Sentinel-1 Validation — 2024 + 2025 seasons

Event-level: model warned (mean_prob ≥ 0.10) in 72 h before SAR-confirmed flood. 99 flood events across 115 Sentinel-1 overpasses · 2025 out-of-sample F1 = 0.885.

Communes — peak risk

Window breakdown

Run metadata

1. Overview

This engine produces a 72-hour probabilistic flood forecast on a 250 m grid (1,956 cells) covering Dangkor District, Phnom Penh. It runs every 6 hours on GitHub Actions using only free/open data. Predictions are physics-based — no machine learning in v1 — and combine a meteorological driver, a hydrological response model with Monte Carlo uncertainty, and regional amplifier signals.

2. Four-Layer Architecture

1. Layer 1 — Meteorological Driver. Multi-model rainfall forecasts (ECMWF IFS, GFS, ICON) pulled from Open-Meteo for 9 sample points on a 3×3 grid across the district, then interpolated to the 250 m grid via Inverse-Distance Weighting (IDW, power = 2).
2. Layer 2 — Hydrological Response. SCS Curve Number runoff per cell, accumulated along a static flow-routing surface derived from Copernicus DEM (GLO-30) + pysheds D8. A Monte Carlo ensemble of N = 30 scenarios perturbs rainfall, curve number, and drainage capacity to produce a flood probability.
3. Layer 3 — Regional Signals. Three external signals provide confidence context: (1) Forecast verification — compares yesterday's Open-Meteo forecast against the observed archive to score model accuracy; (2) JAXA GSMaP NRT — independent satellite rainfall from the joint NASA/JAXA GPM constellation, 0.1° hourly with ~4h latency; (3) GloFAS — Copernicus river discharge forecast (stub, pending integration). JAXA and forecast-verification are displayed as confidence indicators. GloFAS, when active, boosts the regional amplifier to 1.3.
4. Layer 4 — Fusion & Classification. Physics probability is multiplied by a regional amplifier, clipped to [0, 1], then binned into four risk levels.

3. Data Sources

4. Layer 1 — Rainfall (IDW Interpolation)

Sample-point forecasts are interpolated to grid cell i using:

where d_ik is the Euclidean distance from cell i to sample point k in UTM zone 48N (EPSG:32648). Hourly values are aggregated into 6-hour timesteps by summation.

5. Layer 2 — SCS Curve Number Runoff

The NRCS (SCS) Curve Number method (TR-55, 1986) converts rainfall P (mm) to direct runoff Q (mm) for each cell and timestep:

Curve Number is assigned per cell from ESA WorldCover × hydrologic soil group (default group D — clay floodplain for Dangkor).

AMC (Antecedent Moisture Condition) scales CN based on 7-day prior rainfall:

• Dry (< 12.7 mm): factor 0.85
• Normal (12.7–38.1 mm): factor 1.00
• Wet (> 38.1 mm): factor 1.15

6. Flow Routing & Accumulation

Per-scenario re-routing is avoided by pre-computing a static weight surface once (pysheds: DEM fill → D8 flow direction → flow accumulation). Accumulated runoff volume at each cell is:

where V_local = Q · cell_area / 1000 (m³) and A_flow is the upstream contributing cell count.

7. HAND & Inundation

HAND (Height Above Nearest Drainage) is computed by tracing each cell downstream along D8 until a drainage cell (flow accumulation ≥ 100) is reached; the vertical difference is the HAND value.

A cell is flagged as flooded in a timestep if either condition holds:

1. HAND < 2.0 m AND V_acc > drainage_capacity (500 m³)
2. Cumulative local runoff > local_storage (500 m³) — bathtub fallback

8. Monte Carlo Ensemble (N = 30)

Each scenario samples a model + perturbs inputs:

• Rainfall: lognormal, σ ≈ 10 %
• Curve Number: normal, σ ≈ 10 %
• Drainage capacity: normal, σ ≈ 20 %

9. Layer 4 — Fusion & Classification

α_regional = 1.0 by default, boosted to 1.3 when GloFAS indicates active flood conditions. In v0.2, JAXA GSMaP and forecast-verification signals are displayed as confidence indicators only and do not alter α_regional — active amplification is reserved for v1.0 pending re-validation.

10. Commune Aggregation

Each grid cell is spatially joined to one of Dangkor's 13 sangkats (GADM level-3). For each commune:

plus the percent of commune area at L3+ and L4 — used for public-facing messaging and Telegram alerts.

11. Sentinel-1 SAR Validation

The engine was validated against two full monsoon seasons (2024 + 2025) of Copernicus Sentinel-1 SAR imagery processed in Google Earth Engine. Each S1 overpass was converted to a per-cell flood flag using VV-band backscatter change detection (−3 dB anomaly threshold vs. dry-season baseline, JRC permanent-water mask applied). This produced 224,940 cell-image observations across 115 overpasses and 1,956 grid cells.

Metric choice: Cell-level instantaneous accuracy is the wrong measure for a 72-hour early-warning system, because the model predicts a 72-hour peak probability while SAR captures one instantaneous snapshot. The correct question is: "On days when SAR confirmed significant flooding, did the model issue a warning in the 72 hours before?"

The 2025 season was fully out-of-sample (no thresholds were tuned on it) and returned F1 = 0.885 — confirming the model generalises across monsoon years. Warn threshold: district mean_prob ≥ 0.10 in any run within 72 h of the overpass.

12. Limitations

• No live river-gauge assimilation (Mekong/Bassac levels enter only via GloFAS, stubbed in v1).
• Flow routing uses a static weight surface, not dynamic hydraulics — no backwater or levee-overtopping physics.
• Urban drainage infrastructure is represented only by a lumped capacity parameter.
• 13 missed events (of 99) cluster in brief early-season convective storms where rainfall intensity exceeds 72h average signal.
• Forecasts are probabilistic and model-based — this is NOT an official warning.

See LIMITATIONS.md for the full list.