Similarly to our previous study (Zhao et al., 2022), a time series of remotely sensed LST was used to drive the MGPM. Specifically, we used a modified MODIS LST product developed by Zou et al. (2014, 2017), which partially accounts for surface influences such as snow cover, vegetation, and cloud presence through a cloud-gap filling algorithm and calibration with three AWS observations from representative permafrost regions with distinct surface types – alpine steppe, alpine meadow, and alpine desert – in the central QTP. Validation showed strong agreement between the modeled and observed LST, with R2 values ranging from 0.91 to 0.93 and RMSE values around 3 °C. Further evaluation at the TSH AWS site in the WKL region during the 2016–2018 observation period confirmed the product's reliability, with an R2 greater than 0.90 and an RMSE of 2.09 °C, demonstrating its effectiveness in capturing spatial variations in LST across the QTP.

In this study, we further refined the Zou et al. (2017) product to reconstruct historical LST data prior to 2003, extending the dataset back to 1980 using machine learning approaches. Three statistical models were employed: least squares linear regression (LR; Xing et al., 2023), random forest regression (RFR; Breiman, 2001), and multiple linear regression (MLR; Jiao et al., 2023). The LR model assumes a long-term linear relationship between air temperature (AT) and LST. For the RFR and MLR models, eight auxiliary variables known to influence LST were incorporated: AT, precipitation (Pre), skin temperature (ST), soil temperature in the top 0–10 cm (ST_1), fractional cloud cover (CFC), surface net radiation budget (SRB), leaf area index (LAI), and digital elevation model (DEM). Detailed descriptions of these variables, including their spatial resolution and data sources, are provided in Table 1. The main steps for reconstructing monthly LST from 1980 to 2022 are as follows.

All input variables were resampled to a spatial resolution of 1 km × 1 km using the nearest-neighbor method to match the resolution of the LST_Zou dataset. Monthly averages were then computed from the available data, which varied in terms of temporal resolution across sources. Missing values were filled through spatial interpolation using nearby data points. Notably, the latest downscaled AT and Pre data provided by Qin et al. (2022) extend only through 2019. To fill the data gap for 2020–2022, statistical downscaling was applied to AT and Pre from the CN05.1 dataset, following the method described by Su et al. (2016). CN05.1 is a gridded dataset developed by the China Meteorological Administration (CMA), offering daily meteorological variables at a spatial resolution of 0.25°. For more details, refer to Wu and Gao (2013). Additionally, since the earliest available satellite-based LAI data begin in 1982, values for 1980–1981 were filled by assuming no change and using the average LAI from the period 1982–1986.

Data from 2003 to 2019 were used for model training. In the LR model, AT was used as the sole input variable, with LST_Zou as the target output. For the MLR and RFR models, eight auxiliary variables (see Table 1) were used as inputs, with LST_Zou again serving as the output variable. The dataset was randomly partitioned into 10 subsets: 10 % of the samples were reserved for validation, and the remaining 90 % were used for training. This process was repeated 2000 times to ensure robustness. Model performance was evaluated using four metrics: R2, RMSE, MAE, and bias (Zhao et al., 2022). Details are provided in Fig. 2.

The monthly values of the eight auxiliary variables from 1980 to 2022 (see Table 1) were used as inputs into the trained LR, MLR, and RFR models from step (2). This enabled the generation of a continuous monthly LST time series starting from 1980.

In our modeling framework, we incorporated detailed thermophysical characterization of the subsurface based on measurements from 15 boreholes with observations across the WKL permafrost survey area, with depths ranging from 15 to 59 m. Core samples, field observations, and borehole logs (Li et al., 2012; Zhao and Sheng, 2019) indicate that ground ice content in the WKL region varies between 5 % and 50 %, depending on the type of Quaternary sediment. Higher ice contents are typically found in fine-grained glaciofluvial and lacustrine sediments due to enhanced segregation ice formation, while coarse-grained alluvial and colluvial deposits generally show lower ice content. Vertically, ice-rich layers are consistently observed near the upper boundary of permafrost, typically between 2 and 3 m depth. Ice content tends to increase slightly between 3 and 10 m and remains relatively stable below 10 m (Zhao et al., 2010a; Zhao and Sheng, 2019; Zou et al., 2024).

Depth-specific thermophysical parameters (thermal conductivity and heat capacity) for each stratigraphic class were estimated by calibrating modeled permafrost temperature and thaw depth against borehole observations. Calibration was performed using a numerical inverse modeling approach that minimizes the difference between simulated and observed ground temperatures by adjusting the thermal properties (Marchenko et al., 2024; Nicolsky et al., 2017). This method is detailed in Nicolsky et al. (2007, 2017), with examples of soil thermal property optimization provided in Zhao et al. (2022) and Marchenko et al. (2024).

Site-level stratigraphic and thermophysical data were spatially upscaled using vector-based geomorphological classification maps of western China at a 1:1 000 000 scale (Zhou and Cheng, 2019). Five common stratigraphic classes in the WKL region – i.e., glarosional, alluvial plain, eolian, colluvial valley, and lacustrine deposits – were identified. A summary of major Quaternary deposits is provided in Table 2, and their spatial distribution (Fig. 1b) is based on Zhou et al. (2007), gridded at 1 km × 1 km resolution to match our simulation scale.

The model resolution is determined by the input datasets. The computational domain covers the entire 55 669 km² WKL permafrost survey area, with a spatial resolution of 1 km × 1 km and a monthly temporal resolution. Following Zhao et al. (2022), each grid cell extends 100 m vertically, divided into 282 layers, with thicknesses ranging from 0.05 m at the top 4 m to 0.5 m at greater depths.

For each modeling grid cell, the ground thermal regime is simulated using site-specific stratigraphy and a time series of LST as the upper boundary condition. At the lower boundary (100 m depth), a Neumann condition is applied to represent geothermal heat flux, set at a constant value of 0.0724 W m⁻². This value is derived from measurements obtained from a 700 m deep borehole near the WKL permafrost region (Hu et al., 2000). To estimate a realistic initial soil temperature profile, a model spin-up is conducted using climate forcing from the early simulation years. Steady-state conditions are considered to be achieved when the temperature difference at all soil layers between consecutive annual cycles is less than 0.0001 °C. This equilibrium profile is then adopted as the initial condition for the transient simulation.

After excluding lake and glacier-covered areas, simulations were conducted for 47 284 grid cells. To improve computational efficiency, a spatial clustering approach was adopted following Cable et al. (2016), grouping grid cells based on similarities in climate forcing and soil thermal properties. Instead of simulating each grid cell individually, clusters were used as representative units.

To characterize the upper boundary LST forcing, a harmonic function was used to fit the time series (Sun et al., 2019). The fitted coefficients (initial annual mean temperature, trend, annual amplitude, and phase angle) were used to group climate forcing into distinct clusters. These were then combined with five soil thermal property classes (see Table 1), resulting in 13 248 unique input combinations for the WKL region. This approach reduced the number of simulations to just 28.02 % of the total grid cells, remarkably lowering computational demand. Similar cluster-based methods have been successfully applied in Canada (Zhang et al., 2013, 2014), Alaska (Cable et al., 2016), and the Swiss Alps (Fiddes et al., 2015) permafrost zones.

To assess long-term changes in the permafrost thermal regime over the last 43 years, key diagnostics were extracted from the modeled vertical soil temperature profiles down to 50 m. These include the mean annual ground temperature at 15 m depth (MAGT15m), which corresponds to the depth of zero annual amplitude (ZAA) on the QTP (Jin et al., 2008; Zhao et al., 2010b), and the temperature at the top of permafrost (TTOP). Additional depths were also evaluated against available borehole observations.

The active-layer thickness (ALT) was estimated using linear interpolation to locate the maximum depth of the 0 °C isotherm during the annual thawing period (Liu et al., 2020). Following Zhao et al. (2022) and Wu et al. (2018), a grid cell was classified as permafrost if the maximum annual ground temperature at any depth within the upper 50 m remained at or below 0 °C for 2 consecutive years. Cells where the minimum annual ground temperature dropped to ≤ 0 °C within 50 m depth during the same period were identified as seasonally frozen ground. Cells not meeting either condition were classified as unfrozen ground.

The TSH comprehensive observatory is located in the central–northern part of the WKL permafrost survey area (see Fig. 1a). The Quaternary deposits in this region are primarily lacustrine, consisting of fine-grained sediment from an ancient lake that dried up during the Lower Pleistocene (Li and Li, 1991). Since October 2015, an AWS at TSH has continuously recorded key meteorological variables, including hourly air temperature at 2, 5, and 10 m heights; relative humidity; shortwave and longwave radiation (both upward and downward); wind speed; and precipitation. Additionally, ground temperatures have been recorded automatically since 2010 from a 59 m deep borehole (ZK015; 35.36° N, 79.54° E; see Fig. 1b) at depths of 3, 6, 10, and 20 m (Zhao et al., 2021a). LST at TSH is estimated using continuous radiation measurements (since October 2015) and by applying the Stefan–Boltzmann law (see Hu et al., 2024 for details), providing a robust reference for validating satellite-derived LST and ground thermal modeling (see Sect. 4.1.1).

Beyond the TSH observatory, 27 boreholes have been drilled across the WKL region to monitor the ground thermal regime. These boreholes range in depth from 7.5 to 33 m and are distributed across various geomorphic units, soil types, and vegetation zones, covering elevations from 4200 to 5200 m a.s.l. (see Fig. 1b). Detailed descriptions are available in Zhao and Sheng (2019) and Li et al. (2012). A total of 15 of these boreholes are instrumented with thermistor sensors (accuracy ± 0.1 °C) placed at depths of 3, 6, 10, and 20 m (Zhao et al., 2021a), and manual ground temperature measurements have been conducted at 1- to 2-year intervals since 2010 during annual field investigations using a digital multimeter. In this study, data from these 15 boreholes were used for model calibration and validation. The remaining boreholes were used to support the spatial modeling of permafrost distribution, serving as reference points for identifying permafrost presence or absence.

During field investigation in September 2010, when seasonal thaw depths reach their annual maximum, a ground-penetrating radar (GPR) was used to manually measure thaw depth at 45 sites, most of which were located near boreholes (see Fig. 1b). The methodology and results are comprehensively described in Zhao and Sheng (2019). After removing duplicate measurements within the same 1 km × 1 km grid cells, a total of 25 unique thaw depth measurements were retained for model validation.

We implemented and compared the three algorithms described in Sect. 3.2.1 to identify the optimal model for reconstructing monthly LST data from 1980 onward. The validation results are presented in Fig. 2. Most data points in the scatterplots cluster closely along the 1:1 line, indicating a strong positive correlation (R2 > 0.90) and good agreement between LST_Zou and the estimated LST values. The LR model produced a mean absolute error (MAE) of 2.05 °C and an RMSE of 2.61 °C. The MLR model showed moderate improvement, with lower errors (MAE = 1.16 °C, RMSE = 1.55 °C). However, the RFR model yielded the best performance, achieving the lowest error metrics (MAE = 0.87 °C, RMSE = 1.26 °C).

Figure 3 compares the mean annual cycle of LST estimates from the three statistical models (LR, MLR, and RFR) with ERA5 Land, LST_Zou, and in situ observations from the TSH AWS over the period 2016–2018. All datasets exhibit a similar seasonal cycle consistent with the in situ data. However, both LST_Zou and ERA5-Land LST exhibit a systematic cold bias, particularly during the summer months of July, August, and September. The LST values estimated by all three statistical models help reduce this bias to varying extents, with the RFR model performing best. Despite this improvement, a residual cold bias in LST_Zou remains apparent during the same period. Overall, the RFR model-generated LST time series closely matches in situ observations and demonstrates sufficient accuracy for use in subsequent ground thermal modeling. Therefore, the RFR-derived monthly LST was adopted as the input forcing in the following simulation analyses.

Figure 4 shows the regional average of annual LST anomalies relative to the 1980–2022 mean. The results reveal a consistent positive trend of +0.40 °C per decade over the WKL region during this period. Interdecadal analysis highlights a remarkable warming trend in the mid-1980s, which then slowed slightly from the 2000s, during which LST deviations were relatively smaller. In the last decade, only positive anomalies were recorded, with 2016 exhibiting the largest positive deviation (+1.45 °C) compared to the 1980–2022 climate average.

To further assess regional LST anomaly patterns in WKL, Fig. 5 shows decadal deviations from the 1980–2022 mean. In the 1980s, most of the region (63.25 %) exhibited negative anomalies between -1.5 and -0.5 °C, with only 0.46 %, mainly at high elevations, falling below -1.5 °C. The 1990s shown a sharp warming, with 90.95 % of the area shifting to near-normal levels (-0.5 to 0 °C). By the 2000s, warming intensified: 83.78 % of WKL showed positive anomalies (0 to 0.5 °C), and 4.34 % exceeded 0.5 °C. Between 2011 and 2022, warming became more pronounced, with 63.97 % of the region being above 0.5 °C and some high-altitude zones surpassing 1.0 °C.

To validate the model's representation of large-scale ground thermal conditions, simulation outputs were compared with available in situ datasets (Fig. 1; Sect. 3.3), including MAGT10m (MAGT at 10 m) measurements from 15 sites in 2010, ALT data from 11 sites, thaw depth observations from 25 sites, and four historical permafrost distribution maps spanning different periods.

The comparison between observed and modeled MAGT10m results at 15 permafrost boreholes shows that 93.3 % (14 out of 15) of the data points cluster closely around the best-fit line, with deviations within ±0.25 °C (Fig. 6a). The analysis indicates strong overall agreement between measured and modeled MAGT10m for temperatures above -1 °C, with errors of 0.10 °C or less. However, for MAGT10m below -2 °C, the model shows a slight cold bias, particularly in areas with lacustrine sediments in the lowland regions of central WKL, where ground temperatures vary drastically due to complex local conditions (Fig. 6b). Despite this, the deviations between observed and simulated temperatures remain within 0.3 °C. Overall, the comparison suggests that the MGPM effectively replicates the measured MAGT10m, capturing the spatial variability in the validation area, with a correlation coefficient of r = 0.98 (p<0.01), and achieving an MAE and RMSE of 0.12 and 0.15 °C, respectively.

The scatterplots and spatial maps comparing measured and modeled ALT at 11 sites and thaw depths at 25 sites are shown in Fig. 7. The comparisons indicate that the model generally captures the range of ALT across the WKL region effectively. At 72.7 % of the sites (8 out of 11), the simulated ALT values closely match the observations, with deviations within ±0.25 m of the measurements (Fig. 7a). Notably, for the eolian sediment class, characterized by relatively shallow ALT around 2 m, the model performs exceptionally well, showing minimal bias (≤ 0.05 m), which suggests that the modeling approach is well suited to these conditions. However, the model underestimates ALT by approximately 0.25 m in lacustrine sediments near lake areas, where measured ALT exceeds 3 m (Fig. 7c).

A similar pattern is modeled for thaw depths: 91.3 % of the modeled values (21 out of 23) fall within ±0.25 m of the observations (Fig. 7b). For thaw depths greater than 3 m, the model tends to underestimate values, with the largest discrepancies of up to 0.5 m occurring in northern marginal permafrost zones (Fig. 7c).

Overall, despite slightly larger biases (> 0.25 m) at a few locations, the model effectively captures the spatial variability of ALT and thaw depth across the major geomorphological units of the WKL region. It yields an r of 0.96 for ALT and 0.94 for thaw depth, with corresponding MAE and RMSE values of 0.13 m and 0.16 m for ALT and 0.16 m and 0.18 m for thaw depth.

Figure 8 compares four representative frozen-soil type maps of the WKL region with the corresponding outputs from MGPM simulation outputs. In this analysis, 28 boreholes (see details in Fig. 1 and Sect. 3.3) are used as reference points to evaluate the accuracy of permafrost and seasonally-frozen-ground distributions. The results show that, while the maps by Li and Cheng (1996) and Wang et al. (2006) capture the general presence of permafrost across WKL, they fail to accurately delineate areas of seasonally frozen ground (Fig. 8a–c). Notably, these two maps show remarkable discrepancies in northeastern WKL, where they indicate continuous permafrost, while our simulation identifies seasonally frozen ground (Fig. 8i–j).

In contrast, the maps by Cao et al. (2023) and Zou et al. (2017), along with our simulation results, display a more accurate spatial pattern of frozen-ground types, correctly identifying nearly all permafrost and seasonally-frozen-ground locations, except for a single site near lakes in the southern WKL (Fig. 8c–d, g–h). However, small mismatches remain: compared to our simulations, Cao et al. (2023) and Zou et al. (2017) overestimate the extent of seasonally frozen ground by 1.84 % and 1.61 %, respectively, designating certain areas as seasonally frozen where our model indicates permafrost (Fig. 8k–l). Additionally, our simulation identifies about 0.61 % (Cao) and 0.58 % (Zou) of the central lowland region as seasonally frozen ground, whereas both maps categorize these areas as permafrost (Fig. 8k–l).

To investigate how the thermal state of permafrost evolves under ongoing climate change, it is first necessary to understand its initial conditions. Fig. 9 presents the modeled baseline distribution of MAGT15m, TTOP, and ALT for the year 1980. The results reveal pronounced spatial variability in the ground thermal regime across the WKL permafrost survey area. MAGT15m decreases dramatically with elevation, with the warmest average values of around 0.5 °C being simulated in the central low-elevation zones (below 4800 m a.s.l.) and the coldest, below -10 °C, being found in high-elevation areas (around 6000 m a.s.l.). Moderate variations in MAGT15m are also modeled across different soil stratigraphic classes. The coldest average MAGT15m, approximately -3.5 °C, occurs in eolian sediments, while the warmest, about -1 °C, is found in alluvial plain deposits.

A similar spatial pattern is evident in the modeled TTOP, although TTOP values are generally slightly lower than those of MAGT15m across the region (Fig. 9d–f). Likewise, ALT shows a strong elevation dependency. In lower-elevation areas (below 5400 m a.s.l.), ALT typically ranges from 2.5 to 3.0 m, with some localized zones exceeding 3.0 m. At higher elevations, ALT decreases progressively, dropping below 1.0 m, and approaches 0 m in areas above 6000 m a.s.l., where perennially frozen conditions prevail. ALT also varies markedly across stratigraphic classes: the alluvial class exhibits the greatest average ALT, while the glarosion class shows the shallowest (Fig. 9i).

Figure 10 shows the simulated interdecadal changes in MAGT15m, TTOP, and ALT across the WKL permafrost region from 1980 to 2022. From the 1980s to 1990s, MAGT15m remained relatively stable in 62.4 % of the region (±0.3 °C; Fig. 10a). A clear warming trend emerged from the 1990s to 2000s, with MAGT15m rising in 67.2 % of the area and localized increases exceeding 1.8 °C (Fig. 10b). From the 2000s to 2010–2022, warming became more variable, and 47.1 % of the region experienced cooling, with decreases of up to -1.8 °C (Fig. 10c). Overall, from 1980 to 2022, 58.6 % of the region warmed (up to +1.8 °C), while 25.5 %, mainly in central WKL, cooled, with decreases below -1.8 °C (Fig. 10d).

TTOP followed a similar trend. The largest increase occurred between the 1990s and 2000s, when 86.7 % of the region warmed, and 16.4 % showed increases above 0.8 °C (Fig. 10e). From the 2000s to 2010–2022, 70.5 % of the region continued to warm (up to +1.8 °C; Fig. 10f–g). Over the full period, 81.7 % of the region experienced a TTOP increase, with 17.2 % warming by over 1.3 °C. However, a small central area (∼ 7.4 %) showed declines ranging from -0.3 to -1.3 °C (Fig. 10h).

ALT increased most significantly between the 1980s and the 1990s, with 74.2 % of the region showing growth of 0.1–1.5 m and with some areas exceeding 1.5 m (Fig. 10i). From the 1990s to 2000s, 58 % of the region continued to increase, although 7.6 % modeled a decrease of -0.3 to -1.0 m (Fig. 10j). From the 2000s to 2010–2022, 59 % of the area experienced ALT increases, while 0.97 % showed sharp declines beyond -0.8 m (Fig. 10k). Overall, ALT increased by an average of 0.17 m across WKL from 1980 to 2022, with 83.1 % of the region warming and 16.9 %, mainly central, cooling, in some places by more than -0.8 m (Fig. 10l).

Figure 11 illustrates the interdecadal variations of MAGT15m, TTOP, and ALT across different elevation zones and soil stratigraphic classes. Overall, the modeled MAGT15m showed minor fluctuations and a slight upward trend from the 1980s to 2010–2022. The most noticeable increase occurred at the highest elevations (5600–6000 m a.s.l.), though changes remained less pronounced than those in TTOP (Fig. 11a–b). MAGT15m showed no remarkable differences across soil classes (Fig. 11d). In contrast, TTOP exhibited a clear warming trend across most soil classes, except in alluvial sediments. ALT increased remarkably with elevation from the 1980s to 2000s (Fig. 11c) and showed substantial variability among soil classes. The largest ALT increase (> 0.17 m) occurred in alluvial and lacustrine sediments, while the smallest (0.11 m) was in glarosion sediments (Fig. 11f).

Table 3 shows permafrost aggradation and degradation in response to climate variability across the WKL permafrost area from 1980 to 2022. Based on the initial simulation for the 1980s, approximately 82.27 % of the WKL area was underlain by permafrost, with 55.58 % occurring in eolian stratigraphy and 67.9 % occurring at elevations between 4800 and 5600 m a.s.l.

Permafrost extent remained unchanged from the 1980s to 1990s. A slight decline of 0.15 % was simulated between the 1990s and 2000s, followed by a 0.44 % increase from the 2000s to 2010–2022. These changes were primarily concentrated in low-elevation areas below 4800 m a.s.l. and in regions with alluvial plain sediments (Table 3). Overall, the simulations indicate that permafrost extent in WKL has remained relatively stable over the last 43 years.

Accurate representation of soil properties is critical for modeling water and heat transport in frozen soils at both global and regional scales (Dai et al., 2019; Lawrence et al., 2008; Harp et al., 2016; Hu et al., 2023a). However, most soil datasets used in models are based on data from seasonally frozen regions, and there remains a notable lack of coverage in the permafrost areas of the QTP, particularly for deeper soil layers (Hengl et al., 2017; Li et al., 2014, 2015; Shangguan et al., 2013). Westermann et al. (2017) addressed similar limitations in the Siberian permafrost region by using geomorphological classification maps to parameterize large-scale patterns of ground thermal properties such as sediment type, ground ice content, and surface characteristics.

In this study, we adopted a comparable approach by applying an existing stratigraphic classification map, gridded at 1 km × 1 km resolution, to represent the spatial distribution of sediment types, ground ice, and surface properties in the WKL region. These classifications were then used to parameterize subsurface properties in our model. However, small-scale heterogeneity in ground conditions may introduce considerable variability into the ground thermal regime, which cannot be resolved at the 1 km × 1 km resolution. Moreover, variability within each sediment class (Table 1) can result in biased model outputs.

To quantify the aforementioned model parameter uncertainty, we conducted a one-at-a-time sensitivity analysis (Fig. 12) using three representative boreholes located in stable permafrost, unstable permafrost, and seasonally frozen ground (see Table 4). Key model parameters were perturbed by ±10 % to evaluate their effects on permafrost thermal regime (MAGT15m and ALT). The results show that, among all parameters, upper boundary temperature (i.e., surface forcing) exerted the strongest influence on MAGT15m, though the absolute impact was modest, around ±0.15 °C in seasonally frozen ground and ≤±0.1 °C in permafrost areas. ALT showed similarly limited sensitivity, varying by ∼ ±0.1 m in stable permafrost and ±0.05 m in unstable zones. Soil thermal conductivity and water and/or ice content had a more pronounced effect on ALT, particularly in unstable permafrost, where a 10 % change could lead to a ±0.10–±0.15 m variation. In contrast, soil heat capacity had minimal influence on both MAGT15m and ALT.

The above analysis indicates that the model demonstrates robustness in relation to parameterization uncertainties and that uncertainties associated with stratigraphy have a limited effect on overall performance. Although stratigraphic classification and spatial variability inevitably introduce some degree of uncertainty, our approach is well supported by field measurements and observed thermal properties. Despite these limitations, we are confident that the model accurately represents the key thermal characteristics of each sediment class – key factors for simulating permafrost dynamics. Continued improvements in subsurface datasets, particularly in permafrost regions, will be essential for improving model performance in future applications.

The model assumes equilibrium initial conditions based on the first year's climate forcing, implying a stable land–atmosphere heat exchange prior to 1980. While this setup does not capture transient ground temperature states at that time, its influence diminishes over time. Sensitivity analysis also shows that initial conditions have a limited impact (e.g., moderate in seasonally frozen ground (∼ ±0.12 °C) and negligible in permafrost areas). Moreover, simulated permafrost temperatures, ALT, and thaw depth align well with observations and benchmark maps, suggesting that initialization uncertainties have minimal impact on long-term results.

The current MGPM configuration does not include some processes such as soil water convection or lateral heat and water fluxes, which can significantly affect ground thermal regimes, especially near taliks, waterbodies, and permafrost margins (Boike et al., 2015; Bense et al., 2012; Sjöberg et al., 2016; Kurylyk et al., 2016). As a result, the model may not fully capture thermal dynamics in areas with strong lateral fluxes, such as sharp mountain ridges or zones near lakes. Nevertheless, the model performs well in simulating ground temperature and ALT in the WKL region, suggesting that one-dimensional heat conduction captures the dominant thermal processes in this area.

In addition, the subsurface thermal model that is the MGPM uses satellite-derived LST as the upper boundary condition, which does not explicitly account for snow and vegetation canopy effects, potentially introducing uncertainties in densely vegetated areas. However, in the permafrost regions of the QTP, snow cover is typically thin (∼ 3 cm) and short-lived (lasting less than a day per event), and vegetation is sparse, with less than 10 % cover in the west (Wu and Zhang, 2008; Che et al., 2008; Wang et al., 2016; Zou et al., 2017; Orsolini et al., 2019; Yan et al., 2022). Under these conditions, the thermal offset between ground surface temperature (GST) and LST is minimal (Hachem et al., 2012). While thin snow cover may briefly cool the surface due to high albedo and rapid melt (Zhang, 2005), it contributes little thermal insulation, and its effect is likely to be negligible over the decadal timescale of our study. Still, the model's limitations highlight the need for further validation, especially regarding hydrogeological influences on permafrost thermal regimes and improved representation of surface heterogeneity in future developments.