2050 Blizzard Disaster Risk Forecast Dataset Development of Yili Region of China

CUI Jing¹  DAI Xiaoai^2,3*  LIU Yan^4,5*

1. College of Earth and Planetary Sciences, Chengdu University of Technology, Chengdu 610059, China;

2. College of Geography and Planning, Chengdu University of Technology, Chengdu 610059, China;

3. State Key Laboratory of Geohazard Prevention and Geoenvironment Protection, Chengdu University of Technology, Chengdu 610059, China;

4. Institute of Desert Meteorology, China Meteorological Administration, Urumqi 830002, China;

5. Field Scientific Experiment Base of Akdala Atmospheric Background, China Meteorological Administration, Urumqi 830002, China

Abstract: The Yili Region of Xinjiang, located in the inland region of Central Asia, experiences complex and variable climatic conditions. Blizzard disasters rank among the most prevalent natural hazards in this area, posing serious threats to regional ecology, agriculture, animal husbandry, and residents’ livelihoods. However, current assessments of systematic risks associated with blizzard disasters in this region are insufficient, and there is a lack of risk prediction data for future scenarios, limiting the refinement of disaster prevention and mitigation decisions. Using the Random Forest model, the authors integrated meteorological elements (including air temperature, snow cover, and wind speed) with topographic and geomorphological factors to simulate and evaluate annual blizzard disaster risks for the Yili Region during the historical period (2000–2020) and under the SSP2-4.5 scenario for 2050. Blizzard disaster risk is defined as the probability of significant socioeconomic losses caused by blizzard events within a specific geographic context. All risk values are normalized to the [0,1] range, with higher values indicating greater risk. The dataset is archived in .tif format with missing values represented as –9999. It features a spatial resolution of 500 m and comprises 22 data files totaling 78.5 MB (compressed into one package with 16.8 MB). This dataset supported the completion of the first author’s Master of Engineering thesis.

Keywords: blizzard disaster; Yili Region; Random Forest; risk assessment; future scenario; 2050; Master of Engineering thesis

DOI: https://doi.org/10.3974/geodp.2026.02.05

Dataset Availability Statement:

The dataset supporting this paper was published and is accessible through the Digital Journal of Global Change Data Repository at: https://doi.org/10.3974/geodb.2025.11.07.V1.

1 Introduction

Xinjiang Uygur Autonomous Region is one of the regions in China prone to frequent blizzard disasters, where snow disasters caused by blizzards result in varying degrees of losses almost every year^[1–3]. Such extreme weather events not only inflict direct economic losses but also exert profound impacts on regional socio-economic development, human activities, and the ecological environment^[4]. The Yili Region has become a key area for blizzard disaster research due to its distinctive geographical and climatic conditions. Nestled between the northern and southern branches of the Tianshan Mountains, the Yili River Valley features a humid continental climate that generates extremely heavy snowfall^[5]. Characterized by long winters and intense snowfall, this region has significantly greater snow depth and higher snowfall frequency than other parts of Xinjiang, making it a high‑risk zone for blizzard disasters and severely disrupting local production and residents’ daily lives^[2,5–7].

According to the precipitation intensity criteria issued by the National Meteorological Center, a 24-hour snowfall accumulation of ≥10 mm is classified as a blizzard^[8]. Blizzard disaster risk is generally defined as the probability that blizzard events will result in substantial socio‑economic losses (including casualties, housing damage, and declines in agricultural and livestock production) under specific geographical and environmental conditions. Existing research data indicate that a total of 81 blizzard events were recorded in the Yili Region between 2000 and 2020. These events not only resulted in casualties and housing damage but also severely impacted local agricultural and livestock production, establishing blizzards as a significant risk factor constraining regional ecology and economic development^[5]. Against the backdrop of global climate change, extreme weather events is likely to increase further in the future. To address this, there is an urgent need to construct a long-term, high-precision dataset for assessing blizzard disaster risks to support the improvement of regional risk prevention and control systems. However, current systematic risk assessment research on blizzard disasters in the Yili Region remains inadequate, with a particular scarcity of risk prediction data for future scenarios. This situation makes it challenging to meet the decision-making requirements for refined disaster prevention and mitigation.

A comprehensive assessment of blizzard disaster risk requires the consideration of multidimensional factors, including meteorology, topography, snow cover, and socio- economic conditions^[9,10]. The complexity of this assessment process imposes higher requirements on research methodologies and data integration. A review of relevant domestic and international research indicates that current methods for snow disaster risk assessment are shifting from traditional statistical analysis towards the integration of machine learning, and from single data sources towards multi-source data integration. Domestic studies have predominantly relied on traditional statistical methods such as Analytic Hierarchy Process and Logistic regression, with a focus on pastoral and plateau regions. These studies typically construct risk regionalization frameworks based on disaster-causing factors, disaster-prone environments, and disaster-bearing bodies^[11–15]. In contrast, international research places greater emphasis on methodological innovation and framework development. For instance, Yang, et al.^[16] actively explored machine learning algorithms, utilizing the XGBoost model to identify high-risk areas, while Lee, et al.^[17] established a DPSIR (Driving forces- Pressure-State-Impact-Response) assessment framework incorporating socio-economic factors. Xu, et al.^[18] applied Copula functions to improve the precision of hazard assessment. Furthermore, international studies have more extensively integrated multi-source remote sensing data, such as MODIS and GRACE^[19,20]. However, the statistical analyses or empirical models commonly adopted in traditional research often struggle to capture the complex nonlinear relationships among multiple factors. Machine learning algorithms, with their core advantages in handling complex data and uncovering potential correlations, can effectively overcome this limitation. As a classic ensemble learning algorithm in the field of machine learning, the Random Forest (RF) model offers high accuracy, resistance to overfitting, and strong interpretability. With the rapid development of artificial intelligence in recent years, it has been widely applied in various evaluation studies^[21].

This study integrates multi-source data and employs the RF model to quantitatively assess the historical blizzard disaster risk in the Yili Region. Furthermore, by incorporating future data under the SSP2-4.5 scenario, it predicts and generates blizzard disaster risk data for 2050. The 2050 projections focus on the Shared Socioeconomic Pathway 2—medium radiative forcing scenario (SSP2-4.5), which represents moderate socio-economic development and medium climate mitigation efforts. This scenario aligns relatively well with the current development trends in Yili^[22] and can provide a reasonable reference for short-to-medium term disaster prevention planning.

2 Metadata of the Dataset

The metadata of the Blizzard disaster risk assessment dataset in 2050 based on the simulation model from 2000 to 2020 in Yili Region, Xinjiang, China^[23] is summarized in Table 1. It includes the dataset full name, short name, authors, year of the dataset, temporal resolution, spatial resolution, data format, data size, data files, data publisher, and data sharing policy, etc.

Table 1  Metadata summary of the Blizzard disaster risk assessment dataset in 2050 based on the simulation model from 2000 to 2020 in Yili Region, Xinjiang, China

3 Methods

The development process of this dataset is shown in Figure 1. First, based on recorded blizzard events in Yili Region, historical blizzard disaster points were extracted. By integrating multi-source data—including meteorological, snow cover, topographic, socio-economic, and land use factors—12 key influencing factors were identified. Subsequently, training samples were generated based on the historical blizzard disaster points, and the RF machine learning algorithm was adopted for model training. After the model learned the complex relationships between the various factors and the disasters, it ultimately output the annual blizzard disaster risk data of Yili Region from 2000 to 2020, as well as the risk data for 2050 under the SSP2-4.5 scenario.

Figure 1  Flowchart of the dataset development

3.1 Data Preprocessing

3.1.1 Extraction of Blizzard Disaster Point Data

Compiling the snow disaster data for the year 2000 from the Encyclopedia of Meteo­rological Disasters in China—Xinjiang, along with the snow disaster data (2001–2020) from the Department of Civil Affairs, we obtained 81 blizzard events in the Yili Region from 2000 to 2020. Based on this data, disaster points were further selected. For the disaster clearly confined to a county seat, considering that the administrative centers of county seats were usually representative, with relatively fixed and easily determinable locations, the administrative center of the county was selected as the disaster point. For areas with a more ambiguous impact scope, the number of randomly selected points was determined based on factors such as the area size, terrain complexity, and land use type distribution, to ensure that these points could reasonably reflect the disaster distribution in the area. For example, in the northern mountainous areas or eastern pastoral areas of a certain county, due to the lack of clear location information for disaster points, combined with the characteristics of topography, geomorphology and land use types, 51 representative blizzard disaster points were ultimately extracted from the original data.

3.1.2 Extraction of Indicator Factors

First, based on the mechanisms of snow disaster formation and relevant studies^[25,26], combined with the topographic, meteorological, snow cover, and socio-economic characteristics of the Yili Region, a total of 12 disaster-impacting factors were selected. These factors include elevation, slope, aspect, annual average temperature, annual average wind speed, annual average snow depth, annual number of days with snow depth greater than 10 cm, population density, GDP per unit land area, livestock inventory at the beginning of the year per unit land area, and land use type. Second, the foundational data used for extracting these impact factors were collected and sorted out. The data information and sources are presented in Table 2. Subsequently, the factor data were calculated, and the process is as follows:

Annual average temperature and wind speed were calculated based on daily raster data. The daily temperature data for 2019–2020 were reconstructed via the ANUSPLIN interpolation algorithm^[27] utilizing observations from 15 meteorological stations in the Yili Region. According to the method proposed by the China Meteorological Administration, where snowfall is identified when daily precipitation occurs alongside temperatures below 3 °C^[28]. The annual number of blizzard days was extracted by integrating daily precipitation and temperature data. For the historical period, data on livestock inventory at the beginning of the year per unit land area were obtained from the Xinjiang Statistical Yearbook. To project these figures for 2050, the framework was constructed using a RF regression model that incorporated factors such as temperature, precipitation, population, and GDP^[29,30]. The historical annual average snow depth and the annual number of days with snow depth greater than 10 cm were calculated based on daily raster data. For future projections, surface snow water equivalent data from the ACCESS-ESM1-5 model’s r1i1p1f1 ensemble of the Coupled Model Intercomparison Project Phase 6 (CMIP6) were used^[31]. The data for the historical simulation period (1981–2014) and the projection period (2025–2058) were cropped into 34-year equal-length sequences corresponding to common and leap years, so as to reduce interannual fluctuations. Subsequently, using the long-term series of daily snow depth dataset in China as the reference data, bilinear interpolation was applied to resample the data to a consistent 0.25°×0.25°spatial resolution, maintaining spatial continuity and ensuring spatial matching for subsequent bias correction^[31]. Following this, localized calibration was performed by incorporating the density of snow in the Xinjiang region^[32], converting surface snow water equivalent to snow depth to reduce systematic bias. The future data were then corrected using historical deviations (CMIP6 historical simulation data minus reference data)^[31] to ensure the physical rationality and accuracy of the data. Furthermore, population density, GDP per unit land area, and land use type were treated as annual variables and utilized directly after preprocessing. Among these, the codes of land use type data from different sources were unified. In addition, slope and aspect were extracted from DEM data.

In view of the differences in spatiotemporal resolution and format among the multi-source data, all data have been co-registered to the Albers coordinate system, and all factors were ultimately resampled to a 500-m resolution and converted into the .tif raster format.

3.2 Construction Method of the Assessment Model

This study employs the RF algorithm to construct a blizzard disaster risk assessment model for the Yili Region. As an ensemble learning method, the RF generates multiple training subsets through Bootstrap sampling, randomly selects features for node splitting in each decision tree, and ultimately determines the classification results through a voting mechanism^[33]. This method can effectively handle the complex nonlinear relationships in meteorological and geographical data and has demonstrated good performance in small-scale disaster risk assessments^[21].

Table 2  Statistical table of basic data information and sources

In the specific implementation process of the model, key features were first extracted from the pre-processed blizzard-related meteorological and geographical data—such as snowfall, temperature, wind speed, and topographic information—through feature engineering. Subsequently, Bootstrap Sampling was employed to generate B training subsets D_b from the original dataset D, where each subset was obtained via random sampling with replacement, maintaining a sample size consistent with the original dataset. In this study, B was set to 200. The model was trained employing five-fold cross-validation to optimize parameters^[16,34], with 70% of the samples allocated for training and 30% for validation. For each training subset D_b, a decision tree T_b was constructed. During the splitting of each node in the tree,  features were randomly selected from all P features to constitute a candidate feature subset, and the optimal split point was chosen based on the principle of minimizing Gini impurity to reduce the risk of overfitting. For a sample x to be predicted, each decision tree generates a prediction probability p_b (x), and the average probability of blizzard disaster occurrence is finally obtained by synthesizing the prediction results of all decision trees through a weighted voting mechanism:

                                                                                              (1)

For high-altitude areas above 3,000 m (which are sparsely populated regions in both historical and future periods, with population density below 1 person/km²), a risk down-weighting process was implemented based on population distribution characteristics. The corrected risk value is defined as:

                                                                    (2)

Where H(x) is the elevation of the sample point, (x)∈[_min,1] is the down-weighting coefficient based on normalized population density, and _min is the preset minimum weight threshold.

During the model validation phase, the Receiver Operating Characteristic Curve (ROC) and the Area Under the Curve (AUC) were used to evaluate the model’s discriminative ability^[35]. After the accuracy met the requirements, blizzard disaster risk assessment data for 2000–2020 were generated based on historical observation data. Furthermore, by integrating future data under the SSP2-4.5 scenario, blizzard disaster risk data for 2050 were predicted and generated. Finally, all output results were normalized to the [0,1] interval.

4 Data Results and Validation

4.1 Dataset Composition

This dataset, archived in a folder named “Blizzard_Risk_Yili2000-2020&2050”, contains the annual blizzard disaster risk data for the historical period (2000–2020) and the under the SSP2-4.5 scenario in 2050. The annual data files for the historical period (2000–2020) are named “BlizRisk_yyyy.tif” (where yyyy represents the specific year), and the future scenario data file for 2050 is named “BlizRisk_2050_SSP2-4.5.tif”.

All data feature a spatial resolution of 500 m and are in .tif format, comprising 22 files with a total data volume of 78.5 MB (16.8 MB when compressed). The data values represent a normalized risk index ranging from 0 to 1, where higher values indicate a greater risk of blizzard disasters.

4.2 Data Results Analysis

As shown in Figure 2, influenced by topography and geomorphology, the blizzard disaster risk in the Yili Region exhibited significant spatial differentiation from 2000 to 2020. Specifically, the risk was relatively low in mountainous areas such as the Narat Mountains in the south, the Keguqin Mountains in the north, and the Wusun Mountains in the central part. In contrast, the risk in the north and south river valley areas was significantly higher than that in the mountainous areas due to dense population and easy snow accumulation. The spatial distribution of risk was relatively stable over the 21 years, with certain fluctuations in local areas. For example, the risk in the southern river valley areas gradually decreased after 2013.

As shown in Figures 3 and 4, under the SSP2-4.5 scenario, the overall blizzard disaster risk in the Yili Region is projected to increase by 2050. The risk increase is more significant in the southern valley areas, as well as in the northwestern and central mountainous regions. Compared with the average risk level during 2000–2020, the areas with increased risk primarily rose from the interval of 0.20–0.30 to the 0.36–0.63. The area with risk values above 0.63 changed little, with only a slight increase in risk observed in small parts of the northern and southern river valley areas (Figure 5).

4.3 Data Validation for 2000–2020

The underlying assumption for validating this dataset is that if the RF model can accurately

Figure 2  Annual spatial distribution maps of blizzard disaster risk in the Yili Region (2000–2020)

Figure 3  Spatial distribution map of blizzard disaster risk in the Yili Region under the
SSP2-4.5 scenario for 2050

Figure 4  Spatial distribution map of the normalized difference in blizzard disaster risk between 2050 and the historical average in the Yili Region

simulate historical blizzard disaster risks, it can also reliably predict future blizzard disaster risks. As illustrated in Figure 6, the model validation results indicate that the RF model exhibits favorable predictive performance, with an average AUC value of 0.760,1± 0.108,8.

Compared with existing research on blizzard disasters in the Yili Region^[9], the spatial risk patterns in this dataset show high consistency with those studies, indic­ating strong agreement in the spatial characteristics of risk distribution. Tem­porally, the annual blizzard disaster risk results were standardized using Z-scores to measure their deviation from the multi-year mean in terms of standard deviations. A threshold of Z-score >1.5 was used to identify high-risk abnormal areas; this threshold corresponds to a one-tailed probability of approximately 6.68% under a standard normal distribution and can effectively differentiate between regular fluctuations and high-risk states^[36,37]. Combined with the condition of a risk value >0.5, abnormal areas that significantly deviate from the mean and have a relatively high-risk level were identified. By sorting these abnormal areas in descending order based on their proportion, it was ultimately determined that the years 2005, 2001, 2010, 2000, 2003, 2012, and 2014 ranked among the top seven for the proportion of abnormal areas during the historical period of 2000–2020. Their proportions fell within the 1.8% to 6.0% range, significantly exceeding the average risk level of the historical period, identifying them as high-risk abnormal years. Combined with the blizzard disaster loss data for the Yili Region from 2000 to 2020 (including population affected, fatalities, house collapses, crop damage, and large livestock deaths), a disaster loss index was calculated to analyze the interannual distribution of the occurrence frequency of snow disaster intensity levels^[5]. The years identified as having relatively severe disasters were 2003, 2005, 2006, 2010, 2011, 2012, and 2014, yielding a precision of 71.4% and a recall of 71.4%. Among them, although the risk value in 2010 was not the highest, the simulated high-risk areas highly overlapped with the areas of high population and economic exposure, resulting in the most severe actual disaster losses. It can be concluded that the high-risk years identified by the model are highly consistent with the actual disaster loss index, which verifies the effectiveness of the model in the temporal dimension.

5 Discussion and Conclusion

Based on the RF model, this study integrated multi-source data, including meteorological, topographic, snow cover, and socio-economic factors. It first simulated the historical blizzard disaster risk in the Yili Region for the period 2000–2020, and subsequently predicted and generated the blizzard disaster risk for this region under the SSP2-4.5 scenario for 2050. From 2000 to 2020, the blizzard disaster risk in the Yili Region exhibited a spatial distribution pattern of “high risk in river valley areas and low risk in the northern and southern mountainous areas”. Under the SSP2-4.5 scenario for 2050, the overall blizzard disaster risk in the Yili Region shows an increasing trend, with a more significant risk increase in the central and northwestern mountainous areas, as well as the southern valley. Model validation demonstrates good predictive performance (AUC=0.760,1±0.108,8), and the high-risk years identified by the model are highly consistent with actual disaster records, demonstrating the strong applicability and reliability of the model. Based on the above results, the following recommendations for disaster prevention and mitigation are proposed: First, for river valley areas with persistently high risks, the existing disaster prevention and mitigation systems should be maintained, and monitoring, material reserves, and emergency response mechanisms should be consolidated; Second, for areas with concentrated agriculture and animal husbandry and increasing risks, such as the northern parts of Zhaosu County and Tekes County, the layout of disaster prevention facilities and the reserve of disaster-resistant materials should be optimized to enhance disaster resilience capacity.

This dataset continues and refines the overall distribution pattern of “high risk in the central valley and low risk in the northern and southern mountainous areas” revealed by our team’s previous research[1], further validating the spatial differentiation characteristics of blizzard disaster risk in the Yili Region. Previous studies, designed to meet the disaster prevention deployment needs of the Scientific Expedition Office, employed geographic data and comprehensive evaluation methods based on county-level administrative units. The assessment results presented the characteristics of risk agglomeration at the county level, facilitating direct integration with regional disaster prevention work. On this basis, the present study incorporates historical disaster point data and leverages the powerful nonlinear relationship capture capability of the RF algorithm to achieve a more refined spatial expression of risk distribution, reflecting our team’s continuous optimization of blizzard disaster risk assessment methodologies.

Utilizing the RF model, this study builds upon the strengths of integrating multi-source data and capturing complex nonlinear relationships, incorporates future climate scenarios and population down-weighting processing, thereby enhancing the regional applicability and forward-looking nature of the assessment. Nevertheless, certain uncertainties exist in the future projections: On one hand, due to the limitation of using a single CMIP6 model, the simulation biases inherent in the model itself may propagate to the assessment results. The subsequent work needs to further reduce uncertainty by integrating multi-model ensembles. On the other hand, there still remains room for improvement in terms of model parameter optimization and the setting of multiple future scenarios. This dataset supported the completion of the first author’s Master of Engineering thesis.

Author Contributions

Cui, J. agreed with and adopted the overall design scheme for dataset development proposed by Dai, X. A. and Liu, Y., collected and processed the data, and wrote the data paper; Dai, X. A. and Liu, Y. reviewed the data and revised the data paper.

Conflicts of Interest

The authors declare no conflicts of interest.

References

[1]        Wang, Q. X., Liu, W. P., Li, Y. Y., et al. Different regional animal husbandry in Xinjiang snow disaster losses time-frequency characteristics [J]. Journal of Glaciology and Geocryology, 2015, 37(4): 905–915.

[2]        Wei, J. J., Li, N., Wan, Y., et al. Mechanisms and causes of typical snowstorms in Urumqi [J]. Arid Land Geography, 2022, 45(5): 1381–1391.

[3]        Yang, Q., Qin, L., Gao, P., et al. Prediction of maximum snow depth based on RBF network in severe snow disaster area of Xinjiang [J]. Desert and Oasis Meteorology, 2024, 18(1): 89–95.

[4]        He, W. X., Hao, X. H., Liu, F. G., et al. A multi-source information fusion model construction and risk grading study of snowstorm risk assessment on roads in Ili Region [J]. Remote Sensing Technology and Application, 2025, 40(1): 177–191.

[5]        Huo, H., Liu, Y., Maiwulaxia, M. Spatiotemporal distribution characteristics and impact assessment of snow disasters in the Ili Region of Xinjiang from 1990 to 2020 [J]. Arid Land Geography, 2024, 47(11): 1828–1840.

[6]        Wang, X., Chu, C. J., Mou, H. Spatial pattern and interannual variation characteristics of snow disaster in Xinjiang [J]. Arid Zone Research, 2020, 37(6): 1488–1495.

[7]        Wang, X. Q., Lu, X. Y., Ma, Y., et al. Study on snow disaster assessment method and snow disaster regionalization in Xinjiang [J]. Journal of Glaciology and Geocryology, 2019, 41(4): 836–844.

[8]        General Administration of Quality Supervision, Inspection and Quarantine of P. R. China, National Standardization Administration. Grade of precipitation (GB/T 28592—2012) [S]. Beijing: Standards Press of China, 2012.

[9]        Huang, Y. Y, Dai, X. A., Liu, Y., et al. Research on blizzard disaster risk zoning in Ili Region, Xinjiang [J]. Meteorological, Hydrological and Marine Instruments, 2025, 42(2): 104–107.

[10]     Ma, X. F., Huang, X. D., Deng, J., et al. Comprehensive risk assessment of snow disasters in Qinghai Province [J]. Acta Prataculturae Sinica, 2017, 26(2): 10–20.

[11]     Ding, Y. L., Sun, X. R., Gao, Y., et al. Assessment and division of forest snow disaster risk in Inner Mongolia [J]. Arid Zone Research, 2019, 36(2): 486–493.

[12]     Delegerima, Li, Y. P., Meng, X. F., et al. Study of the risk evaluation of snow disaster in pastoral areas of Xilingol League, Inner Mongolia [J]. Journal of Glaciology and Geocryology, 2020, 42(4): 1353–1362.

[13]     Chen, H. J., Yang, J. P., Ding, Y. J., et al. Quantitative assessment of snow risk about livestock in the Qinghai-Xizang Plateau [J]. Journal of Catastrophology, 2022, 37(2): 102–110.

[14]     Li, F., Hou, G. L., E, C. Y., et al. Township unit-based risk assessment of snowstorm hazard in Guoluo Prefecture of Qinghai Plateau [J]. Journal of Natural Disasters, 2014, 23(6): 141–148.

[15]     Miao, A. M., Wang, H. X., Lu, Z. Y.. Risk regionalization of blizzard disaster in Shanxi Province based on GIS [J]. Chinese Agricultural Science Bulletin, 2016, 32(20): 133–140.

[16]     Yang, Z. Y., Gokon, H., Yu, Q. Machine learning-based identification and assessment of snow disaster risks using multi-source data: insights from Fukui Prefecture, Japan [J]. Progress in Disaster Science, 2025, 26: 0100426.

[17]     Lee, H., Kim, D., Chung, G. Classification of risk levels for snow damage estimation considering socioeconomic factors in South Korea [J]. Applied Water Science, 2024, 14(11): 1–17.

[18]     Xu, Q. X., Huang, F. Y., Mou, S. H., et al. Snow disaster hazard assessment on the Tibetan Plateau based on copula function [J]. Sustainability, 2023, 15(13):10639.

[19]     Singh, D., Zhu, Y., Liu, S. Y., et al. Exploring the links between variations in snow cover area and climatic variables in a Himalayan catchment using earth observations and CMIP6 climate change scenarios [J]. Journal of Hydrology, 2022, 608: 127648.

[20]     Sahu, R., Ramsankaran, R., Bhambri, R., et al. Evolution of supraglacial lakes from 1990 to 2020 in the Himalaya-Karakoram Region using cloud-based Google Earth Engine platform [J]. Journal of the Indian Society of Remote Sensing, 2023, 51(12): 2379–2390.

[21]     Wang, Q. L., Ma, X. X., Liu, X. X., et al. Mountain torrent disaster risk assessment method and application based on Random Forest [J]. Yellow River, 2022, 44(4): 63–66,73.

[22]     Bureau of Statistics of Ili Kazak Autonomous Prefecture. Statistical Yearbook of Ili Kazak Autonomous Prefecture on the 2024 National Economic and Social Development [EB/OL]. (2025-04-09). https://www. xjyl.gov.cn/xjylz/c112816/202504/67c6814847cb40a0a33773fb1df5466b.shtml.

[23]     Cui, J., Dai, X. A., Liu, Y. Blizzard disaster risk assessment dataset in 2050 based on the simulation model from 2000 to 2020 in Yili Region, Xinjiang, China [J/DB/OL]. Digital Journal of Global Change Data Repository, 2025. 0https://doi.org/10.3974/geodb.2025.11.07.V1.

[24]     GCdataPR Editorial Office. GCdataPR data sharing policy [OL]. https://doi.org/10.3974/dp.policy.2014.05 (Updated 2017).

[25]     Yu, D., Du, H. R., Guo, J., et al. Research of snow disaster risk assessment in the Three-River-Source Area based on multi-source data [J]. Advances in Meteorological Science and Technology, 2023, 13(4): 34–40.

[26]     Zhuang, X. C., Zhou, H. K., Wang, L., et al. Evaluation and cause study on the snow disasters in Pastoral Areas of Northern Xinjiang [J]. Arid Zone Research, 2015, 32(5): 1000–1006.

[27]     Qian, Y. L., Lv, H. Q., Zhang, Y. H. Application and evaluation of daily meteorological element interpolation method based on ANUSPLIN software [J]. Journal of Meteorology and Environment, 2010, 26(2): 7–15.

[28]     Liu, D., Zhang, L. J., Jiang, S. Y., et al. Hazard prediction and risk regionalization of snowstorms in Northeast China [J]. Physics and Chemistry of the Earth, Parts A/B/C, 2020, 116: 102832.

[29]     Zhou, J., Niu, J., Wu, N., et al. Annual high-resolution grazing-intensity maps on the Qinghai-Tibet Plateau from 1990 to 2020 [J]. Earth System Science Data Discussions, 2023, 16(11): 5171–5189.

[30]     Zhang, X. Z., Li, M., Wu, J. S., et al. Alpine grassland aboveground biomass and theoretical livestock carrying capacity on the Tibetan Plateau [J]. Journal of Resources and Ecology, 2022, 13(1): 129–141.

[31]     Chen, T., Gao, G., Du, X. H., et al. CMIP6 projected changes in snow cover in the Qinghai-Tibet Plateau [J]. Scientia Geographica Sinica, 2024, 44(5): 901–910.

[32]     Wang, H. D., Zhang, X. L., Xiao, P. F., et al. Towards large-scale daily snow density mapping with spatiotemporally aware model and multi-source data [J]. The Cryosphere, 2023, 17(1): 33–50.

[33]     Yang, S., Li, D. Y., Yan, L. X., et al. Landslide susceptibility assessment in high and steep bank slopes along Wujiang River based on Random Forest model [J]. Safety and Environmental Engineering, 2021, 28(4): 131–138.

[34]     Huang, C. Y., She, D. X., Liu, X. B., et al. Predicting future meteorological drought risk in mainland China using random forest model [J]. Journal of Hydrology: Regional Studies, 2025, 61: 102633.

[35]     Liang, Z., Wang, C. M., Duan, Z. J., et al. A hybrid model consisting of supervised and unsupervised learning for landslide susceptibility mapping [J]. Remote Sensing, 2021, 13(8): 1464.

[36]     Ruijsch, D., van Mourik, J., Biemans, H., et al. Thrive or wither: exploring the impacts of multiyear droughts on vegetation [J]. Journal of Geophysical Research: Biogeosciences, 2025, 130(7): e2025JG008992.

[37]     Wang, X. X., Wu, Y. H., Luo, M. L., et al. Exports of organic matter, phosphorus and nitrogen from Sichuan Basin: a critical region regulating water quality of the Upper Yangtze River, China [J]. Journal of Hydrology, 2025, 662: 133889.