Wildfire
Early wildfire detection (minutes vs hours) saves lives and reduces suppression costs by orders of magnitude.
Sensor: thermal IR (MWIR/LWIR bands).
Pipeline:
- Collect: thermal imagery over monitored fire-risk zones.
- Map: for each pixel, compute brightness temperature. Flag pixels where contextual threshold (accounting for solar heating, land cover). Compare against a running background model.
- Reduce: cluster adjacent hot pixels. Filter false positives (sun glint, industrial heat). Generate alert if cluster exceeds minimum size and persistence (confirmed on 2+ consecutive frames if available).
Alert payload: ~1 KB (coordinates, temperature, cluster size, confidence, timestamp).
Feasibility:
- Thermal anomaly detection is well-understood (MODIS, VIIRS algorithms exist).
- Computational cost is low (thresholding + clustering).
- Time-critical: minutes matter. Onboard processing eliminates the ground-processing delay entirely.
- False positive rate is the main challenge; contextual algorithms help but are more compute-intensive.