines agroforestry as ‘individual trees on cropland or pasture or mixed crops (including trees)’. Forest data were derived from the Hansen Global Forest Change v1.11 dataset15 at 30 m, with forests defined by >30% tree cover and then refined to reflect only natural and planted forests by excluding agroforestry and plantations13. From these refined forest data, the average annual deforestation rate between the years 2015 and 2023 for natural and planted forests was calculated within 1-km pixels, which served as the dependent variable for propensity score matching. Covariates included biophysical, socio-economic and land-cover variables known to influence deforestation or agroforestry in Southeast Asia (Supplementary Table 2).
Causal inference and post-hoc analyses
Propensity score matching was used to estimate the impact of agroforestry on deforestation rates. Pixels of 1-km resolution with agroforestry (>5%) and without were matched 1:1 to their nearest neighbour without replacement based on propensity scores; matching was performed within each subnational region across Southeast Asia to capture unaccounted-for confounding. The matching process aimed to balance covariates between treatment and control groups by achieving a standardized mean difference (SMD) < 0.25 indicating sufficiently low bias and thus a valid counterfactual. Successful matching occurred in 81 out of 90 subnational regions, with most covariates showing SMD < 0.25 (Supplementary Table 3), confirming effective balancing. Marginal effects were estimated using g-computation to derive the average treatment effect23. A sensitivity analysis tested pixel sizes from 1 km to 12 km, with consistent treatment effects and statistical significance observed up to 6 km, demonstrating the robustness of the 1-km resolution (Supplementary Table 4 and Supplementary Fig. 1).
To investigate underlying factors influencing treatment effects, we performed post-hoc analyses in the form of random forest and multiple linear regressions. For these post-hoc analyses, the following covariates with Pearson’s r > 0.8 against other covariates were removed to prevent multicollinearity: elevation, long-term mean annual temperature, subsoil pH, subsoil organic carbon and travel time to nearest port. The dependent variable was the average treatment effect from propensity score matching for 38 regions (P < 0.05) with the same covariates as independent variables. The random forest model trained on 300 trees had an out-of-bag R2 of −0.18, with variable importance computed via permutation24 (Supplementary Fig. 2). The multiple linear regression yielded an adjusted R2 of −0.34, with no significant covariates (P < 0.05) (Supplementary Table 5).
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Responses