Van et al. (2025) Tree-Based Regressor Comparison for Burn Severity Mapping: Spatially Blocked Validation Within and Across Fires

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-11-19
• Authors: Linh Nguyen Van, Giha Lee
• DOI: 10.3390/rs17223756

Research Groups

[Information not provided in the paper text.]

Short Summary

This study benchmarks six tree-based machine learning models for predicting post-fire burn severity from satellite data, evaluating their generalization capabilities both within and across ten U.S. wildfires to provide practical recommendations for rapid severity mapping.

Objective

• To benchmark the performance and spatial generalization capabilities of six tree-based regression models (Decision Tree, Random Forest, Extra Trees, Bagging, Gradient Boosting, AdaBoost) for predicting wildfire severity from Landsat surface reflectance data.

Study Configuration

• Spatial Scale: Ten individual wildfire events across the western United States.
• Temporal Scale: Post-fire period, focusing on timely mapping of burn severity after a wildfire event.

Methodology and Data

• Models used: Decision Tree (DT), Random Forest (RF), Extra Trees (ET), Bagging, Gradient Boosting (GB), AdaBoost (AB).
• Data sources: Landsat surface reflectance data (satellite-derived spectral features) and ground-based Composite Burn Index (CBI) metrics.

Main Results

• Under within-fire spatial generalization (Leave-One-Cluster-Out, LOCO):
  • Random Forest (RF) performed best with R² = 0.679, MAE = 0.397, RMSE = 0.516.
  • Extra Trees (ET) was statistically comparable with R² = 0.673, MAE = 0.393, RMSE = 0.518.
  • Bagging followed closely with R² = 0.668, MAE = 0.402, RMSE = 0.525.
• Under cross-fire transfer (Leave-One-Fire-Out, LOFO):
  • Extra Trees (ET) demonstrated the best transferability with R² = 0.616, MAE = 0.450, RMSE = 0.571.
  • Gradient Boosting (GB) followed with R² = 0.564, MAE = 0.479, RMSE = 0.606.
  • Random Forest (RF) achieved R² = 0.543, MAE = 0.490, RMSE = 0.621.
• Tree ensembles, particularly ET and RF, are competitive for rapid severity mapping with minimal tuning. RF is recommended for individual fires with local calibration, while ET is preferred for transferability to unseen fires.

Contributions

• Provides a comprehensive benchmarking of six tree-based regression models for post-fire burn severity mapping, specifically evaluating their generalization performance across different spatial domains (within-fire and cross-fire).
• Offers practical guidance for selecting appropriate machine learning models based on the specific application need: Random Forest for local calibration within a single fire and Extra Trees for improved transferability to new, unobserved fires.

Funding

[Information not provided in the paper text.]

Citation

@article{Van2025TreeBased,
  author = {Van, Linh Nguyen and Lee, Giha},
  title = {Tree-Based Regressor Comparison for Burn Severity Mapping: Spatially Blocked Validation Within and Across Fires},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17223756},
  url = {https://doi.org/10.3390/rs17223756}
}