Flood sensor signals: what six satellite products see, how often, and whether it lines up with reported floods

Generated 2026-09-03 by floodsig html from the processed observation tables in this repository; every number and figure is computed at build time. Part of AIGLE (LIST · OCHA Centre for Humanitarian Data). FloodScan is AER-licensed: derived numbers here are for internal use.

Executive summary

Six satellite flood products were compared over four country cases (Cameroon 2024, Niger 2024, Bangladesh 2024, Mozambique 2026): 37 first-level administrative units, 50,341 unit-day observations, 13 reported flood events giving 39 reported province-events as the yardstick, plus two independent impact anchors (IDMC displacement records, CERF allocations). The question is not which product is best but whether combining products that fail in different ways buys enough continuity of evidence to act on — and the data say: partly, and mostly through one product.

• Combined detection is high but leans on GFM. At least one product flagged 33 of 39 province-events (85 %, 95 % CI 69–97 %). GFM (S-1) alone reaches 22/39; the other five products add 11 province-events between them, and after step 4 of the ladder no product adds a detection. Remove GFM and the union falls to 67 %; the three dating products alone reach 79 %, the three confirming products 51 %.
• Continuity of looks is a different story from continuity of detection. A usable look (≥ 50 % of the province) from at least one product existed on the reported start day for 100 % of province-events and within a week for 100 % — but that is the daily microwave products; the three dating products alone give 90 % on day 0.
• Optical products go blind precisely when it floods. The median province has a good-look rate of 0 % for Cameroon 2024 MODIS, Mozambique 2026 MODIS, Mozambique 2026 VIIRS. MODIS' good-look share inside versus outside reported flood windows: Cameroon 2024 25 % vs 5 %; Niger 2024 49 % vs 85 %; Bangladesh 2024 7 % vs 43 %; Mozambique 2026 38 % vs 24 %. Cloud is reported honestly as 'not observed' (the code divides by visible land), so this shows up as missingness, not as dry ground.
• Latency. Median first-flag latency from the reported start, dating products: FloodScan +16 d, GFDS +0 d, GFM (S-1) +12 d. For the flash-flood province-events, 10 product-hits occurred across 6 province-events; the reference flash flood (Feni–Cumilla, Aug 2024) is a worked example below.
• The tiering is monotone but weakly separated. Share of clusters anchored by a report: strong 47 %, moderate 25 %, weak 12 %. No rain data exist in the repository, so the rain-precursor half of a calibration is not possible.
• The GFM tile-blur caveat, sized: the median province receives only 20 % (Cameroon 2024), 36 % (Niger 2024), 15 % (Bangladesh 2024), 43 % (Mozambique 2026) of its GFM signal from inside its own borders.
• 63 moderate/strong clusters match no report (GFM (S-1) 23, MODIS 20, GFDS 9, CYGNSS 6, VIIRS 4, FloodScan 1 by peaking product). They cluster in rainy seasons; they are 'not in the reports', not false alarms, and the compiled list has known gaps (Mozambique March–June 2026 has none).
• Data changed since the published report. FloodScan now carries the full daily record for Cameroon 2024, Niger 2024, Mozambique 2026 (team DB), replacing the 90-day public window — but without the day-of-year baseline, so its z is an in-period median/MAD; 24 of 29 FloodScan unit series and 25 of 36 MODIS series sit on the absolute MAD floor, where 'z ≥ 2' means '≥ 0.4 % of the unit', an area threshold rather than a statistical one. The headline verdict and hit counts nevertheless reproduce the published ones exactly (Appendix A).
• Independent impact anchors agree with the signals, and adjudicate part of the unexplained set. Against IDMC displacement records (an independent yardstick) the union flags 23/25 province-events (92 %, 95 % CI 75–100 %). Of the 63 moderate/strong clusters with no compiled report, 8 overlap an IDMC displacement record in the same province (Bangladesh 2024 0/29, Cameroon 2024 5/19, Mozambique 2026 0/11, Niger 2024 3/4). The 6 CERF response allocations inside the case periods came a median 36 days after the first dating-product flag in the affected provinces; the 2 anticipatory allocations moved money CERF-NER-24-RR-1398 11 d after the reported start of the event it anticipates and CERF-MOZ-26-RR-1513 2 d before the reported start of the event it anticipates.

1. Project goal

Anticipatory action for floods releases funding on pre-agreed triggers, and those triggers are almost always a forecast river discharge at a gauge. That leaves flash floods — a disproportionate share of flood deaths and displacement — essentially uncovered: forecasts at that timescale barely exist, and global hydrological models are unreliable in many of the places OCHA works. This repository is the observational leg of that problem. If a flash flood cannot be forecast, can it at least be seen reliably, quickly and consistently enough from satellites to (a) open an action window and (b) build a multi-year record of events against which a rainfall-to-flood model could later be calibrated?

The unit of analysis is the first-level administrative unit (ADM1, official OCHA p-code) and the yardstick is a compiled list of reported floods — deliberately imperfect: reports are dated by publication, describe impact rather than water, and miss events. Throughout, a satellite signal with no matching report is "not in the reports", not a false alarm. Two independent impact anchors — IDMC displacement records and CERF allocation dates — are used in Section 6 to test that framing.

The six products, and what they cannot do

They were chosen because they fail in different ways: two passive-microwave products see through cloud but at 10 km; one radar product sees 20 m detail but only on pass days; two optical products attempt daily but are blind under cloud; one GNSS-reflectometry product sees through cloud and canopy but composites 30 days. The analytical consequence is an asymmetry the method must respect: FloodScan, GFDS and GFM can date an onset; CYGNSS, MODIS and VIIRS can only confirm or extend one, because compositing and cloud make their first-detection dates unreliable by weeks.

2. The challenge

• Revisit is not observation. Three products are nominally daily, one is daily-attempted, one observes only on Sentinel-1 passes (median good-look rate 23 %, 23 %, 41 %, 50 % across the cases), and one is a 30-day composite that is always "complete". The relevant quantity for a decision is how stale the latest usable look is on the day the decision is taken.
• Missingness is structured. Cloud removes the optical products when it rains, i.e. when it floods; GFDS only exists at monitored sites (Niamey has 560 km² of monitored area); FloodScan does not exist for Bangladesh at all. A day a product could not see must never be read as "dry".
• The products' normals differ. GFDS ships a static, non-seasonal normal; GFM ships none, so the repo builds a per-tile same-month climatology from earlier years' scene metadata; the others get an in-period median/MAD with an absolute floor. "z = 3" is therefore not one statement across columns.
• Scale mismatch. GFM's numbers are per 300 km tile, so a flood elsewhere in a tile bleeds into a province; a flash flood in one district is diluted across a division.
• The yardstick is soft. 13 reported events, 39 province-events, windows of [start − 10 d, end + 21 d]; known gaps in the compiled list: Mozambique March-June 2026: no events compiled (sources checked stop at the 23 March 2026 response report); Bangladesh July 2024 northern floods: dates/divisions from secondary sources only; Cameroon North region (Benue) 2024: mentioned in some reports, not pinned down — not listed.

3. Data

Cases

What each product's value is, and how its z was computed

The observation table stores one row per (product, unit, day) with value (the product's own flood indicator), observed_fraction (share of the unit actually seen) and, for GFDS and GFM, a product-side z. Everything else is derived in signals.py. The table below states what the code actually does, including two places where the code and its own documentation disagree.

Summary statistics per case and product

Revisit and coverage, in one vocabulary

Good look = at least 50 % of the province observed that day; partial look = at least 20 %; typical gap = median days between consecutive good looks; longest blind stretch = longest run of days without a good look. Medians across provinces. (The published report used "days observed" and "days with a good look" for the same quantity.)

Cameroon 2024

Niger 2024

Bangladesh 2024

Mozambique 2026

Impact datasets added as independent anchors

Described and used in Section 6: IDMC GIDD disaggregated displacement records (2008–2025, filtered to storm and flood disaster records on GIDD's typed hazard fields — landslides excluded, counts and rules there — and coordinates mapped to ADM1 here) and CERF allocations (2006–2026, country level, plus the project-level export that carries disbursement dates and the anticipatory-action marker). Neither was used to compile the report list.

Missing data: how much, where, and whether it is random

4. Exploratory analysis

One block per case: the shared-scale z grid (products × provinces), a province explorer, and the evidence matrix; then three worked examples and the signals that match no report. Reading guide for the grids: dashed line z = 2 (headline), dotted z = 3 (strong); yellow / orange / red bands are the method's weak / moderate / strong episodes; grey bands are reported windows; a break in a line is a day without a usable observation.