FireFlash SpaceSherpa FireFlash
Initializing...
Fire Detections
High confidence
Nominal
Low confidence
Lightning Proxy
Thunderstorm warning centroid
Drought
D0 Abnormally Dry
D1 Moderate
D2 Severe
D3 Extreme
D4 Exceptional
Welcome to FireFlash
Spot the lightning that becomes wildfire — before the fire grows.
FireFlash watches NASA satellites and weather feeds in real time, then ranks every fire detection by how likely it is to become a serious wildfire. The riskiest events float to the top.
Left panel
Ranked list of every fire happening now. Click any row to dive in.
Map (center)
Live fires plus drought, fire-weather alerts, and recent thunderstorms.
Detail (right)
Why the AI flagged this fire. Plain-English score plus the data behind it.
Try this: Click any row in the left panel. The map flies to it and the right panel explains the risk. Hover over labels for what each metric means.
EXTREMEact now
HIGHmonitor
MODERATEwatchful
LOWroutine
MINIMALlog only
Version Notes
SpaceSherpa FireFlash release history
v2.5 April 2026
  • Phase 2e XGBoost model (v5.0, 21 features, 70 trees) live: AUC-ROC 0.9549, AUC-PR 0.7096, Brier 0.0217
  • Cumulative training feature stack: USDM drought, SRTM 30m terrain (elevation, slope, aspect), NLCD fuel receptivity, ERA5 reanalysis weather (humidity, wind, temp, precip)
  • Detail panel now shows live elevation, slope, aspect, weather inhibitor reason, and SHAP per-feature attribution from /api/explain
  • Snow/winter detector: tier auto-downgrades EXTREME -> LOW for active snow forecasts; pre-inhibitor tier preserved for transparency
  • Drift surface and ground-truth feedback endpoints added (/api/drift, /api/outcome)
v2.1 April 2026
  • XGBoost ML model (53 trees, 25 features) deployed to Cloudflare Worker edge inference
  • Calibrated probability via 63-breakpoint isotonic regression (test AUC-ROC 0.9221)
  • Trained on 2.17M FPA FOD lightning fires + 137M GOES-16 GLM flashes (2018-2020)
  • ML prediction shown alongside heuristic for A/B comparison in detail panel
  • Pure-JS tree walker with float32 parity: max divergence vs Python 1e-6
  • Worker API now reports model metadata at /api/health
v2.0 April 2026
  • Added GLM lightning proxy layer (Severe Thunderstorm Warning centroids)
  • Added comprehensive User's Guide with embedded research
  • Added Version Notes overlay
  • Added 4 Literature Review reference popups
  • Lightning proximity scoring factor integrated into heuristic engine
  • Enhanced UI with guide and reference system
  • Footer status bar expanded with lightning feed indicator
v1.0 April 2026
  • Initial release
  • NASA FIRMS VIIRS fire detection integration
  • NWS Red Flag Warning overlay
  • US Drought Monitor overlay
  • Heuristic ignition risk scoring (0-100)
  • Dark theme MapLibre GL JS interface
  • Real-time 5-minute auto-refresh
SpaceSherpa FireFlash
See the spark before the fire
What is FireFlash?

FireFlash predicts where lightning is most likely to start a wildfire. It watches NASA satellite fire detections, drought conditions, weather alerts, and lightning indicators in real time. For every fire it sees, an AI model trained on 137 million historical lightning flashes ranks how likely that detection is to grow into a serious wildfire.

The riskiest events float to the top of the list on the left. Click any one to see exactly why the model flagged it.

How To Use This Tool (60 seconds)

1. Look at the left panel. It lists every active fire detection, ranked by ignition risk. The colored badge on each card is the score (0 to 100); the higher and redder, the worse.

2. Filter if you want. Use the tabs at the top of the list (All / Critical / High / Hi-Conf Fires) to narrow what you see.

3. Click any card. The map zooms to that fire and the right panel opens with a plain-English summary at the top, then the data behind the score.

4. Hover anything you don't recognize. Almost every label has a tooltip explaining what it means.

5. Adjust the time window. Top-right buttons (1h / 6h / 24h / 48h) control how far back the satellite hits go. VIIRS satellites only pass overhead about every 12 hours, so 24h is the most useful default during quiet periods.

Reading The Risk Tiers

Every detection lands in one of five tiers. Plain language:

TierScoreWhat it means
EXTREME80-100Strongest set of ignition signals the model can detect. Pre-position resources; alert local responders.
HIGH60-79Several risk factors line up. Worth monitoring closely.
MODERATE40-59Mixed signals. Stay aware; no immediate action needed unless conditions shift.
LOW20-39Active hotspot, but surrounding conditions don't favor ignition right now. Routine monitoring.
MINIMAL0-19Weak risk signals. Log only.

If the model says EXTREME but the weather where you are is snowing or raining hard, the system automatically downgrades the tier and shows the reason (e.g. "active snow"). The original tier is still visible for transparency.

What The Map Colors Mean

Fire dots are color-coded by satellite confidence: red = high confidence, orange = nominal, yellow = low. They are not severity; they are how sure the satellite is the hotspot is real.

Drought layer uses the official US Drought Monitor scale: yellow (D0 abnormally dry) → amber → orange → red → dark red (D4 exceptional). Drier ground = more likely to ignite.

Lightning zones are cyan circles on counties under active severe-thunderstorm warnings — a real-time proxy for "lightning was just here."

Fire-weather alerts shade counties under active National Weather Service Red Flag Warnings or Fire Weather Watches. Toggle layers on/off in the top-left of the map.

How It Works

Data Pipeline: FireFlash ingests four primary data feeds in real time. FIRMS active fire detections arrive via CSV API from the VIIRS S-NPP satellite. NWS fire weather alerts (Red Flag Warnings, Fire Weather Watches) load via GeoJSON from api.weather.gov. The US Drought Monitor provides weekly categorical drought severity as GeoJSON. Lightning proxy data comes from NWS Severe Thunderstorm Warning centroids, standing in for direct GOES-16 GLM data until full integration is complete.

Scoring Algorithm: Each fire detection is scored on a 0-100 scale across seven weighted factors:

FRP 20
Temp 20
Conf 15
Drought 25
RFW 15
N 5
Maximum point allocation per scoring factor (total: 100 pts + lightning proximity bonus)
FactorMax PointsHow It Works
Fire Radiative Power20Higher FRP indicates more intense thermal output. Scaled at 0.2 pts per MW, capped at 20.
Brightness Temperature20Temperatures above 320K contribute up to 20 pts. Higher temps suggest active combustion.
Detection Confidence15VIIRS confidence class: high = 15, nominal = 8, low = 3.
Drought Severity25D0 = 3, D1 = 8, D2 = 15, D3 = 22, D4 = 25 pts. Drought-stressed landscapes sustain ignition.
Red Flag Warning15Full 15 pts if the detection falls within an active NWS Red Flag Warning polygon.
Nighttime Detection5Night fires detected by satellite are more likely to be real (fewer false positives from solar reflection).
Lightning Proximity+10 bonusIf a Severe Thunderstorm Warning centroid lies within ~1 degree, up to 10 bonus pts are added.

Tier Classification:

Critical 80-100 High 60-79 Moderate 40-59 Low 20-39 Minimal 0-19
Data Sources
NASA FIRMS (Fire Information for Resource Management System)
VIIRS S-NPP Satellite | 375m resolution | ~3hr latency | CSV API
Near-real-time active fire detections from the Visible Infrared Imaging Radiometer Suite aboard the Suomi NPP satellite. Provides fire radiative power, brightness temperature, confidence class, and precise geolocation for each thermal anomaly.
NWS Weather Alerts
api.weather.gov | Real-time GeoJSON | Red Flag Warnings + Fire Weather Watches
Active fire weather alerts from the National Weather Service. Red Flag Warnings indicate critical fire weather conditions: low humidity, strong winds, and dry fuels. FireFlash uses the polygon geometry to determine which detections fall within alert zones.
US Drought Monitor (USDM)
Multi-agency product | Weekly updates | Categorical D0-D4 | GeoJSON
A collaborative product of NOAA, USDA, and the National Drought Mitigation Center. Provides categorical drought severity from D0 (Abnormally Dry) through D4 (Exceptional Drought). Drought-stressed landscapes have depleted soil moisture and dry fuels, increasing the probability that any ignition source will sustain combustion.
GLM Lightning Proxy (Severe Thunderstorm Warnings)
NWS api.weather.gov | Centroid extraction | Proxy for GOES-16/17 GLM
Currently, FireFlash uses the centroids of active Severe Thunderstorm Warning polygons as a proxy indicator for lightning activity. This provides a coarse but operationally useful signal about where convective storms with lightning potential are occurring. Future integration: Direct GOES-16/17 Geostationary Lightning Mapper (GLM) data via AWS S3, providing individual flash event geolocation at ~8km resolution with 20-second update cadence.

Data Tiering Architecture:

TierSourcesResolutionCadence
Global SynopticMODIS NDVI, SMAP soil moisture9-25 kmDaily to 3-day
Regional DynamicVIIRS, MODIS LST, Sentinel-2375m - 1 kmDaily
High-Value LocalSentinel-2, Landsat, ECOSTRESS20-70m3-5 days
Real-Time Active FireVIIRS/GOES thermal, GLM2 km (GLM), 375m (VIIRS)10-30 min
What Makes This Novel?
1
Pre-Ignition Focus
Most fire detection systems are post-ignition: they alert after fires are already burning. FireFlash is pre-ignition focused, identifying WHERE fires are most likely to START and SUSTAIN. This enables preventive resource positioning rather than reactive response.
2
Landscape Condition Integration
Integrates satellite-derived landscape condition (soil moisture proxies via drought, vegetation stress) with active fire intelligence. Drought severity acts as a fuel-readiness indicator, amplifying risk scores in areas where even small ignition sources can sustain combustion.
3
Research-Grounded Scoring
Heuristic scoring inspired by composite fire risk indices from peer-reviewed research. The weighted combination approach draws from Lit Review 2 Section 7, where composite indices weight soil moisture 25-30%, LFMC 25-30%, drought 15-20%, temperature 10-15%, and vegetation 10-15%.
4
Resolution Paradox Bridge
Bridges the "resolution paradox" (Lit Review 2 Section 9): uses coarse global data for synoptic awareness while flagging areas needing fine-scale assessment. Aligns with the operational finding that "most fire management does not require intra-daily temporal frequency."
5
Addressing the Deployment Gap
Fewer than 5-10% of ML fire prediction models have been deployed operationally (Lit Review 1 Section 7.1). FireFlash starts with a transparent heuristic approach that can be validated and incrementally enhanced with ML, avoiding the common trap of building models that never leave the lab.
6
Prevention-Focused DSS
Prevention-focused decision support systems offer strategic advantages in cost-effectiveness and community resilience (Lit Review 3 Section 3.4). By shifting the decision point upstream of ignition, FireFlash enables actions that are orders of magnitude cheaper than suppression.
Research Foundation

Four comprehensive literature reviews were conducted in April 2026 to inform the design, scoring methodology, data architecture, and user experience of FireFlash. Together, they synthesize findings from over 150 peer-reviewed studies spanning remote sensing, wildfire prediction, decision support systems, and AI/ML architectures. Click each review below to explore the full summary.

01
Lightning-to-Ignition Prediction
50+ studies (2019-2026) on holdover fires, dry lightning, LFMC thresholds, ML model performance
›
02
Satellite-Derived Pre-Fire Landscape Condition
7 thematic areas covering FAW, SMAP, Sentinel-2 LFMC, ECOSTRESS, composite indices
›
03
Wildfire Decision-Support Systems and UX
8 research areas: ethical visualization, IoT alerting, digital twins, cognitive load, DSS design
›
04
AI/ML Architecture for Prevention-Focused Fire Prediction
XGBoost, Random Forest, CNN-LSTM, transfer learning, SHAP analysis, class imbalance handling
›
Future Roadmap
Planned Enhancements
  • Direct GOES-16 GLM Integration via AWS S3 for individual lightning flash geolocation at ~8km resolution
  • XGBoost/Random Forest ML Model replacing heuristic scoring (87%+ accuracy target per Lit Review 1)
  • Sentinel-2 LFMC Layer at 20m resolution for live fuel moisture content mapping
  • ECOSTRESS Vegetation Stress integration for water use efficiency and canopy temperature
  • Holdover Fire Prediction for smoldering detection window (35% detected within 1-6hrs)
  • Multi-Day Temporal Risk Trending for identifying escalating risk corridors
Literature Review 1
Lightning-to-Ignition Prediction
April 2026 | 50+ studies (2019-2026) | Holdover fires, dry lightning, ML models, LFMC thresholds
Key Findings
H
Holdover Fire Detection Windows
35% of lightning-ignited fires are detected within 1-6 hours; 28% are detected at 6-24 hours; 14% persist for multiple days before detection. This "holdover" phenomenon is a critical operational challenge, as fires can smolder undetected in duff layers and organic soils for extended periods before flaming combustion begins.
D
Dry Lightning Multiplier
Dry lightning (cloud-to-ground strikes with less than 2.5mm of accompanying precipitation) increases ignition probability 1.5-3x compared to wet lightning. The absence of precipitation means the initial ignition is not extinguished, while the convective conditions that produce dry thunderstorms often include strong winds and low humidity.
L
LFMC Threshold Effects
Live Fuel Moisture Content below 60-70% increases fire likelihood 1.8-fold (up to 2.5x in shrublands). LFMC represents the ratio of water to dry matter in living vegetation, providing a direct measure of fuel flammability. Sentinel-2 derived LFMC achieves R2=0.83 in grasslands.
M
ML Model Performance
XGBoost achieves 87-88% accuracy with AUC 0.90-0.91 for lightning ignition prediction. However, fewer than 5-10% of ML fire prediction models have been operationally deployed, highlighting the gap between research performance and real-world utility.
Top 8 Predictive Variables

Ranked by relative importance score across multiple studies:

Temperature
98
FFMC
95
Precipitation
92
DMC
85
Wind Speed
82
NDVI
78
Dry Lightning
75
LFMC
72
Critical Features Summary
FeatureImportanceMeasurementResolution
TemperatureVery HighWeather station / reanalysisPoint / 0.25 deg
Fine Fuel Moisture CodeVery HighFWI System calculationPoint-based
Precipitation (72h)HighWeather station / radarPoint / 1 km
Duff Moisture CodeHighFWI System calculationPoint-based
Wind SpeedHighWeather station / modelPoint / 3 km
NDVIModerate-HighMODIS / Sentinel-2250m - 10m
Dry Lightning FlagModerate-HighLightning network + precipEvent-based
Live Fuel MoistureModerate-HighSentinel-2 inversion20m
Literature Review 2
Satellite-Derived Pre-Fire Landscape Condition
April 2026 | 7 thematic areas | FAW, SMAP, LFMC, ECOSTRESS, composite indices
Key Findings
F
Fraction of Available Water (FAW)
FAW below 0.50 indicates plant stress. FAW below 0.20 represents extreme drought conditions. FAW below 0.10 marks the driest conditions where ignition sustainability is highest. This metric integrates soil moisture, rooting depth, and soil water holding capacity into a single indicator of landscape water availability.
S
SMAP Soil Moisture
SMAP soil moisture at 9km resolution is adequate for regional fire weather forecasting but insufficient for event-level prediction. The coarse resolution means individual fire starts cannot be predicted from SMAP alone, but regional drying trends provide critical synoptic context.
L
Sentinel-2 LFMC Performance
Sentinel-2 derived LFMC at 20m resolution achieves R2=0.83 in grasslands, R2=0.43 in forests, and R2=0.21 in shrublands. The vegetation-type dependency means a single inversion model cannot serve all biomes; ensemble approaches with biome-specific calibration are needed.
E
ECOSTRESS as Leading Predictor
ECOSTRESS Water Use Efficiency (WUE) is the leading predictor in deep learning wildfire models. Nighttime Land Surface Temperature anomalies are surprisingly more predictive than daytime values, likely because nighttime thermal signatures better capture soil and canopy moisture deficits without solar heating confounds.
Composite Index Weighting

Recommended weight allocation for multi-factor fire risk indices:

Soil Moisture
25-30%
LFMC
25-30%
Drought Index
15-20%
Temperature
10-15%
Vegetation Index
10-15%
Data Tiering for Operational Implementation
TierSourcesResolutionCadenceUse Case
Global SynopticMODIS NDVI, SMAP9-25 kmDaily to 3-dayRegional risk assessment, seasonal trend
Regional DynamicSentinel-2, MODIS LST, VIIRS375m - 1 kmDailyActive monitoring, resource staging
High-Value LocalSentinel-2, Landsat, ECOSTRESS20-70m3-5 daysTargeted assessment, LFMC mapping
Real-Time Active FireVIIRS/GOES thermal100-375mHourly to continuousImmediate detection and response
Literature Review 3
Wildfire Decision-Support Systems and UX
April 2026 | 8 research areas | Ethical visualization, IoT, digital twins, cognitive load
Key Findings
E
Ethical Principles for Wildfire Visualization
Edgeley et al. identified 5 ethical principles from a study of 101 participants: accuracy over aesthetics, uncertainty transparency, avoiding alarm fatigue, equitable information access, and cultural sensitivity in risk communication. FireFlash incorporates these through transparent scoring breakdowns and tiered risk classification.
I
IoT Early Warning Performance
IoT-based fire detection systems achieve alert latency under 10 seconds with communication range exceeding 5 km. These systems complement satellite observations by providing continuous ground-truth data, though they are limited to areas with sensor deployment.
C
CNN Detection Accuracy
Convolutional Neural Network-based fire detection achieves 94.7% accuracy and 92.3% precision in image-based classification tasks. Combined with satellite imagery, CNN approaches can identify thermal anomalies missed by traditional threshold-based detectors.
D
Digital Twin Impact
Digital twin deployment in fire management achieved a 42.8% reduction in detection latency and 38.8% decline in system downtime. These virtual replicas enable scenario planning, resource optimization, and predictive maintenance of monitoring infrastructure.
A
Automation Bias Risk
Automation bias is a documented risk in fire DSS. Systems should provide interpretable explanations, not binary alerts. FireFlash addresses this by showing the full scoring breakdown with individual factor contributions, enabling operators to apply domain knowledge alongside algorithmic assessment.
T
Traffic-Light Risk Communication
Traffic-light categorical risk levels enhance decision speed and reduce cognitive load. Color-coded severity tiers (the approach used by FireFlash) outperform continuous numeric scales for rapid operational decision-making under time pressure.
Pre-Ignition vs Post-Ignition DSS Comparison

Pre-Ignition DSS (FireFlash)

  • Forecast risk before fires start
  • Data: weather forecasts, drought, fuel moisture, lightning
  • Temporal horizon: hours to days ahead
  • Decision: where to pre-position resources
  • Output: risk zones, probability surfaces
  • Value: prevention, cost avoidance
  • Metric: fires prevented or caught early

Post-Ignition DSS (Traditional)

  • Respond to fires already burning
  • Data: active fire detections, smoke, perimeter
  • Temporal horizon: minutes to hours behind
  • Decision: where to deploy suppression
  • Output: fire perimeters, spread models
  • Value: containment, structure protection
  • Metric: acres burned, structures saved
Literature Review 4
AI/ML Architecture for Prevention-Focused Fire Prediction
April 2026 | XGBoost, Random Forest, CNN-LSTM, transfer learning, SHAP analysis
Key Findings
X
XGBoost Performance
XGBoost achieves 85.81% out-of-sample accuracy on lightning-ignited wildfires, with best-in-class results reaching 88.8% accuracy and AUC 0.90-0.91. Its gradient-boosted decision tree architecture handles the heterogeneous feature types (categorical, continuous, spatial) common in fire prediction datasets.
R
Random Forest Baseline
Random Forest achieves greater than 85% accuracy with AUC exceeding 0.90 on harmonized 500m predictors. Its ensemble of decorrelated decision trees provides robust performance with built-in feature importance ranking, making it an excellent interpretable baseline model.
C
CNN-LSTM for Temporal Patterns
CNN-LSTM hybrid architectures achieve AUC-ROC of 0.926 for daily fire danger prediction. The CNN component extracts spatial features from satellite imagery while the LSTM captures temporal sequences in weather and fuel moisture data, enabling the model to learn multi-day drying patterns that precede ignition events.
T
Transfer Learning Across Regions
Transfer learning achieves mean transfer AUC of 0.85 across Mediterranean climate regions. Models trained in one fire-prone region can be fine-tuned for another with limited local data, reducing the need for region-specific training datasets. This is critical for extending predictions to data-sparse areas.
S
SHAP Feature Importance
SHAP (SHapley Additive exPlanations) analysis reveals temperature contributes approximately 30%, precipitation approximately 25%, and vegetation indices 15-20% of model importance. These interpretability tools are essential for building trust with fire managers who need to understand why a model flags a particular area.
I
Class Imbalance Challenge
Ignition events represent only 1-3% of observations in training datasets (approximately 100:1 negative-to-positive ratio). SMOTEENN (SMOTE + Edited Nearest Neighbors) outperforms standard SMOTE for fire prediction by both synthesizing minority samples and cleaning noisy majority examples. Isotonic regression and Platt scaling are critical for probability calibration.
Model Performance Comparison

Benchmark accuracy/AUC across architectures for wildfire prediction:

CNN-LSTM (AUC)
92.6%
XGBoost (Acc.)
88.8%
Random Forest (Acc.)
85%+
Transfer Learning (AUC)
85.0%
CatBoost (Acc.)
84.4%
LightGBM (Acc.)
84.2%
Logistic Reg. (Acc.)
76-82%
SHAP Feature Importance Distribution

Average SHAP value contribution to model output across studies:

Temperature
~30%
Precipitation
~25%
Vegetation Index
15-20%
Soil Moisture
10-15%
Wind / Other
10-20%
Operational Deployment Considerations
ChallengeSolutionImpact
Class imbalance (100:1)SMOTEENN resamplingImproved recall without precision collapse
Probability calibrationIsotonic regression / Platt scalingReliable confidence scores for operators
Regional generalizationTransfer learning with fine-tuningMean AUC 0.85 across regions
InterpretabilitySHAP values per predictionOperator trust and override capability
Temporal dependencyLSTM / sequence modelsCaptures multi-day drying patterns