Pandey et al. (2025) Permafrost distribution modeling using remote sensing and machine learning technique in the Garhwal Himalaya, India

Identification

• Journal: Environmental Earth Sciences
• Year: 2025
• Date: 2025-12-27
• Authors: A. C. Pandey, AFM Tariqul Islam, Chandra Shekhar Dwivedi, Bikash Ranjan Parida, Alexey Maslakov, E. S. Koroleva
• DOI: 10.1007/s12665-025-12739-7

Research Groups

• Department of Geoinformatics, School of Natural Resource Management, Central University of Jharkhand, Ranchi, India
• Faculty of Geography, Lomonosov Moscow State University, Moscow, Russia
• Yamal-Nenets autonomous okrug (region) Arctic Research Center, Salekhard, Russia

Short Summary

This study modeled permafrost distribution in the Garhwal Himalaya, India, by integrating topo-climatic variables and rock glacier inventories using Binary Logistic Regression, Random Forest, and Extreme Gradient Boosting, demonstrating high accuracy and identifying elevation and mean temperature of the warmest quarter as key predictors.

Objective

• Evaluate the predictive performance and classification accuracy of Binary Logistic Regression (LRM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) models for permafrost distribution.
• Analyze the relative importance of predictor variables contributing to permafrost development in mountainous regions.

Study Configuration

• Spatial Scale: Garhwal Himalaya, Uttarakhand, India (11,235 km²), located between 30° 44’ 16’’ N to 31° 25’ 59’’ N and 78° 48’ 28’’ E to 80° 23’ 47’’ E. Elevation ranges from 1048 meters to 7795 meters above sea level. The study primarily focuses on the Alaknanda River valley.
• Temporal Scale:
  • Mean Annual Air Temperature (MAAT): 1970–2000
  • Mean Annual Ground Temperature (MAGT): 2000–2016
  • Land Surface Temperature (LST): 2000–2020
  • Mean Snow Cover (SC): 2000–2020
  • Potential Incoming Solar Radiation (PISR): November to March (snow-free season)
  • Maximum Temperature of the Warmest Quarter (MTOWQ): 1970–2000
  • Rock glacier imagery: Snow-free periods in 2017, 2019, and 2022

Methodology and Data

• Models used:
  • Binary Logistic Regression (BLR)
  • Random Forest (RF)
  • Extreme Gradient Boosting (XGBoost)
• Data sources:
  • Dependent Variable: 268 rock glaciers (247 active, 21 relict) mapped using high-resolution imagery from Sentinel-2 and Google Earth Pro (Digital Globe, SPOT datasets).
  • Independent Variables (Climatic & Topographic):
    • Mean Annual Air Temperature (MAAT): WorldClim version 2 (1970–2000), 1 kilometer resolution, resampled to 30 meters.
    • Mean Annual Ground Temperature (MAGT): Obu et al. (2019) product (2000–2016), 1 kilometer resolution, resampled to 30 meters.
    • Land Surface Temperature (LST): MOD11A2.061 (2000–2020), 1 kilometer resolution, processed via Google Earth Engine (GEE), resampled to 30 meters.
    • Mean Snow Cover (SC): MOD10A1.006 (2000–2020), 500 meter resolution, processed via GEE, resampled to 30 meters.
    • Potential Incoming Solar Radiation (PISR): Derived from ASTER DEM (~30 meters) for November to March, in kilowatt-hours per square meter (kWh/m²).
    • Maximum Temperature of the Warmest Quarter (MTOWQ): WorldClim version 2.1 (1970–2000).
    • Elevation, Slope, Slope Aspect: ASTER Global DEM (Version 3) (~30 meters).

Main Results

• A total of 268 rock glaciers were mapped, comprising 247 active and 21 relict forms.
• Active rock glaciers are predominantly found at elevations between 4200 meters and 5200 meters above sea level, associated with mean annual air temperatures (MAAT) below 0 °C (especially below -4 °C) and potential incoming solar radiation (PISR) of 1000–1400 kWh/m².
• Relict rock glaciers are mainly located at elevations between 4000 meters and 4800 meters above sea level, with PISR values of 500–1000 kWh/m².
• A strong negative Pearson correlation (r = -0.935) was observed between MAAT and elevation.
• Logistic Regression Models (LRM) achieved classification accuracies of:
  • LRM-MAAT: 94.8% (Area Under the Curve (AUC) = 0.905)
  • LRM-MAGT: 91.8% (AUC = 0.844)
  • LRM-SC: 92.05% (AUC = 0.762)
  • LRM-LST: 91.4% (AUC = 0.695)
• Ensemble machine learning models outperformed LRMs in testing accuracy:
  • Random Forest: 97.6% accuracy (R² = 0.85, F1-score = 0.883, F2-score = 0.907)
  • Extreme Gradient Boosting (XGBoost): 97.0% accuracy (R² = 0.84, F2-score = 0.903)
• Feature importance analysis for RF and XGBoost identified elevation and Mean Temperature of the Warmest Quarter (MTOWQ) as the most dominant predictors of permafrost distribution. Mean Annual Ground Temperature (MAGT) and Land Surface Temperature (LST) also contributed moderately, while snow cover, aspect, slope, and PISR had minor influence.
• Permafrost probability is highest at elevations above approximately 4200 meters above sea level, where MAAT ranges from -8 °C to -16.9 °C and MAGT from -1 °C to -5.17 °C.

Contributions

• This study provides a refined methodology for mapping permafrost in data-sparse regions by integrating remote sensing and machine learning techniques.
• It is among the first studies in the Himalayas to employ extreme air temperature thresholds and validate their correlation with Mean Annual Ground Temperature (MAGT) for permafrost modeling.
• The research demonstrates the potential of incorporating rock glacier inventories and a comprehensive set of climatological and topographic variables into predictive modeling frameworks for the central Himalaya.
• The generated permafrost maps offer critical baseline information to support infrastructure planning, hazard assessment, and climate adaptation strategies in the fragile mountain environment.
• The established modeling framework is replicable for similar studies in other data-scarce, high-altitude regions.

Funding

Not explicitly stated in the paper.

Citation

@article{Pandey2025Permafrost,
  author = {Pandey, A. C. and Islam, AFM Tariqul and Dwivedi, Chandra Shekhar and Parida, Bikash Ranjan and Maslakov, Alexey and Koroleva, E. S.},
  title = {Permafrost distribution modeling using remote sensing and machine learning technique in the Garhwal Himalaya, India},
  journal = {Environmental Earth Sciences},
  year = {2025},
  doi = {10.1007/s12665-025-12739-7},
  url = {https://doi.org/10.1007/s12665-025-12739-7}
}