{"resourceId":"mdpi-fire-goes-deep-learning-wildfire-segmentation-small-fire-omission-20261011","versions":[{"version":"external-bd74f1e160b42e9f6e9921f94089092db103a5d45c5d853987b2cf314324e8a1","resource":{"id":"mdpi-fire-goes-deep-learning-wildfire-segmentation-small-fire-omission-20261011","title":"Peer-reviewed GOES deep-learning study cuts false positives to 8.48% but still misses 39.91% of small fires and does not show earlier discovery","organization":"University of Nevada, Reno","sector":"Wildfire detection research","geography":"Contiguous United States and Hawaii","publishedAt":"2026-10-11","publicationDate":"2026-10-11","eventDate":null,"sourceName":"Fire (MDPI)","sourceLabel":"Fire journal paper on GOES wildfire segmentation with deep learning","sourceUrl":"https://www.mdpi.com/2571-6255/9/10/445","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["infrastructure","operating-model"],"finding":"A peer-reviewed paper in MDPI's journal Fire (vol. 9, issue 10, article 445, published Oct 11, 2026) by University of Nevada, Reno researchers and a Cloudfire Inc. co-author tests a two-step deep-learning method that infers VIIRS-like fire maps and brightness temperature from 5-minute GOES satellite imagery, on 257 U.S. wildfires (2018-2024) with fire-event-level train/validation/test splits (81 test fires, 1,001 test snapshots) and three random seeds. Against an autoencoder baseline, average precision rose from 0.105 to 0.368 and fire-region temperature error fell 72.1% (81.33 K to 22.69 K). Image-level false positives on non-fire images fell to 8.48%, near the operational GOES FDC product's 6.36%, with lower overall omission (31.40% versus 40.35%). But small fires, 37% of the test set, were missed 39.91% of the time. Evaluation used 1-degree regions already centered on known fires, and the authors state that earlier identification of previously unknown fires is not demonstrated; end-to-end latency averaged about 3.4 minutes, mostly NOAA processing. Why now: newly published independent evaluation that qualifies vendor and agency claims about AI satellite wildfire detection in recent editions.","sledRelevance":"Gives fire agencies and state wildfire programs concrete limits to apply when vendors pitch satellite AI detection: good for tracking known large fires every few minutes, weak on small fires and unproven for first discovery.","evidence":"Peer-reviewed study with event-level splits, three seeds and comparison to the operational GOES FDC product; retrospective, known-fire regions only.","architectureImplications":"Satellite-imagery pipeline polling NOAA archives with parallax correction and GPU inference; NOAA ground processing dominates latency.","governanceImplications":"Agencies should treat outputs as statistical estimates, not observations, and keep human confirmation for any dispatch decision.","securityPrivacyImplications":"Uses public satellite data; limited privacy concerns, but dependency on NOAA data availability is an operational risk.","caveats":"Trained to reproduce the VIIRS product, so it inherits VIIRS omissions; partially synthetic regression target; monitoring of known fires only, not discovery; CONUS and Hawaii only; small-fire omission remains high; data available only on request; one co-author is employed by a company in the field; published eight days before this edition. Opened via WebFetch because box curl returned 403.","streamIds":["emergency-services"],"roles":{"sales":"For sales, this paper is a credibility tool: it shows what satellite AI does well (updating maps of known fires within about five minutes) and what it does not (finding new small fires). Use it to set expectations with wildfire programs and to steer conversations toward layered detection with cameras and reports. Never present satellite AI as replacing first-report systems.","engineering":"For engineering, note the evaluation design: event-level splits avoid leakage, but regions were centered on known fires, so discovery performance is untested. Latency is dominated by NOAA processing, not the model. Any proof of value should test full-domain detection, small-fire omission in local fuels, and how outputs integrate with existing camera and CAD alerts.","delivery":"For delivery, frame satellite AI outputs as situational-awareness layers for active incidents, labeled as estimates. Write procedures that require confirmation before resources move, and measure how often the layer changes decisions. Revisit the use case if the vendor cannot show performance on small fires in your region."},"roleDetails":{"sales":{"customerProblem":"Wildfire programs need frequent updates on fire location and intensity, but satellites trade resolution for revisit time.","stakeholders":["State wildfire agency","Fire intelligence officers","Emergency management GIS lead"],"discoveryQuestions":["Is your need first discovery or tracking known fires?","What small-fire miss rate is acceptable?","How do you combine camera, satellite and 911 reports today?"],"valueHypothesis":"Satellite AI could add near-real-time tracking of known fires as one layer in a multi-sensor picture.","potentialEngagement":"Evaluation design for satellite AI tools using local historical fires, including small-fire omission tests.","unsupportedClaims":["Earlier detection of new fires","Reliable small-fire detection","Use as observation for fire modeling"]},"engineering":{"fit":"Near-real-time mapping layer for already-identified fires from GOES imagery.","architectureIntegration":"Polling of NOAA GOES archives, orthorectification, GPU inference and GIS output into incident tools.","prerequisites":["GPU compute","GOES data access","GIS integration","Local validation fires"],"constraints":["Trained on U.S. fires only","Inherits VIIRS reference biases","Latency bound by NOAA processing"],"security":"Protect integrity of the processing pipeline and label outputs as model estimates.","proofOfValue":["Small-fire omission rate on local fires","Full-domain false-alarm rate","Latency from scan to incident map"]},"delivery":{"work":"Pilot satellite AI as a labeled tracking layer for active incidents.","dependencies":["Data feeds","GIS team","Incident management teams"],"ownership":"State wildfire intelligence or GIS unit.","skillsAdoption":"Analysts learn the layer's error modes and communicate uncertainty to commanders.","governanceCheckpoints":["Outputs labeled as estimates","Human confirmation before dispatch","Post-season accuracy review"],"acceptanceCriteria":["Measured small-fire omission on local fires reported before expansion","False-alarm rate on non-fire scenes within agreed threshold"],"risks":["Treating estimates as observations","Missed small fires","Vendor overclaiming discovery speed"]}},"retrievedAt":"2026-10-19T22:07:56-05:00","enrichedAt":"2026-10-19T22:07:56-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Requires analysts who understand model uncertainty; no direct workforce change.","procurementImplications":"Buyers should require local validation metrics, especially small-fire omission and full-domain false alarms.","operatingModelImplications":"Satellite AI supplements, not replaces, cameras and public reports; humans confirm before acting.","sourceVerification":{"responseId":"webfetch-mdpi-fire-9-445-2026-10-19","openedUrl":"https://www.mdpi.com/2571-6255/9/10/445","referenceExcerpt":"Earlier identification of previously unknown fires is not demonstrated in this study.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}