Tang et al. (2025) REBAI: Development and validation of a novel indicator for burned area detection using Sentinel-2 images

Identification

• Journal: Ecological Indicators
• Year: 2025
• Date: 2025-12-11
• Authors: Mingxiu Tang, Xiufang Zhu, Chuanwu Zhao, Junying Song, Chang Xiao, Rui Guo
• DOI: 10.1016/j.ecolind.2025.114501

Research Groups

• State Key Laboratory of Remote Sensing and Digital Earth, Beijing Normal University, Beijing, China
• Key Laboratory of Environmental Change and Natural Disasters of Chinese Ministry of Education, Beijing Normal University, Beijing, China
• Hangzhou International Innovation Institute, Beihang University, Hangzhou, China
• China Mobile Jiutian Artificial Intelligence Technology Co., Ltd, Beijing, China

Short Summary

This study develops and validates the novel Red-Edge Burned Area Index (REBAI) using Sentinel-2 data, demonstrating its superior and stable performance in accurately detecting small-scale and subtle burned areas across diverse global forest ecosystems compared to existing spectral indices.

Objective

• To formulate REBAI by synergistically utilizing red-edge and shortwave infrared bands to quantify post-fire changes in vegetation structure and function.
• To validate REBAI's spectral discrimination capability between burned and unburned areas.
• To evaluate REBAI's effectiveness in identifying small burned patches and areas exhibiting subtle spectral changes.
• To explore REBAI's potential as a high-quality input feature for machine learning models to enhance burned area mapping.

Study Configuration

• Spatial Scale: Twelve diverse forest burned areas across 6 continents (ranging from 0.07 km² to 73.71 km² in size, with 8 sites under 1.07 km²) and four grassland burned areas were studied. Sentinel-2 imagery was resampled to a uniform 10 m spatial resolution.
• Temporal Scale: Two Sentinel-2 image scenes (pre-fire and post-fire) were acquired for each study area, with pre-fire images selected within three months or the corresponding seasonal period of the preceding year. Sentinel-2 offers a 5-day global revisit period.

Methodology and Data

• Models used:
  • Bare Soil Index (BSI) for vegetation masking.
  • Red-Edge Burned Area Index (REBAI) formulation.
  • 16 unsupervised thresholding algorithms (e.g., Otsu’s, K-Means, Maximum Entropy) for burned area extraction, combined with a voting mechanism (14 of 16 algorithms).
  • Jeffries-Matusita (JM) distance and Common Isolation Values (CIV) for separability analysis.
  • Recursive Feature Elimination with Cross-Validation (RFECV) using Random Forest (RF) as the base estimator for feature importance assessment.
  • Random Forest (RF) classifier for comparison and exploration of REBAI as an input feature.
  • Wilcoxon signed-rank test for statistical significance of accuracy differences.
• Data sources:
  • Sentinel-2 Multi-Spectral Instrument (MSI) Level-2A surface reflectance products (COPERNICUS/S2SRHARMONIZED) from Google Earth Engine (GEE).
  • Manual visual interpretation and delineation of burned areas for reference data (200 burned and 200 unburned forest sample points per study area, totaling 4800 points).

Main Results

• REBAI, formulated using Sentinel-2 bands B4, B5, B6, B8A, and B12, demonstrated superior spectral separability between burned and unburned areas, achieving a JM distance consistently greater than 1.9 and an average CIV of 0.001.
• REBAI ranked highest in feature importance among 10 spectral bands and 13 common indices, with an average rank of 2.08 and a standard deviation of 1.26 across 12 diverse forest study areas.
• Burned area detection using REBAI achieved high and stable accuracy across all 12 forest study areas, with an average Overall Accuracy (OA) of 96.60 % ± 3.58 %, Precision of 92.97 % ± 11.27 %, Recall of 87.12 % ± 8.07 %, and F1 score of 0.8917 ± 0.0702.
• REBAI significantly outperformed other common indices (e.g., NDVI, NBR, EVI, VRI, SAVI, NBR+) in F1 score (p < 0.05) and exhibited greater robustness and stability across varying environmental conditions and burn characteristics.
• REBAI's performance was comparable to or superior to a Random Forest classifier trained on the same spectral bands, with higher average OA, Precision, and F1 score.
• Preliminary analysis suggested REBAI's transferability to grassland burned areas, yielding an average F1 score of 0.9138 across four case studies.
• Bi-temporal analysis using REBAI differences improved F1 score by 13.12 % and Recall by 26.70 % in an exploratory case study, and integration with machine learning models also showed F1 score improvements of up to 14.72 %.

Contributions

• Introduction of REBAI, a novel, physically-based spectral index specifically designed for burned area detection using Sentinel-2 red-edge, red, and shortwave infrared bands.
• Demonstrated superior accuracy and stability of REBAI in detecting small-scale and subtle spectral burns across a globally diverse set of forest ecosystems, outperforming 13 widely used indices.
• Established REBAI as a robust and computationally efficient standalone tool for rapid burned area mapping and a high-quality, discriminative feature for enhancing machine learning models.
• Provided preliminary evidence for the broad applicability and transferability of REBAI across different vegetation types, including grasslands.

Funding

• National Natural Science Foundation of China [grant number 42192583]

Citation

@article{Tang2025REBAI,
  author = {Tang, Mingxiu and Zhu, Xiufang and Zhao, Chuanwu and Song, Junying and Xiao, Chang and Guo, Rui},
  title = {REBAI: Development and validation of a novel indicator for burned area detection using Sentinel-2 images},
  journal = {Ecological Indicators},
  year = {2025},
  doi = {10.1016/j.ecolind.2025.114501},
  url = {https://doi.org/10.1016/j.ecolind.2025.114501}
}