A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires, July 2020–August 2026
- 1California High School, San Ramon, CA, 94583, USA
- 2Department of Biological and Agricultural Engineering, University of California, Davis, CA, 95616, USA; Corresponding author.
- 3Department of Biological and Agricultural Engineering, University of California, Davis, CA, 95616, USA
Abstract
This article describes an open image dataset for active-fire segmentation in satellite imagery. It contains 2148 image-mask pairs from 25 California wildfires, acquired from 23 July 2020 to 21 August 2026. Each image is a 512 × 512-pixel, three-channel composite of Sentinel-2 Level-2A bands B12, B11 and B8A at 20 m sampling. The same linear rendering is applied to every image. Masks distinguish background, SWIR-rule active fire and invalid observations. They were generated from shortwave-infrared brightness and near-infrared contrast, followed by two steps of neighborhood growth. The release includes chip-level metadata and a partition with 18 training, three validation and four test wildfires. Each wildfire belongs to only one partition. Of the image pairs, 841 contain active-fire labels, which occupy 0.0766% of all grid cells. An analyst reviewed 233 test chips without seeing the rule masks. These annotations provide a separate assessment of label agreement at chip and connected-component levels. The accompanying code includes a ResNet-34 U-Net with checkpoint and probability-threshold selection on validation data. The images and masks provide a fixed input for comparing segmentation methods on wildfires excluded from training. Dataset version 1.0.0 is deposited on Zenodo; the preparation and training software is available on GitHub.
Versions
Related
- Learning Curve and Detection Accuracy of Sentinel Lymph Node Biopsy for Endometrial Cancer Staging: Experience from an ESGO & GESEA Certified Center
- Benchmark Readiness of Open-access Musculoskeletal Radiograph Datasets for Foundation Model Segmentation: A Systematic Audit
- Adaptive Post-Processing Recovers Most of the Gap to nnU-Net v2 in Head and Neck GTV Segmentation: A Paired Three-Arm HECKTOR 2025 Benchmark
- Predicting MCI-to-Alzheimer’s Conversion: A Systematic Benchmark of Classical, Machine Learning, and Deep Survival Models
- Timing and Context Features in Machine Learning Classification of Inter-Patient ECG Heartbeats on the MIT-BIH Benchmark
BibTeX
@article{mitra2026sentinel2,
title = {{A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires, July 2020–August 2026}},
author = {Shreyan Mitra and Mohammadreza Narimani and Parastoo Farajpoor},
journal = {Data in Brief},
year = {2026},
month = aug,
doi = {10.1016/j.dib.2026.113357},
url = {https://doi.org/10.1016/j.dib.2026.113357},
note = {CC BY},
}
RIS
TY - JOUR
TI - A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires, July 2020–August 2026
AU - Shreyan Mitra
AU - Mohammadreza Narimani
AU - Parastoo Farajpoor
PY - 2026
DA - 2026/08/01
JO - Data in Brief
DO - 10.1016/j.dib.2026.113357
UR - https://doi.org/10.1016/j.dib.2026.113357
AB - This article describes an open image dataset for active-fire segmentation in satellite imagery. It contains 2148 image-mask pairs from 25 California wildfires, acquired from 23 July 2020 to 21 August 2026. Each image is a 512 × 512-pixel, three-channel composite of Sentinel-2 Level-2A bands B12, B11 and B8A at 20 m sampling. The same linear rendering is applied to every image. Masks distinguish background, SWIR-rule active fire and invalid observations. They were generated from shortwave-infrared brightness and near-infrared contrast, followed by two steps of neighborhood growth. The release includes chip-level metadata and a partition with 18 training, three validation and four test wildfires. Each wildfire belongs to only one partition. Of the image pairs, 841 contain active-fire labels, which occupy 0.0766% of all grid cells. An analyst reviewed 233 test chips without seeing the rule masks. These annotations provide a separate assessment of label agreement at chip and connected-component levels. The accompanying code includes a ResNet-34 U-Net with checkpoint and probability-threshold selection on validation data. The images and masks provide a fixed input for comparing segmentation methods on wildfires excluded from training. Dataset version 1.0.0 is deposited on Zenodo; the preparation and training software is available on GitHub.
SN - 2352-3409
ER -