New thermal imaging dataset could transform wildfire detection from the air

A new dataset is poised to accelerate machine learning applications in aerial wildfire monitoring. Scientists have compiled FLAME 3, combining synchronized thermal and visual imagery from unmanned aerial vehicles (UAVs) captured during prescribed fires—offering researchers the raw material to build smarter fire detection and assessment tools.

Why thermal imaging matters for fire management#

Radiometric thermal cameras are increasingly affordable for UAVs and can measure per-pixel temperature across a scene. This capability is far more powerful than traditional approaches, which convert raw thermal readings into visible RGB images using arbitrary color scales. Yet the technology has remained largely underutilized in wildfire research, chiefly because high-quality training datasets have been scarce.

The FLAME 3 effort directly tackles this bottleneck by providing the first large-scale, publicly available collection of side-by-side visual spectrum and radiometric thermal imagery captured at real fires. The dataset includes radiometric thermal images stored as TIFFs—a format that preserves precise temperature data—alongside aerial photographs and nadir (straight-down) thermal plots from six separate burn sites.

Streamlined data pipeline#

Beyond the imagery itself, the researchers have developed an automated processing workflow that simplifies the journey from field collection to neural network input. This pipeline handles synchronization, alignment, and preparation steps that would otherwise consume substantial engineering effort, lowering barriers for teams wanting to build fire-detection models.

Access and next steps#

A subset of FLAME 3 is available on Kaggle for computer vision applications, while the complete dataset with all six burns is accessible to researchers upon request. The team hopes the release will catalyze a new wave of AI-driven tools for wildfire detection, segmentation, and real-time assessment—tasks that could become far easier as models gain access to richer thermal data than previously possible.