📄 earth_science

Global non-cropland soil organic carbon loss risk under multi-level warming scenarios

This study employs a coupled process-data framework to reveal that global non-cropland soil organic carbon stocks face increasing net losses under multi-level warming scenarios, with the most significant declines occurring in forest ecosystems and high-latitude regions driven by the imbalance between temperature-accelerated decomposition and carbon inputs.

Original authors: Keqiang Wang, Zipeng Zhang, Xiangyu Ge, Lijing Han, Jingzhe Wang, Jianli Ding, Lei Zhang

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Keqiang Wang, Zipeng Zhang, Xiangyu Ge, Lijing Han, Jingzhe Wang, Jianli Ding, Lei Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Beneath our feet lies a vast, invisible reservoir that holds more carbon than the atmosphere and all the world's forests combined. This reservoir is soil organic carbon, a complex mixture of decaying plant matter, roots, and microscopic life that has accumulated over thousands of years. It is the earth's way of storing energy, and its stability is crucial for the climate. When this carbon stays locked in the ground, it helps keep the planet cool. But when it breaks down and escapes into the air as carbon dioxide, it acts like a blanket, trapping heat and accelerating global warming. Scientists have long known that rising temperatures can speed up the breakdown of this organic matter, but the full picture of how different parts of the world will react has remained hazy. Most previous studies focused on farmland, where human hands constantly till and fertilize the soil, or on small, controlled experiments that are hard to scale up to the size of a continent. The vast, wild lands—forests, grasslands, and shrublands that cover most of the planet—have been less understood in this context.

A team of researchers set out to map the future of these wild soils under a warming world. They wanted to know exactly how much carbon might be lost if the planet gets hotter, and where those losses would hit hardest. To do this, they built a digital model that combines two powerful ways of thinking about the earth. First, they used a process-based model, which acts like a biological calculator, simulating how soil microbes eat and break down organic matter based on temperature and moisture. Then, they paired this with a machine learning tool, a type of computer program that learns patterns from massive amounts of real-world data. By feeding this hybrid system data from nearly 50,000 soil sampling points across the globe, they could simulate what would happen to the soil's carbon if temperatures rose by different amounts, ranging from a mild increase to a severe one. They specifically excluded farmland to focus on the natural, non-agricultural landscapes that dominate the Earth's surface.

The results paint a clear and concerning picture. Under every warming scenario they tested, the soil in these wild ecosystems lost carbon. There was no scenario where the soil gained enough carbon to offset the losses caused by the heat. As the warming intensified, the losses grew deeper and faster. In the most extreme scenario they modeled, where temperatures rose significantly, the global stock of carbon in the top layer of soil dropped by nearly 18 petagrams. To put that in perspective, that is roughly equivalent to the total amount of carbon released by all human activities over several years, but this time coming from the ground itself. The losses were not spread evenly; they were concentrated in the forests of the Northern Hemisphere, particularly in the vast woodlands of northern North America and Eurasia, as well as in high-altitude regions like the Tibetan Plateau. These areas, which currently store immense amounts of carbon, appear to be the most sensitive to rising heat.

The study also revealed that not all landscapes react the same way. Forests suffered the greatest carbon losses, followed by grasslands and then shrublands. This makes sense when looking at the mechanics of the soil: forests have rich, deep layers of organic matter that are highly active when warmed, leading to a rapid release of carbon. In contrast, the driving forces behind these changes were identified as the amount of plant growth and the average temperature. When plants grow more, they put more carbon into the soil, but when the air gets warmer, the microbes that break down that carbon work faster, releasing it back into the air. In the simulations, the speed of decomposition outpaced the ability of plants to replace the lost carbon. The researchers found that soil properties, such as the ability of clay minerals to hold onto carbon, played a role, but they could not stop the overall trend of loss in a warming world.

While the simulations showed a consistent pattern of loss, the researchers noted that the uncertainty in their predictions was higher in the coldest, most remote parts of the world, where data is scarce. This suggests that while the general trend is clear, the exact amount of carbon lost in specific high-latitude regions could be even larger than estimated. The study did not claim to have solved the mystery of soil carbon, but it provided a much sharper map of the risk. It demonstrated that the natural balance of the world's wild soils is fragile. As the planet warms, these vast reservoirs are likely to become sources of carbon rather than sinks, creating a feedback loop where warming causes more carbon release, which in turn causes more warming. The findings serve as a stark reminder that the stability of the global climate depends not just on what we emit, but on how the earth's own systems respond to the heat we have already unleashed.

Technical Summary: Global Non-Cropland Soil Organic Carbon Loss Risk under Multi-Level Warming Scenarios

Problem Statement
Soil organic carbon (SOC) constitutes the largest terrestrial carbon reservoir, yet its response to global warming remains insufficiently understood due to complex interactions between climate, ecosystems, and soil properties. While previous studies have utilized field experiments and process-based models, these approaches face limitations in spatial extrapolation and scalability, particularly for non-cropland ecosystems (forests, grasslands, shrublands) which dominate the terrestrial surface. Existing research often focuses on static SOC mapping or specific agricultural contexts, leaving a gap in understanding how global non-cropland SOC stocks respond to varying intensities of warming across different land cover types and climate zones. There is a need for a framework that integrates mechanistic process understanding with spatial prediction capabilities to assess SOC dynamics under multi-gradient warming scenarios.

Methodology
This study developed a coupled "process-data integration" framework combining the RothC process model with a Random Forest (RF) machine learning algorithm. The methodology proceeded through the following steps:

Data Collection and Preprocessing: The study utilized 49,244 global soil sample points from the World Soil Information Service (WoSIS). Data were standardized to a 0–20 cm depth interval using equal-area spline functions. The study domain was restricted to global terrestrial non-cropland areas, explicitly excluding croplands, wetlands, bare ground, and water bodies to isolate temperature-driven responses from management or hydrological confounders.
Scenario Design and Process Simulation: Seven temperature scenarios were established: one baseline (0% warming) and six warming scenarios (T5–T30). Warming magnitudes were calculated as 5% to 30% increments of the difference between monthly maximum and mean temperatures. The RothC model was run at each sampling point for a one-year period under these scenarios to simulate SOC turnover. The model was initialized to a steady state over 1,000 years and then forward-run using historical climate data (1981–2014) with temperature perturbations applied. Carbon inputs were adjusted based on Net Primary Productivity (NPP) data.
Spatial Modeling: The difference in SOC between each warming scenario and the baseline ( $\Delta$ SOC) was calculated. To address heavy-tailed distributions and outliers, the target variable underwent quantile truncation. A Random Forest model was then trained using 15 environmental covariates (climatic, soil, topographic, and vegetation factors) to predict the spatial distribution of $\Delta$ SOC globally.
Analysis and Uncertainty: Model performance was evaluated using $R^2$ , RMSE, and MAE. SHAP (Shapley Additive Explanations) analysis was employed to interpret feature importance and marginal effects. Uncertainty was assessed using Quantile Regression Forests (QRF) to generate prediction intervals.

Key Contributions

Process-Data Fusion Framework: The study established a novel coupling of the RothC process model with Random Forest, balancing the mechanistic constraints of carbon cycling with the spatial extrapolation capabilities of machine learning.
Multi-Gradient Warming Assessment: Unlike binary or single-scenario studies, this work comprehensively captured SOC responses across a continuous spectrum of warming intensities (5% to 30% of the temperature range) for global non-agricultural lands.
Spatially Explicit Non-Cropland Focus: The research specifically targeted the often-overlooked non-cropland terrestrial ecosystems, providing a granular assessment of SOC loss risks in forests, grasslands, and shrublands.

Key Results

Global Net Loss: Under all warming scenarios, global terrestrial non-cropland SOC in the 0–20 cm layer exhibited a net loss. As warming intensity increased from T5 to T30, global SOC stocks decreased by 7.58 to 17.60 Pg, with average $\Delta$ SOC values ranging from −40.23 to −92.73 g C m⁻².
Land Cover Sensitivity: Forest ecosystems experienced the largest absolute losses (up to 11.88 Pg under T30), followed by grasslands and shrublands. Forests demonstrated higher sensitivity to warming disturbances compared to other land cover types.
Spatial Heterogeneity: Negative SOC responses were concentrated in mid-to-high latitude and high-altitude regions, including northern North America, northern Eurasia, and the Qinghai–Tibet Plateau. Continental climate zones showed the most severe losses.
Driving Factors: SHAP analysis identified Net Primary Productivity (NPP) and Mean Annual Temperature (AMT) as the dominant predictors. This indicates that SOC responses are regulated by the balance between ecosystem carbon inputs (NPP) and temperature-driven decomposition processes. Soil properties related to carbon stabilization (e.g., CEC) also played a significant role.
Model Performance: The Random Forest model demonstrated stable predictive capability across scenarios, with $R^2$ values between 0.50 and 0.54. The results showed a smoother, more continuous temperature response curve compared to the variability observed in isolated warming experiments.

Significance and Claims
The authors claim that this study demonstrates that warming-induced SOC responses are regulated by the dynamic balance among carbon inputs, decomposition losses, and soil carbon stabilization processes. The coupled framework is presented as a methodological advancement that improves the integration of mechanistic constraints with spatial prediction, offering new insights into global SOC responses to climate warming.

The study concludes that the proposed framework supports regional carbon management under changing climate conditions by revealing an accelerating trend in SOC loss under strong warming scenarios. It highlights that non-cropland ecosystems, which constitute the majority of the terrestrial surface, are vulnerable to carbon loss as warming intensifies, particularly in high-latitude and continental regions. The authors position their work as a step toward shifting from static SOC mapping to scenario-based response assessments, providing a reference for future studies on SOC-climate feedbacks.

More like this