New Map Reveals Permafrost Loss Across the Qinghai-Tibet Plateau

High on the Qinghai-Tibet Plateau, the ground itself is quietly retreating. A research team led by Yuhong Chen and Zhuotong Nan has produced the most precise snapshot yet of where frozen ground persists across the vast region known as the Third Pole, and their findings confirm what climate scientists have long feared: the plateau’s permafrost is shrinking. The new study, published in The Cryosphere, delivers a 1-kilometer-resolution permafrost distribution map for the year 2020, showing that permafrost now covers approximately 1.038 million square kilometers, or 39.35 percent of the plateau. That figure represents a decline of 48,000 square kilometers, a 1.82 percent loss, compared with a baseline map for 2010. While the percentage may sound modest, the affected area is larger than the Netherlands and Denmark combined, and the consequences ripple far beyond the plateau’s borders.

The Qinghai-Tibet Plateau is often called the Asian Water Tower because it feeds the headwaters of the Yangtze, Yellow River, Indus, Mekong, and Ganges. Its frozen ground acts as a vast subterranean reservoir and thermal regulator, influencing hydrology, carbon cycling, ecosystem stability, and the safety of critical infrastructure such as the Qinghai-Tibet Railway. Permafrost, defined as ground that remains below zero degrees Celsius for at least two consecutive years, underlies roughly a quarter of the exposed land surface in the Northern Hemisphere, but the plateau hosts the largest permafrost extent in mid- and low-latitude regions. Because the plateau’s permafrost is relatively warm and thermally unstable, it is especially vulnerable to warming. Between 1960 and 2015, the region’s annual mean temperature rose at 0.33 degrees Celsius per decade, more than double the global average, making the plateau a bellwether for cryospheric change worldwide.

What makes the new map scientifically significant is its temporal specificity. Most existing permafrost maps of the plateau reflect long-term historical averages, pooling observations across multiple decades to compensate for sparse ground data. That approach produces a composite picture that smooths over the rapid climatic shifts of recent years, creating a mismatch for applications that need precise, period-specific benchmarks, from calibrating land surface models to assessing engineering risk. The 2020 map, by contrast, captures the current thermal state of the ground, offering researchers and engineers a reference that reflects conditions as they actually are rather than as they averaged out over the twentieth century.

Building the map required the team to solve a stubborn data problem. The mapping framework is based on an extended ground surface frost number model, a semi-physical approach that calculates the probability of permafrost occurrence from the balance between surface freezing and thawing loads. The model’s central driver is the frost number F, defined as the ratio of the ground surface freezing index to the sum of the freezing index and an empirically scaled thawing index. When F exceeds 0.5, freezing dominates and permafrost can persist; when it falls to or below 0.5, seasonal thaw wins out and the ground is classified as seasonally frozen ground or non-frozen ground. The freezing and thawing indices were derived from five years of MODIS satellite land surface temperature observations covering 2016 through 2020, a window chosen to satisfy the two-year permafrost definition while minimizing bias from interannual climate anomalies.

Raw satellite skin temperatures, however, are not the same as ground surface temperatures. Cloud cover frequently obscures the plateau, leaving extensive gaps in the MODIS record, so the team applied a sophisticated gap-filling method that reconstructs missing pixels using the geometric relationship between the sun, cloud position, and satellite viewing angle. They then corrected the satellite-derived thawing index for the thermal buffering effects of vegetation, using a multilinear regression that incorporates vegetation greenness and latitude at sixteen-day intervals. No equivalent correction was needed for the freezing index, because winter snow on the plateau is thin, ephemeral, and spatially discontinuous, producing negligible insulation at regional scale.

The thorniest challenge involved a dimensionless empirical soil parameter, denoted E, which captures how local soil thermal properties and moisture conditions modify the ground thermal regime. In the team’s earlier 2010 map, E was calibrated against extensive field surveys conducted that year. No comparable surveys were repeated in 2020, so the researchers turned to a space-for-time substitution strategy, a well-established technique in Earth system science in which spatial gradients stand in for temporal change. The hypothesis was that although the specific value of E at any location may shift with moisture or surface cover, its statistical relationship with environmental drivers such as topography, soil texture, vegetation, and precipitation remains stable over a decadal interval. Using the high-quality 2010 E field as a training target, they tested two transfer learning approaches: a physics-informed neural network and a Random Forest regression.

The comparison produced a cautionary tale for the era of machine learning in geoscience. The neural network achieved 95.26 percent accuracy within its training subregions but overfit severely when extrapolated across the entire plateau, generating unrealistic spatial artifacts and E values scattered across an implausibly wide range. The Random Forest approach, by contrast, generalized robustly. Among four sampling strategies evaluated with five-fold cross-validation, k-means clustering sampling with 5,000 training points performed best, achieving a root mean square error of 0.048 and explaining 74.83 percent of the variance in the 2010 E field. An ensemble analysis of sixteen model configurations confirmed the stability of the predictions: the coefficient of variance of E remained below 5 percent across 98.5 percent of the plateau, and estimated total permafrost area varied only within a narrow band of roughly 7,000 square kilometers.

Validation against 109 independent borehole records yielded an overall accuracy of 0.84 and a Cohen’s Kappa of 0.58, outperforming two existing 2020-period maps, one based on the temperature at the top of permafrost and another on mean annual ground temperature. Notably, the new map excelled at distinguishing seasonally frozen ground, achieving a true negative rate of 69.2 percent compared with just 34.6 percent for the mean-temperature-based map. This matters because existing products tend to overestimate permafrost extent, frequently misclassifying seasonally frozen ground as permafrost. In several transition zones, including the Gaize area, the Altun-Kunlun mountain margin, and the headwaters of the Yangtze, ground observations consistently favored the new simulation. At one meteorological station where the reference maps indicated permafrost, in-situ measurements showed a thawing index of roughly 2,004 degree-days far exceeding a freezing index of about 1,332 degree-days, a robust indicator of seasonally frozen ground that supported the team’s classification.

The decadal comparison revealed two distinct degradation hotspots with different underlying mechanisms. In the central plateau, particularly the transition zones of the Qiangtang Plateau, the dominant change was the conversion of permafrost to seasonally frozen ground, accounting for 7.41 percent of the total changed area. Degraded zones there showed an amplified increase in the ratio of thawing to freezing indices, more than double that of adjacent stable permafrost, pushing already marginal ground with mean annual temperatures close to zero across the thaw threshold. The close spatial correspondence between simulated degradation and the distribution of thermokarst lakes identified from Sentinel-2A imagery, along with known engineering defect points along the Qinghai-Tibet Railway and Highway, provides independent evidence that the mapped changes reflect real physical instability. In the southern and southeastern margins, flanking the Hengduan, Himalayan, and Gangdise mountains, the dominant transition was instead the conversion of seasonally frozen ground to non-frozen ground, which accounted for 39.62 percent of the total change and signals a vertical ascent of the frozen ground’s lower limit.

Perhaps the most intriguing finding concerns soil moisture. Although the plateau experienced a general wetting trend during the study decade, regions where the soil parameter increased, indicating rising degradation potential, received substantially smaller precipitation increases than regions where the parameter declined. Because water’s high specific heat capacity and latent heat of fusion buffer ground temperatures, wetter soils resist atmospheric warming while drier soils respond readily to it. Even under a warm-wet climate trajectory, spatial heterogeneity in moisture can therefore drive divergent permafrost fates across the plateau. The new map, publicly available through Figshare and the National Cryosphere Desert Data Center, offers engineers a tool for identifying high-risk infrastructure zones, gives ecologists a temporally specific baseline for monitoring the headwaters of Asia’s great rivers, and provides modelers with the period-specific benchmark they have lacked. As the Third Pole continues to warm at more than twice the global rate, this frozen snapshot of 2020 may prove to be one of the most valuable baselines the cryospheric community possesses.

Subject of Research: Permafrost distribution mapping and degradation on the Qinghai-Tibet Plateau

Article Title: A 2020 permafrost distribution map of the Qinghai-Tibet Plateau

Article References: A 2020 permafrost distribution map of the Qinghai-Tibet Plateau. (n.d.). https://doi.org/10.5194/tc-20-5697-2026

Image Credits: AI Generated

DOI: 10.5194/tc-20-5697-2026

Keywords: permafrost, Qinghai-Tibet Plateau, The Cryosphere, frost number model, remote sensing, machine learning, Random Forest, climate change, frozen ground, Third Pole, thermokarst, borehole validation