Methodology
How the score is built and what it refuses to claim
Indices are heuristic. Predictive accuracy and occurrence-probability calibration are not established, and every number says which version made it.
The versioned scientific model cards provide current formulas, synthetic arithmetic examples, coverage and limitations. Indices are heuristic; predictive accuracy and occurrence-probability calibration are not established. Physical version 5, historical version 4 and news attention version 1001 are separate contracts.
Data Pipeline
Every event passes through an 11-stage normalization pipeline before it appears on the platform. This ensures consistent, comparable data regardless of the original source format.
- 01Source IngestionTiered polling of registered sources; availability varies by source
- 02Format NormalizationConvert heterogeneous formats to unified schema
- 03Coordinate ValidationVerify lat/lng, anti-meridian normalization
- 04Temporal AlignmentNormalize timestamps to Unix epoch milliseconds
- 05Type ClassificationMap source categories to 16 standard disaster types
- 06Severity MappingCompute severity (critical/high/medium/low) with type-specific thresholds
- 07Cross-Source DeduplicationRepresentative-based clustering to merge same events from different sources
- 08Population ImpactGaussian decay model for exposure estimation using city population data
- 09Cascade Detection20 interaction models for secondary risk analysis
- 10CalamityScore ComputationComposite 0-100 score with 4 weighted components
- 11Quality GateMinimum data completeness check for archival and SEO indexing
CalamityScore (0-100)
The physical CalamityScore is a heuristic priority index. Equal values across hazards do not imply equal losses or likelihood. It combines four weighted components:
Type-specific magnitude (e.g., Richter scale, wind speed, AQI, fire radiative power)
Normalized distance-weighted urban contribution; not unique people
Rule-based relative weights of secondary-hazard scenarios
Recorded archive reference where covered, otherwise identified current-batch fallback
Scores range from 0 to 100; their boundaries are index levels, not damage estimates. A legacy data-support index (0–1) combines source weights and field availability; it is not predictive accuracy. Current evidence dimensions identify source support, completeness and freshness separately.
CalamityScore formula
CalamityScore = clamp(round(w_i * I(e) + w_p * P(e) + w_c * C(e) + w_h * H(e)), 0, 100)
where:
- I(e) = intensity score (0-100), type-specific metric mapping
- P(e) = normalized urban exposure index, using the recorded city dataset and distance profile
- C(e) = Cascading risk: max(probValue) * 60 + min(count,4)/4 * 40
- H(e) = recorded archive-reference transform when covered, otherwise current-batch intensity anomaly fallback
- Weights: w_i=0.4, w_p=0.3, w_c=0.2, w_h=0.1
Confidence = source_reliability × 0.6 + 0.2(if pop data) + 0.1(if cascades) + verification_boost + 0.05(base), capped at 1.0
where source_reliability is data-driven (0.65-0.95), verification_boost = 0.1 if 3+ sources, 0.05 if 2 sources
Intensity score table by disaster type
| Type | Primary metric | Score mapping |
|---|---|---|
| Earthquake | Source magnitude | v5: M≤4→0; M4–9→20×(M−4); M≥9→100 |
| Fire | Active pixel count / FRP (MW) | Pixels: ≥715→95, ≥358→80, ≥72→60, ≥15→40; not hectares |
| Cyclone | Wind (kt, Saffir-Simpson) | Cat5=100, Cat4=85, Cat3=70, Cat2=55, Cat1=40 |
| Flood | GDACS alert level | Red=85, Orange=50, else=20 |
| Volcano | Alert level (VONA) | Red=95, Orange=60, else=30 |
| Air Quality | PM2.5 (µg/m³), recorded averaging window | Shared EPA breakpoint mapping; instantaneous proxy is not official daily AQI |
| Tsunami | Wave height (m) | 5m+=100, 2m+=80, 0.5m+=55 |
| Solar Storm | NOAA G/S/R scale | L5=100, L4=80, L3=60, L2=40 |
| Radiation | uSv/h | 10+=95, 3+=70, 1+=45 |
| Asteroid | Torino/Palermo scale | T8+=95, T5+=80, T2+=55, P>0=60 |
| Avalanche | European danger level | L5=95, L4=75, L3=55, L2=35 |
| River Flood | NWIS flood category | Major=90, Moderate=60, Minor=35 |
Calamity Forecast Index (CFI)
CFI is a heuristic 0–100 activity and scenario index for selected country and hazard pairs. It is not a calibrated probability or a validated seven-day forecast. Reproduction requires the recorded model version and input vintage, not only today’s archive.
Relative rule-based attention to secondary hazards
Episodes per reporting day in a 90-day window, log-scaled and bounded
Global calendar-month prior, not locally calibrated seasonality
Short-term deviation from the expected event rate for the region
Recent recorded media attention level, not acceleration
Scores are classified into five levels: critical (80–100), high (60–79), elevated (40–59), moderate (20–39), low (0–19).
CFI computation details
CFI = clamp(0.30 × cascade + 0.25 × return_period + 0.20 × seasonal + 0.15 × trend + 0.10 × news, 0, 100)
Successful outputs retain computation time and a two-hour validity interval. Polling does not guarantee a new successful assessment. The archive baseline is normalized by reporting days.
Selection: at most 100 countries with recorded activity. Missing pairs are unassessed. Observed activity, target scenario and localization are separate channels.
Country Risk Rating
Selected countries receive an activity grade (A–E) and event-severity-based operational status. The grade does not measure structural vulnerability or coping capacity. Expired or missing validity is not current low risk.
Aggregate severity of active events weighted by CalamityScore
Scenario count and maximum relative weight
Cumulative urban index; overlapping contributions are not unique people
Scaled mean CFI, sharing inputs with the other rating terms
Delta detection: the engine compares the previous and current computed state. Thresholds and cooldowns control emitted deltas. Recorded component changes explain direction; a dominant event is not proof of causation.
Rating grades and operational status
Activity index grades
- A: Index 0–20; not proof of no significant event.
- B: Index 21–40.
- C: Index 41–60.
- D: Index 61–80.
- E: Index 81–100; not a quantified loss estimate.
Operational status
- Green: No higher status threshold met by the assessed records.
- Yellow: Recorded severity counts meet the yellow rule.
- Orange: Recorded severity counts meet the orange rule.
- Red: Recorded severity counts meet the red rule; not a prediction of deterioration.
Selected scope: at most 100 countries. Ratings have their own computation and validity interval; do not borrow the CFI timestamp.
Population Exposure Index (PEI)
Urban PEI is a distance-weighted city population index. It does not count affected or unique people and omits places outside the recorded city coverage. Zero means no covered urban contribution, not that nobody is present. Separate Italy-only gridded population uses WorldPop 2025 and available ShakeMap MMI≥5 footprints; see its model card.
PEI = Σ (city_pop × decay(distance) × source_confidence)
Gaussian decay(d) = exp(−3(d/R)²) within the heuristic radius R
source_confidence ∈ [0.7, 1.0] based on monitoring source reliability
Decay profiles by disaster type
| Type | Decay profile | Rationale |
|---|---|---|
| Earthquake | Gaussian | Heuristic circle; no ShakeMap field in the urban PEI |
| Fire | Linear | Heuristic circle; not a mapped burn perimeter |
| Cyclone | Gaussian | Heuristic circle; not a simulated wind field |
| Flood / River Flood | Step / Linear | Heuristic circle; not a floodplain footprint |
| Volcano | Gaussian | Heuristic circle; not an ash transport model |
| Tsunami | Linear | Heuristic circle; not tsunami propagation or run-up |
| Air Quality | Gaussian | Heuristic circle; not an atmospheric plume simulation |
| Radiation | Gaussian | Heuristic circle; not measured dispersion |
| Landslide / Avalanche | Step | Heuristic circle; not mapped runout |
| Disease / Drought / Weather | Step | Heuristic radius from reporting context |
City database: 33,352 cities (pop ≥ 15,000) from GeoNames. KD-tree spatial index for O(log n) nearest-neighbor queries. Source confidence: single-sensor 0.7, 2-3 sensors 0.85, official agency 1.0.
Cross-Source Verification
Legacy verification labels summarize source support. Source labels and co-reporting do not establish editorial or measurement independence:
At least three recorded source labels support the cluster; independence is not established by this count.
Two recorded source labels support the cluster; this is not a calibrated confidence estimate.
Single official primary agency (USGS, NOAA, etc.). Reliable but unconfirmed.
Single non-primary source. Treat with caution.
Magnitude reconciliation
When multiple sources report different magnitudes for the same earthquake, we compute a weighted average:
M_reconciled = Σ(w_i × M_i) / Σ(w_i)
where w_i = 1.5 for primary agencies (USGS, EMSC, etc.), 1.0 for other sources.
Magnitude uncertainty = standard deviation across reported values. Exposed as source_agreement.magnitude in the API.
Location history and experimental loss proxies
Local archive summaries describe recorded activity in the available sample. They do not establish complete observation years, actuarially calibrated return periods, economic losses or insurance suitability. Legacy return-period, Poisson and loss-ratio fields are experimental assumptions; missing coverage must remain unassessed. They are not part of the CFI formula. See the model contracts and validation limitations.
Source Reliability Scoring
Source support combines assigned category priors and available co-reporting statistics. Latency and volume are availability signals, not evidence of measurement accuracy. These weights have not been calibrated as predictive confidence.
Reliability computation
Base reliability by category:
- Official agency (USGS, NOAA, etc.): 0.95
- Scientific network (ISC, IRIS, etc.): 0.88
- Automated sensor: 0.75
- Crowdsourced (SafeCast, PurpleAir): 0.65
Adjusted by: cross-match rate (events confirmed by other sources), reporting latency (faster = bonus up to +0.05), and event volume (more events = more trustworthy, up to +0.05).
Legacy support blending: base × (1 − w) + observed-support term × w; the result is not accuracy.
Cascade Detection
Our cascade engine models 20 interactions between disaster types using a Directed Acyclic Graph (DAG) with max depth 3. Rules assign scenario weights from event characteristics and geographic assumptions. These weights are not calibrated conditional probabilities. Chain weights attenuate multiplicatively (0.7 at depth 1, 0.5 at depth 2+). The six families of interaction, drawn, are on the cascade page.
Examples:
- Earthquake → Landslide (amplifier: 0.3 × (M-4), clamped [0.5, 2.0])
- Earthquake → Tsunami (for submarine M7.0+ events)
- Cyclone → Flood (rule-based scenario, not a precipitation simulation)
- Volcanic eruption → Air quality degradation (rule-based scenario, not simulated dispersion)
- Fire → Debris flow (post-fire hydrophobic soil, up to 6 months)
Complete H1-H20 interaction table
| ID | Trigger | Target | Base weight | Radius | Window | Reference |
|---|---|---|---|---|---|---|
| H1 | Earthquake M7+ | Tsunami | 0.30 | 500km | 1h | Kanamori 1972 |
| H2 | Earthquake M5+ | Landslide | 0.25 | 150km | 24h | Keefer 1984 |
| H3 | Earthquake M6+ | Flooding | 0.10 | 100km | 48h | Seed & Idriss 1971 |
| H4 | Earthquake M7.5+ | Volcanic unrest | 0.05 | 200km | 30d | Manga & Brodsky 2006 |
| H5 | Volcano + glacier | Lahar | 0.60 | 80km | 48h | Major & Newhall 1989 |
| H6 | Volcano coastal | Volcanic tsunami | 0.08 | 300km | 6h | Paris et al. 2014 |
| H7 | Volcano VEI4+ | Air quality | 0.70 | 1000km | 72h | Robock 2000 |
| H8 | Cyclone Cat1+ | Flooding | 0.80 | 300km | 72h | Emanuel 2005 |
| H9 | Cyclone Cat2+ | Landslide | 0.30 | 200km | 72h | Caine 1980 |
| H10 | Cyclone landfall | Tornado outbreak | 0.40 | 500km | 48h | McCaul 1991 |
| H11 | Flood severe | Landslide | 0.20 | 50km | 72h | Iverson 2000 |
| H12 | Landslide major | River blockage | 0.35 | 100km | 7d | Costa & Schuster 1988 |
| H13 | Fire rule threshold | Debris flow | 0.45 | 30km | 180d | Cannon et al. 2001 |
| H14 | Fire major | Air quality | 0.75 | 200km | 72h | Reid et al. 2016 |
| H15 | Solar storm G3+ | Infrastructure | 0.50 | Global | 72h | Pulkkinen et al. 2017 |
| H16 | Solar storm S3+ | Radiation | 0.40 | Global | 48h | Shea & Smart 2012 |
| H17 | Solar storm R3+ | Comms disruption | 0.50 | Global | 24h | Berdermann et al. 2018 |
| H18 | Tsunami | Coastal flooding | 0.90 | 10km | 6h | Synolakis & Bernard 2006 |
| H19 | Flood tropical | Disease outbreak | 0.15 | 100km | 14d | Watson et al. 2007 |
| H20 | Avalanche L4+ | Burial risk | 0.50 | 5km | 24h | Schweizer et al. 2003 |
Spatial conditions: 82 coastal bounding boxes, 146 active volcanoes (36 glacier + 110 notable), tropical band 23.5°S-23.5°N. Chain decay: depth 1 = 0.7×, depth 2+ = 0.5× multiplicative attenuation.
SEO Quality Gate
Not all events generate individual pages. Our quality gate requires a minimum of 5 data points (type, coordinates, timestamp, severity, source, city, score, title) and type-specific severity thresholds to ensure each published page provides meaningful, valuable content.
| Type | Minimum Threshold |
|---|---|
| Earthquake | M ≥ 2.5 |
| Air Quality | AQI ≥ 100 |
| Satellite heat detection (NASA FIRMS) | ≥ 100 MW of radiative power, or ≥ 20 pixels |
| Wildfire declared by an agency | Severity ≥ high |
| Asteroid | Always indexed |
| All others | Severity ≠ low |
Cross-Source Deduplication
When the same event is reported by multiple agencies (e.g., USGS, EMSC, INGV for the same earthquake), our representative-based clustering algorithm merges them while preserving the most data-rich instance. Same-source events are never merged to avoid collapsing genuinely distinct events.
Dedup algorithm details
Algorithm: Representative-Based Clustering
- Grid-cell spatial index (latitude-corrected longitude neighbor search)
- For each event, search neighbor grid cells within type-specific radius
- Match candidates: same type, different source, within temporal + spatial + magnitude thresholds
- Select representative: highest priority source (629 priority entries across 218 sources)
- Merge metadata from all cluster members into the representative
Default thresholds per type:
- Earthquake: 50km, 10min, 0.5 magnitude tolerance
- Fire: 25km, 6h (satellite revisit time)
- Cyclone: 300km, 12h (advisory cycle)
- Flood/River Flood: 100km, 24h
- Air Quality: 30km, 3h (sensor reporting)
16 regional overrides for seismically dense regions (Japan, Italy, Indonesia, etc.) where events cluster more tightly.
Data Accuracy & Scientific Metadata
Event data accuracy depends on the upstream monitoring source. We do not modify source measurements (magnitudes, wind speeds, etc.), we normalize, deduplicate, and enrich. Our confidence metric reflects the completeness of available data, not the accuracy of the source measurement itself.
Where available, events include scientific metadata: magnitude type (Mw, ML, Mb), measurement uncertainties (location, depth, magnitude error), processing status (automatic vs. reviewed), and timestamp provenance (event time, report time, ingest time). This metadata is sourced from USGS, EMSC, and other agencies and exposed via the API for authenticated users.
References
- Gill, J.C. & Malamud, B.D. (2014). Reviewing and visualizing the interactions of natural hazards. Reviews of Geophysics, 52(4), 680-722.
- Keefer, D.K. (1984). Landslides caused by earthquakes. Geological Society of America Bulletin, 95(4), 406-421.
- Kanamori, H. (1972). Mechanism of tsunami earthquakes. Physics of the Earth and Planetary Interiors, 6(5), 346-359.
- Wald, D.J. et al. (1999). Relationships between PGA, PGV, and MMI in California. Earthquake Spectra, 15(3), 557-564.
- Emanuel, K. (2005). Increasing destructiveness of tropical cyclones over the past 30 years. Nature, 436, 686-688.
- Cannon, S.H. et al. (2001). Wildfire-related debris flow from a hazards perspective. USGS Fact Sheet 103-01.
- Iverson, R.M. (2000). Landslide triggering by rain infiltration. Water Resources Research, 36(7), 1897-1910.
- Costa, J.E. & Schuster, R.L. (1988). The formation and failure of natural dams. GSA Bulletin, 100(7), 1054-1068.
- Manga, M. & Brodsky, E.E. (2006). Seismic triggering of eruptions in the far field. Annual Review of Earth and Planetary Sciences, 34, 263-291.
- Pulkkinen, A. et al. (2017). Geomagnetically induced currents: Science, engineering, and applications readiness. Space Weather, 15(7), 828-856.
- Synolakis, C.E. & Bernard, E.N. (2006). Tsunami science before and beyond Boxing Day 2004. Phil. Trans. R. Soc. A, 364, 2231-2265.
- Holland, G.J. (2008). A revised hurricane pressure-wind model. Monthly Weather Review, 136, 3432-3445.
How to Cite
Platform
Calamity.live (2026). Multi-Source Natural Disaster Monitoring Platform with Cascading Hazard Detection. https://calamity.live
Dataset
Calamity.live Event Archive (2026). Available at: https://calamity.live/events
BibTeX
According to Calamity.live, a platform that aggregates real-time data from 80 scientific monitoring sources across 16 disaster types, the physical priority index combines intensity (40%), urban PEI (30%), scenario weight (20%) and recorded reference or batch anomaly (10%). It is a heuristic 0–100 index, not a probability or comparable loss scale. Hazard literature motivates interactions; it does not validate the chosen numeric weights.
Disclaimer: This platform provides algorithmic aggregation of scientific monitoring data for informational purposes. It is not a replacement for official emergency warnings from national meteorological services, civil protection agencies, or other authorized sources. Always follow local emergency services guidance.