Skip to content
Back to projects
March 2026

BAM: Self-Supervised Burn Area Mapping

A self-supervised Earth Engine workflow that combines temporal burn detection and machine learning to map wildfire burned areas at 30-metre resolution without manually annotated training data.

GEEPythonMLGeoAIRS & GIS
BAM framework for wildfire burned area mapping visualization
On this page

Mapping burned areas without hand-labeled training data

BAM is a self-supervised workflow that turns pre- and post-fire Landsat observations into 30-metre burned-area maps. The project addresses a practical bottleneck: detailed manual labels are often unavailable when a fire occurs, while coarse products can miss fragmented burn patterns.

The framework combines a physically motivated spectral change signal with machine-learning refinement in Google Earth Engine. An open application and source repository make the research accessible for exploration and adaptation. It trains a classifier for each scene using automatically generated labels; it is not a model that skips training altogether.

How the workflow works

  1. Prepare the observations. Landsat 8/9 top-of-atmosphere imagery is processed with Simplified Atmospheric Correction (SREM). Pre- and post-fire observations are screened using the workflow’s exclusion masks.
  2. Find a conservative training signal. Change in the Automated Temporal Burn Index (ATBI), based on near-infrared and shortwave-infrared reflectance, is separated using adaptive Otsu thresholding. This produces scene-specific pseudo-labels without manual annotation.
  3. Refine the delineation. A Gradient Tree Boosting classifier uses spectral, textural, and topographic predictors to produce a burn-probability layer and a binary burned-area map at 30 metres.

The separation between label generation and refinement is important: the temporal index supplies an initial signal, while the feature stack helps describe the spatial and spectral setting. Automatically generated labels can still contain errors, so removing manual annotation does not remove the need for independent assessment.

What the evaluation shows

The study evaluates 15 wildfire events across six continents. It reports a mean F1-score of 0.994 under internal spatially blocked assessment against pseudo-labels. The blocks are 3 × 3 kilometres, with a ten-fold assessment design. This is evidence of internal consistency, not an independently measured global ground-truth accuracy of 99.4%.

A separate assessment uses 3-metre PlanetScope imagery for the 2024 Park Fire. After aggregation to 30 metres, the reported relationship between BAM’s probabilistic output and burn fraction has R² = 0.80. This independent check concerns one event and should not be generalized to every ecosystem or fire.

Comparisons with the MODIS MCD64A1 product show up to 93% more mapped burned area in fragmented landscapes. That quantity describes a difference in mapped area. It is not a resolution improvement, and additional mapped area alone does not demonstrate that every extra pixel is correctly classified.

Working with the outputs

The binary map supports delineation and area summaries. The probabilistic output preserves variation that can be useful for examining mixed burned and unburned pixels. Treat the score according to the study’s validation, rather than assuming it is a universally calibrated probability or a direct measurement of burn fraction.

For a new application, document the event boundary, observation dates, masks, and comparison data. Inspect the original imagery alongside the results. Retain the configuration used to create the map so that an area estimate can be reproduced and differences between events can be explained.

Appropriate uses and limitations

BAM is useful for research on fire footprints, fragmented burns, and post-fire landscape assessment where detailed training labels are scarce. Cloud, smoke, acquisition timing, and surface changes unrelated to fire can affect the available signal. Its default cropland exclusion is a conservative wildland-mapping choice; agricultural burning and fires at the cropland–wildland boundary require particular attention to that mask.

Rapidly available imagery supports timely analysis, but the framework does not guarantee a fixed operational delivery time. Wider independent validation remains necessary when transferring the workflow to a new setting or using its outputs for consequential decisions.

Explore and reproduce the work

Open the BAM application to explore the project, and use the source repository for implementation and configuration details. The paper explanation provides the research context; the full paper documents the processing, evaluation, and limitations. Cite the published study when using the research workflow.

Case study updated .

Related Projects

Interested in collaborating or have questions about this project?

Get in Touch