BAM: A Physics-Informed Self-Supervised Framework for Near-Real-Time Wildfire Burned Area Mapping from Multi-Source Earth Observation

The operational monitoring of global wildfires is frequently constrained by the latency of supervised training data generation and the radiometric challenges of distinguishing charred biomass from natural dark surface features. To address these limitations, this study presents the Burn Area Mapper (BAM), a fully automated, physics-informed machine learning framework implemented in Google Earth Engine. The framework employs a Physics-Informed Self-Supervision paradigm that eliminates the need for manual annotation by leveraging combustion stoichiometry to generate scene-specific training labels. Central to this approach is the novel Automated Temporal Burn Index (ATBI), which uses multiplicative scaling of NIR and SWIR bands to enforce a strict bimodal distribution, enabling an adaptive Otsu algorithm to isolate high-confidence burn signals automatically. These physics-derived labels are refined using a Gradient Tree Boost (GTB) classifier that incorporates textural and topographic features. We validated the framework across 15 wildfire events spanning 6 continents, achieving a global mean F1-score of 0.994 without regional retraining. Comparative benchmarking against the MODIS MCD64A1 product revealed that BAM captures up to 93% more burned area in fragmented landscapes. Furthermore, validation against 3-meter PlanetScope imagery (R2 = 0.80) confirms that BAM's probabilistic output effectively proxies sub-pixel burn fraction, resolving fine-scale fire fronts and unburned refugia invisible to coarse-resolution systems. By automating the production of high-precision burn maps in near-real time using robust atmospheric inversion (SREM), BAM offers a scalable, open-source solution for disaster response and post-fire ecological assessment.