Burnt Area Maps (BAM): Methods & Applications

• Burnt Area Maps (BAM) are geospatial products that classify land into burned and unburned areas using binary segmentation and change detection methods.
• BAM integrates data from sensors like Landsat and Sentinel-2, applying techniques such as bi-temporal change detection and unsupervised extraction for detailed fire mapping.
• These maps support ecological assessment and operational response by providing spatially explicit burned-area delineation and dynamic fire evolution monitoring.

Searching arXiv for the cited BAM papers to ground the article in current literature. Burnt Area Maps (BAM) are geospatial products that represent wildfire-affected land as burned versus unburned space, most commonly as binary rasters or delineated perimeters. In current remote-sensing literature, BAM encompasses several closely related formulations: post-fire binary segmentation from optical imagery, bi-temporal change detection from pre- and post-fire acquisitions, high-resolution unsupervised extraction when labels are unavailable, SAR-based delineation under cloud and smoke, dynamic burned-area evolution derived from active-fire propagation, and even prediction of a fire’s final burned-area footprint before the event is over (Cambrin et al., 2024, Luces et al., 2023, Paugam et al., 30 Oct 2025, Pang et al., 2024). The field is therefore defined less by a single sensor or model class than by a common objective: spatially explicit characterization of burn extent at the scale, latency, and semantic granularity required for ecological assessment, operational response, or fire-behavior analysis.

1. Conceptual scope and output representations

In its canonical form, BAM is a binary segmentation problem in which each pixel is assigned to burned or unburned classes. This formulation is explicit in Sentinel-2 burned-area delineation benchmarks such as CaBuAr, where the target is a binary raster map derived from CAL FIRE perimeters, and in FLOGA and related Sentinel-2 change-detection work, where the output is a pixel-wise burned/unburned mask conditioned on pre- and post-fire imagery (Cambrin et al., 2024, Sdraka et al., 2023). Closely allied formulations treat BAM as change detection rather than single-date segmentation, so that the model learns fire-induced change directly from paired acquisitions rather than inferring it from a post-fire scene alone (Sdraka et al., 2023, Seydi, 9 Sep 2025).

The scope of BAM has broadened beyond retrospective optical mapping. Some studies define the output as a predicted final burned-area raster for an ongoing wildfire, using early fire progression plus environmental drivers to forecast the event-scale binary extent (Pang et al., 2024). Others define burned area dynamically as the cumulative area reached by a reconstructed fire front through time, producing arrival-time maps, evolving perimeters, Burn Area time series, and Fire Growth Rate instead of a single static post-fire scar (Paugam et al., 30 Oct 2025). A further extension appears in CYGNSS-based workflows, where BAM is treated as an active-fire support product generated during the event, rather than only after the event, and then coupled to downstream forecasting and decision-support systems (Roy-Singh et al., 8 Aug 2025).

These variants do not eliminate the classical post-fire burn-scar product; they reframe it. A plausible implication is that “BAM” now names a family of products spanning binary scar delineation, temporally updated burned-extent monitoring, and predictive fire-footprint mapping. The common denominator remains a spatial burned/unburned representation, but the observation model, latency target, and semantics of “burned” vary across retrospective, dynamic, and predictive settings.

2. Observational basis and reference data

BAM has been built from a wide range of Earth-observation modalities. Landsat remains central because it provides 30 m multispectral imagery, a long historical archive, and NIR/SWIR bands suited to burn-scar discrimination; Landsat-based work includes scene-level U-Net mapping in Chile, semi-automatic BFAST-driven annual burn mapping from Landsat-7, and the global 30 m annual GABAM 2015 product implemented in Google Earth Engine (Mancilla-Wulff et al., 2023, Tecuapetla-Gómez et al., 2019, Long et al., 2018). Sentinel-2 supports higher-detail event-scale mapping with multispectral 10–20 m inputs, including paired pre- and post-fire imagery in CaBuAr, FLOGA, multitask delineation benchmarks, AlphaEarth-based embedding workflows, and foundation-model adaptation studies (Cambrin et al., 2024, Sdraka et al., 2023, Arnaudo et al., 2023, Seydi, 9 Sep 2025, Shibli et al., 6 May 2026).

Moderate-resolution sensors remain important where temporal density is decisive. MODIS underlies products such as MCD64A1 and Fire_cci, is used in regional impact studies, and forms the low-spatial, high-temporal branch of BAM-MRCD