The Xinjiang-Xizang transportation corridor， comprising the G219 Highway and the planned Xinjiang-Xizang Railway， is an important strategic route connecting Xinjiang and Xizang， and its construction and operation are of great significance for promoting regional economic development. The corridor traverses glacier-concentrated areas in the West Kunlun Mountains， the Gangdise Mountains， and the northern slope of the Himalayas， where glacier retreat， glacial lake expansion， and surging glacier activity may trigger hazards such as glacial lake outburst floods （GLOFs）， ice avalanches， and debris flows， posing potential threats to transportation infrastructure. However， systematic studies on glacier change characteristics and their engineering impacts along the Xinjiang-Xizang transportation corridor remain limited. To address this gap， this study comprehensively analyzed glacier changes along the corridor during 2000—2020 using multi-source remote sensing data and products， including glacier velocity， mass balance， and surging glacier inventories， and further evaluated their potential impacts on the planned Xinjiang-Xizang Railway. The results showed that glaciers along the Xinjiang-Xizang transportation corridor exhibited a significant retreat trend overall， although pronounced spatial heterogeneity existed. Between 2009 and 2020， the number of glaciers increased from 6 224 to 6 698， whereas the total glacier area decreased by 801.92 km2， and the retreat and fragmentation of large glaciers led to an increase in the number of small glaciers. Glacier area loss mainly occurred in the southeastern part of the corridor， where glaciers in the Gyirong region experienced area reductions exceeding 60%， while glaciers along the Hekang-Ritog section exhibited relatively small area changes and even slight increases in some areas. Over the past two decades， the regional glacier velocity decreased from 8.66 m⋅a-1 to 3.31 m⋅a-1 overall， indicating a gradual weakening of glacier motion. Around 2000， high-velocity glaciers were mainly concentrated along the Hekang-Ritog and Burang-Zhongba sections， but after 2010， glacier velocities in most areas declined to below 5 m⋅a-1， and glacier motion tended to slow down with reduced spatial variability. The average glacier mass balance along the corridor during 2000—2020 was （-0.10±0.02） m w.e.⋅a-1， indicating an overall mass loss. Glaciers along the Hekang-Ritog section exhibited mass gain， whereas from Ritog toward the southeastern part of the corridor， glacier mass balance gradually shifted from positive to negative， with the most severe mass loss occurring along the Zhongba-Gyirong section. The Burang-Tingri section exhibited an average mass balance of -0.46 m w.e.⋅a-1 and was characterized by a large number and extensive area of glacial lakes， making it a high-risk region that deserves particular attention during the future construction of the Xinjiang-Xizang Railway. Continuous glacier melting not only promoted the formation and expansion of glacial lakes but may also increase the likelihood of GLOFs and related hazards. Therefore， long-term monitoring and quantitative risk assessment were necessary for this region. This study identified eight surging glaciers that may potentially affect transportation infrastructure along the corridor， among which glaciers 5Y644J0011， 5Y644J0005， 5Y642Q0004， and 5Y642Q0027， located along the Hekang-Ritog section， exhibited terminus advances during 2009—2020， with advance rates ranging from 4.35 m⋅a-1 to 51.71 m⋅a-1. In particular， glacier 5Y642Q0027 showed continuously increasing terminus velocities in recent years， reaching a maximum of 80 m⋅a-1 and exhibiting pronounced surge characteristics. Moreover， this glacier was located only 3.29 km from the G219 Highway， indicating a relatively high hazard potential. Future efforts should focus on strengthening the dynamic monitoring of the aforementioned surging glaciers and avoiding their direct impact zones during the route selection and engineering construction of the Xinjiang-Xizang Railway. In conclusion， this study systematically reveals the overall patterns and regional differences of glacier changes along the Xinjiang-Xizang transportation corridor over the past two decades and preliminarily identifies two sections， namely Hekang-Ritog and Burang-Tingri， as potential high-risk sections where glacier changes may significantly affect the planned Xinjiang-Xizang Railway. The findings not only provide scientific support for the planning and construction of the Xinjiang-Xizang Railway and glacier hazard risk assessment， but also offer a case reference for studying the impacts of alpine glacier changes on major engineering projects under climate change.
Under the background of ongoing climate change， the Qinghai-Xizang Plateau is experiencing pronounced glacier retreat and accelerated permafrost degradation， leading to the release of substantial amounts of meltwater. These processes have exerted significant impacts on regional hydrological systems， particularly on lake water volume changes. Although previous studies have investigated the roles of glacier meltwater and permafrost meltwater separately， a systematic understanding of their combined effects on lake dynamics remains insufficient. This knowledge gap constrains the accurate assessment of future hydrological responses and water resource evolution under continued climate warming. To address these issues， this study focused on the Aksai Chin river basin located along the Xinjiang-Xizang transportation corridor， a region characterized by the coexistence of extensive glaciers and ice-rich permafrost. Multi-source remote sensing datasets were employed to quantitatively assess the contributions of glacier meltwater and ground ice meltwater to regional lake water changes. Specifically， Sentinel-1 SAR data acquired from 2019 to 2021 were used to derive time-series permafrost deformation based on advanced InSAR techniques. The retrieved deformation signals were further analyzed to characterize long-term subsidence and seasonal dynamics of permafrost. Based on these deformation results， a subsurface ground ice meltwater estimation model was established to quantify the release rate of meltwater caused by permafrost degradation. In addition， glacier meltwater release within the river basin was estimated using glacier surface elevation change data spanning the period from 2015 to 2020. By integrating these datasets， this study provided a comprehensive assessment of the relative contributions of glacier meltwater and ground ice meltwater to lake water volume changes at the basin scale. Furthermore， this integrated approach helped reduce the uncertainty associated with single-source estimation methods and improved the reliability of basin-scale water balance assessments in complex cryosphere environments. The results indicated that the total water volume change in the Aksai Chin river basin was approximately 0.0193 Gt⋅a-1 during 2015—2020. The permafrost region exhibited an overall subsidence trend， with an average deformation rate of about -2.00 mm⋅a-1， accompanied by a mean seasonal amplitude of approximately 3.70 mm. These deformation characteristics reflected active permafrost degradation processes and pronounced seasonal thaw-freeze cycles. The estimated release rate of ground ice meltwater was approximately （-0.0043±0.0012） Gt⋅a-1， contributing about 22.28% of the observed lake water changes. In comparison， glacier meltwater release was significantly higher， with an estimated annual mean rate of approximately -0.046 Gt⋅a-1， which was about 10.70 times greater than that of ground ice meltwater. The contribution of glacier meltwater to lake water volume change reached approximately 238.34%， exceeding 100%. This apparent over-contribution suggested that glacier meltwater not only compensated for lake water losses driven by other hydrological components， but also dominated the overall water balance. This phenomenon may be associated with a reduction in net basin precipitation or enhanced evaporation in the river basin， indicating that glacier meltwater played a dominant role in regulating regional lake dynamics under current climatic conditions. Moreover， the significant differences in the magnitudes of glacier and permafrost contributions highlighted the importance of considering multiple cryosphere components when assessing hydrological responses in cold regions. Neglecting any process may lead to biased interpretations of water balance mechanisms and affect the scientific understanding of future trends under climate change scenarios. Overall， the findings of this study provide new insights into the coupled impacts of glacier retreat and permafrost degradation on hydrological processes in high-altitude cold regions. The integrated methodological framework developed in this study offers an effective approach for quantifying the respective contributions of multiple meltwater sources in the cryosphere. These findings have important implications for deepening the understanding of water resource evolution and for predicting future hydrological responses along the Xinjiang-Xizang transportation corridor under continued climate warming.
The Xinjiang-Xizang transportation corridor， consisting of National Highway G219 and the planned Xinjiang-Xizang Railway， is an important engineering corridor connecting Xinjiang and Xizang. The corridor traverses high-mountain cryospheric regions including the West Kunlun Mountains， Gangdese Mountains， Himalayas， and Nyainqêntanglha Mountains. Under the combined effects of climate warming， glacier retreat， permafrost degradation， tectonic activity， and fluvial erosion， the threat of glacier-related hazards has become increasingly pronounced. Although glacial–periglacial hazards， such as ice avalanches， rock-ice avalanches， and freeze-thaw-induced landslides， differ from glacial lake outburst floods （GLOFs） in terms of dominant materials and triggering mechanisms， the two hazard types may form cascading processes. Unstable ice or slope materials entering a glacial lake can generate impact waves and disturb the stability of the lake system or moraine dams. Previous regional assessments have generally investigated these two types of hazards separately and have not adequately considered the external disturbance posed by potentially unstable source materials surrounding glacial lakes in GLOF susceptibility assessment. Therefore， this study focused on the Xinjiang-Xizang transportation corridor and its adjacent areas， integrated glacier-related hazard inventories with topographic， geological， meteorological， and hydrological data， and analyzed the spatial distribution patterns of glacial-periglacial hazards and GLOFs. The certainty factor （CF） and random forest （RF） models were used to assess the susceptibility to glacial-periglacial hazards. The susceptibility derived from the better-performing model was then used to represent potential disturbances from unstable source materials surrounding glacial lakes and incorporated as one of the indicators in an analytic hierarchy process （AHP） framework for further GLOF susceptibility assessment. Spatial analysis results of hazards showed that more than 90% of historical GLOFs occurred within 2.24 km of modern glaciers， mainly distributed at elevations above 4 590 m and on relatively gentle slopes. In contrast， glacial-periglacial hazards were concentrated at elevations of 3 776~4 590 m， on slopes steeper than 17.16°， and near drainage networks， with higher occurrence frequencies on south- and southeast-facing slopes. In the glacial-periglacial hazard susceptibility assessment， the RF model achieved an AUC of 0.916 on the independent test dataset， exceeding the value of 0.831 obtained by the CF model and exhibiting stronger spatial discrimination. Areas classified as low， moderate， high， and very high susceptibility accounted for 68.59%， 23.31%， 6.37%， and 1.73% of the study area， respectively. High and very high susceptibility zones were mainly concentrated from the eastern Himalayas to the Nyainqêntanglha Mountains， while elevation， distance to glaciers， and distance to drainage networks were the three factors with the greatest contributions. The GLOF susceptibility assessment classified 934 glacial lakes as low susceptibility， 804 as moderate susceptibility， 363 as high susceptibility， and 78 as very high susceptibility. A total of 441 lakes were classified as high or very high susceptibility， accounting for 20.2% of all assessed lakes， and were mainly distributed in the Nyainqêntanglha Mountains and along the northern side of the eastern and central Himalayas. Among the 72 historical outburst lakes that could be spatially matched with lakes having complete attribute information， 54 were classified as high or very high susceptibility， corresponding to a historical-event hit rate of 75.0%. Within the 20 km buffer zone of the planned railway， high and very high glacial-periglacial hazard susceptibility zones together accounted for 11.0% of the area， with the most prominent segment being approximately 90 km in the southern part of the Rutog section. Screening of potential GLOF flow paths based on river network data further identified 12 railway sections potentially affected by GLOFs， including three along the Lhaze-Tingri section， two along the Saga-Zhongba section， six along the Zhongba-Burang section， and one in the eastern part of the Zanda section. Enhanced monitoring of upstream glacial lakes is therefore recommended for these sections. Overall， this study statistically characterizes the spatial distribution patterns of glacial-periglacial hazards and GLOFs， assesses the susceptibility of both hazard types， and preliminarily identifies railway sections potentially exposed to glacier-related hazards along the planned Xinjiang-Xizang Railway. The findings provide a scientific basis for route planning and disaster risk reduction along the Xinjiang-Xizang transportation corridor.
Located along the western margin of the Qinghai-Xizang Plateau， the Xinjiang-Xizang transportation corridor traverses high-relief mountainous areas， including the Kunlun， Karakoram， Gangdise， and Himalayan mountains. The proposed Xinjiang-Xizang Railway and the existing G219 highway face potential hazards from accelerated glacier retreat under global warming， including ice avalanches and glacial lake outburst floods. However， baseline data on glacier thickness and ice volume remain scarce， topographic controls on glacier distribution are poorly understood， and glaciers within upstream catchments along the railway have not been systematically investigated. To address these issues， this study applied a laminar-flow-based ice thickness model using the Randolph Glacier Inventory （RGI 6.0）， integrating glacier surface velocities （50 m resolution）， ASTER GDEM （30 m）， and near-surface air temperature （1 km） to derive pixel-wise ice thickness and calculate volumes. Valley shape factors were iteratively updated， and a minimum slope threshold （4°） was set to avoid singularities. Model parameters were regionally calibrated using 4 579 glaciers （77.83% of the total area）， and uncertainty was estimated via error propagation considering errors from velocity， flow parameters， shape factors， density， and slope gradients， yielding a relative volume uncertainty of 18.63%. Additionally， sampling points were placed at 1 km intervals along the planned railway alignment， and ArcGIS hydrological tools were used to delineate upstream catchments. Glaciers potentially affecting the railway were identified through spatial overlay analysis. The results showed that the 5 971 modeled glaciers （covering an area of 4 944.74 km2） had a total ice volume of （259.104±48.271） km3， with major concentrations in the Hekang-He’an （43.80%）， Rutog （16.76%）， Burang （15.53%）， and Gyirong-Lhaze （19.62%） segments. Glaciers at elevations of 5 200~6 200 m accounted for 90.84% of the total volume， glaciers on slopes of 10°~25° accounted for 78.68%， and north-， northeast-， and northwest-facing glaciers contributed 34.64%， 22.87%， and 12.61%， respectively， demonstrating the strong constraint of topography on glacier distribution. A total of 43 glaciers were identified within the upstream catchments along the planned railway， with a total volume of （5.703±1.062） km3 （only 2.20% of the corridor total）， predominantly in the Hekang-Rutog segment. Among these， 18 glaciers larger than 1 km2 （volume 5.294 km3） required sustained high-resolution monitoring. This study provides the first comprehensive baseline data on glacier volume for the corridor， clarifies the controlling effects of topographic factors on glacier distribution， and precisely identifies key glacier groups within high-risk catchments along the railway. The findings fill critical data gaps in ice thickness inversion and offer direct baseline data and decision support for hazard assessment， engineering protection design， and future climate adaptation strategies for the Xinjiang-Xizang Railway. This study provides practical guidance for the safe construction and operation of major linear infrastructure in high-mountain environments.
The mountain glacier-proglacial lake system is a key component of the cryospheric mass and energy cycle. Simulating the future evolution of proglacial lakes within this system is of great scientific significance for revealing the eco-environmental evolution in glacier-peripheral regions， promoting sustainable water resource utilization， and optimizing disaster risk mitigation and adaptive management. However， existing studies on future glacial lake simulation and prediction primarily focus on potential future glacial lakes and ice-dammed lakes. A common limitation of such proglacial lake estimation methods is that， by using glacier boundaries as a static constraint， they fail to account for the impact of proglacial lakes on the ablation of the glacier terminus in contact with them. Meanwhile， existing one-dimensional shallow ice approximation models exhibit limitations in simulating proglacial lake evolution， as they can only provide one-dimensional length information. This restricts their further application and makes it difficult to characterize the two-dimensional spatiotemporal dynamics of lake expansion. Therefore， employing higher-order or more efficient ice flow models is critical for simulating the two-dimensional expansion of proglacial lakes. To address this issue， this study constructed a simulation framework centered on the instructed glacier model （IGM） and coupled with terminal calving processes， utilizing multi-source remote sensing observations， topographic data， and ice thickness data. This study addressed the following scientific questions： （1） how the future two-dimensional expansion of high-risk proglacial lakes can be simulated； and （2） as the lakes expand， how the potentially hazardous slopes around the margins of the glacier catchments will evolve. The Jiemayangzong Glacier-proglacial lake system along the Xinjiang-Xizang transportation corridor was selected as the study area to simulate its evolution under future climate scenarios. Potential hazard source slopes around the proglacial lake within the glacier catchment were further identified using established expert empirical criteria. Multi-source remote sensing observations （1990s—2025） showed that the proglacial lake expanded by 0.66×106 m2 （+86.84%）： accompanied by a retreat of the lake-ice interface of approximately （1 188.68±66.98） m—a process primarily controlled by the recession of the glacier terminus. These findings offered essential insights for future simulations of proglacial lake expansion. A retrospective analysis of the historical evolution of the Jiemayangzong proglacial lake （2009—2025） using the developed simulation scheme showed good consistency between the simulated glacier terminus retreat and remote sensing observations （R 2=0.96， RMSE=31.4 m）： confirming the reliability of the proposed simulation scheme. Future climate scenario projections indicated that the proglacial lake area at the terminus of the Jiemayangzong Glacier was expected to reach its peak around 2047， at which time its area and volume would increase to 1.86 km² （+12.54%） and 0.067 km³ （+30.25%）： respectively， relative to current levels. A high-elevation hanging glacier on the eastern side of the proglacial lake （area≈10.47×10⁴ m²， with an average thickness of 17.2 m） was identified as a potential high-risk， high-potential-energy source. In addition， the rapidly expanding high-elevation glacial lake above the eastern side of the proglacial lake was identified as a high-risk source， with an expansion rate of 4 969.79 m2·a-1 over the past five years and an estimated current average water depth of approximately 11.62 m. The two-dimensional simulation of future proglacial lake evolution conducted in this study provides important methodological references and a scientific basis for the dynamic prediction and risk assessment of high-risk proglacial lakes along transportation corridors. Furthermore， this study contributes to identifying critical parameters of proglacial lake expansion （such as maximum area and timing of peak attainment） and their driving mechanisms， deepening the understanding of the interaction mechanisms between proglacial lakes and glaciers， and thereby offering scientific support for glacier-related disaster prevention and control as well as sustainable water resource management.
The coastal-inland transition zone of Princess Elizabeth Land in East Antarctica links the marine-influenced coastal environment of Prydz Bay with the interior ice sheet and constitutes an important part of the Chinese PANDA traverse and inland expedition route. The Grove Mountains， located on the inland side of this transition zone， are characterized by a mosaic of blue-ice areas and exposed rock， pronounced topographic relief， and strong surface heterogeneity. However， due to the long-term lack of continuous and stable ground-based meteorological observations in this region， its near-surface meteorological characteristics and their differences from those of the coastal environment remain insufficiently understood. Based on the first complete annual cycle of hourly continuous observations from the Grove Mountains automatic weather station during 2023—2024， long-term observational records at Zhongshan Station， and ERA5 reanalysis data， this study comparatively investigated the near-surface meteorological characteristics， long-term variations， and extreme events at the Grove Mountains and Zhongshan Station， thereby providing a basis for understanding the near-surface meteorological characteristics and their spatial differences across the coastal-inland transition zone of East Antarctica. To enable direct comparison with Zhongshan Station， wind speeds observed at different heights at the Grove Mountains were extrapolated to the standard 10 m level. The Weibull distribution was applied to characterize wind speed probability distributions， and percentile-based thresholds were used to identify extreme wind and extreme temperature events. The results revealed a pronounced coastal-inland climatic gradient between the Grove Mountains and Zhongshan Station. During 2023—2024， the annual mean air temperature at the Grove Mountains was -24.9 ℃， substantially lower than the -10.0 ℃ recorded at Zhongshan Station， while the corresponding mean air pressures were 787 hPa and 984 hPa， respectively. In terms of the wind regime， the mean 10 m wind speed at the Grove Mountains reached 13.3 m⋅s-1， markedly higher than the 6.7 m⋅s-1 observed at Zhongshan Station. Wind direction at the Grove Mountains was highly concentrated within the NE-ENE sector， indicating a persistent and stable katabatic wind regime. In contrast， wind directions at Zhongshan Station were distributed across a broader sector， and wind speed variability was greater， indicating stronger influence from synoptic-scale systems， coastal thermal contrasts， and local topography. Evaluation against observations indicated that ERA5 reproduced air temperature and air pressure variations at both sites reasonably well， although it tended to underestimate high wind speeds. Based on ERA5， the long-term analysis for 1991—2024 showed that warming was most pronounced in spring at both sites. The spring warming rate at the Grove Mountains reached ［0.60 ℃⋅（10a）-1］， exceeding the corresponding rate of ［0.47 ℃⋅（10a）-1］ at Zhongshan Station， whereas no significant long-term trends were observed for air pressure or wind speed. Analysis of extreme events further indicated that the occurrence frequency of extreme wind events at Zhongshan Station during 2002—2024 exhibited pronounced interannual variability but no significant long-term trend， with the highest frequency occurring in winter. Composite analyses of the 500 hPa circulation fields showed that these extreme wind events were generally associated with intensified low-pressure systems over Prydz Bay and enhanced northeasterly mid-tropospheric flow. Although the observational period at the Grove Mountains was relatively short， extreme wind events there also mainly occurred under strengthened large-scale dynamical forcing， and the associated low-pressure systems were stronger， more extensive， and located farther east than those linked to Zhongshan Station. In addition， the frequencies of extreme high-temperature and extreme low-temperature events at Zhongshan Station did not show significant long-term trends during the study period. Overall， this study establishes a preliminary meteorological baseline for the Grove Mountains based on the first complete annual cycle of continuous observation data and reveals the main characteristics of near-surface climatic differentiation across the coastal-inland transition zone of East Antarctica. These results provide observational support for inland expedition planning and for the evaluation and improvement of Antarctic reanalysis products and numerical models.
Accurate estimation of snow depth in seasonal snow cover regions is of great significance for regional water resource assessment， hydrological process modeling， and snow disaster early warning. The Songhua River Basin， located in Northeast China， is a typical seasonal snow cover region where snow processes profoundly influence the regional hydrological cycle and agricultural production centered on the black soil zone. However， the complex terrain conditions and diverse land cover types within the basin pose substantial challenges for passive microwave remote sensing-based snow depth estimation—factors such as vegetation scattering， topographic heterogeneity， and variations in snow grain size collectively increase the uncertainty in interpreting microwave signals. In recent years， machine learning methods have demonstrated considerable potential in capturing the nonlinear relationships between snow depth and multi-source explanatory variables. Nevertheless， systematic evaluations of different ensemble learning models for this basin remain limited， and the relative contributions of various influencing factors to estimation accuracy require further quantitative investigation. To address these gaps， this study aims to： （1） construct a multi-dimensional feature dataset by integrating AMSR-2 passive microwave brightness temperature data， ERA5-Land reanalysis data， SRTM elevation data， and MODIS land cover data； （2） systematically compare the snow depth estimation performance of four models—XGBoost， CatBoost， LightGBM， and Random Forest （RF）； and （3） quantitatively reveal the contribution mechanisms of various influencing factors to estimation accuracy through feature importance analysis. The Songhua River Basin （119°52′~132°31′ E， 41°42′~51°38′ N）， covering a total area of approximately 56.12×104 km²， was selected as the study area. Multi-source data from ten snow seasons spanning 2013 to 2023 were collected and processed. A total of 24 feature variables were extracted and classified into four categories： 13 brightness temperature-related variables （including multi-frequency channels and polarization difference combinations）， four meteorological and thermal variables （air temperature， snow density， soil temperature， and snow layer temperature）， three terrain factors （elevation， slope， aspect， and roughness）， as well as land cover type and geographic location variables. The dataset was chronologically divided into a training set （2013—2021， 69 771 samples） and an independent test set （2022—2023， 18 422 samples） to simulate real-world forecasting scenarios. Model performance was evaluated using four metrics： R²， RMSE， MAE， and MRE. The results showed that the four models achieved comparable generalization performance on the test set， with R² ranging from 0.6661 to 0.6738 and RMSE ranging from 4.44 to 4.50 cm. XGBoost exhibited the best overall performance （R 2=0.6738， RMSE=4.44 cm）， LightGBM achieved the best error stability （MAE=2.85 cm）， while Random Forest showed the highest training accuracy but suffered from significant overfitting. Feature importance analysis revealed that brightness temperature differences were the core predictors across all models， with 18.7V/36.5H and 18.7H/89.0V contributing 34% and 10% in XGBoost， respectively. Soil temperature also demonstrated moderately high and stable importance （6%~12%）. Longitude ranked highly across all models， reflecting the dominant role of moisture transport from the Sea of Japan in shaping the east-west gradient of snow depth distribution in the basin. All models exhibited systematic underestimation， with bias increasing with snow depth. In the shallow snow region （≤5 cm）， the MRE exceeded 130%， indicating that the low RMSE in this range is primarily due to the small snow depth values rather than high predictive accuracy. In the deep snow region （>25 cm）， LightGBM achieved the best relative accuracy （MRE=32.92%）， while RF performed the worst. Land cover type significantly affected estimation accuracy： cropland yielded the highest accuracy （RMSE： 3.91~4.04 cm）， followed by grassland （4.17~4.62 cm）， while forested areas produced the lowest accuracy （6.81~7.48 cm）， mainly due to the attenuation of microwave signals by forest canopies. Further analysis indicated that the cross-land cover RMSE differences were largely driven by the mean snow depth characteristic of each type， rather than by land cover properties alone. In addition， the XGBoost model significantly outperformed the ERA5-Land reanalysis snow depth product （R 2=0.283， RMSE=6.588 cm）， highlighting the advantages of the data-driven approach at the regional scale. This study integrates multi-source data fusion with machine learning and provides a reliable method for snow depth estimation in the Songhua River Basin. The findings reveal the key predictive roles of brightness temperature differences and soil temperature， elucidate the moderating effect of land cover type on estimation accuracy， and identify the limitations of current models in forested areas and under deep snow conditions. Future research should focus on developing microwave radiative transfer models suitable for forest-covered regions and systematically evaluating the spatial generalization capability of the models， so as to further enhance the applicability and robustness of snow depth estimation in complex terrain environments.
Accurately monitoring the spatiotemporal distribution of snow depth （SD） is essential for understanding global climate change， surface energy balance， and regional hydrological cycles. This is particularly critical in arid and semi-arid regions such as Xinjiang， China， where seasonal snowmelt serves as a vital freshwater resource. Currently， machine learning algorithms are widely used for large-scale SD inversion based on satellite remote sensing data due to their robust non-linear fitting capabilities. However， these data-driven approaches are often criticized as “black box” models that lack physical interpretability and struggle to capture the complex metamorphism of snow cover. Conversely， while physical snow process models can accurately simulate the physical mechanisms of internal snow cover evolution， their application at large spatial scales is frequently hindered by high computational costs， strong dependence on high-precision meteorological forcing data， and complex requirements for localized parameter calibration. To balance the physical robustness of physical models and the scalability of data-driven methods， this study proposed a novel SD inversion algorithm： the coupled snow process model-random forest （CSPM-RF）. This method integrated the physical constraints of a snow process model into a machine learning framework to enhance the accuracy and physical consistency of SD estimation. Specifically， the snow thermal model （SNTHERM）， driven by uncalibrated GLDAS meteorological forcing products， was utilized to simulate the physical snow processes across the Xinjiang region from 2013 to 2021. Although using uncalibrated meteorological forcing products with default parameters limited the absolute simulation accuracy， the model effectively extracted critical microphysical parameters—including the weighted mean grain size （Davg） and snow density （rho） of the snow profile， as well as the simulated snow depth （prior_SD）. These parameters served as invaluable “physical prior features.” These physical priors were then combined with optical remote sensing data （MODIS normalized difference snow index and fractional vegetation cover）， passive microwave data （AMSR2 brightness temperature differences）， and topographic features to drive the random forest inversion algorithm. To optimize the input feature space， the Shapley Additive Explanations （SHAP） method was employed to evaluate feature importance and eliminate redundant variables. Based on 125 775 valid daily observations from 105 meteorological stations （stratified into 69 training stations and 36 independent testing stations）， the validation results demonstrated the superior performance of the CSPM-RF algorithm. （1） On the independent test set， the model achieved a coefficient of determination （R 2） of 0.715 and a root mean square error （RMSE） of 4.811 cm. An ablation study indicated that incorporating SNTHERM-derived prior features improved the R 2 by 6.9% and reduced the RMSE by 7.6% compared to a purely data-driven model， confirming that physical priors could enhance the model’s generalization ability even when derived from uncalibrated meteorological data. （2） The model exhibited robust performance under shallow to moderate snow conditions （SD≤30 cm）， which dominate the study area， but showed systematic underestimation in deep snow regions due to passive microwave signal saturation effects. Spatially， the optimal accuracy was achieved in the mid-latitude Tianshan Mountains and mid-altitude zones （1 000~2 000 m）. Land cover analysis revealed that croplands and grasslands yielded the highest inversion accuracy. （3） Temporally， the model demonstrated interannual stability and successfully reproduced the macroscopic snow distribution pattern of “Three Mountains and Two Basins.” Trend analysis indicated that the regional SD remained relatively stable during the study period. （4） Furthermore， compared with an existing pure data-driven snow depth product （RFSD） over the overlapping period of 2013—2020， the CSPM-RF model showed a significant advantage. It achieved a higher overall accuracy （R 2 of 0.737 vs. 0.606， RMSE of 4.803 cm vs. 5.874 cm） and effectively mitigated systematic biases caused by snow metamorphism， exhibiting notably better performance under deep snow conditions （SD>30 cm） and in complex terrains such as the Tianshan region and mid- and high-altitude zones. In conclusion， this study confirms that prior features simulated by a physical snow process model driven by uncalibrated meteorological products can provide critical physical constraints for machine learning algorithms. By bridging the gap between physical mechanisms and data-driven methods， the CSPM-RF algorithm offers a valuable， high-precision new approach for operational large-scale snow depth monitoring in arid and semi-arid regions.
Against the background of global climate warming， permafrost regions such as the Qinghai-Xizang Plateau are experiencing accelerated warming and continuous permafrost degradation. Coupled with the increasing demand for wide embankments and complex structures in high-grade highway construction， conventional settlement evaluation criteria based on single indicators， static judgments， or empirical thresholds face severe challenges regarding applicability， accuracy， and predictive capability. Scientific， rational， and dynamically adaptive criteria for road settlement risk assessment are not only an important basis for design optimization， operational safety assurance， and distress prevention and control in cold-region road engineering but also key support for maintenance strategy formulation， engineering resource allocation， and life-cycle management. This paper presents a systematic review of road settlement risk assessment methods in permafrost regions， covering the evolution， indicator systems， technical support， applicable conditions， existing limitations， and optimization directions of relevant assessment approaches. Based on core domestic and international literature， this study systematically reviews the development of road settlement assessment research in permafrost regions and focuses on summarizing and comparatively analyzing five mainstream categories of methods： field and remote sensing monitoring， empirical statistical and curve-fitting， theoretical analytical， numerical simulation， and intelligent data-driven and deep learning methods. Field monitoring and remote sensing techniques can directly acquire information on subgrade deformation， ground temperature， moisture conditions， and surface settlement， thus constituting an important basis for identifying road distresses and validating assessment models. Empirical statistical and curve-fitting methods require relatively few parameters and are convenient for engineering application， making them suitable for settlement trend description and short-term prediction. Theoretical analytical methods help reveal the intrinsic relationships among thaw settlement， consolidation deformation， and subgrade thermal stability， thereby providing mechanistic support for establishing evaluation criteria. Numerical simulation methods can couple temperature fields， moisture migration， stress-deformation， and phase-change processes， and are applicable to analyzing road settlement evolution under complex working conditions. Intelligent data-driven and deep learning methods show considerable potential in multi-source data processing， nonlinear relationship mining， and regional risk identification. By examining the technical characteristics， advantages， limitations， and applicable scenarios of these methods， this paper reveals their underlying logic， applicability boundaries， and key constraints. The review indicates that the existing methodological system for road settlement risk assessment in permafrost regions still faces evident bottlenecks. A clear mismatch in capabilities remains within the assessment system， and field monitoring， empirical statistical methods， and theoretical analytical approaches have not yet formed an effective complementary framework. Multi-method collaboration and multi-source information fusion remain insufficient， restricting the extension from local settlement identification to regional risk discrimination. In addition， assessment results remain disconnected from engineering decision-making， and a dynamic early-warning and decision-making system serving proactive prevention and control throughout the full life cycle is still lacking. To overcome these methodological bottlenecks， this paper discusses future development directions for road settlement risk assessment in permafrost regions. Future studies should strengthen integrated multi-source monitoring and establish long-term， continuous， and coordinated space-air-ground observation systems. They should promote the deep integration of mechanistic and data-driven models to improve the physical interpretability and predictive efficiency of assessment methods. Greater attention should also be paid to characterizing the spatial distribution of settlement and regional heterogeneity so that the assessment scale can be extended from local cross-sections to entire routes and regional domains. Furthermore， real-time early warning and response should be advanced by establishing a dynamic risk warning and maintenance decision-making system for road safety during the operation period. Meanwhile， the probabilistic representation of assessment results should be enhanced by fully considering uncertainties associated with permafrost environments， engineering parameters， and climate change. This review provides a theoretical reference for developing a collaborative assessment framework for road settlement in permafrost regions and offers methodological support for promoting the transition of road settlement risk assessment from static analysis to dynamic prediction， from single-criterion judgment to integrated decision-making， and from passive treatment to proactive prevention and control.
Permafrost degradation in the Qinghai-Xizang engineering corridor has intensified because of climate warming and human activities. Consequently， retrogressive thaw slumps （RTSs） occur frequently， posing potential threats to the safe operation of major linear infrastructure and the regional ecological environment. This study used multi-temporal high-resolution remote sensing to investigate the spatial distribution， temporal changes， and influencing factors of RTSs along the Qinghai-Xizang Highway （G109）， providing a basis for hazard identification， monitoring， and engineering risk prevention. The study area comprised a 60 km buffer zone on either side of G109. Gaofen-series remote sensing images with a spatial resolution of 2 m， acquired during the active-layer thawing seasons of 2016， 2019， and 2025， were manually interpreted to establish three RTS inventories. Changes in slump number， area， spatial pattern， and expansion characteristics were analysed. Elevation， slope gradient， slope aspect， active-layer thickness， ground-ice content， fault density， distance to roads， and the normalized difference vegetation index （NDVI） were selected as influencing factors. Spatial statistics and the area-based frequency ratio （FR） method were applied to evaluate the relative enrichment of RTS areas in different classes of influencing factors. Meanwhile， annual mean air temperature and annual precipitation during 1971—2020， together with land surface temperature data during 1981—2018， were used to describe the regional climatic and thermal background associated with RTS activity. RTSs exhibited significant spatial clustering and complex morphology. In 2025， 918 RTSs with a total area of 35.09 km2 were identified， and they were concentrated mainly in the Fenghuoshan-Wudaoliang section. Fenghuoshan was the core area of RTS development， containing 603 RTSs covering 25.07 km2， equivalent to 65.69% of all mapped RTSs and 71.42% of their total area. From 2016 to 2025， the number of RTSs increased from 557 to 918， and their total area increased from 13.71 km2 to 35.09 km2， representing an increase of 155.95%. Area growth was stage-dependent. From 2016 to 2019， it resulted from both new RTS formation and expansion of existing RTSs. From 2019 to 2025， expansion of existing slump patches accounted for 86.14% of the area increase. Multi-temporal imagery also showed that existing slumps experienced headwall retreat， lateral and frontal expansion， and local patch coalescence. This indicated that after RTSs formed， their evolution was characterized primarily by the continuous expansion of existing slumps. RTSs were preferentially enriched at elevations of 4 700~4 900 m， on gentle-to-moderate slopes of 4°~16°， and on north-， northeast-， and northwest-facing slopes. High FR values were also observed in areas with active-layer thicknesses of less than 2 m， medium-to-high ground-ice content， and moderate vegetation cover， particularly where NDVI was 0.3~0.4. These findings identify shallow， ice-rich permafrost as a key material condition for RTS development. Relatively high fault density and moderate vegetation cover may further regulate local hydrothermal conditions， whereas distance to roads did not show a clear near-road enrichment pattern at the 60 km buffer scale. From 1971 to 2020， annual mean air temperature and annual precipitation increased at rates of 0.22 ℃/decade and 10.13 mm/decade， respectively. Land surface temperature generally exhibited an increasing trend during 1981—2018， although the trend showed marked spatial differences among corridor sections. These changes provide a regional environmental context for enhanced RTS activity， although the strength of these associations and the underlying causal relationships require further quantitative verification. For established RTSs， exposure and continued thawing of ground ice in the headwall are the dominant material conditions for headwall retreat and boundary expansion， while external hydrothermal conditions mainly affect the intensity and rate of expansion. Risk assessment should consider not only new RTSs but also the sustained expansion of existing ones. Priority monitoring should focus on the Fenghuoshan-Wudaoliang section and on headwall retreat， boundary expansion， and patch coalescence in actively expanding RTSs.
As an important ecological security barrier in western China， the Qilian Mountains host extensive permafrost that is highly sensitive to global climate change. In recent years， thaw-related hazards driven by rising temperatures and ground-ice melting have become increasingly prominent. As a typical thermokarst landform， retrogressive thaw slumps （RTSs） not only profoundly affect the regional ecological environment， hydrological processes， and carbon cycle but also pose potential threats to engineering facilities and human safety. Although preliminary investigations have been conducted， in-depth research on long-term deformation monitoring and multi-factor driving mechanisms of typical RTSs in the Qilian Mountains remains necessary. This study focused on representative RTSs in the Kongkeli area of the Qilian Mountains. By integrating multi-source remote sensing data with field investigations， the spatiotemporal evolution characteristics and driving factors of RTSs in the Kongkeli area from 2017 to 2023 were systematically analyzed. First， Sentinel-2 optical images were processed using super-resolution techniques to a 2 m resolution， and the annual boundaries and areas of RTSs were extracted through manual visual interpretation. Second， the horizontal surface displacement field was obtained using the co-registration of optically sensed images and correlation （COSI-Corr） method. Simultaneously， the small baseline subset interferometric synthetic aperture radar （SBAS-InSAR） technique was applied to 200 Sentinel-1A radar images to obtain time-series cumulative deformation and annual average deformation rates. Furthermore， soil samples were collected from the K3 RTS and its surrounding areas to analyze the influence of local environmental factors， such as soil moisture content， porosity， organic matter content， and ground-ice content， on RTS development. Finally， continuous wavelet transform （CWT）， cross-wavelet transform （XWT）， and wavelet coherence （WTC） methods were employed to explore the periodicity and lag effects of surface deformation in relation to monthly average temperature and precipitation. SBAS-InSAR analysis revealed that the maximum cumulative deformation in the Kongkeli area reached -319 mm between 2017 and 2023， with an annual average deformation rate of approximately -43 mm⋅a-1. Spatially， intense deformation zones were primarily concentrated near the headwall， and their distribution highly coincided with the expansion boundaries identified through optical interpretation， validating the retrogressive erosion characteristics from a physical deformation perspective. Based on sampling data， the development of RTSs in the Kongkeli area was strongly controlled by local environmental factors. The soil in this area exhibited high moisture content （up to 67.45%）， high porosity （up to 0.74）， and high ground-ice content （0.34 m3⋅m-3）. These physicochemical properties， combined with gentle slopes of 3° to 10°， provided favorable conditions for the initiation and progression of RTSs. The results showed that the Kongkeli RTS group exhibited a significant expansion trend during the monitoring period， with the total area nearly doubling. The area of the largest RTS， K3， increased from approximately 2×104 m2 to over 4.4×104 m2， and the newly identified K6 RTS in 2021 was also in an active expansion stage. Regarding deformation characteristics， RTS evolution followed a three-stage pattern of “stability—intense change—deceleration，” with displacement dominated by the north-south （downslope） direction and maximum horizontal displacements concentrated between 3 m and 4 m. Additionally， wavelet analysis revealed a significant response pattern of surface deformation to climatic factors. The deformation time series showed strong coherence with monthly average temperature and precipitation on a 12-month annual cycle， with correlation coefficients exceeding 0.9. However， there was a lag of approximately 45 days in the response of surface deformation to climatic fluctuations， likely due to the time required for heat conduction into the active layer and soil moisture migration. In conclusion， the evolution of RTSs in the Kongkeli area of the Qilian Mountains results from the combined effects of permafrost degradation under climate warming， local topography， and soil physicochemical properties. This study reveals the complex spatiotemporal dynamic processes of RTSs， providing an important case study and scientific evidence for the identification， change analysis， and ecological protection regarding thermokarst hazards in the permafrost regions of the Qilian Mountains.
The unfrozen water content in frozen soils changes drastically in the near phase transition zone and is significantly affected by pore structure characteristics. Therefore， revealing the pore-scale occurrence and distribution patterns of unfrozen water in the near phase transition zone and establishing a reliable prediction model for unfrozen water content are of great engineering significance for frozen soil engineering， geological disaster prevention， and infrastructure construction in cold regions. In this study， a series of low-field nuclear magnetic resonance tests were conducted under multiple working conditions to obtain the temperature-dependent variation characteristics of unfrozen water content in fine sand in the near phase transition zone. The pore radius distribution and the variation characteristics of unfrozen water in different pore sizes were inverted and analyzed， and a CatBoost machine learning prediction model for unfrozen water content in the near phase transition zone was established. The results showed that： （1） the temperature-dependent variation patterns of unfrozen water content in fine sand in the near phase transition zone were significantly affected by the initial water content. Samples with high initial water content exhibited exponential and sharp changes in unfrozen water content with temperature variation， whereas samples with low initial water content showed linear and gradual variation characteristics. （2） Under low initial water content， the unfrozen water in fine sand in the near phase transition zone was mainly distributed in small pores. Under high initial water content， the unfrozen water in medium and large pores froze rapidly， resulting in a corresponding increase in equivalent pore radius， and its variation patterns were consistent with those of unfrozen water content. （3） The CatBoost-based prediction model for unfrozen water content in fine sand showed high accuracy within the ranges of initial water content， dry density， and temperature covered in this study. The coefficient of determination R 2 was 0.98. The model significantly outperforms the traditional exponential empirical model， especially under extreme water-content conditions.
Driven by climate warming， the thickening of the active layer and the thinning or even disappearance of permafrost have significant impacts on global ecosystems， climate systems， infrastructure， and human societies. Therefore， accurate monitoring of the current status of permafrost regions is of great importance. With the advancement of modern technology， the number of techniques for monitoring permafrost in permafrost regions has increased. Compared with traditional detection techniques， ground-penetrating radar （GPR）， with its advantages of being non-invasive， rapid， and high-resolution， has gradually become an effective tool for studying permafrost regions. However， existing reviews mostly focus on cold regions in general or cover various geophysical methods in permafrost research， lacking a systematic summary specifically addressing the research progress of GPR in permafrost regions. To address this gap， this study conducts a literature search using “GPR” and “permafrost” as keywords in Web of Science， and “探地雷达” （ground-penetrating radar） and “多年冻土” （permafrost） in the China National Knowledge Infrastructure （CNKI）， identifying 427 relevant publications from 1998 to 2024. Based on this analysis， this study introduces the basic detection principles of GPR and systematically reviews its application methods and research progress in estimating active layer thickness and soil moisture content， as well as in identifying ground ice in permafrost regions. In terms of active layer thickness detection， extensive studies have been conducted， and GPR technology is relatively mature. However， challenges remain in estimating the moisture content of the active layer， such as the current reliance on empirical models for inversion， which limits accuracy. Furthermore， the propagation of GPR signals is influenced by factors including soil conductivity and mineral composition， which complicates precise estimation. Regarding ground ice detection using GPR， the dielectric constant and electrical conductivity of frozen soil are affected by ice content， unfrozen water， salinity， and other factors. This leads to variable electromagnetic wave propagation characteristics， thereby affecting the effectiveness of GPR detection. Additionally， subsurface heterogeneities such as ice wedge networks and thaw slumps can cause multiple reflections， increasing the difficulty of data interpretation. Therefore， this review facilitates a deeper understanding of GPR applications in permafrost studies and provides guidance for future method optimization and research directions.
Freeze-thaw cycles can induce pore structure reorganization， weaken interparticle contacts， and promote internal damage accumulation in red clay， thereby reducing its strength and adversely affecting the stability of soil structures in seasonally frozen regions. To explore an environmentally friendly stabilization approach， red clay from the Yunnan Plateau was synergistically modified using locust bean gum （LBG） and rice straw fiber （RSF）. Unconfined compressive strength （UCS） tests， freeze-thaw cycle tests， X-ray diffraction （XRD）， and scanning electron microscopy （SEM） were conducted to investigate the mechanical response and microscopic mechanisms of LBG-RSF-modified red clay. The LBG and RSF contents ranged from 0% to 4% and 0% to 0.8%， respectively. The specimens were subjected to 0， 2， 4， 6， and 8 freeze-thaw cycles， with each cycle consisting of freezing at -10 ℃ for 12 h and thawing at 15 ℃ for 12 h. The results showed that both LBG and RSF improved the UCS of red clay， although their effects differed. For LBG-only treatment， UCS increased continuously with increasing LBG content， reaching 239 kPa at 4% LBG， which was 7.74 times that of untreated soil. For RSF-only treatment， UCS initially increased and then decreased with increasing RSF content， reaching a maximum value of 97.12 kPa at 0.6% RSF. The stress-strain curves indicated that LBG primarily increased the peak strength， whereas RSF mitigated post-peak stress reduction and improved deformation compatibility. The specimen modified with 4% LBG+0.6% RSF exhibited the best overall mechanical performance， with a UCS of 282.78 kPa， which was 9.17 times that of untreated soil and 18.3% higher than that of the specimen modified with 4% LBG alone. Freeze-thaw cycles reduced the UCS of all specimens， with the major strength loss occurring during the first two cycles. After eight freeze-thaw cycles， the UCS values of untreated soil， soil modified with 0.6% RSF， soil modified with 4% LBG， and soil modified with 4% LBG+0.6% RSF were 10.87， 28.33， 112.57， and 179.19 kPa， respectively. The UCS of the synergistically modified specimen remained 16.48 times that of untreated soil， and its compressive strength loss rate was 36.63%， which was 28.13 and 34.2 percentage points lower than those of untreated soil and soil modified with 0.6% RSF alone， respectively. Macroscopic failure observations showed that untreated specimens developed pronounced longitudinal cracking， surface erosion， edge spalling， and overall loosening after repeated freeze-thaw cycles， whereas the LBG-RSF-modified specimens maintained relatively intact profiles， with only localized spalling and shear cracks. XRD results indicated that LBG addition did not alter the basic mineral composition of red clay. SEM observations showed that LBG formed gel networks between soil particles and filled pores， thereby enhancing interparticle cementation， whereas RSF restrained crack propagation through bridging， anchoring， and reinforcement effects. Their combination formed a stable “gel-fiber-soil” composite skeleton， which improved microstructural integrity and resistance to freeze-thaw damage. These findings provide a reference for material selection and mix design in the green stabilization of red clay used in subgrades， slopes， and shallow foundation soil in the Yunnan Plateau and similar seasonally frozen regions.
Interannual variability （IAV） of soil temperature is a key indicator of the soil thermal regime and an essential parameter for understanding how the land surface responds to ongoing climate change. In high-latitude regions， permafrost plays a critical role in regulating land-atmosphere energy exchange and carbon storage. Therefore， accurately characterizing soil temperature IAV is fundamental for assessing ecosystem stability and climate feedbacks. However， large uncertainties remain among widely used reanalysis products， particularly in their ability to represent soil temperature and its IAV across different seasons， soil depths， and permafrost types. This study systematically evaluated the performance of six widely used reanalysis datasets （ERA5-Land， ERA5， GLDAS-Noah， CFSR/CFSv2， FLDAS， and MERRA-2） in reproducing observed soil temperature and its IAV at depths of 40 cm and 160 cm across the Eurasian high-latitude permafrost region. The evaluation was based on long-term station observations from 116 meteorological monitoring stations. The analysis focused on multiple aspects， including seasonal differences in biases， vertical structural characteristics， and differences between continuous and discontinuous permafrost regions， with the aim of identifying systematic biases and limitations in existing reanalysis products. The results indicated that all evaluated reanalysis products exhibited substantial uncertainties in representing soil temperature and its IAV， with the uncertainties being most pronounced during winter and spring. These periods corresponded to enhanced snow insulation effects and weakened land–atmosphere coupling， and these processes were not consistently and accurately represented across models. Spatially， the uncertainties were particularly pronounced in continuous permafrost regions， where strong ground-snow-atmosphere interactions and complex thermal regimes may increase model sensitivity. Most reanalysis products showed a systematic cold bias in soil temperature across both shallow and deep layers. This cold bias was generally accompanied by an overestimation of IAV， suggesting that the amplitude of interannual fluctuations was amplified in model representations to some extent. Among the six reanalysis datasets， ERA5， ERA5-Land， and MERRA-2 demonstrated relatively better performance in capturing observed IAV patterns， whereas CFSR/CFSv2 exhibited the largest deviations in both magnitude and variability. However， despite these relative differences， none of the individual reanalysis products could accurately reproduce the observed temporal evolution characteristics of IAV， including its long-term trends and stage-dependent transitions. To address these limitations， a multi-product merged soil temperature dataset was developed using a Taylor skill score-constrained weighting approach. This method assigned weights to individual products based on their ability to reproduce the observed variability structure， thereby maximizing consistency with station observations while reducing systematic errors. The resulting merged dataset integrated the complementary advantages of individual reanalysis products and provided a more robust representation of both mean soil temperature states and IAV across space， time， and depth. Validation results showed that the merged dataset significantly improved performance compared to all individual products. The Taylor skill scores were consistently higher across different permafrost types， soil depths， and seasons. In particular， improvements were most pronounced in the representation of IAV， with the root mean square error （RMSE） of IAV generally decreasing to within 1 ℃. These improvements indicated that multi-product fusion effectively reduced structural uncertainties inherent in individual reanalysis systems and enhanced the reliability of soil thermal variability estimates. Based on the merged dataset， a long-term analysis revealed that soil temperatures across the Eurasian high-latitude permafrost region exhibited a persistent warming trend over the past four decades. This warming signal was observed consistently across both shallow and deep soil layers， although the magnitude of warming showed clear spatial and vertical differences. In addition to the long-term warming trend， the IAV of soil temperature exhibited a clear stage transition in the early 1990s. After this transition， IAV in continuous permafrost regions showed a sustained increase， indicating enhanced interannual thermal fluctuations and reduced thermal stability. In contrast， discontinuous permafrost regions were characterized by more episodic and less coherent variation characteristics. Vertical structural analysis further revealed that IAV was generally more pronounced in deeper soil layers （160 cm） than in shallow layers （40 cm）， indicating that thermal anomalies can propagate downward and accumulate in the subsurface. These findings indicate emerging soil thermal instability in continuous permafrost regions under ongoing climate warming. The combined effects of the increasing IAV and persistent warming trends suggest that permafrost thermal regimes are becoming more dynamic and less stable over time. The merged dataset developed in this study provides data support for future studies investigating responses of the land surface to climate change over the permafrost region and their implications for regional and global climate feedbacks.
In the engineering construction of saline soil in cold regions， the coupling effect of sulfate salt crystallization expansion and pore water freezing expansion is a key factor inducing vertical frost jacking of pile foundations and threatening the long-term safety of structures. To reveal the mechanism by which the coupling effect of salt expansion and frost heave influences the vertical frost jacking characteristics of pile foundations in saline soil areas， this study comprehensively considered the combined effects of salt crystallization expansion and water freezing expansion under low-temperature environments. Based on the deformation compatibility and stress transfer mechanism at the pile-soil interface， a vertical mechanical calculation model for a single pile in sulfate saline soil under the combined jacking action of salt expansion and frost heave was established. Focusing on the evolution of ice crystallization， salt crystallization precipitation， and volume expansion during the cooling process of sulfate saline soil， this study selected key control parameters including ice crystallization rate， salt expansion ratio， frost heave potential， soil elastic modulus， and pile-soil interface contact characteristics， and analyzed the internal relationship between the coupling effect of salt expansion and frost heave and pile-soil interaction. Based on linear elasticity theory and pile-soil coupling deformation compatibility， the constitutive relationship and boundary constraints of the pile-soil interface were introduced. The governing differential equations for vertical displacement and axial force of the pile were established using the separation of variables method， and the analytical solution for the frost jacking displacement of the pile under the coupling effect of salt expansion and frost heave was derived， enabling quantitative description of pile deformation characteristics and internal force distribution patterns. To verify the rationality and accuracy of the theoretical model， the analytical results were compared with published laboratory model test data and field in-situ monitoring data from both domestic and international literature. The results showed that the variation trends of pile displacement and axial force calculated in this study were highly consistent with the test curves， and the numerical deviations were within allowable ranges， demonstrating that the proposed mechanical model and analytical method have good reliability and engineering applicability. This could provide theoretical support for the prediction， risk assessment， and early warning of pile frost jacking disasters in saline soil areas of cold regions. Based on the verified analytical model， a multi-factor parametric analysis was systematically carried out to comprehensively investigate the influences of key parameters such as pile length， soil moisture content， salt content， and porosity on the frost jacking characteristics of a single pile. The results showed that the frost jacking displacement of the pile decreased significantly with increasing pile length. Reasonably increasing pile length could effectively enhance the embedded anchoring effect of the pile foundation in the non-frost-heave stable layer， weaken the uplift effect of the expansion force in the upper frost-heave zone on the pile， and significantly improve the anti-frost-jacking stability of the pile foundation. An increase in soil moisture content strengthened the volume expansion effect of pore water freezing and raised the frost heave ratio of the soil. The expansion deformation was transferred to the pile through friction and extrusion forces at the pile-soil interface， resulting in a synchronous increase in the vertical displacement of the pile. An increase in salt content promoted the crystallization and volume expansion of sodium sulfate， aggravated the lateral extrusion and vertical uplift of the soil around the pile， significantly enhanced the interaction force at the pile-soil interface， and increased the axial force of the pile accordingly. The soil porosity was positively correlated with the axial force of the pile as a whole. An increase in porosity provided sufficient pore space for moisture migration， salt transport， and redistribution， further intensifying the uplift trend of the pile. When the porosity reached the critical threshold of 0.5， the compressibility of the soil skeleton was significantly enhanced， the internal expansion stress was obviously released and dissipated， and both the pile top displacement and pile axial force showed a sharp drop. The research findings can provide a scientific basis and technical reference for the anti-frost-jacking design and construction parameter optimization of pile foundation engineering in saline soil areas of cold regions.
Water-rich sandy slopes are characterized by strong groundwater recharge， high permeability， and susceptibility of soil shear strength to weakening due to changes in water content. Under such conditions， conventional grouting， anchoring， and anti-slide structures are susceptible to dilution， scouring， and local failure caused by groundwater flow， making it difficult to form a continuous and reliable reinforced body. Artificial ground freezing can create a temporary water-blocking and reinforcement structure through low-temperature freezing. However， existing studies have an insufficient understanding of the synergistic evolution of temperature field expansion， frozen wall closure， moisture migration， and frost heave deformation induced by different freezing pipe layouts under slope conditions. In particular， comparative validation of the quincunx-patterned double-row freezing pipe layout in water-rich sandy slopes remains lacking. To address this issue， this study took a water-rich sandy slope as the research object， and combined physical model tests with numerical simulations based on secondary development in FLAC3D. Two freezing pipe layout schemes， namely a single-row layout and a quincunx-patterned double-row layout， were designed to investigate the time-dependent responses of the slope temperature field， frozen wall morphology， moisture migration， and frost heave displacement under these two schemes. Based on similarity theory， the similarity relationships for the artificial freezing slope model tests were established， with a geometric similarity ratio of 1∶5 and a time similarity ratio of 1∶25. Additionally， a thermo-hydro-mechanical coupled numerical model considering latent heat of ice-water phase change， moisture migration， and frost heave effects was developed and validated using experimental monitoring data. The experimental and numerical results showed that the quincunx-patterned double-row freezing pipe layout significantly enhanced the overlapping cooling effects among adjacent freezing pipes. The minimum temperature in the freezing core reached -9 ℃， approximately 2 ℃ lower than that under the single-row layout. The average temperature of the frozen zone was approximately -3.5 ℃， and the time required to reach thermal equilibrium was shortened by approximately 2 d. Under the double-row layout， frozen wall closure occurred at 0.5 d， and a closed frozen wall network was formed at 2 d. The final frozen wall thickness reached approximately 500 mm， which was about 1.1 times that of the single-row condition， and the growth rate of frozen wall thickness increased by 39.7%. The frozen wall closure path exhibited a staged evolution from a V-shaped pattern to a W-shaped pattern and finally to a trapezoidal pattern. The numerical simulation results were generally consistent with the temperature monitoring results obtained from the physical model tests， with errors controlled within 3 ℃. Further analysis indicated that seepage vectors continuously pointed toward the freezing front during the freezing process， indicating synchronization between moisture migration and frozen wall expansion. Frost heave displacement increased most rapidly within 1~3 d of freezing， and the maximum displacement occurred mainly in the shallow frozen zone. The findings provide experimental and numerical support for optimizing freezing pipe layout， controlling freezing duration， and mitigating frost heave deformation in artificial ground freezing reinforcement of water-rich sandy slopes.
Numerous large open-pit coal mines and highway and railway projects under construction are distributed in the cold regions of western China， and a large number of exposed high and steep rock slopes exist along the lines. Such slopes generally develop cracks， joints， and other defects. Under the repeated action of diurnal temperature variations and seasonal freeze-thaw cycles， the frost heave and thaw shrinkage effect of rock mass is significant. In particular， under the dynamic disturbances such as blasting， traffic loads， and mechanical excavation， the internal stress field and freeze-thaw damage of rock mass superimpose on each other， which can easily induce geological disasters such as slope instability and rock collapse. To reveal the dynamic fracture characteristics and degradation mechanisms of rocks with different lithologies under freeze-thaw cycles， this study took red sandstone， black sandstone， and granite from Danba County， Sichuan Province as the research objects， and systematically conducted freeze-thaw cycle tests， ultrasonic tests， nuclear magnetic resonance tests， and drop-weight impact loading tests. （1） The saturated specimens were frozen at -20 ℃ for 12 hours and thawed in water at room temperature for 12 hours in a high-low temperature test chamber. Each stage consisted of 20 cycles， with a total of 100 cycles. （2） The longitudinal wave velocity and transverse wave velocity of the specimens under different numbers of freeze-thaw cycles were obtained by ultrasonic testing， from which the dynamic elastic modulus and dynamic Poisson’s ratio were then calculated. （3） The transverse relaxation time （T 2） spectrum distribution and pore size distribution characteristics of the specimens were obtained by nuclear magnetic resonance technology， and the damage evolution patterns of the rock meso-structure during the freeze-thaw process was quantitatively characterized. （4） Dynamic loading was applied to the single cleavage triangle （SCT） specimen using the drop-weight impact test device. A numerical model for calculating the stress intensity factor was established using the ABAQUS finite element software， and the dynamic fracture initiation toughness of the rock under different numbers of freeze-thaw cycles was solved. The results showed that： （1） the freeze-thaw resistance of rock was mainly controlled by its pore structure characteristics and mineral composition. Red sandstone was rich in calcite cement with strong water sensitivity， and had a high content of primary micropores， resulting in the most significant freeze-thaw damage. Black sandstone had relatively lower cement content and micropore proportion， and its damage degree was moderate. Granite was an igneous rock with crystalline connections among its minerals and a dense structure， exhibiting the strongest frost resistance. （2） With increasing number of freeze-thaw cycles， the porosity of the three rock types showed an upward trend， while the ultrasonic wave velocity decreased accordingly. After 100 freeze-thaw cycles， the porosity increase and wave velocity decrease of red sandstone were the largest， and its damage was the most severe. The damage of granite was the weakest， and that of black sandstone was between the two. （3） Freeze-thaw cycles had a significant deterioration effect on the dynamic fracture characteristics of the three rock types. Under the same number of freeze-thaw cycles， granite consistently exhibited the highest crack propagation velocity and dynamic fracture toughness， followed by black sandstone， while red sandstone had the lowest values. With increasing number of freeze-thaw cycles， the crack propagation velocity and dynamic fracture toughness of the three rock types showed a nonlinear decreasing trend and gradually tended to stabilize. From the perspective of combining meso-pore structure evolution with macro-fracture mechanical response， this study systematically revealed the dynamic fracture deterioration mechanism of rock types with different lithologies under freeze-thaw cycles， clarified the controlling role of pore structure characteristics and mineral composition on rock frost resistance， and established a quantitative relationship between dynamic fracture initiation toughness and porosity. The research findings can provide a theoretical basis for the long-term stability assessment and prevention of rock mass engineering in cold regions under dynamic loading such as blasting and earthquakes.
Freeze-thaw cycling is a major cause of continuous degradation of rock masses in geotechnical engineering projects such as railways， water conservancy facilities， and slopes in cold regions. The mesoscopic evolution of internal pores and micro-fractures in rocks serves as the key basis for revealing the mechanical mechanism of freeze-thaw damage and for quantitatively evaluating the long-term stability of rock masses. CT non-destructive scanning provides an effective means for the visual characterization of rock mesostructures. However， freeze-thaw action induces changes in mineral composition distribution， resulting in significant noise in CT images， highly non-uniform grayscale distributions， weak gradients at micro-fracture edges， and the multi-scale nested coexistence of pores and interconnected fractures. Traditional single-filtering methods tend to smooth fracture details， while global （or local） threshold segmentation methods are prone to over-segmentation， under-segmentation， and topological distortion of fractures. Furthermore， existing three-dimensional reconstruction algorithms cannot accurately describe mesoscopic damage evolution characteristics throughout the entire freeze-thaw process. To address these issues in CT three-dimensional reconstruction， this study used freeze-thaw red sandstone as the research object and constructed a refined CT identification and quantitative characterization system for three-dimensional damage in freeze-thaw rocks by integrating multi-scale feature enhancement and adaptive segmentation theory. First， a hybrid preprocessing algorithm combining non-local means （NLM） filtering and wavelet multi-scale decomposition was proposed. This algorithm adaptively suppressed imaging noise and artifacts while hierarchically enhancing edge features of fractures at different scales， thereby resolving the conflict between noise reduction and detail preservation. Second， the ALRG adaptive segmentation algorithm， which coupled local grayscale statistics with region growing， was developed. By integrating multiple criteria—including global histogram priors， dynamic local threshold correction， gradient consistency constraints， and spatial connectivity verification—this algorithm achieved accurate identification of the rock matrix， isolated pores， and micro-fractures. Finally， gradient-weighted vertex interpolation （GWVI） was introduced to improve the marching cubes （MC） three-dimensional reconstruction algorithm and correct the interpolation bias of isosurface vertices. This was combined with mesh simplification and feature-preserving smoothing to reduce staircase-like geometric artifacts in the three-dimensional models. The method was validated using freeze-thaw cycling tests and high-precision CT scanning experiments on red sandstone. Results showed that the porosity calculation error of the proposed method was less than 1.5%. The mean Dice similarity coefficient for pore and fracture segmentation reached 0.77±0.06， and the average Hausdorff distance was （2.37±0.31） μm. Segmentation accuracy and structural topological fidelity were significantly superior to those of traditional algorithms such as Otsu and Niblack. Based on high-precision three-dimensional models， the entire mesoscopic damage evolution process was quantitatively analyzed. The analysis revealed the evolution patterns of red sandstone under freeze-thaw cycles： initiation of stress concentration in small pores → coalescence and expansion of medium pores → interconnection of fractures forming a network. A quantitative correlation model linking fracture geometric parameters and pore coordination number with the freeze-thaw damage rate was established. It was clarified that the high concentration of ice-induced stress within micropores is the dominant mechanical mechanism controlling fracture initiation in freeze-thaw rocks. The integrated method of multi-scale image processing， adaptive segmentation， and high-precision three-dimensional reconstruction constructed in this study provides reliable digital characterization tools and theoretical support for the mesoscopic analysis of freeze-thaw damage mechanisms in rock masses in cold regions， correlative modeling of macro- and micro-mechanical properties， and long-term prevention and control of engineering frost damage.
To improve the prediction accuracy and efficiency of the compressive strength of fiber-reinforced slag concrete in cold regions， this study established a database containing 705 samples， including 118 self-designed laboratory test records and 587 records collected from published domestic and international literature. These records covered various fiber types and freeze-thaw service conditions with different numbers of freeze-thaw cycles and ambient temperatures. The input parameters included mix proportions， fiber performance characteristics， and typical cold-region environmental conditions， and the predicted compressive strength ranged from 13.35 MPa to 107.9 MPa. After continuous variables were standardized and one-hot encoding was applied to discrete fiber categories， the dataset was randomly split into a training set and a test set at a ratio of 7∶3 to avoid data leakage and ensure fair comparative analysis of each model under identical data conditions. Four mainstream machine learning regression models， including random forest （RF）， support vector regression （SVR）， extreme gradient boosting （XGBoost）， and categorical boosting （CatBoost）， were constructed for comparative analysis. Particle swarm optimization （PSO）， with a population size of 5 and a maximum number of iterations of 30， was combined with 10-fold cross-validation to optimize hyperparameters. This method effectively reduced the significant subjectivity induced by manual parameter tuning and fully exploited the inherent prediction potential of each model. The optimal hyperparameters obtained via PSO were as follows： PSO-CatBoost used 500 iterations， a tree depth of 6， and a learning rate of 0.003； PSO-XGBoost used a maximum tree depth of 6， a minimum child weight of 4.11， and a kernel scale of 1.06747； PSO-SVR used a penalty coefficient of 6.75 and an insensitive loss coefficient of 0.001； and PSO-RF used 119 decision trees and a minimum leaf size of 1. The performance of each optimized model on the test set was as follows： PSO-XGBoost achieved a coefficient of determination （R 2） of 0.9267， a mean absolute error （MAE） of 3.92 MPa， and a root mean square error （RMSE） of 5.31 MPa； PSO-RF achieved an R 2 of 0.9251， an MAE of 4.03 MPa， and an RMSE of 5.52 MPa； PSO-SVR achieved an R 2 of 0.8855， an MAE of 7.02 MPa， and an RMSE of 8.73 MPa； and PSO-CatBoost achieved an R 2 of 0.8593， an MAE of 7.56 MPa， and an RMSE of 9.91 MPa. Among all models， PSO-CatBoost achieved the best fitting performance on the training set， with a training R 2 of 0.9793， while PSO-XGBoost showed the best generalization ability for unseen engineering samples collected from actual construction projects. SHAP interpretability analysis was further used to reveal the internal prediction mechanism of the machine learning models and address the black-box limitation of data-driven algorithms. The analysis results showed that curing temperature， curing age， water-cement ratio， fiber length， fiber elastic modulus， cement content， and water content were the core factors determining the compressive strength of concrete. The findings provide reliable data support and systematic theoretical support for fiber-reinforced slag concrete widely used in cold-region engineering， enabling rapid strength prediction and mix proportion optimization design.
Icing on expressways in mountainous areas of Guizhou during winter frequently threatens traffic safety and operational efficiency. The rapid expansion of Guizhou’s expressway network in recent years has created an urgent demand for icing risk forecasting and early warning. However， a refined early warning model for icing risk on expressways under complex mountainous conditions is still lacking， and there is no graded warning technique that addresses different time periods. This study aims to establish an icing risk early warning model for expressways in mountainous areas of Guizhou by integrating multiple influencing factors， thereby providing a scientific basis for traffic management decisions. First， using low-temperature hazard grade data from the 2022 meteorological disaster risk survey in Guizhou， together with the longitude and latitude information of bridges and tunnels with lengths≥500 m， the road icing risk zoning developed in 2015 was revised to better adapt to climate change and reflect the influence of highway structural features. Second， six factors （air temperature， precipitation， icing risk zoning， traffic travel intensity， highway structural features， and driving time periods） were selected. A weighted analytic hierarchy process （WAHP） was adopted to construct a short-term （12 h） icing risk warning model. A sixth-order judgment matrix was built based on expert scoring， and the weight of each factor was calculated. Third， using hourly air temperature and pavement temperature data from the Baina Tunnel traffic meteorological station for the winters of 2018 and 2021—2023， a Logistic regression model was employed to establish the probabilistic relationship between ambient air temperature and pavement temperature≤0 ℃， yielding a nowcasting （0~3 h） icing probability warning model. Finally， the two models were integrated and applied. Two icing events of different types were selected for case validation using traffic control records， and model performance was evaluated using indicators such as spatial overlap ratio， hit rate， and false alarm rate. The results showed that in the revised road icing risk zoning， very high-risk and high-risk zones accounted for 31.44% of the total area of Guizhou Province and were concentrated in the high-altitude areas of central and western Guizhou. In the mountainous areas of Guizhou， a daily minimum air temperature>1 ℃ or a daily minimum ground temperature>1.5 ℃ could be regarded as the threshold for essentially no icing conditions. When the ambient air temperature was between -3.5 ℃ and -0.9 ℃， the icing probability entered a phase of rapid increase， with -2.2 ℃ being the sensitive peak point of probability variation. For the short-term warning model， all actual traffic control sections were fully covered by the model outputs classified as “higher risk” or above， achieving a spatial overlap ratio of 100%. For warning sections classified as “higher risk” and above with a total length exceeding 50 km， the hit rate reached 100%， and the false alarm rate was 39.02%. The WAHP-based short-term warning model could effectively predict the spatial extent and intensity levels of expressway icing， while the Logistic regression-based nowcasting model could accurately capture the initiation time and intensity evolution of icing events. The short-term model is suitable for proactive traffic control decisions， and the nowcasting model supports dynamic emergency scheduling. This technique fills the gap in refined icing risk warnings for the complex terrain of the Yunnan-Guizhou Plateau and provides an operational tool for winter traffic safety management. Its methodological framework can be extended to similar mountainous environments in southern China， offering important reference value for improving highway operational safety under adverse weather conditions.
Hydropower projects in Xizang are entering the development stage. However， engineering ice parameters required for these projects remain scarce. Ice formation is governed by local meteorological and hydrological conditions in winter， and determining engineering ice parameters with specific return periods requires long-term observational data. Consequently， even if the collection of meteorological， hydrological， and ice-condition data for hydropower projects in winter begins now， these data cannot directly serve as a basis for engineering ice parameters during the planning and preliminary design stages of hydropower projects. To address this issue， this study used historical meteorological data， along with high-frequency， continuous meteorological， hydrological， and ice measurements from the Pangduo Hydropower Station during a single winter， to develop a methodology for assessing the start date and end date of negative air temperature and ice thickness in winter across various return periods at the study site. First， daily average temperature records from 28 national meteorological stations in Xizang spanning from the 1950s to 2024 were collected， yielding three winter meteorological and phenological indicators： the start date of negative air temperature， the end date of negative air temperature， and negative accumulated temperature. Following Mann-Kendall tests and linear trend analysis on the data， probability distributions of these three indicators at the 28 stations for different return periods were derived by fitting P-III frequency curves. Subsequently， key meteorological parameters required for the planning and preliminary design of ice-resistant structures in hydropower stations—namely， the earliest start date of negative air temperature， the latest end date of negative air temperature， and the maximum negative accumulated temperature for 50- and 100-year return periods—were extracted. Trivariate Logistic regression was then employed to establish statistical relationships between these extreme values and geographical factors （elevation， latitude， and longitude） for different return periods. The resulting model enabled the interpolation of these values at arbitrary locations across Xizang. Finally， based on field measurements from the Pangduo hydropower station， a modified Stefan ice-thickness equation based on negative accumulated temperature was derived， incorporating ice-surface sublimation and the inhibitory effect of sub-ice water temperature. By integrating this modified equation with maximum negative accumulated temperature values for various return periods， corresponding ice thicknesses were calculated. This approach successfully assesses local ice thickness across different return periods by combining historical meteorological data with high-frequency， continuous in-situ measurements. It is straightforward and practical， effectively meeting requirements for planning and preliminary design stages of hydropower projects. However， for the detailed design stage of conventional hydropower projects or pumped-storage hydropower stations， supplementary investigations are recommended to account for localized microclimates and the impacts of frequent water pumping and release cycles.
Accurate runoff prediction is an essential scientific foundation for flood risk prevention and control， optimal allocation of water resources， and ecosystem protection， as its accuracy and reliability directly determine the effectiveness of water management decisions. However， runoff is influenced by multiple factors， including climate variability， human activities， and underlying surface conditions， exhibiting strong nonlinearity， non-stationarity， and complexity that pose significant challenges to accurate prediction. Existing runoff prediction methods generally rely on long-term， high-quality in situ hydrometeorological data. Yet， in many regions globally， particularly in developing countries， hydrological observation stations are sparse and data are severely scarce， making traditional approaches that depend on in situ data for modeling and validation difficult to implement. To systematically address the triple dilemma of “data scarcity， accuracy deficiency， and interpretability deficit”， this study developed an integrated prediction framework combining multi-source public datasets， a Bayesian long short-term memory Encoder-Decoder （BDL-LSTM-ED） model， and SHapley Additive exPlanations （SHAP）. The framework leveraged publicly available gridded products， including the China Meteorological Forcing Dataset version 2 （CMFD V2）， the fifth generation European Centre for Medium-Range Weather Forecasts reanalysis （ERA5）， and the Global Land Evaporation Amsterdam Model （GLEAM）， to enable hydrological modeling in regions with missing or insufficient ground-based observations. This study conducted systematic validation in the upper reaches of the Shule River Basin， a typical data-scarce alpine inland river basin on the northeastern margin of the Qinghai-Xizang Plateau. Testing phase results showed that model performance was highly consistent with that of the training phase， confirming applicability and robustness under data-scarce conditions. Among point prediction models， the LSTM-ED architecture exhibited superior predictive capability， achieving a Nash-Sutcliffe efficiency （NSE） of 0.923 and a root mean square error （RMSE） of 14.619 m3⋅s-1 for 1-day-ahead runoff prediction， substantially outperforming the baseline LSTM model （NSE=0.831， RMSE=21.581 m3⋅s-1）. This confirmed the feasibility of achieving high-accuracy runoff simulations in river basins lacking meteorological observations by relying solely on multi-source public datasets. Building on this foundation， the BDL-LSTM-ED model was constructed by implementing Bayesian inference via Monte Carlo dropout （100 stochastic forward passes）. This probabilistic framework maintained acceptable prediction accuracy （1-day NSE=0.906） while providing explicit uncertainty quantification through 95% prediction intervals constructed from the mean and standard deviation of 20 forward passes. SHAP-based interpretability analysis revealed the model’s ability to identify physically meaningful， temporally varying contributions of key hydroclimatic drivers across different prediction horizons. Overall， historical runoff emerged as the most influential predictor， followed by minimum temperature and soil moisture， which ranked higher than precipitation and evapotranspiration， reflecting the snowmelt-dominated runoff generation regime of the study basin. The temporal evolution of feature importance further revealed intrinsic patterns of hydrological response. For 1-day-ahead runoff prediction， importance was dominated by historical runoff and recent thermal conditions （minimum temperature and evapotranspiration）. For 2~3-day-ahead runoff prediction， the importance of soil moisture increased. For 4~7-day-ahead runoff prediction， precipitation became the dominant driver. This multi-scale driver evolution pattern demonstrated that the model learned feature-target relationships highly consistent with hydrological theory， significantly enhancing its physical credibility and interpretability. Limitations of this study include the observed under-coverage of prediction intervals during extreme flood events， suggesting that the current Bayesian framework may not fully propagate input data uncertainties or effectively distinguish between aleatoric and epistemic uncertainty. Future research directions include introducing heteroscedastic output structures or quantile regression approaches to explicitly decouple different types of uncertainty， constructing multi-source uncertainty quantification frameworks， embedding physical constraints to enhance extreme event simulation， and employing transfer learning or generative data augmentation techniques to improve model training in data-scarce regions. In conclusion， this study demonstrates that integrating public data products， Bayesian deep learning， and explainable artificial intelligence provides an effective， reliable， and interpretable runoff prediction tool for data-scarce river basins.
Zhuonai Lake， located in the Hoh Xil region of the Qinghai-Xizang Plateau， is a typical alpine closed lake. In recent decades， under the combined effects of climate warming， regional humidification， and extreme hydrological disturbances， Zhuonai Lake has experienced significant fluctuations in water extent and notable changes in its lakeshore ecology. In particular， the 2011 outburst event dramatically altered the hydrological processes of regional lakes， shoreline morphology， and surrounding vegetation patterns， making Zhuonai Lake a typical case for understanding the response mechanisms of alpine closed lakes to long-term climate change and abrupt disturbance events. Therefore， this study aims to systematically reveal the long-term evolution characteristics of the lake area， the response patterns of lakeshore vegetation， the major climatic driving factors， and the potential risk of future instability in Zhuonai Lake. To achieve these objectives， this study integrated multi-source remote sensing imagery， climate reanalysis data， and auxiliary snow and ice datasets to construct a long time-series analytical framework for 1994—2024. Landsat TM/ETM+/OLI and Sentinel images were employed to extract lake boundaries and monitor vegetation dynamics. To improve the accuracy of water body identification in the complex alpine environment， three water indices—NDWI， MNDWI， and AWEIsh—were jointly used， combined with the Otsu automatic thresholding method， to extract lake extent. Fractional vegetation cover （FVC） was retrieved using the NDVI-based pixel dichotomy model， and a 500 m buffer zone outward from the annual lake shoreline was defined as the lakeshore area for vegetation response analysis. In addition， ERA5-Land data were used to extract annual mean temperature， annual precipitation， and potential evapotranspiration， while glacier and snow cover data were incorporated to evaluate the contribution of cryospheric water supply. To ensure the robustness of the analytical results， Pearson， Spearman， and Kendall correlation analyses were further used to systematically examine the relationships among lake area， vegetation cover， precipitation， temperature， and snow/ice cover. The results showed that the evolution of Zhuonai Lake from 1994 to 2024 can be divided into five stages： a high-level equilibrium stage （1994—2001）， a sustained expansion stage （2002—2011）， an abrupt outburst-collapse stage （2011—2012）， a low-level decline stage （2012—2021）， and an initial recovery stage （2022—2024）. During the high-level equilibrium stage， the lake area remained generally stable， indicating a dynamic balance between inflow and evaporation. During the sustained expansion stage， the lake area increased continuously and reached 266.63 km2 before the 2011 outburst. The 2011 outburst event caused a sharp reduction of approximately 110.90 km2 in lake area， representing a decrease of approximately 41.6%， and rapidly shifted the lake from a closed storage-dominated state to a low-level post-disturbance state. In the subsequent decade， the lake remained at a persistently low level. Although a recovery trend has emerged since 2022， the recovery rate is substantially lower than the pre-outburst expansion rate. Linear fitting results showed that the annual expansion rate before the outburst was 1.1528 km2⋅a-1， whereas the recovery rate after 2022 was only 0.715 km2⋅a-1， indicating that the outburst event significantly weakened the hydrological recovery capacity of the lake. Lakeshore vegetation exhibited a strong coupled response to fluctuations in lake area. At the regional scale， vegetation conditions in most parts of the study area remained generally stable. Areas of vegetation improvement were mainly distributed along the newly formed shoreline after the outburst， whereas degraded areas were primarily concentrated on the exposed lakeshore beach zones following rapid water retreat. Within the 500 m lakeshore buffer zone， FVC showed a clear phased variation sequence： a gradual increase before the outburst， an anomalous peak in 2012， a prolonged low-level period afterward， and synchronous recovery since 2022. The anomalously high FVC value observed in 2012 was mainly associated with the exposure of wet lakeshore beaches and the resulting spectral interference， rather than indicating a significant improvement in ecological quality. As the lake area increased again after 2022， lakeshore vegetation also recovered synchronously. This indicates that lakeshore vegetation is highly sensitive to lake hydrological fluctuations and may have developed an eco-hydrological positive feedback mechanism during the recovery stage. Correlation analysis further demonstrated that lake area was significantly positively correlated with precipitation and FVC but negatively correlated with annual mean temperature. These results indicate that precipitation is a major factor driving lake expansion and recovery， whereas rising temperatures may inhibit lake recovery by enhancing evapotranspiration and exacerbating leakage associated with permafrost degradation. Snow/ice cover also contributed to lake recharge， but its effect exhibited obvious nonlinear characteristics. In particular， the finding that rank correlations were stronger than linear correlations suggests that the contribution of cryospheric meltwater may only become pronounced when snow and ice conditions reach specific thresholds. Therefore， the long-term variation of Zhuonai Lake is essentially the result of the combined effects of water supply from precipitation and snow/ice meltwater， and losses from evaporation， seepage， and post-outburst outflow. In conclusion， over the past three decades， Zhuonai Lake has experienced a complex hydro-ecological evolution process characterized by pre-outburst expansion， abrupt collapse， prolonged recession， and limited recent recovery. The 2011 outburst event fundamentally reshaped the lake’s hydrological regime， shoreline structure， and vegetation distribution， and Zhuonai Lake has not yet recovered to its pre-outburst state. Based on the recent recovery trend， Zhuonai Lake is likely to remain in a relatively low and fragile equilibrium state over the next 10~15 years. However， under continued regional warming and humidification， the risk of a secondary outburst triggered by extreme precipitation events should not be neglected. This study provides important scientific evidence for understanding the long-term evolution patterns of alpine closed lakes， clarifying the ecological response mechanisms of lakeshore vegetation， and supporting hydrological risk assessment， ecological conservation， and disaster early warning on the Qinghai-Xizang Plateau.
The black soil region of Northeast China is one of the country’s most important soybean-producing areas， where crop yield and quality are strongly constrained by soil nitrogen availability. This region is frequently affected by seasonal freeze-thaw cycles， which significantly alter the forms and concentrations of total nitrogen in soils. In this study， soil samples （0~60 cm） were collected from soybean fields in Harbin as the research object. A multi-factorial interactive simulation experiment was conducted to measure total nitrogen content under different soil depths， soil moisture contents， soil temperatures， and freeze-thaw frequencies. The experiment was designed to simulate changes in soil nitrogen under winter freeze-thaw conditions. The results revealed the following： （1） the freeze-thaw processes significantly affected the total nitrogen content of soybean field soils. After 1~7 freeze-thaw cycles， the total nitrogen content in each soil layer showed increases of 0.05 to 0.99 g⋅kg-1 compared with that before freezing and thawing. However， the surface soil required fewer freeze-thaw cycles to reach its peak than the deeper soils. （2） Under different soil temperature treatments， total nitrogen content increased as the freezing temperature decreased， reaching its maximum under the -30 ℃ freeze-thaw condition. （3） After low-temperature treatment， total nitrogen content in the 0~20 cm and 20~40 cm soil layers under 15% moisture content was higher than that under 10% moisture content. The opposite trend was observed in the 40~60 cm deep soil layer. This study demonstrates that freeze-thaw processes play a regulatory role in the total nitrogen content of soybean field soils in the black soil region， providing a scientific basis for optimizing nitrogen fertilizer management and reducing nitrogen loss. These findings contribute significantly to the sustainable development of agriculture in the region.
The Yulong Snow Mountain and its surrounding areas， as a typical monsoonal maritime glacier region， are important for studying glacier dynamics and the coupled carbon-water mechanisms of ecosystems. The terrestrial carbon cycle governs the formation of terrestrial ecosystem productivity and modulates the global dynamic energy balance， thereby reflecting the functional and structural characteristics of terrestrial ecosystems. Gross primary productivity （GPP） accounts for the largest proportion of global terrestrial carbon flux and serves as a key indicator for assessing the productivity of terrestrial ecosystems. Against the backdrop of global climate change， the frequency， intensity， and duration of droughts have increased， significantly impacting the vegetation growth of terrestrial ecosystems. Research has found that drought can continue to affect vegetation growth even after its cessation， referred to as the cumulative and lag effects of drought on vegetation growth. However， existing research primarily focuses on the drought response of individual land use types， with limited in-depth investigations into the cumulative and lag effects of diverse land use types on drought in the Yulong Snow Mountain and its surrounding areas. This study utilized scPDSI data （Climatic Research Unit， University of East Anglia） and MODIS GPP products （National Aeronautics and Space Administration， NASA） to investigate the spatiotemporal variations of GPP across different vegetation types in the Yulong Snow Mountain and its surrounding areas from 2001 to 2020， as well as their cumulative and lag responses to scPDSI. In addition， the Pearson correlation coefficient was employed to quantify the impact of drought on the lag and cumulative effects of vegetation GPP， as well as to clarify the corresponding time span of these effects. The study also investigated how different vegetation types adapted their GPP to drought conditions over time. The results showed that： （1） the average vegetation GPP in the Yulong Snow Mountain and its surrounding areas over the past two decades was 1 328.48 g C⋅m-2. The average GPP values for cropland， grassland， forest land， and shrubland were 1 202.67 g C⋅m-2， 1 129.84 g C⋅m-2， 1 460.30 g C⋅m-2， and 1 638.83 g C⋅m-2， respectively. From 2001 to 2020， GPP in the study area showed an increasing trend with interannual fluctuations. The rate of GPP increase varied across different land use types， with shrubland exhibiting the most significant increase， followed by forest land， grassland， and cropland. （2） In the study area， areas where scPDSI and GPP were positively correlated at different lag scales accounted for 44.94% of the total vegetation area， whereas the area with a negative correlation accounted for 55.06%. The lag effect of drought on the GPP of the four land use types was primarily long-term （9~12 months）. In the study area， Yulong Snow Mountain area contained only three land use types： cropland， grassland， and forest land. The average lag effect of drought in this area was 5.99 months. （3） In the study area， the area where cumulative scPDSI was negatively correlated with GPP accounted for 60.73% of the total vegetation area， while the area with a positive correlation accounted for 39.27%. The cumulative effect of drought on the GPP of the four land use types was mainly characterized by long-term accumulation （9~12 months）. Among these， shrubland， cropland， and grassland also exhibited large areas of short-term accumulation （1~4 months）. The average duration of the cumulative effect of drought in the Yulong Snow Mountain area was 6.30 months. These findings provide valuable insights into the ecosystem dynamics of the Yulong Snow Mountain area and highlight the nuances of GPP responses to climate variability， which can inform future conservation and management strategies. This study provides an important theoretical foundation for effectively guiding ecological restoration and construction efforts in typical low-latitude maritime glacial regions， contributing to vegetation management strategies amid the ongoing impacts of climate change.
Snow is an essential component of the cryosphere and significantly influences the response to global climate change and the stability of cold-region ecosystems. Accurate and timely monitoring of snow cover is therefore of great significance for understanding climate change， regional hydrological cycles， and ecological balance. The recent launch of China’s new-generation polar-orbiting satellite， Fengyun-3F （FY-3F）， provides a new opportunity for snow cover monitoring through multi-source data fusion. To meet the demand for daily-scale snow cover monitoring in the cold regions of the plateau， a passive microwave snow identification algorithm applicable to the Qinghai-Xizang Plateau was developed using brightness temperature data from the Microwave Radiation Imager （MWRI） onboard FY-3F， combined with snow cover products derived from the Medium Resolution Spectral Imager （MERSI） and land surface classification data. Passive microwave data from February to April 2024 were selected to systematically analyze the microwave radiation characteristics of typical land surface types. Using spatial clustering methods， four key indicators for different underlying surfaces—TB18V-TB36V， TB18V-TB18H， TB22V-TB89V， and TB89V—were extracted. Classification thresholds for various indicators were determined through statistical analysis， and a regionally adaptive microwave-based snow identification algorithm was established. The reliability and accuracy of the proposed algorithm were validated using ground-based meteorological station measurements and MODIS daily cloud-free snow cover products from December 2023 to February 2024. The results demonstrated that in cross-validation using MODIS daily cloud-free snow products， the algorithm achieved stable agreement indicators， with the Kappa and Matthews correlation coefficient （MCC） indicating consistent classification performance. The F1 value， representing the harmonic mean of precision and recall， ranged from 0.75 to 0.85， and the average ROC AUC and PR AUC values both exceeded 0.90， indicating strong discriminative ability and consistency. The validation accuracy based on meteorological station measurements was 77.9%， with a missed detection rate of 28.4% and a false alarm rate of 19.0%. Overall， the results confirm that the proposed algorithm exhibits high robustness and reliability under various evaluation indicators， providing an effective framework for daily-scale snow monitoring over the Qinghai-Xizang Plateau and offering valuable methodological support for hydrological modeling， water resource management， and climate change research in high-altitude regions.
Precise identification of moraine-dammed glacial lakes on the Qinghai-Xizang Plateau is essential for glacial hydrology research， glacial lake outburst flood risk assessment， and cryospheric hazard monitoring. The sub-meter optical imagery provided by China’s Gaofen （GF） satellite series has created favorable conditions for fine-scale mapping. However， achieving automated and high-precision identification under the complex terrain and moraine deposition background of the Qinghai-Xizang Plateau still faces three major challenges： （1） pronounced spectral confusion among mountain shadows， dark moraine materials， and dark water bodies， which leads to significant false detections when using traditional thresholding methods and general semantic segmentation models in complex backgrounds； （2） the continuous spectral-textural gradient formed by thin ice cover， wet snow along shorelines， and turbid shallow water， which easily causes boundary localization errors and geometric discontinuities of mapped objects； and （3） radiometric and textural differences between different sensors， such as GF-2 and GF-7， which weaken the model’s cross-domain generalization capability. To address these issues， this study proposed a CAU-Net deep learning model specifically designed for high-precision identification of moraine-dammed glacial lakes on the Qinghai-Xizang Plateau. Built upon the U-Net architecture， the model incorporated a coordinate attention mechanism into the encoding-decoding process to enhance spatial position awareness and contextual discrimination， and introduced deformable convolution to improve geometric adaptability to irregular shorelines and narrow water channels. A high-mountain glacierized region in the Himalayas was selected as the study area. Experimental results based on GF-2 imagery showed that the proposed method achieved a mean Intersection over Union （MIoU） of 95.50%， an F1-score of 97.70%， an overall accuracy （OA） of 98.54%， and a precision of 97.86% on the test set， outperforming mainstream models such as U-Net， DeepLabV3， and PSPNet. Further cross-sensor experiments indicated that， even without additional fine-tuning， the model still maintained good object integrity and boundary geometric consistency on GF-7 imagery， demonstrating strong cross-sensor generalization capability. The findings suggest that deep learning models structurally tailored for complex backgrounds and weak-boundary scenarios can provide an effective technical pathway for the high-precision identification of moraine-dammed glacial lakes supported by China’s Gaofen satellite imagery.
Conventional snow depth observation methods suffer from three major limitations that restrict their ability to meet the operational requirements of meteorological services. Manual snow depth measurements， although regarded as the benchmark method， are inherently discontinuous in time and sparse in spatial distribution due to high labor costs and environmental constraints. Ground-based sensors， including ultrasonic snow gauges and laser rangefinders， require frequent maintenance and calibration， yet still face fundamental difficulties such as uneven underlying surfaces， sensor drift， and the influence of snow particle size and surface moisture on measurement accuracy. Satellite-based retrievals， while providing broad spatial coverage， are constrained by coarse spatiotemporal resolution， with errors typically exceeding 10 cm， which is far beyond the acceptable tolerance for site-specific operational applications. To address these shortcomings， this study proposed an intelligent visual recognition method for snow depth observation based on a pinwheel-shaped convolution （PConv）-enhanced YOLO11 model. The method utilized the existing Tianlian intelligent observation platform， a camera-based system deployed by the Beijing Meteorological Service at 20 national meteorological stations. These cameras periodically captured high-resolution images of a fixed snow gauge installed at each station and provided the visual input for the entire pipeline. The method established a fully automated snow depth observation pipeline comprising four functional modules： snow cover detection， snow gauge detection， scale line detection， and snow depth calculation. The first three modules were deep learning models trained on independently constructed datasets derived from the same image sources but annotated differently according to their respective tasks. The fourth module calculated snow depth through a physically based formula using pixel-coordinate outputs from the scale line detector. The snow cover detection module performed image-level binary classification， distinguishing images with snow cover from those without. This module served as the gateway of the entire pipeline. If no snow was detected， the process terminated immediately to avoid invalid computation. If snow was present， the subsequent modules were activated. The snow gauge detection module employed object detection to localize the snow gauge within the complex background. By accurately identifying the bounding box of the snow gauge， this module extracted the region of interest， effectively eliminating interference from surrounding objects such as branches， fences， buildings， and other background clutter. The scale line detection module performed small-target detection to identify individual scale marks on the snow gauge. Given that each scale mark occupied only a few to several tens of pixels in the image， this represented a challenging task that demanded high sensitivity to fine-grained features. To address this， this study replaced the standard convolutional layers of YOLO11 with PConv modules， in which pinwheel-shaped convolution kernels were superimposed within 2×2 local neighborhoods to enlarge the local receptive field and enhance the model’s ability to capture features progressively from edge to center， thereby improving detection performance on small， elongated targets. A dual-assignment strategy was further applied to remove overlapping detection boxes， ensuring that each scale mark was uniquely detected. To train these modules， three datasets were constructed： a snow cover detection dataset （3 604 images with binary labels）， a snow gauge detection dataset （3 561 images with bounding box annotations）， and a scale line detection dataset （3 082 images with fine-grained bounding boxes for each individual scale mark）. Collectively， these datasets contained more than 10 000 images. All annotations were performed manually by trained personnel， with annotation accuracy exceeding 99%， thereby providing reliable ground truth for model training. The datasets covered more than 10 diverse scenarios， including sunny， cloudy， snowy， nighttime with supplementary lighting， backlighting， and tilted gauge conditions， ensuring model robustness under complex outdoor environments. Experimental results demonstrated that the three modules achieved precision values of 97.8%， 99.9%， and 99.8%， and recall values of 97.4%， 99.8%， and 99.8%， respectively. Ablation studies confirmed that under otherwise identical conditions， replacing standard convolutions with PConv significantly improved precision from 99.4% to 99.8% and recall from 99.2% to 99.8% for the scale line detection task. To further evaluate the applicability of the method in operational applications， three independent snowfall events in Beijing were selected for validation： a heavy snowstorm on December 13—14， 2023， a moderate-to-heavy snow event on December 10—11， 2023， and a localized snowstorm on January 17—18， 2026. Hourly comparisons between the model-retrieved snow depths and manual intensive observations yielded mean absolute errors of 1.060 cm， 0.867 cm， and 1.285 cm， with average correlation coefficients of 0.843， 0.943， and 0.830， respectively. The results demonstrated that the proposed method could stably track the complete evolution of snow cover from initial accumulation to continuous growth under different snowfall intensities， durations， and illumination conditions， with errors maintained within the centimeter level. The proposed method demonstrates several significant advantages. It achieves centimeter-level accuracy in ground snow depth measurement， provides continuous 24/7 temporal coverage at minimal operational cost， and maintains robust performance across diverse snow conditions and illumination variations. Compared with existing approaches， it simultaneously offers snow presence/absence discrimination and quantitative snow depth retrieval within a single lightweight framework based on YOLO11， which can be directly deployed on edge devices with end-to-end latency below 60 ms. The modular architecture ensures system maintainability and allows independent optimization of each component. Given that the method leverages the existing Tianlian camera infrastructure without requiring additional sensors or specialized underlying surface treatment， it substantially reduces hardware and maintenance costs while greatly improving the temporal continuity and spatial density of snow depth observations， making it a practical and reliable solution for operational deployment at meteorological stations. Overall， the intelligent visual recognition method proposed in this study provides a high-accuracy， low-cost， easy-to-deploy， and all-weather technical solution for automated operational snow depth observation， demonstrating practical value for large-scale deployment and application within meteorological observation networks. Meanwhile， the three high-quality annotated datasets constructed in this study provide data support for future related algorithmic research.
