Risk-Aware Early Wildfire Detection from Satellite Imagery Guided by Fire Weather Indices

Overview

Wildfires are becoming more severe, so detection systems need to be both fast and reliable. Although weather satellites can scan California every few minutes, public wildfire alerts are often delayed 20–30 minutes or more, allowing fast-moving fires to grow significantly before response begins. Most satellite detection methods rely on fixed thermal thresholds and do not account for underlying fire weather risk. This project develops a risk-aware wildfire detection framework that combines meteorological fire danger modeling with real-time GOES satellite imagery to speed up alerts while limiting false alarms. First, daily Fire Weather Index (FWI) values for California (2015–2024) were computed and standardized (FWI_z). These were linked to CAL FIRE ignition records to model fire start probability using logistic regression. Higher fire danger significantly increased ignition likelihood. Raising the threshold from FWI_z ≥ 0 to FWI_z ≥ 3 improved precision from 0.548 to 0.750 while reducing alert frequency from 21.9% to 2.0%, quantifying the tradeoff between sensitivity and operational burden. Second, GOES-18 thermal imagery from the 2025 Palisades Fire was processed using brightness temperature differencing and object-level grouping. The detector achieved an F1 score of 0.735. External validation on GOES-16 imagery from the 2018 Woolsey Fire yielded an F1 score of 0.565, detected the first labeled fire within 30 minutes, and produced zero false alarms during five hours of pre-fire imagery, demonstrating cross-satellite generalization. A risk-aware adaptive threshold policy dynamically adjusted detection sensitivity based on fire danger. Under elevated conditions (FWI_z ≈ 2), detection delay decreased from 30 minutes to 0 minutes without introducing pre-ignition false alarms. If this approach had been implemented during the Palisades Fire, it would have detected the fire about 30 minutes sooner; under simple spread-rate assumptions, that time advantage corresponds to roughly 2.5–23 km² (~600–5,700 acres) of avoided potential exposed area, illustrating why reducing latency matters in real incidents.

Competition history

• CSEF 2026 Environmental Engineering (Track 2) (Senior Division) · Entry S-12-07

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Environmental Engineering projects
• CSEF Environmental Engineering projects
• Every CSEF 2026 project