EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Regional Snow Cover Fraction Prediction in Xinjiang Using a Terrain-Aware Graph Neural Network

Yun Gao, Xusheng Zhang, Xueqing Hou, Jiangming Chai

Atmosphere · 07.11.2026 · 17 (11) · 1101

DOI: 10.3390/atmos17111101 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Atmosphere, online erschienen am 07.11.2026. Verfasst von Yun Gao und 3 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Science and Technology Development Plan for the “Two Zones” in the Xinjiang Uygur Autonomous Region, 2026 Xinjiang Uygur Autonomous Region Natural Science Foundation General Project.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Snow cover fraction (SCF) reflects the regional snow cover changes in arid mountainous areas. We developed a terrain-aware graph convolutional network–long short-term memory network (GCN-LSTM), where 9 meteorological stations in Xinjiang serve as graph nodes and 15 monthly meteorological variables as dynamic node features, while altitude, slope, and aspect define a fixed terrain information graph. The overall study period spanned March 2000–June 2022; the model uses a 12-month input window and predicts regional snow cover fraction in Xinjiang one month in advance under strict time series partitioning. On the 30-month test set (January 2020–June 2022), the model’s mean seed-level root mean square error (RMSE) was the lowest point estimate (0.0386); paired block-bootstrap intervals for ensemble-prediction differences from monthly climatology and ridge included zero. The model followed the pronounced seasonal cycle, but its five-seed ensemble had a deseasonalized anomaly R2 of −0.009, providing limited evidence of predictive skill beyond seasonality. The terrain-weighted graph had lower test RMSE than the geographic graph but similar error to the terrain-binary graph; intervals for most other graph comparisons crossed zero. Radiation accounted for 31.28% of total grouped SHAP (Shapley Additive Explanations) attribution and Lag-1 meteorological inputs for 36.24% of lag attribution. Removing the Moisture group during retrained ablation increased ensemble test RMSE by 0.0033. Different rankings indicate that attribution analysis and ablation analysis capture different aspects of model dependencies. Across seven retrospective annual outer Test folds (84 months), the ensemble did not show a stable improvement over monthly climatology (pooled RMSE 0.03896 vs.0.03763); terrain information remains a spatial prior whose broader predictive value requires further evaluation.

Förderung

  • National Natural Science Foundation of China
  • Science and Technology Development Plan for the “Two Zones” in the Xinjiang Uygur Autonomous Region
  • 2026 Xinjiang Uygur Autonomous Region Natural Science Foundation General Project

Zitieren

APA

Gao, Y., Zhang, X., Hou, X., & Chai, J. (2026). Regional Snow Cover Fraction Prediction in Xinjiang Using a Terrain-Aware Graph Neural Network. Atmosphere, 17(11), 1101. https://doi.org/10.3390/atmos17111101

Vancouver

Gao Y, Zhang X, Hou X, Chai J. Regional Snow Cover Fraction Prediction in Xinjiang Using a Terrain-Aware Graph Neural Network. Atmosphere. 2026;17(11):1101. doi:10.3390/atmos17111101

BibTeX
@article{Gao2026,
  author = {Gao, Yun and Zhang, Xusheng and Hou, Xueqing and Chai, Jiangming},
  title = {Regional Snow Cover Fraction Prediction in Xinjiang Using a Terrain-Aware Graph Neural Network},
  journal = {Atmosphere},
  year = {2026},
  volume = {17},
  number = {11},
  pages = {1101},
  doi = {10.3390/atmos17111101}
}
Kostenlos starten →