Detection
Fire detected
No fire
Circle size ∝ fire coverage
This is a proof-of-concept dashboard for a Sentinel-2 based active wildfire monitoring service. In a production deployment, satellite imagery would be acquired on demand from SentinelHub or the Copernicus Data Space over a configured zone of interest, processed onboard or in the cloud, and alerts would be issued in near real-time.
All imagery comes from the BC Wildfire Sentinel-2 Dataset (Kaggle) — 516 paired RGB and SWIR preview tiles (256 × 256 px) from British Columbia, Canada, acquired between April and July 2025. Fire-labelled tiles were selected by overlapping Sentinel-2 acquisitions with NASA FIRMS hotspot detections; no-fire tiles were randomly sampled within BC boundaries.
Active fire pixels are identified using a two-step SWIR test based on Pierre Markuse's QuickFire algorithm: first a brightness threshold on B11 + B12, then a spectral-shape filter requiring B12 / B11 ≥ 1.4 (active flames emit more strongly in the longer SWIR band than cold bright surfaces). Four confidence levels are assigned based on the SWIR sum. No cloud or water masking is applied to these pre-processed preview images. Accuracy on this dataset: precision 0.95, recall 0.93, F1 0.94.
The Fire Detection overlay stamps confidence-level colors directly onto the natural-color RGB: dark red (low) → yellow (medium) → orange (high) → red (very high). The SWIR composite shows B11 · B12 · B08 as R · G · B — fire hotspots glow yellow-orange.