Interannual Dynamics of Boreal Forest Wildfires (2020–2024) Derived from MODIS-Constrained Sentinel-2 Burned-Area Mapping

Boreal forests are increasingly affected by wildfires, creating a strong need for consistent annual burned-area datasets to support carbon-budget and climate studies. Existing annual burned-area products covering the entire boreal forest are mainly derived from Landsat imagery at 30 m resolution and may still suffer from omission and temporal misallocation in annual mapping. Here, we present BS2BAM, an automated burned-area mapping framework that integrates MODIS fire products and Sentinel-2 time-series imagery for annual mapping across the boreal forest. MODIS active-fire and burned-area products were first used to define wildfire candidate zones (WCZs). Burned and unburned samples automatically generated from the 2023 Canadian fire season were then used to train a random forest model, which estimated burn probability from Sentinel-2 spectral indices within the WCZ. The resulting probability maps were further refined through region growing to produce annual burned-area maps on a 10 m output grid for 2020–2024. Independent validation showed that, relative to Fire_GFL, BS2BAM reduced omission error from 30.82% to 15.98% and improved the Dice coefficient from 73.39% to 86.73%. Spatiotemporal statistics for 2020–2024 indicate that BS2BAM better captured the 2021 Siberian and 2023 Canadian wildfires than Fire_GFL and GABAM. BS2BAM provides a consistent high-resolution annual burned-area dataset for investigating interannual wildfire dynamics in boreal forests and supporting related carbon-cycle and climate studies.