The research, led by Dhia Zalfa Zahira, Anjar Dimara Sakti, and Ketut Wikantika of Institut Teknologi Bandung, goes beyond simply counting lost trees. It tracks how much carbon those forests actually stored — both above ground in trunks, branches, and leaves, and below ground in root systems — and how that storage changed year by year. Most previous studies of the region offered static snapshots for a single year or focused only on aboveground biomass. By modelling both aboveground and belowground biomass carbon across five time points, the team has produced one of the most dynamic pictures yet of a tropical carbon reservoir in decline.

Defining what counts as an intact forest was itself a technical feat. The researchers started with the Hansen tree cover dataset derived from Landsat imagery, flagging any pixel with at least 30 percent canopy cover in 2000. They then filtered out oil-palm plantations using a global planting-year dataset and removed cropland using the MODIS land cover product. From that baseline, they iteratively subtracted three drivers of loss — stand-replacing deforestation, new cropland, and new oil-palm planting — at five-year intervals, generating intact forest maps for 2000, 2005, 2010, 2015, and 2020. This multi-layer filtering matters because oil-palm plantations, though tree-covered, are not ecologically intact forests; they typically replace natural forest and store far less carbon.

With the forest maps in hand, the team turned to a Random Forest regression model, an ensemble machine learning method that aggregates predictions from hundreds of decision trees. Thirteen environmental predictors fed the model, spanning vegetation indices such as NDVI and NIRv, productivity metrics including gross and net primary productivity, leaf area index, and the fraction of absorbed photosynthetically active radiation, along with climate variables like land surface temperature and precipitation, and topographic features including elevation, slope, and a topographic wetness index. All layers were processed at 300-metre resolution and aggregated to 1.5-by-1.5-kilometre grids to reduce noise and pixel misalignment.

A key design choice set this model apart: rather than training only on forest pixels, the researchers trained across the full gradient of Southeast Asian vegetation — forests, shrublands, grasslands, and cropland alike — using more than two million grid cells from 2010. This ensures the model learned the entire spectrum of carbon density, reducing bias at forest edges and recently disturbed areas. The data was split 80-20 into training and testing sets, with a Kolmogorov-Smirnov test confirming that both subsets covered the region geographically without spatial bias. Against the NASA reference biomass dataset, the model achieved a testing R² of 0.899 for aboveground carbon and 0.874 for belowground carbon, indicating strong predictive skill with limited overfitting.

Independent validation against the European Space Agency’s Climate Change Initiative Biomass product — built from radar observations by Sentinel-1, Envisat, and the ALOS satellites — showed moderate but consistent agreement, with R² values of 0.645 and 0.632 for the two carbon components. Among the thirteen predictors, the fraction of absorbed photosynthetically active radiation emerged as the single most influential variable, which makes physical sense: how much solar energy a canopy captures directly governs how much carbon it can fix into biomass. Land surface temperature and precipitation followed, while the classic vegetation indices contributed surprisingly little, apparently limited in their sensitivity to structural biomass in dense tropical canopies.

The carbon accounting that followed is sobering. In 2000, intact forests in the region held an estimated 26.18 megatonnes of aboveground biomass carbon and 6.61 megatonnes belowground, a combined total of roughly 32.8 megatonnes. By 2020, those figures had fallen to 15.96 and 3.96 megatonnes respectively — a drop of more than 39 percent in twenty years. The steepest losses occurred between 2010 and 2015, when forest area shrank by nearly 384,000 square kilometres and biomass carbon declined at a rate of about 1,496 megatonnes per year in total. The hardest-hit landscapes were Kalimantan and Sumatra in Indonesia, Sarawak and Sabah in Malaysia, and cropland-driven loss zones in northern Thailand, Myanmar, and parts of Papua.

To translate those ecological losses into terms policymakers cannot ignore, the team ran an indicative economic valuation under two carbon pricing scenarios. At the current average price across ASEAN+3 countries — a modest 6.45 US dollars per tonne of CO₂ equivalent — the region’s intact forest carbon was worth about 776 billion dollars in 2000, falling to roughly 471 billion by 2020. Under the globally recommended price of 75 dollars per tonne, aligned with social cost of carbon estimates, the value dropped from about 9.02 trillion dollars to 5.48 trillion. The authors are careful to stress these are illustrative, non-market figures, not credit-eligible carbon trading values, but they make the scale of the liability unmistakable: a loss of nearly 4 trillion dollars in indicative carbon assets in two decades.

The findings land at a delicate moment for climate policy. Southeast Asia harbours nearly 15 percent of the world’s tropical forests, and their continued erosion threatens not only regional biodiversity but global mitigation commitments under the Paris Agreement, REDD+ frameworks, and Sustainable Development Goal targets on sustainable forest management. There is also a real risk of a tipping dynamic: as intact forests degrade, they can flip from carbon sinks to net carbon sources, emitting more than they absorb and accelerating the very warming that stresses them. The spatially explicit maps produced by this study could help governments prioritise conservation zones where avoided deforestation yields the greatest climate benefit, and design community-based forest management schemes that recognise indigenous peoples as the primary stewards of these carbon reservoirs.

The study is candid about its limitations. The 2000 baseline may include some old plantations that the Hansen dataset cannot distinguish from natural forest; the model assumes predictor relationships stable across two decades; and the absence of explicit spatial block cross-validation means the reported accuracy figures may be somewhat optimistic. The economic valuation, by design, omits additionality, leakage, and monitoring requirements. Still, the core signal — a 59 percent collapse in intact forest area and a 39 percent loss of biomass carbon, tracked with low reported uncertainty — is difficult to dispute. Future work integrating radar-based biomass data and country-specific carbon accounting could sharpen the picture further. What this study makes clear is that Southeast Asia’s intact forests are not merely an ecological treasure; they are a quantifiable, and rapidly depreciating, planetary asset.

Subject of Research: Remote sensing and random forest modelling of intact forest biomass carbon dynamics and carbon economics in Southeast Asia from 2000 to 2020

Article Title: Intact forest carbon dynamics and indicative carbon economy in Southeast Asia: A remote sensing and random forest approach

Article References: Zahira, D. Z., Sakti, A. D., & Wikantika, K. (2026). Intact forest carbon dynamics and indicative carbon economy in Southeast Asia: A remote sensing and random forest approach. Environmental Advances, 26, Article 100758. https://doi.org/10.1016/j.envadv.2026.100758

Image Credits: AI Generated

DOI: 10.1016/j.envadv.2026.100758

Keywords: intact forests, Southeast Asia, biomass carbon, remote sensing, random forest, deforestation, oil palm, carbon pricing, REDD+, aboveground biomass, belowground biomass, machine learning