Preamble

ARID LAND GEOGRAPHY
Vol. 48 No. 8
August 2025

Predicting and Analyzing Glaciers on the Qinghai-Xizang Plateau Using a Random Forest Model
ZHANG Yiming¹,², TANG Yulei³, FENG Junbo¹
¹ Civil-Military Integration Center, China Geological Survey, Chengdu 610036, Sichuan, China
² College of Earth and Planetary Sciences, Chengdu University of Technology, Chengdu 610059, Sichuan, China
³ Center for Geophysical Survey, China Geological Survey, Langfang 065000, Hebei, China

Abstract: Glaciers represent a key focus of natural resource monitoring on the Qinghai-Xizang Plateau, and their investigation holds significant implications for understanding climate change in the region. This study integrates multi-source data to construct a random forest model (R² = 0.72), generating an annual 1 km-resolution glacier prediction dataset for the plateau from 2000 to 2020. The analysis reveals: (1) Spatial distribution patterns: Glaciers predominantly occur on slopes of 0°–40° (accounting for 97.92% of total area) and at elevations of 4000–7000 m (comprising 99.38% of total area), with greater coverage on northern versus southern slopes and western versus eastern slopes. (2) Spatiotemporal change characteristics: Between 2000 and 2020, glaciers exhibited a significant overall retreat trend. Spatially, pronounced changes concentrate along the plateau margins, gradually weakening toward the interior where minor variations dominate. (3) Regional variations: Glaciers in the Himalaya and Nyainqentanglha mountains showed significant retreat, those in the Karakoram Mountains displayed slight retreat, while the Kunlun Mountains exhibited concurrent patterns of slight advancement and retreat.

Keywords: glaciers; prediction and analysis; random forest model; spatial distribution; spatiotemporal changes; Qinghai-Xizang Plateau

Introduction

Glaciers, often termed "mountain solid reservoirs," constitute vital components of water resources. The Qinghai-Xizang Plateau ranks among the most climate-sensitive regions globally, with its towering topography driving climatic patterns across eastern China and exerting significant influence on hemispheric and global climate systems. Research demonstrates that climate change substantially impacts plateau glaciers, which in turn affect atmospheric water vapor and radiation cycles, altering surface runoff, lake levels, groundwater, and sea levels across the Northern Hemisphere with profound consequences for human societies. Consequently, dynamic glacier monitoring is essential.

Since the 21st century, numerous scholars have investigated spatiotemporal glacier change patterns across the Qinghai-Xizang Plateau. Current research methodologies include numerical modeling, geodetic surveys, satellite remote sensing, and field observations. Numerical models simulate hydrological processes and physical mechanisms but typically yield only regional totals without detailed distribution or boundary information, limiting their generalizability. Geodetic methods, such as digital elevation model differencing, measure surface elevation or velocity changes. Satellite remote sensing employs imagery and GIS techniques for glacier mapping through visual interpretation or automated classification. While visual interpretation offers high accuracy, it remains labor-intensive and susceptible to subjective misclassification. Classical computer-assisted methods—including ratio thresholding, snow cover indexing, and supervised/unsupervised classification—demonstrate effectiveness but lack universal applicability. Recent innovations encompass multi-scale image segmentation, neural networks, deep learning, and object-oriented interpretation. However, these techniques lack ground validation, while sparse observation stations constrain field-based approaches.

This study proposes a novel statistical modeling approach to predict glacier distributions. Given that glacier evolution responds to integrated effects of climate change, geographic setting, and human activity, establishing probabilistic relationships between environmental variables and glacier coverage enables prediction using remote sensing, GIS, and machine learning. This approach expands monitoring coverage and frequency while leveraging long-term equilibrium between glaciers and environmental factors without requiring explicit dynamic coupling mechanisms.

1.1 Study Area Overview

The Qinghai-Xizang Plateau, Asia's interior highland, represents Earth's highest and China's largest plateau, spanning over 2.5 million km² and known as the "Roof of the World." It constitutes the most extensive, highest-elevation, and lowest-temperature mid-low latitude mountain glacier region globally. Major glacierized ranges include the Kunlun, Karakoram, Himalaya, Nyainqentanglha, and Hengduan mountains, among others [FIGURE:1]. Plateau-wide glaciers total 49,000 km² in area, 4,500 km³ in volume, and 20,000 individual glaciers. Influenced by westerlies and South Asian monsoons, these glaciers comprise maritime, subcontinental, and extreme continental types, playing crucial roles in national water resources and ice-water cycling.

1.2.1 Data Sources

This study integrates multiple datasets [TABLE:1]: glacier distribution data from the National Tibetan Plateau Data Center; annual normalized difference vegetation index (NDVI) data from Terra and Aqua MODIS products; water yield modulus from the Chinese Academy of Sciences Resource and Environmental Science Data Center; temperature and precipitation gridded data (0.5°×0.5°) from the China Meteorological Administration; land use data from GlobeLand30 and ESA datasets (300 m resolution); and population density from the Institute of Geographic Sciences and Natural Resources Research. Temperature data comprise monthly maxima, minima, and means; precipitation data are monthly totals. All covariates underwent data cleaning, resampling via weighted averaging, co-Kriging interpolation, and bilinear interpolation to create a structured covariate database spanning 2000–2020.

1.3.1 Variable Selection

Variable relative importance serves as a key criterion in random forest modeling. This study collected potential covariates (temperature, precipitation, NDVI, grassland, permanent snow/ice, elevation, water yield modulus, bare land, population, slope/aspect, land use, other vegetation) and quantified their importance through iterative permutation [FIGURE:2]. Variables with <5% relative importance were eliminated as redundant (water yield modulus, bare land, population, land use, slope/aspect, other vegetation), retaining only meaningful predictors.

Temperature emerged as the dominant factor (56% model weight), followed by NDVI (16%), grassland (12%), permanent snow/ice (9%), and precipitation (7%). Temperature and NDVI contributed most substantially, with permanent snow/ice, precipitation, and grassland collectively accounting for 28%, while other factors contributed marginally.

1.3.2 Model Construction

Random forest, a parallel ensemble learning algorithm developed by Breiman and Cutler, constructs classifiers with multiple decision trees through random attribute selection. Studies demonstrate its effectiveness for complex multi-source spatial data prediction. Model training employed bootstrap resampling to construct regression trees using selected variables (temperature, precipitation, NDVI, grassland, permanent snow/ice). Each tree was built to terminal nodes containing single data points, with predictions averaged across all trees. The final model comprised 500 trees, achieving near-optimal performance with parallel processing (parallelism=32, concurrency=64). Initial training required 232 hours, optimized to 2.5 hours per prediction.

1.3.3 Data Prediction

Post-modeling predictions yielded annual glacier data for all years. To minimize interference from transient snow and lake ice, we identified advantageous grid cells where predicted glacier coverage exceeded 50% of grid area, designating these as valid training samples. A convolution operation refined predictions:

Let $f[m,n]$ represent the input image with horizontal index $m$ and vertical index $n$, and $g[x,y]$ the convolution kernel. The output $h[m,n]$ is:

$$h[m,n] = \sum_{x}\sum_{y} f[m-x,n-y] \cdot g[x,y]$$

where $x$ and $y$ are displacement parameters. By adjusting kernel coefficients, convolution results asymptotically approached observed values. Applying this operation to all annual predictions produced the 2000–2020 spatiotemporal glacier distribution dataset. Area errors between predicted and observed values were 5.77% (2017), 5.23% (2020), and 6.70% (2001), with overall error of 1.76%, confirming reliability.

1.3.4 Model Validation

Ten-fold cross-validation assessed model accuracy and prevented overfitting. Training samples were randomly partitioned into ten subsets, with nine used for training and one for validation in each iteration. All values were log-transformed for validation. The model achieved coefficient of determination (R²) = 0.72 [FIGURE:3]. Additional metrics included root mean square error (RMSE = 1.52), mean fractional bias (MFB = 0.15), and mean fractional error (MFE = 0.29), demonstrating robust predictive performance.

2.1 Spatial Distribution Characteristics of Glaciers on the Qinghai-Xizang Plateau

Glacier distribution is governed by integrated climatic and topographic controls, particularly low temperatures, abundant precipitation, high elevations, and gentle slopes that provide favorable mass accumulation and storage conditions. The 2020 prediction data reveal that glaciers concentrate in western ranges (Kunlun, Karakoram), southwestern-southern ranges (Gangdise, Himalaya), southern ranges (Nyainqentanglha), southeastern ranges (Hengduan), central ranges (Tanggula), and northeastern ranges (Qilian), decreasing from southwest to northeast. The Kunlun, Karakoram, and Himalaya account for over half of total glacier area.

Topographic analysis of 2020 data using ArcGIS shows slope, aspect, and elevation constraints [TABLE:2]. Slope analysis [FIGURE:4] indicates 36.46% of glaciers occur on 10°–20° slopes, 27.61% on 20°–30°, and 23.68% on 0°–10°, with only 2.08% on slopes >40°. Collectively, 97.92% of glacier area lies within 0°–40° slopes, consistent with Himalayan patterns. Aspect analysis [FIGURE:5] reveals northern slopes host the largest proportion (16.00%), followed by northeastern (15.27%). Eastern, southern, and southeastern slopes show similar coverage (~13.3%), while western slopes are minimal (9.85%). Overall, northern slopes exceed southern slopes, and western slopes exceed eastern slopes in glacier coverage.

Elevation analysis [FIGURE:6] demonstrates that 69.99% of glaciers occupy 5000–6000 m, 19.98% occupy 6000–7000 m, and 11.10% occupy 4000–5000 m, totaling 99.38% within 4000–7000 m. Areas above 7000 m comprise only 0.35% due to temperature and topographic limitations. Continental glacier termini can descend to 5100 m, while maritime glaciers may reach 3000 m.

2.2.1 Overall Spatiotemporal Glacier Change Trends (2000–2020)

The Mann-Kendall test, which constructs standard normal distribution statistic Z to assess trends, was applied to quantify glacier changes. For the entire plateau, the 2000–2020 data show significant retreat (Z = -4.21, P < 0.01) [TABLE:3]. Spatially, Z-values reveal that minor retreat dominates (54.14% of area), uniformly distributed across the region. Significant retreat concentrates in southeastern margins (34.07%), while minor advancement occurs in northwestern margins (21.01%). No-change areas are minimal (2.31%), and significant advancement is negligible (0.17%) [FIGURE:7, FIGURE:8].

2.2.2 Typical Glacier Change Trends

Different glacier types and sizes respond variably to climate change, creating regional heterogeneity. To characterize this, we examined four major ranges: Himalaya, Karakoram, Kunlun, and Nyainqentanglha.

Himalaya Mountains: Time-series analysis reveals significant retreat (Z = -5.32, P < 0.01) from 2000–2020. Spatially, significant retreat dominates (43.84%), with minor retreat (32.67%) and no-change areas (9.61%) distributed uniformly. Minor advancement appears locally in central and eastern areas (21.01%), while significant advancement is negligible (2.31%) [FIGURE:9].

Karakoram Mountains: Known for concentrated glacier development outside polar regions, this range shows no significant trend (Z = -1.42, P > 0.05). However, strong minor changes prevail (97.55%), with minor retreat dominant in central-eastern areas (89.55%) and minor advancement in central regions (8.00%). No-change areas are sparse (2.39%), and significant retreat is minimal (0.06%) [FIGURE:10].

Kunlun Mountains: Spanning eastern and western sections, this range shows no significant trend (Z = 1.21, P > 0.05). Minor changes dominate (94.00%), with minor advancement concentrated in the western section (59.61%) and minor retreat distributed uniformly (33.95%). No-change areas are relatively extensive (6.26%), while significant retreat and advancement are negligible (0.14% and 0.04%, respectively) [FIGURE:11].

Nyainqentanglha Mountains: Exhibiting maritime characteristics in the east and continental features in the west, this range shows significant retreat (Z = -3.15, P < 0.01). Significant retreat dominates (90.57%), with minor retreat locally in the east (7.42%). Significant and minor advancement are negligible (1.78% and 0.18%), and no-change areas are minimal (0.05%) [FIGURE:12].

3 Conclusions

This study innovatively integrates multi-source remote sensing, meteorological, geographic, and environmental data to develop a large-scale glacier distribution simulation method. Key achievements include:

1. A random forest glacier coverage model (R² = 0.72) built upon a comprehensive covariate database, producing a 1 km-resolution annual glacier distribution dataset for 2000–2020.

2. Spatial distribution patterns: 97.92% of glaciers occur on 0°–40° slopes and 99.38% at 4000–7000 m elevations, with predominant north, east, and south aspect distribution and secondary west aspect coverage.

3. Spatiotemporal changes: 2000–2020 data reveal significant overall retreat. Spatially, pronounced changes concentrate along plateau margins, weakening toward the interior where minor variations dominate.

4. Regional variations: Himalaya and Nyainqentanglha glaciers show significant retreat; Karakoram glaciers exhibit minor retreat; Kunlun glaciers display concurrent patterns of minor advancement and retreat.

References

[1] Zhou Yushan. Investigation of glacier mass balance in the Qinghai-Tibet Plateau and its surroundings based on multi-source remote sensing data[J]. Geography and Geo-Information Science, 2019, 35(4): 142.

[2] Gu Yunyang. Monitoring the changes of glacier mass balance in the source region of the Yangtze River from 2000 to 2020 with the imaging geodetic method[D]. Changsha: Central South University, 2023.

[3] Wang Ninglian, Yao Tandong, Xu Baiqing, et al. Spatiotemporal pattern, trend, and influence of glacier change in Tibetan Plateau and surroundings under global warming[J]. Bulletin of Chinese Academy of Sciences, 2019, 34(11): 1220-1232.

[4] Wang Yingzheng, Li Jia, Wu Lixin, et al. Using remote sensing images to monitor the glacier changes in Qilian Mountains during 1987—2018 and analyzing the impact factors[J]. Journal of Glaciology and Geocryology, 2020, 42(2): 344-356.

[5] Ji Qin, Zhang Cuilan, Ding Yuekai, et al. Glacier monitoring in Qomolangma Nature Reserve based on multi-source remote sensing data[J]. Arid Land Geography, 2023, 46(10): 1591-1601.

[6] Zhang Yu. Glacier change in Tanggula Mountain and its response to climate fluctuation from 1980 to 2020[D]. Dalian: Liaoning Normal University, 2023.

[7] Tian Mengqi, Duan Keqin, Shi Peihong. Glacier changes on the Qinghai-Tibet Plateau based on Google Earth Engine: A case study of the Purog Kangri glacier[J]. Scientia Geographica Sinica, 2023, 43(6): 943-951.

[8] Zhang Meng. Research on the method of extracting glacier area information based on the new index using Landsat imagery[D]. Xi'an: Northwest University, 2020.

[9] Yan Lili, Wang Jian. Study of extracting glacier information from remote sensing[J]. Journal of Glaciology and Geocryology, 2013, 35(1): 110-118.

[10] Zhang Yili. Data sets of Qinghai-Tibet Plateau boundary[EB/OL]. [2019-06-11]. https://portal.casearth.cn/dataDetail/64b8866cd6067c9150a4a7ad.

[11] Liu Shiyin, Ding Yongjian, Li Jing, et al. Glaciers in response to recent climate warming in western China[J]. Quaternary Sciences, 2006, 26(5): 762-771.

[12] Liu Shiyin, Yao Xiaojun, Guo Wanqin, et al. The contemporary glaciers in China based on the Second Chinese Glacier Inventory[J]. Acta Geographica Sinica, 2015, 70(1): 3-16.

[13] Liao Meiyu, Fang Xiuqin, Jiang Xinyuan, et al. Spatiotemporal characteristics of land use/cover changes in the Yellow River Basin over the past 40 years[J]. Journal of Soil and Water Conservation, 2024, 38(2): 165-177.

[14] Ye Qinghua, Cheng Weiming, Zhao Yongli, et al. A review on the research of glacier changes on the Tibetan Plateau by remote sensing technologies[J]. Journal of Geo-Information Science, 2016, 18(7): 920-930.

[15] Ye Qinghua. Glacier coverage data on the Tibetan Plateau in 2017 (TPG2017, Version1.0)[EB/OL]. [2019]. https://data.tpdc.ac.cn/zh-hans/data/9777a877-8bee-492c-75a4.

[16] Ye Q H, Zong J B, Tian L D, et al. Glacier changes on the Tibetan Plateau derived from Landsat imagery: Mid-1970s—2000—2013[J]. Journal of Glaciology, 2017, 63(238): 273-287.

[17] Svetnik V, Liaw A, Tong C, et al. Random forest: A classification and regression tool for compound classification and QSAR modeling[J]. Journal of Chemical Information and Computer Sciences, 2003, 43(6): 1947-1958.

[18] Zhou Jihua, Lai Liming, Chen Qiao-e, et al. Multi-temporal vegetation mapping of the Tibetan Plateau via machine learning model simulation and experiential knowledge[J]. Chinese Science Bulletin, 2025, 70(1): 134-144.

[19] Wright M N, Ziegler A. Ranger: A fast implementation of random forests for high dimensional data in C++ and R[J]. Journal of Statistical Software, 2017, 77(1): 1-17.

[20] Zhang Yili, Li Bingyuan, Zheng Du. Datasets of the boundary and area of the Tibetan Plateau[J]. Acta Geographica Sinica, 2014, 69(Suppl. 1): 65-68.

[21] Ding Yuekai, Liu Rui, Zhang Cuilan, et al. Remote sensing monitoring of glacier and glacial lake changes in Yairu Zangbo Basin, Himalayas[J]. Arid Land Geography, 2022, 45(6): 1870-1880.

[22] Ye Qinghua, Wu Yuwei. Glacier coverage data on the Tibetan Plateau in 2001 (TPG2001, Version1.0)[EB/OL]. [2018]. https://data.tpdc.ac.cn/zh-hans/data/4fc1e6b1-5f3adc4fde19.

[23] Ye Qinghua. Glacier coverage data on the Tibetan Plateau in 2013 (TPG2013, Version1.0)[EB/OL]. [2018]. https://data.tpdc.ac.cn/zh-hans/data/38d741e0-d732a01b8063.