Data flow, indices and scoring models
Reference document for the territorial risk processing architecture: satellite sources, product levels, spectral indices, cloud masking, fire weather, hydrological and atmospheric composition models, spatio-temporal aggregation and the multi-criteria scoring model.
The system takes observations from 9 European and American operational sources (Copernicus Data Space Ecosystem, EUMETNET, NASA LANCE), harmonises them in a common spatial model (UAT / NUTS-3, EPSG:4326 ↔ EPSG:3857), applies cloud masking based on the SCL band, calculates derived spectral indices (NDVI/NDMI/NBR/dNBR/NDWI) through statistical aggregation over the area polygon (Sentinel Hub Statistical API / openEO), and combines the indicators into a 0-100 composite score across four modules, normalised and prioritised through a weighted multi-criteria model. Data is refreshed every night (cron 03:00 UTC), with full provenance traceability for each indicator.
1. Processing architecture
2. Data sources and product levels
All sources are operational and standardised at European level. Spatial resolutions, revisit cycles and NRT (near-real-time) latencies determine the maximum cadence of the indicators.
| Source | Operator | Product | Resolution | Revisit | Bands/var. | Latency |
|---|---|---|---|---|---|---|
| Sentinel-2 MSI | ESA / Copernicus | L2A (BOA, atm. corrected) | 10 / 20 / 60 m GSD | ≈5 days (2 satellites) | B2-B12, SCL | ≈ 24-72 h |
| Sentinel-3 SLSTR | ESA / Copernicus | LST L2 | 1 km (thermal) | < 1 day | S7-S9 (3.7-12 µm) | ≈ 24 h |
| Sentinel-1 SAR (C) | ESA / Copernicus | GRD / SLC | 5×20 m (IW) | ≈6-12 days | VV+VH | ≈ 24 h |
| VIIRS / MODIS (FIRMS) | NASA LANCE | Active Fire (I-band 375 m) | 375 / 1000 m | ≈4×/day | TIR anomalies | NRT < 3 h |
| OpenAQ | OpenAQ (open data) | Ground station measurements (including RNMCA) | point (stations) | hourly | PM2.5/PM10/O₃/NO₂ | ≈ 1-2 h |
| GloFAS / EFAS (CEMS) | ECMWF / JRC | River discharge (hydrological) | ≈0.05° grid | daily | river discharge | ≈ 24 h |
| MET Norway / Yr | Met.no (Norwegian Meteorological Institute) | Locationforecast 2.0 (numerical model) | ≈1-11 km | hourly | T, RH, wind, precip | NRT |
| MeteoAlarm / ANM | EUMETNET / ANM | CAP-1.2 warnings | NUTS-3 (county) | on issue | 13 types × 4 levels | NRT |
| CLMS (Corine / HRL) | EEA / Copernicus | Land cover, HRL forest | 10-100 m | annual/6 years | CLC classes | batch |
3. Derived spectral indices
Calculated from Sentinel-2 L2A bands (bottom-of-atmosphere reflectance, BOA). The B# notation follows the ESA convention.
| Index | Formula | Interpretation | Range |
|---|---|---|---|
| NDVI | (B8 − B4) / (B8 + B4) | Vegetation vitality / photosynthetically active biomass | [-1, 1] |
| NDMI | (B8 − B11) / (B8 + B11) | Vegetation water content (water stress) | [-1, 1] |
| NBR | (B8 − B12) / (B8 + B12) | Normalized Burn Ratio (pre/post fire) | [-1, 1] |
| dNBR | NBR(pre) − NBR(post) | Burn severity (USGS classes) | [-2, 2] |
| NDWI | (B3 − B8) / (B3 + B8) | Surface water content / water bodies | [-1, 1] |
4. Cloud masking (Sentinel-2 Scene Classification)
Every pixel is classified (SCL band, 12 classes). Clouds, shadows, cirrus and snow are excluded; only clear pixels are kept. The most recent scene with at least 50 valid pixels in the polygon is selected.
5. Fire weather component (Canadian FWI System)
The fire weather index used by EFFIS is derived from 6 sub-components, propagated from temperature, relative humidity, wind and accumulated precipitation.
6. Thematic sub-models
Pluvial risk from forecast precipitation (3 days), soil saturation (volumetric moisture 0-7 cm) and accumulated precipitation (7 days). The fluvial component brings in GloFAS/EFAS river discharge (≈5 km grid) wherever the polygon intersects the river network.
A combined index from the soil moisture anomaly and the precipitation deficit, calibrated on the CEMS Combined Drought Indicator scale. Used as a risk multiplier for fire and vegetation.
PM2.5, PM10, O₃ and NO₂ measured by ground stations (official networks, including Romania's RNMCA), converted into the European Air Quality Index (EAQI, 1-6). The smoke index is estimated from the PM2.5 level in a fire context. Where there is no station nearby, the values are estimated.
Land Surface Temperature from Sentinel-3 SLSTR (thermal channels 10.8/12 µm, split-window), used as a proxy for surface dryness and heat stress on vegetation.
7. Scoring model (multi-criteria)
Each indicator is normalised to [0, 1] through linear scaling on agronomic or meteorological thresholds, then combined through a weighted sum per module. The result is projected onto the 0-100 scale and classified into 5 classes.
| Module | Weights (illustrative) |
|---|---|
| 🔥 Fire | FWI 0.30 · NDMI⁻¹ 0.20 · drought 0.20 · LST 0.10 · fires 0.15 · forest 0.05 |
| 🌿 Vegetation | NDVI 0.35 · NDMI 0.25 · LST 0.15 · drought 0.15 · season 0.10 |
| 💨 Air | PM2.5 0.45 · PM10 0.25 · O₃ 0.20 · NO₂ 0.10 (EAQI scaling) |
| 💧 Water | forecast precip. 0.50 · soil saturation 0.30 · recent precip. 0.20 |
8. Standardisation, aggregation and cadence
EPSG:4326 (WGS84) for storage; EPSG:3857 for rendering; GeoJSON / OGC Simple Features geometries.
STAC for cataloguing, openEO + Sentinel Hub Statistical API for zonal aggregation, CAP-1.2 for warnings, OAuth2 client credentials.
mean of the indicators over the UAT bounding box (NUTS-3 for warnings), weighted by clear pixels.
nightly refresh (cron 03:00 UTC); official warnings every 6 hours; traceability per indicator (real time / modelled).
every value carries its source, timestamp and status (measured / modelled / interpolated).
42 administrative units (41 counties and the Municipality of Bucharest), extendable to UAT level (≈3,200).
9. Validation and limitations
- • Optical indices (NDVI/NDMI) depend on cloud cover; the most recent clear scene is used, which can introduce a variable latency in cloudy periods.
- • GloFAS river discharge has a coarse resolution (≈5 km) and does not capture small torrents; the main component of the Water module is pluvial.
- • The scores are a decision support and prioritisation tool, not an official warning system. The authoritative source for warnings remains ANM / INHGA / IGSU.
- • The model weights are calibrated empirically and can be adjusted for each beneficiary; they are not a regulatory standard.
10. References and standards
Indicative document. The detailed technical specification (ATBD, pseudocode, normalisation thresholds and validation protocol) is available on request, under a confidentiality agreement.