RSSEMLabUniversity of Toronto Mississauga
Mosaic of 30 m map panels for one Yukon wildfire showing burn severity, elevation and driver variables

Research Products  /  Open dataset

Subarctic Wildfire Severity

Open data, a trained XGBoost model and code from a 35-year analysis of the climate, vegetation and topographic drivers of wildfire severity in the central Yukon, Canada.

Access the dataset Read the paper View the code
70Fires larger than 50 km², 1986–2020
67,008 km²Study area, central Yukon
18 of 52Predictor variables retained
0.70Validation pseudo r², XGBoost

Overview

Subarctic ecosystems are experiencing earlier, longer and more intense wildfire seasons, and large fires damage the environment, release stored carbon and force residents to relocate. This product supports a study of what drives wildfire severity in the Klondike Plateau and Yukon Plateau-Central ecoregions of the Yukon, Canada, based on the 70 fires larger than 50 km² recorded between 1986 and 2020.

Burn severity measured from Landsat imagery was combined with climate, topographic and forest land cover data at a common 30 m resolution, and an XGBoost regression model, interpreted with SHAP (Shapley Additive Explanations) values, was trained on the selected variables. The model reached a validation pseudo r² of 0.70.

This product can support the study and management of wildfire in subarctic Canada, where remote locations and sparse populations make forest management difficult. The data, model and code can be used to examine how pre-fire vegetation, surface dryness, wind and terrain combine to drive severe burning and to inform practices that reduce the habitat loss and carbon release caused by severe fires, as well as to provide a basis for similar studies in other subarctic regions.

Vegetation

Pre-fire vegetation

The most important variable: pixels with more vegetation before the fire burned more severely. Barren land and coniferous forest were linked to higher severity, broadleaf forest to lower.

Climate

Surface dryness

Skin reservoir content, evaporation from vegetation transpiration and soil temperature ranked second to fourth. Less canopy water, more evaporation, warmer soil and higher wind speed pointed to higher severity.

Topography

Exposed terrain

Ridges and hills with more visible sky and higher wind exposure burned more severely than flat land, valleys and gullies.

Figures

Maps showing the Yukon within Canada, the Klondike Plateau and Yukon Plateau-Central ecoregions, and the elevation of the study area
Study areaKlondike Plateau and Yukon Plateau-Central ecoregions, with elevation at 30 m. Figure 1 from Nelson et al. (2026), CC BY 4.0.
Chart of mean absolute SHAP values for the 18 model variables, led by pre-fire vegetation, and box plots of SHAP values for each land cover class
Variable importanceMean absolute SHAP value of each variable (A) and SHAP values by land cover class (B). Figure 5 from Nelson et al. (2026), CC BY 4.0.
Nine map panels for one fire: location, elevation, dNBR, pre-fire NBR, visible sky, wind exposition index, evaporation from vegetation transpiration, soil temperature and skin reservoir content
Drivers for one fireSeverity (dNBR) and important climate, vegetation and topographic variables at 30 m for a randomly selected fire. Figure 6 from Nelson et al. (2026), CC BY 4.0.
Density plot of observed versus predicted dNBR for the validation set, with a red 1:1 line
Model fitObserved versus predicted dNBR in the validation set; the red line is 1:1. Figure 4 from Nelson et al. (2026), CC BY 4.0.

Figures from Nelson, D. M., He, Y., & Moore, G. W. K. (2026), Big Earth Data, © 2025 The Author(s), published under a CC BY 4.0 licence. The banner image is adapted from panels B–I of Figure 6 (cropped, rearranged, panel labels removed and the white background recoloured).

What’s included

  • wildfire_dataset.csv · 1.86 GBThe wildfire table, one row per observation, with 50 columns: dNBR, pre-fire NBR, land cover, topographic indices and ERA5-Land climate variables
  • xgb_wildfire.json · 1.81 GBThe trained XGBoost regression model in XGBoost JSON format: 20,000 trees built on the 18 selected variables
  • gpu_shap_df_lc.csv · 523 MBSHAP values in long format: row ID, variable, variable value and SHAP value
  • Data Column Codes.xlsx · 9.9 kBFull names of the column codes (for example, 23_clm is skin reservoir content)
  • Code · 53 kBPython scripts for data retrieval (fire severity, ERA5 fire danger indices, topography, land cover), processing and downscaling, feature selection, GPU model tuning, training and inference, and plotting
  • Versions and licencesDataset and model: version 1.0.0, 21 May 2025, CC BY 4.0 · code: v1.0.0, 21 May 2025, GNU GPL v3.0
Dataset and model on Zenodo Code on Zenodo Code on GitHub

Cite

Nelson, D. M., He, Y., & Moore, G. W. K. (2026). Driver analysis of subarctic wildfire severity over a 35-year period. Big Earth Data, 10(1), 319–345. https://doi.org/10.1080/20964471.2025.2558408

Nelson, D., He, Y., & Moore, G. W. K. (2025). Data and Model Supplement: Driver Analysis of Subarctic Wildfire Severity over a 35-year Period (Version 1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.15446613

Nelson, D., He, Y., & Moore, G. W. K. (2025). Code Supplement: Driver Analysis of Subarctic Wildfire Severity over a 35-year Period (v1.0.0) [Software]. Zenodo. https://doi.org/10.5281/zenodo.15485182

Please cite the article when using the data, model or code.

Contact

Questions about the data, the model or collaboration on wildfire research are welcome. Please get in touch with the laboratory.

For other enquiries, see the lab contact page.

Prof. Yuhong He

Principal Investigator, RSSEM Lab
Department of Geography, Geomatics and Environment, University of Toronto Mississauga

Office
DV3271, William G. Davis Building
Address
3359 Mississauga Road, Mississauga, ON L5L 1C6
Email about this dataset