🔥 A Sentinel-2 Benchmark for Deep-Learning Active-Fire Segmentation

Shreyan Mitra, Mohammadreza Narimani* (corresponding author), and Parastoo Farajpoor

Data in Brief · Volume 69 · Article 113357 · 2026 · Open access CC BY 4.0

Department of Biological and Agricultural Engineering · University of California, Davis

Read the open-access paper Download the dataset (Zenodo) Code on GitHub

Why This Dataset

An active-fire mask locates ongoing combustion in a satellite scene. It is not a burn-scar map. Sentinel-2 shortwave infrared at 20 m can reveal those cores, but comparing segmentation models requires a shared rendering, a documented label rule, and a split that does not leak the same wildfire into training and test.

This open Data in Brief article by Mohammadreza Narimani (corresponding author), Shreyan Mitra, and Parastoo Farajpoor releases a fixed archive of 2,148 image–mask pairs from 25 California wildfires (23 July 2020–21 August 2026). Images are 512 × 512 three-channel composites of Sentinel-2 Level-2A bands B12, B11, and B8A with a single linear rendering. Masks code background (0), SWIR-rule active fire (1), and invalid observations (255).

Core story: a rare-class, incident-disjoint Sentinel-2 benchmark so researchers can train and compare active-fire segmentation methods on wildfires held out from training—without re-querying Earth Search each time.
2,148Image–mask chips
25California wildfires
512 × 51220 m SWIR chips
841Chips with fire labels
0.0766%Fire pixels in the archive
CC BY 4.0Open data and paper

Incident-Disjoint Splits

Partitions are assigned by wildfire name: 18 / 3 / 4 fires for train / validation / test (1,470 / 246 / 432 chips). Castle, McCash, and Bobcat are reserved for validation; Slater, Hopkins, CALDWELL, and Windy are reserved for testing. Every chip from a fire stays in one set, so grouped evaluation is possible on events the model has never seen.

Mohammadreza Narimani Sentinel-2 active-fire dataset overview showing 2148 chips, incident-disjoint train validation test split, and rare-class fire pixel prevalence
Figure 1. Dataset overview: chips by partition, incident-disjoint wildfire split, and rare-class prevalence.
Map of 25 California wildfires in the Mohammadreza Narimani Sentinel-2 active-fire benchmark colored by train, validation, and held-out test split
Figure 2. California fires in the archive, colored by the published hold-out split.

SWIR Labeling and Example Chips

Labels come from shortwave-infrared brightness and near-infrared contrast, then two steps of neighborhood growth. Hand-drawn regions of interest show why B11 and B12 separate active flame from vegetation, char, and bright smoke. True-color RGB often hides the same cores under smoke.

Fire labels occupy only 431,427 grid cells. Median fire area on a positive chip is 79 pixels; most connected clusters are small hotspots. Accuracy is therefore a poor headline metric—precision, recall, and overlap on rare positives matter more.

Mohammadreza Narimani Sentinel-2 SWIR band contrast for active-fire labeling with box plots of B8A B11 B12 and a composite scene of flame, char, vegetation, and smoke
Figure 3. Spectral contrast used for algorithmic active-fire labeling (B8A, B11, B12).
Example Sentinel-2 chips from Mohammadreza Narimani California active-fire dataset showing true-color RGB, SWIR composite, and active-fire masks for York, Castle, Sugar, and Hopkins fires
Figure 4. Example chips: true-color, SWIR composite (B12–B11–B8A), and the SWIR-rule fire mask.
Mohammadreza Narimani figure on rare-class geometry in Sentinel-2 active-fire chips showing fire area per positive chip and small hotspot cluster sizes
Figure 5. Rare-class geometry: why overall accuracy is misleading when fire pixels are sparse.

Analyst Review and Reference Model

An analyst reviewed 233 of 432 test chips without seeing the rule masks. Chip-level agreement was 89.7% and rule-component confirmation was 88.5%. Those annotations stay separate from the training labels so users can score agreement independently.

The companion repository includes dataset construction code and a ResNet-34 U-Net reference run with validation-based threshold selection. Dataset version 1.0.0 is pinned on Zenodo; software is MIT-licensed on GitHub. Use those versions if you want comparisons tied to this paper.

Mohammadreza Narimani 40-epoch ResNet-34 U-Net reference run showing training and validation loss and validation fire-class IoU for Sentinel-2 active-fire segmentation
Figure 6. Reported 40-epoch reference run: training/validation loss and validation fire-class IoU.
Mohammadreza Narimani reference-model agreement with SWIR-rule labels and independent analyst review of Sentinel-2 active-fire test chips
Figure 7. Reference-model agreement with rule labels and independent analyst review (233 of 432 test chips).
Mohammadreza Narimani ResNet-34 U-Net predictions on held-out Hopkins and Slater Sentinel-2 chips compared with SWIR-rule fire masks
Figure 8. Held-out predictions for Hopkins and Slater: SWIR composite versus model overlay (magenta) and rule mask (green).
Mohammadreza Narimani bar chart of active-fire pixel counts by California wildfire incident grouped into train, validation, and test partitions
Figure 9. Active-fire pixel-observations by incident, grouped into the published train, validation, and test splits.

Fixed inputs

PNG images and integer masks with chip metadata, so methods can share the same files without reprocessing Sentinel-2 scenes.

Grouped evaluation

Wildfire-level partitions prevent the same incident from appearing in both training and test.

Rare-class design

Documented imbalance, invalid-pixel coding, and analyst review for honest segmentation metrics.

Paper and Citation

Mitra, S., Narimani, M., & Farajpoor, P. (2026). A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires, July 2020–August 2026. Data in Brief, 69, 113357.

DOI: 10.1016/j.dib.2026.113357 · Data: 10.5281/zenodo.22713948 · Open access under CC BY 4.0

Please cite the publisher version when you use this work. The summary on this page is a research card, not a substitute for the article.

Cite and read the paper Zenodo dataset GitHub software

Contact

Mohammadreza Narimani
PhD Candidate, University of California, Davis

📧 mnarimani@ucdavis.edu  |  🎓 Google Scholar