2026-11-03 –, Compagno
Wildfire response needs continuous intelligence from ignition alerts, monitoring, to recovery. This talk presents an open geospatial ML workflow combining weather satellites, active-fire alerts, news, burn-severity analytics, and high-resolution imagery for lifecycle monitoring and damage assessment.
Wildfires are not a single mapping problem. They move through a lifecycle: heat anomalies and ignition alerts, active-fire monitoring, smoke and plume tracking, burn-severity mapping, structure damage assessment, debris removal, and reconstruction. Each stage has different latency, resolution, uncertainty, and decision needs. This talk presents a full-stack wildfire lifecycle intelligence workflow built around open geospatial data, machine learning, and interoperable GIS outputs.
The system starts with public active-fire feeds and thermal/weather satellite observations. NASA FIRMS active-fire detections provide broad hotspot awareness from MODIS and VIIRS. Geostationary weather satellites add high-frequency context for thermal anomalies, smoke movement, and plume evolution where visibility conditions allow. Online wildfire news and incident feeds are used to turn isolated detections into contextualized fire events: what happened, where it happened, what sources support it, and how confident the system is.
| Lifecycle stage | Sensor / feed type | Typical resolution | Coverage / revisit | Provided value | Key caveat |
|---|---|---|---|---|---|
| Early alert | NASA FIRMS MODIS / VIIRS active fire | ~1 km MODIS; 375 m VIIRS | Global near-real-time; U.S./Canada can be faster | Hotspot and fire-event alerts | Detections are not perimeters and can have false positives |
| Rapid monitoring | GOES / ABI weather satellite | 0.5–2 km, band-dependent | Americas; full disk 5–15 min, CONUS ~5 min, mesoscale as fast as 30 sec | Thermal trends, smoke/plume evolution, cloud context | Coarse pixels; clouds, smoke, and view geometry limit detail |
| Planned burn mapping | Sentinel-2 MSI | 10 m / 20 m / 60 m | Global land; about 5-day constellation revisit | Burn scars, vegetation change, fire edge refinement | Optical imagery depends on cloud-free overpass timing |
| Planned burn mapping | Landsat 8/9 OLI/TIRS | 30 m multispectral; 15 m panchromatic | Global; 16 days each, about 8 days combined | Historical baseline, NBR/dNBR, thermal context | Lower temporal resolution than geostationary feeds |
| Frequent optical change | Harmonized Landsat Sentinel-2 | 30 m harmonized products | Global median repeat frequency about 1.4 days in 2025 | Consistent time series for ML and change detection | Still limited by cloud/smoke and optical acquisition |
| Recovery assessment | Very-high-resolution satellite/aerial imagery | 30 cm-class satellite; 1–3 inch aerial where available | Tasked or licensed collections | Parcel-level structure damage, debris cleanup, reconstruction progress | Commercial data should be optional input, not the FOSS dependency |
For active incidents, the workflow uses predictable overpass windows from Sentinel-2, Landsat, and harmonized Landsat/Sentinel products to schedule higher-resolution analysis. When suitable imagery arrives, it computes NBR/dNBR and applies ML models to identify burned areas, severity patterns, edge changes, and candidate damage zones. The outputs are designed as GIS-ready layers rather than screenshots: polygons, points, rasters, confidence scores, timestamps, provenance links, and reviewer notes.
The recovery phase extends the same architecture to very-high-resolution imagery. Recent urban-wildland interface fires, such as the Eaton Fire, show that geospatial intelligence remains important after containment. With licensed commercial satellite or aerial imagery, the workflow can support parcel-level structure damage detection, debris-removal status checks, and reconstruction progress monitoring. The key is to preserve the same open workflow: transparent inputs, auditable model outputs, human review, and export to standard geospatial formats.
The FOSS4G contribution is the architecture and implementation pattern, not a proprietary dashboard pitch. The proposed stack emphasizes open-source and open-standard components: STAC-style cataloging, cloud-optimized GeoTIFFs, PostGIS, GeoParquet, QGIS-compatible review, Python geospatial libraries, and web maps built from interoperable vector and raster services. Commercial imagery can be used as an optional high-resolution input, but the core pipeline, metadata model, and output layers remain portable.
Attendees will leave with a practical model for building wildfire lifecycle intelligence systems that connect alerting, monitoring, burn analysis, damage assessment, and recovery. The main lesson is that resilience comes from connecting the entire wildfire lifecycle through transparent provenance, uncertainty-aware ML, and open geospatial outputs that responders, researchers, and communities can actually use.
Dr. Fuxun Yu has extensive expertise in geospatial AI, remote sensing, and computer vision. He has published 50+ academic papers in top AI, CV, and ML conferences including NeurIPS, ICLR, and ICML, with 1900+ citations. He graduated with his PhD from George Mason University in 2022. After graduation, he joined Microsoft as Principal Research Manager and led the Geospatial Foundational Model (GFM) project at Microsoft.
Rishi Madhok is a geospatial AI and computer vision leader with deep experience across remote sensing and applied ML. He has authored 18+ papers and patents in top AI/CV venues, including NeurIPS, ICLR, ICML, and CVPR. He earned a Master's degree with a focus on computer vision from Carnegie Mellon University. He started his career building perception systems for self-driving vehicles at Uber ATG and later joined Microsoft, where he served as a Principal Applied Science Manager and led geospatial AI initiatives for Microsoft Planetary Computer.