Gridded Evapotranspiration Maps Break Down in the Frozen North, Study Warns

Evapotranspiration, the combined flux of water vapor from soil evaporation and plant transpiration, is one of the central currencies of the Earth system. It links the water cycle to the energy budget and to the carbon cycle, controlling how much moisture returns to the atmosphere, how much energy is partitioned into latent rather than sensible heat, and how ecosystems exchange carbon with the air. For decades, scientists have built gridded evapotranspiration products, datasets that map this flux across the entire globe at regular spatial and temporal intervals, so that climate models, hydrological forecasts, drought monitors and carbon accounting schemes can all draw on a common picture of land-atmosphere exchange. A new study published in Theoretical and Applied Climatology now delivers an uncomfortable message about how reliable that picture really is in one of the planet’s most climate-sensitive zones: the high-latitude cold regions of the Northern Hemisphere.

The research, led by Kaifeng Ma of the Northwest Institute of Eco-Environment and Resources of the Chinese Academy of Sciences together with Junfeng Liu, Chuntan Han and Rensheng Chen, systematically evaluated six typical gridded evapotranspiration products against ground observations from ten sites spread across the high-latitude cold regions of the Northern Hemisphere. The products tested represent the main families of approaches currently in use: EB, a surface energy balance method; CAMELE, a collocation-analyzed multi-source ensemble; ETMonitor, a model driven by multi-source satellite observations; GLDAS, a land data assimilation system; GLEAM, a global land evaporation and soil moisture dataset; and ML, a machine learning based product. Each of these datasets embodies a different philosophy for estimating a flux that cannot be directly measured from space, and each has been validated extensively in temperate and tropical environments. What the team wanted to know was whether that validation record carries over to landscapes where subfreezing temperatures and persistent snow cover dominate for much of the year.

The answer, in short, is no. During the warm season, when soils thaw, vegetation is active and the physics of evaporation resembles the conditions under which most of these products were developed, the picture is mixed but tolerable. The machine learning product and CAMELE performed relatively well, achieving mean correlation coefficients of 0.72 and 0.60 respectively with the ground-based observations. Yet even in the warm season the study found significant spatial heterogeneity among sites, meaning that a product that tracks evapotranspiration skillfully at one location may drift badly at another only a few hundred kilometers away. That kind of spatial inconsistency matters enormously for applications such as basin-scale water balance calculations, where errors at individual grid cells propagate into regional water resource assessments.

The cold season is where the datasets truly fall apart. When snow blankets the ground, soils freeze and the sun sits low on the horizon, the reliability of all six products decreases substantially, and at most sites the correlation coefficient approaches zero. In practical terms, a correlation of zero means the product’s estimates of evapotranspiration carry essentially no information about what is actually happening on the ground. This is a striking result, because winter is precisely the period when the high-latitude water budget is subtle and easily misrepresented. Snow sublimation, the direct conversion of snow and ice into vapor, can remove a meaningful share of the winter snowpack, and blowing snow events can export moisture from a basin entirely. If gridded products cannot reproduce even the timing of these fluxes, then any attempt to close the regional water balance using them alone is built on sand.

To go beyond simple correlation-based evaluation, the team applied a statistical technique known as the three-cornered hat method, or TCH. Originally developed in metrology to estimate the individual instabilities of clocks when no absolute reference is available, the three-cornered hat approach exploits the differences between multiple imperfect datasets to estimate the relative uncertainty of each one without requiring a perfect ground truth. This is an elegant solution to a chronic problem in flux validation: eddy-covariance towers, the standard instruments for measuring evapotranspiration on the ground, are themselves imperfect, particularly in winter, when energy balance closure problems and instrument icing are well documented. By treating the products and the observations as a set of mutually correlated estimates, the TCH analysis allowed the researchers to quantify how much uncertainty each product contributes independently of the others.

The TCH results reinforced the seasonal pattern. Relative uncertainty was lower during the warm season than during the cold season across the product suite, confirming that the winter degradation is not an artifact of one flawed dataset but a systemic weakness shared by energy balance models, satellite-driven products, reanalysis systems, ensembles and machine learning approaches alike. The convergence of evidence from two independent evaluation frameworks, direct comparison with site observations and the model-free uncertainty decomposition, makes the conclusion difficult to dismiss. Whatever the technical lineage of a given product, the combined effects of low temperatures and snow cover push it beyond the regime in which its underlying assumptions were calibrated.

Why should the cold season be so hostile to evapotranspiration estimation? The technical reasons are rooted in process parameterization. Most evapotranspiration algorithms were built around the physics of liquid water: stomatal conductance of leaves, aerodynamic transfer from moist soil, and radiative driving by abundant sunlight. Under snow cover, the relevant processes shift to sublimation from a snow surface, vapor transport through a turbulent boundary layer over ice, and energy exchanges complicated by the high albedo and low thermal conductivity of snow. Field studies have long shown that closing the surface energy balance over homogeneous midwinter snowpack is notoriously difficult even with careful instrumentation, so it is unsurprising that global products struggle. In addition, the machine learning products, however skillful in the training domain, inherit the distributional blind spots of their training data, which are dominated by flux tower sites in warmer, snow-free conditions. When asked to extrapolate to a frozen tundra or taiga landscape in January, they are effectively guessing.

The stakes are rising because the high latitudes are warming faster than almost anywhere else on Earth. The Arctic has warmed nearly four times faster than the globe since 1979, and this amplified warming is reshaping snow cover duration, permafrost extent, vegetation composition and the partitioning of energy at the surface. Boreal peatlands are contributing an increasing share of regional evapotranspiration as the climate warms, and changes in winter sublimation feed directly into snow mass balance and spring runoff, which in turn supply major river systems. Climate models and hydrological projections rely on gridded evapotranspiration datasets both as inputs and as benchmarks. If those datasets are unreliable in the very regions undergoing the most rapid change, then model evaluations, drought indices and carbon budget assessments for the circumpolar North carry hidden errors that no amount of averaging will remove.

The authors are explicit about the practical implication: current mainstream evapotranspiration products should be used with caution in high-latitude cold regions, especially during the cold season. This is not a call to abandon the datasets, which remain valuable tools across most of the planet, but a warning against uncritical application at the climatic margins. The study points toward a clear research agenda: improving process parameterization under low-temperature and snow-covered conditions, expanding the network of winter flux observations in cold regions so that training and validation data reflect the frozen world, and developing products that explicitly represent sublimation and snow-atmosphere exchange rather than treating winter as a low-flux afterthought. Emerging techniques such as structure-from-motion photogrammetry for measuring snow surface sublimation offer promising routes to the ground truth data that winter validation has lacked.

For the broader Earth science community, the study is a reminder that global datasets are only as good as the environments in which they have been tested. The elegant statistics of ensemble products and deep learning can conceal, rather than cure, physical blind spots. As the planet’s frozen margins thaw and shift, the demand for trustworthy evapotranspiration information in the high-latitude Northern Hemisphere will only grow, and closing the gap between what the grids say and what the snow actually does has become one of the more urgent tasks in land surface science.

Subject of Research: Uncertainty in gridded evapotranspiration products across high-latitude cold regions of the Northern Hemisphere

Article Title: Substantial uncertainties in gridded evapotranspiration products in high-latitude cold regions of the Northern Hemisphere

Article References: Ma, K., Liu, J., Han, C., & Chen, R. (2026). Substantial uncertainties in gridded evapotranspiration products in high-latitude cold regions of the Northern Hemisphere. Theoretical and Applied Climatology, 157(10), Article 642. https://doi.org/10.1007/s00704-026-06549-5

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06549-5

Keywords: evapotranspiration, high-latitude cold regions, snow cover, three-cornered hat method, gridded datasets, remote sensing, machine learning, sublimation, climate change, Northern Hemisphere, land-atmosphere exchange, Theoretical and Applied Climatology