Machine Learning Reveals How Warming Penetrates Frozen Soils of Northeast China

Beneath the forests and wetlands of Northeast China lies ground that has been frozen for decades, sometimes centuries, and it is quietly changing. A new study published in Theoretical and Applied Climatology has combined twenty-five years of reanalysis data with an interpretable machine learning framework to answer a deceptively simple question: as the climate warms, which atmospheric and hydrological forces actually drive temperature and moisture changes at different depths within frozen soil? The answer, it turns out, depends strongly on how deep you look, and the findings carry significant implications for permafrost degradation, water resources, and carbon release in one of the world’s most climate-sensitive regions.

The research, conducted by Gaojian Di of the Heilongjiang Province Hydraulic Research Institute and Bo Feng of the Changchun Institute of Technology, focused on the frozen soil region of Northeast China, an area where discontinuous permafrost and seasonally frozen ground interact with a warming boreal climate. Rather than treating the soil column as a single unit, the team examined four distinct soil layers, from the surface down to deeper horizons, capturing the depth-dependent character of climate signals as they propagate downward through the ground. This layered perspective matters because the surface responds almost immediately to atmospheric conditions, while deeper layers integrate and lag behind those signals, sometimes in surprising ways.

To build their analysis, the researchers drew on monthly ERA5-Land reanalysis data spanning 1997 to 2021, a dataset produced by the Copernicus Climate Change Service that blends satellite observations with numerical weather modeling to provide consistent estimates of land-surface conditions. Reanalysis products like ERA5-Land are invaluable in remote, cold regions where ground-based instrumentation is sparse, offering continuous records of soil temperature, soil moisture, radiation, precipitation, snow depth, and runoff across the entire study area. The team analyzed the spatiotemporal evolution of soil temperature and moisture across the four layers, identifying trends and patterns of interannual variability that would be difficult to detect from scattered field stations alone.

The headline result on temperature is strikingly orderly: soils warmed across all four layers during the study period, but the magnitude of that warming generally decreased with depth. This attenuation is physically intuitive, since heat propagating downward from the surface is damped by the thermal inertia of overlying soil, yet documenting the pattern quantitatively across a quarter century provides a benchmark for models of permafrost thermal evolution. The finding aligns with global observations that permafrost is warming worldwide, and it confirms that the frozen ground of Northeast China is following the same trajectory, with surface layers absorbing the brunt of atmospheric warming while deeper horizons warm more slowly but persistently.

Soil moisture told a messier story. Unlike temperature, moisture changes across the four layers displayed a non-monotonic pattern, meaning the direction and magnitude of change did not vary consistently with depth, and the records were accompanied by substantial interannual variability. This complexity reflects the tangled hydrology of frozen ground, where snowmelt timing, thaw depth, infiltration capacity, and evaporation all compete to determine how much liquid water resides in the soil at any given moment. In permafrost regions, the frozen matrix itself impedes drainage, so even modest shifts in thaw duration or precipitation phase can produce outsized and erratic swings in moisture content, complicating both ecological forecasting and engineering planning.

The methodological heart of the study lies in its use of XGBoost, a gradient-boosted decision tree algorithm renowned for its predictive power, paired with SHAP, or SHapley Additive exPlanations, a technique borrowed from game theory that quantifies each input variable’s contribution to a model’s output. Traditional machine learning models are often criticized as black boxes, but the XGBoost-SHAP framework allows researchers to open that box and see precisely which climatic variables explain soil changes at each depth, and how those contributions shift across space and time. This interpretable approach transforms the model from a mere curve-fitting exercise into a scientific instrument for hypothesis testing about the physical drivers of frozen soil dynamics.

For soil temperature, the attribution results revealed a clear depth-dependent hierarchy of drivers. Downward longwave radiation, the infrared energy emitted by the atmosphere toward the surface, made the largest contribution to explaining temperature variations in the surface and shallow soil layers. This makes physical sense: as atmospheric humidity and cloud cover increase under warming, they trap and re-emit more longwave radiation, directly heating the ground surface. For mid-layer soil temperature, however, snow depth emerged as the most significant explanatory variable, highlighting the dual role of snow as both an insulating blanket that shields the ground from extreme winter cold and a reflective surface that moderates spring warming. In the deepest layer, temperature changes were associated with the combined contributions of relative humidity, net surface shortwave radiation, and snow depth, suggesting that deep soil integrates multiple atmospheric pathways rather than responding to any single dominant driver.

The moisture attribution produced an equally compelling finding: total runoff demonstrated a high explanatory contribution across all four soil layers, outranking every other climatic variable. This underscores the tight coupling between surface hydrology and subsurface water storage in frozen terrain, where the generation and routing of runoff govern how water is distributed through the soil column. Following runoff, snow depth, surface air temperature, and total precipitation ranked as the leading contributors to soil moisture variation. The prominence of snow depth in both the temperature and moisture analyses reinforces its status as a master variable in cold-region hydrology, controlling not only the thermal insulation of the ground but also the timing and magnitude of the spring water pulse that recharges soil moisture as thaw proceeds.

Why should readers far from Northeast China care about the thermal state of its frozen ground? The reasons are both global and practical. Frozen soils store vast quantities of organic carbon, and their thaw can release greenhouse gases that amplify warming, a feedback loop of planetary consequence. Regionally, permafrost degradation alters groundwater flow, destabilizes infrastructure such as roads, pipelines, and buildings founded on ice-rich ground, and reshapes ecosystems that boreal forests and wetlands depend upon. Previous studies in the region have documented shrinking permafrost extent, thickening active layers, and shifting hydrological regimes, and the new quantitative attribution adds a crucial piece: it identifies which specific climatic levers are pulling hardest at each depth, information that can sharpen the projections of land surface and permafrost models.

The study also demonstrates a template for future cold-region research. By making the ERA5-Land dataset, the analysis code, and permafrost zoning data publicly available through the Copernicus Climate Data Store and a GitHub repository, the authors have lowered the barrier for other teams to replicate, extend, or challenge their findings. As machine learning continues to permeate the geosciences, the pairing of powerful algorithms with interpretability tools like SHAP offers a way to extract causal insight from big environmental data without sacrificing scientific transparency. For the frozen soils of Northeast China, the message is unambiguous: warming is penetrating the ground layer by layer, snow and radiation are the dominant conductors of that heat, and runoff holds the keys to the region’s soil water future. Understanding those mechanisms in detail is an essential step toward anticipating what a warmer century will do to the ground beneath our feet.

Subject of Research: Depth-dependent responses of frozen soil temperature and moisture to climate change in Northeast China, analyzed with XGBoost-SHAP machine learning

Article Title: Response of multi-layer soil temperature and moisture to climate change in the frozen soil Region of Northeast China based on interpretable machine learning

Article References: Di, G., & Feng, B. (2026). Response of multi-layer soil temperature and moisture to climate change in the frozen soil Region of Northeast China based on interpretable machine learning. Theoretical and Applied Climatology, 157(10), Article 658. https://doi.org/10.1007/s00704-026-06584-2

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06584-2

Keywords: frozen soil, permafrost, Northeast China, soil temperature, soil moisture, climate change, XGBoost, SHAP, interpretable machine learning, ERA5-Land, snow depth, runoff