Annual satellite embeddings show potential to simplify burned-area mapping

The researchers demonstrated that annual Tessera embeddings preserve wildfire disturbance signals well enough to support burned-area mapping with simple models. This reduces downstream dependence on curated fire imagery or dense time-series processing.

For teams monitoring wildfire impacts, annual embeddings could simplify regional burned-area mapping and improve detection of small and medium-sized fires relative to the comparison products tested. False positives in unfamiliar landscapes and weaker late-year detection mean local validation remains important.

The study tests whether annual Earth-observation embeddings retain enough wildfire information to replace fire-specific image preparation in burned-area mapping. Tessera embeddings carried a stronger disturbance signal than AlphaEarth, allowing even linear models to distinguish burned from unburned pixels. Models using a single Tessera embedding matched or exceeded equivalent models using paired pre- and post-fire imagery. Across California, the approach recovered 97% of reference burned area without California fire data in downstream training. Separately, a model trained on US fires transferred without retraining to 88 European fires, achieving an F1 score of 0.88. The practical shift is in processing: dense time-series work moves upstream into embedding production rather than disappearing. This could simplify regional mapping workflows, but the reported results do not establish reliable performance everywhere. Detection weakened for late-calendar-year fires, and regional deployment generated systematic false positives in some unseen landscapes.

Evidence supplied by arXiv.org.