PaperlayerView exact revision on arXiv

Source-linked AI summary

A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires

Shreyan Mitra, Mohammadreza Narimani, Parastoo Farajpoor

arXiv:2609.16199v1eess.IVcs.CVcs.LGeess.SPphysics.geo-ph

TL;DR

Active-fire learning data need explicit records of image selection, label generation and evaluation partitions, while existing resources cover different observation scales. This paper presents paired Sentinel-2 images and algorithmic masks with incident-disjoint partitions and separate analyst review. Across the release, 841 chips contain fire labels, and the reviewed subset reports chip and component agreement; the resource is intended for image-based learning under algorithmic supervision.

  • Problem

    Reusable active-fire learning data require explicit documentation of image selection, label generation and evaluation partitions, while existing resources address different observation scales.

    Source: §2 · p. 1

  • Method

    The paper provides paired images and rule-generated masks, an incident-disjoint partition, and a separate mask-blind analyst review.

    Source: §1 · p. 1, §4.5 · p. 11

  • Results

    841 chips, or 39.2% of all chips, contain at least one active-fire cell; the review yielded 89.7% chip agreement and 88.5% rule-component confirmation.

    Source: §3.4 · p. 3, §4.5 · p. 11

  • Takeaways & Limitations

    The release supports image-based active-fire segmentation research, including comparisons on fires excluded from training and assessment of algorithmic labels.

    Source: §1 · p. 1

    The labels derive from the same optical bands as the inputs, so model agreement measures reproduction of the reference rule; review does not establish field-confirmed boundaries or inter-analyst reliability.

    Source: Limitations · p. 11

Abstract

from arXiv · show

This article describes an open image dataset for developing and evaluating active-fire segmentation methods in satellite imagery. The dataset contains 2,148 image-mask pairs from 25 California wildfires, with acquisitions spanning July 2020 to August 2026. Each image is a 512x512-pixel, three-channel composite derived from Sentinel-2 Level-2A bands B12, B11 and B8A at 20 m spatial sampling. A fixed linear rendering is applied throughout the dataset. Corresponding masks distinguish background, SWIR-rule active fire and invalid observations. The masks were generated from shortwave-infrared brightness and near-infrared contrast, followed by constrained neighborhood growth. The release includes chip-level metadata and an incident-disjoint partition containing 18 training, three validation and four test fires. Among the image pairs, 841 contain active-fire labels; these labels occupy 0.0766% of all grid cells. A mask-blind analyst review covers 233 test chips and provides a separate assessment of the rule-generated labels at chip and connected-component levels. Reference training and evaluation code accompanies the data, including a ResNet-34 U-Net implementation with validation-based checkpoint and threshold selection. The archived images, masks, metadata and review annotations support research on rare-class segmentation, learning from algorithmic labels and transfer across fire incidents. The versioned dataset is deposited on Zenodo, with preparation and reuse software maintained in a public GitHub repository.

1. Value of the Data

The dataset provides shared image-mask inputs for active-fire segmentation, with fire-free scenes and invalid observations explicitly represented. Its incident-disjoint split, sparse labels and separate analyst annotations support method comparison and label assessment.

1. Value of the Data

  • Paired images and masks provide a common segmentation input, including fire-free scenes and an explicit invalid-observation class.Source: §1 · p. 1
  • The incident-disjoint partition supports comparison on fires excluded from training.It is intended for remote-sensing, computer-vision and wildfire researchers.Source: §1 · p. 1
  • Sparse positive labels and varied hotspot sizes support studies of class imbalance, sampling, thresholds and algorithmic supervision.Source: §1 · p. 1
  • Separate analyst annotations support label-agreement assessment and review or relabeling strategies without replacing the rule-generated targets.Source: §1 · p. 1
  • Open preparation and training code documents conventions for rendering, labeling, sampling and evaluation, distinguishing archived-experiment reproduction from updated-dataset construction.Source: §1 · p. 1

2. Background

Active-fire masks describe evidence of ongoing combustion in satellite observations, not eventual burned area. The dataset complements resources at other observation scales with fixed-rendering Sentinel-2 image-mask pairs, incident-level partitions and a documented analyst review.

2. Background

  • An active-fire mask records evidence of ongoing combustion in an observation, not the area eventually burned.Source: §2 · p. 1
  • Reusable Sentinel-2 learning data require records of image selection, label generation and evaluation partitions.The sensor provides shortwave-infrared observations at 20 m sampling for resolving spatial patterns of active fire.Source: §2 · p. 1
  • The dataset complements Landsat-8 active-fire, fire-video and geospatial incident resources with Sentinel-2 pairs, incident-level partitions and analyst review.Its purpose is pixel classification under sparse algorithmic supervision, not a new segmentation architecture; related resources are cited as, and.Source: §2 · p. 1

3. Data Description

The release pairs Sentinel-2 imagery with rule-generated active-fire masks and metadata, organized into incident-disjoint partitions. Its reference evaluation, sparse labels, and separate analyst review provide distinct resources for studying segmentation and label agreement.

3.1. Release organization and partition inventory

  • The archive contains 1,307 fire-free and 841 fire-bearing chips, partitioned by incident into 18 training, three validation and four test fires.The counts describe the archived samples, not fire frequency in California satellite observations.Source: §3.1 · p. 2, Table 2 · p. 2

3.2. Spatial and temporal coverage

  • Chips sample 10.24-km-wide windows within 30.72-km parent windows, producing repeated observations of incident surroundings rather than a continuous statewide mosaic.The mapped perimeters depict incident extents, not instantaneous flame fronts.Source: §3.2 · p. 2, Figure 2 · p. 4

3.3. Image channels, label encoding and metadata

  • The fixed-rendering RGB channels represent Sentinel-2 B12, B11 and B8A at native 20 m spacing; B12 and B11 are SWIR bands and B8A is narrow NIR.A single set of rendering limits applies across chips.Source: §3.3 · p. 3

3.4. Positive-class prevalence and spatial structure

  • Across the release, 431,427 cells carry active-fire labels; test components have a six-pixel median, and summed cells are pixel-observations rather than unique burned-area totals.Components smaller than ten pixels make up approximately 64% of test components but 12% of test fire-labeled pixels; pixel-level and component-level evaluation are distinct.Source: §3.4 · p. 3

3.5. Reference training and label-review outputs

  • At threshold 0.99, the reference model achieves test fire-class IoU 0.837, precision 0.897 and recall 0.926 against SWIR-rule masks; analyst review separately covers 233 test chips.The reference scores exclude invalid observations; the review reports agreement with algorithmic labels at chip and component levels.Source: §3.5 · p. 3

3.6. Incident contributions

  • Active-fire pixel-observations vary by incident with extent, sampled dates, cloud or smoke conditions, and active-front location relative to the sampling window.The inventory supports distinguishing pooled pixel agreement from evaluation that weights incidents equally.Source: §3.6 · p. 4

4. Experimental Design, Materials and Methods

The study defines a reproducible Sentinel-2 sampling and rendering pipeline, generates fire masks with a bounded SWIR hot-target rule, and documents incident-level partitions, analyst review, and a reference segmentation workflow.

4.1. Incident selection and acquisition sampling

  • The sample comprises 25 queried California incidents selected by area and date criteria, with up to 14 Sentinel-2 acquisition dates chosen at approximately even intervals through available scenes.The queried incident sample is not a complete census of California fires; scene-level cloud cover up to 95% was allowed, and cloud and smoke pixels were not masked out.Source: §4.1 · pp. 4–5

4.2. Radiometric conversion and image rendering

  • Images use converted B12, B11, and B8A values on each scene’s 20 m grid, rendered in fixed B12-B11-B8A order with band-specific display limits.Rendering limits are display transformations, not label thresholds; the PNG archive is not a temporally coregistered multiband geospatial stack.Source: §4.2 · pp. 5–6

4.3. Algorithmic label generation

  • The dataset-specific SWIR rule selects high-contrast seed pixels from converted bands, then expands them by at most two four-neighbor steps within candidate pixels.An eight-connected sieve follows growth; valid cells receive fire or background labels, while cells failing the three-band validity test receive 255, not a cloud label.Source: §4.3 · pp. 7–8

4.4. Tiling, sampling and evaluation partitions

  • Incident-name assignment keeps observations from the same incident in one partition, while chip selection retains all fire-bearing chips and subsamples negative chips by integer stride.Incident grouping does not certify a geographic buffer between distinct incidents; users reconstructing geospatial data should retain source identifiers, coordinate systems, and grid transforms.Source: §4.4 · pp. 8–10

4.5. Analyst-review protocol and agreement

  • The mask-blind analyst reviewed fire-stratified test chips using SWIR composites, marked visible active-fire cores, and compared annotations with rule masks after excluding invalid cells.Component confirmation requires any overlap and does not require one-to-one correspondence; its value can depend on annotation width.Source: §4.5 · pp. 10–11
  • The 233-chip review yielded 89.7% chip agreement and 88.5% rule-component confirmation, but these descriptive results are not field-confirmed detection probabilities.Approximately 30% of reviewed rule-positive chips were visually unconfirmed; those chips contained approximately 1% of the reviewed batch’s rule-positive pixels.Source: §4.5 · p. 11

4.6. Reference segmentation configuration

  • The reference model is a ResNet-34 U-Net trained with on-the-fly augmentation and weighted cross-entropy plus soft-Dice loss, with invalid cells excluded from optimization.Copy-paste donors come only from positive training chips, and mask handling preserves categorical values during augmentation.Source: §4.6 · p. 11

4.7. Checkpoint selection and reference evaluation

  • Checkpoint and threshold selection maximize pooled validation fire-class IoU, and reported test scores use that validation-selected operating point.IoU counts are pooled across valid cells, so the reported score is not a mean of per-chip IoUs and measures agreement with algorithmic masks.Source: §4.7 · p. 11

Limitations

The release supports image-based learning from algorithmic supervision, but its labels, sampling, rendered imagery, and evaluation partitions constrain what its results establish.

Limitations

  • Model agreement measures reproduction of the optical-band-derived rule, while the review of 233 test chips does not establish field-confirmed boundaries or inter-analyst reliability.The analyst review is a separate visual assessment, not field confirmation.Source: Limitations · p. 11
  • The sampled incidents are large California fires, and incident holdouts do not guarantee separation of neighboring landscapes or complete perimeter coverage.The fixed parent windows may omit parts of fire perimeters.Source: Limitations · p. 11
  • Table 5 reports the public script’s settings and operating point, but new training runs should record their actual backend, software versions, and saved checkpoint.The distributed settings describe the reference run, not necessarily a subsequent run.Source: Table 5 · p. 12
  • Optical acquisitions do not sample continuous or diurnal fire evolution, and 8-bit PNG channels are not substitutes for radiometric assets or embedded geospatial transforms.Geographic transfer, radiometric retrieval, and operational detection require reconstructed source representations and task-suited independent validation.Source: Limitations · p. 12

Ethics Statements

The work used satellite imagery and publicly accessible incident-perimeter records, with no human participants, animal experiments, or personally identifiable research data.

Ethics Statements

  • The study involved no human participants, animal experiments, or personally identifiable research data; analyst annotations are research data-curation outputs.Source imagery and perimeter attribution are retained in the dataset documentation and figure captions.Source: Ethics Statements · p. 12

CRediT Author Statement

The CRediT statement assigns conceptualization, methodology, software, validation, analysis, investigation, data curation, writing, and visualization primarily to Shreyan Mitra.

CRediT Author Statement

Funding

The research received no specific grant funding.

Funding

  • The research received no specific grant from funding agencies in the public, commercial, or not-for-profit sectors.Source: Funding · p. 12

Data Availability

The dataset and its companion software are publicly available through Zenodo and GitHub, respectively.

Data Availability

  • Version 1.0.0 of the California Sentinel-2 Active-Fire Segmentation Dataset is deposited on Zenodo.The deposit includes the image-mask corpus and associated metadata.Source: Data Availability · p. 12
  • Companion preparation, training, and review software is available in a public GitHub repository.Source: Data Availability · p. 12
  • The data are licensed under CC BY 4.0, while the code is licensed under MIT.Source: Data Availability · p. 12

Related papers

Loading 2609.16199v1…