Deep Soil Heat and Moisture Coupling Drives Forest Carbon Loss, Machine Learning Study Reveals

Deep beneath the forest floor, a silent exchange of carbon is underway, and a new study from China’s Wuling Mountains suggests that the key to predicting it lies not at the surface, but in the coupled behavior of heat and moisture in the deepest soil layers scientists monitored. Using an arsenal of machine learning models trained on a year of high-resolution field measurements, researchers have shown that total soil respiration and its microbial component are governed primarily by hydrothermal interactions in the 20 to 30 centimeter soil layer, while the root-driven fraction remains stubbornly resistant to prediction by environmental variables alone.

Soil respiration, the combined release of carbon dioxide from roots, microbes, and soil fauna, is one of the largest carbon fluxes from terrestrial ecosystems to the atmosphere. Its annual magnitude far exceeds that of fossil fuel emissions, meaning even small shifts in its rate can ripple through global climate-carbon feedbacks. Yet despite decades of study, accurately predicting soil respiration remains a formidable challenge, particularly in heterogeneous landscapes such as subtropical mountainous forests where topography, seasonal climate, and vegetation create a patchwork of conditions.

A team at the Hunan Cili Forest Ecosystem Research Station tackled this problem by measuring soil respiration hourly across four forest types: a pure coniferous forest dominated by Masson pine, a mixed coniferous-broadleaf forest, a citrus orchard representing economic forest, and a secondary broadleaf forest dominated by oak species. Over twelve monthly campaigns spanning June 2024 to July 2025, the researchers collected 6,912 raw collar-level observations, which they aggregated into 2,304 subplot-level values. Using the trenching method, they partitioned total respiration into autotrophic respiration from roots and heterotrophic respiration from microbial decomposition.

With this dataset in hand, the team built machine learning models under four different grouping strategies. A unified approach that pooled all forest types and components performed worst, achieving a test coefficient of determination of just 0.508. Grouping by forest type alone fared no better. But when the data were split by respiration component, with separate models for total, autotrophic, and heterotrophic respiration, performance jumped to 0.795. A fully crossed strategy using twelve sub-models achieved a slightly higher mean of 0.849 but at the cost of greater complexity and overfitting. The component-based approach emerged as the optimal balance of accuracy, generalization, and parsimony.

The researchers then compared seven algorithms, ranging from linear regression to ensemble tree methods including Random Forest, XGBoost, LightGBM, and CatBoost. Under random validation, ensemble models achieved coefficients of determination exceeding 0.90 for total and heterotrophic respiration. However, when the team applied a stricter leave-one-month-out blocked validation, designed to test whether models could generalize to unseen months rather than exploit seasonal patterns, the advantage of these models narrowed substantially. This finding carries an important warning: much of the apparent accuracy of machine learning in ecological prediction may stem from memorizing recurring temporal patterns rather than learning transferable environmental relationships.

CatBoost trained without temporal variables proved the most consistent performer under blocked validation, achieving blocked coefficients of 0.475 for total respiration, 0.544 for heterotrophic respiration, and 0.081 for autotrophic respiration. The team selected this model for ecological interpretation using SHAP analysis, a technique that quantifies each predictor’s contribution to individual predictions. When temporal encodings were included, seasonal variables dominated the importance rankings, but their removal shifted attribution decisively toward environmental predictors.

The SHAP results revealed a striking pattern. For both total and heterotrophic respiration, the interaction between deepest monitored soil temperature and moisture was the single most important predictor, followed by temperature gradient and mid-layer hydrothermal coupling. Two-dimensional partial dependence plots showed that predicted respiration increased sharply once the deep-layer hydrothermal index exceeded a threshold of approximately six, but this positive effect was suppressed when vertical temperature gradients grew large, indicating that strong thermal stratification constrains the hydrothermal response.

A depth-wise ablation analysis confirmed the independent value of deep measurements. Removing all predictors derived from the 20 to 30 centimeter layer reduced the pooled coefficient of determination by 0.167 for total respiration and 0.143 for heterotrophic respiration, while removing shallower layers had negligible effect. This suggests that deep soil conditions carry predictive information that surface measurements alone cannot capture, likely because they integrate whole-profile thermal and moisture dynamics relevant to biological activity at multiple depths.

Autotrophic respiration told a different story. It was most closely associated with deep-layer daytime temperature, temperature gradient, and moisture ratio, but its predictor importance was diffuse and its predictability remained poor across all models and validation protocols. The researchers attribute this to unmeasured plant physiological controls, including photosynthate transport, root phenology, and carbon allocation dynamics, which operate on timescales and through mechanisms not captured by soil environmental sensors alone.

The study’s implications extend beyond the Wuling Mountains. For carbon cycle modelers, the results argue that partitioning soil respiration into components before model fitting, rather than treating it as a single flux, can substantially improve both accuracy and mechanistic insight. For field ecologists, the findings underscore the value of multi-depth soil monitoring, particularly in subtropical systems where vertical heterogeneity is pronounced. And for the broader machine learning community, the work offers a cautionary tale about validation protocols: without temporally blocked testing, models may appear far more skillful than they truly are, masking their reliance on seasonal shortcuts rather than genuine environmental understanding.

Subject of Research: Machine learning prediction of soil respiration components in subtropical mountainous forests

Article Title: Machine learning-based modeling of soil respiration components across four forest types in a subtropical mountainous forest with implications for modeling strategies and hydrothermal controls

Article References: Deng, N., Xiao, Y., Ma, F., & Tian, Y. (2026). Machine learning-based modeling of soil respiration components across four forest types in a subtropical mountainous forest with implications for modeling strategies and hydrothermal controls. Frontiers in Forests and Global Change, 9, Article 1972229. https://doi.org/10.3389/ffgc.2026.1972229

Image Credits: AI Generated

DOI: 10.3389/ffgc.2026.1972229

Keywords: soil respiration, machine learning, autotrophic respiration, heterotrophic respiration, hydrothermal coupling, subtropical forest, SHAP, carbon cycle, CatBoost, soil temperature, soil moisture, model validation