One flood, two satellites, one question
Between 16 and 17 May 2023 the rivers of Romagna broke their banks after two weeks of rain. This story follows one question from the satellite pixel to the economic ledger: what can we actually measure, and what cannot be seen?
The raw material is the Maxar Open Data release for the event: 245 very-high-resolution tiles (0.3–0.7 m), 11 passes from 2015 to 2023, of which a single post-event pass on 23 May 2023. It covers a strip from the coast to the Apennines: 29 municipalities, 588,318 residents. The Bassa Romagna, flooded longest, lies outside it.
Outlined on the map: the municipalities covered by the post-event pass. Scroll to continue; use the arrow keys if you prefer.
The water, as Copernicus saw it
The official footprint comes from the Copernicus Emergency Management Service, which maps flooded areas from satellite radar and optical data within hours. Two activations were needed: EMSR659 (2 May, first wave over the Bassa) and EMSR664 (16 May, the catastrophic wave over Forlì, Lugo, Faenza, Cesena, Ravenna).
delineated flood extent, union of the two activations
polygons; only 23% of the second wave falls inside the Maxar strip
EMS photographs the water at its peak. That matters for the next steps: by the time Maxar flew over, six days later, most of that water was gone.
Before and after
Drag the divider. On the left, the last clean image before the event; on the right, the GeoEye-1 pass of 23 May 2023. The pair is co-registered on the same 0.31 m grid, so what changes is the ground, not the geometry.
On the outskirts of Ravenna the fields turn brown and the watercourses run thick with sediment, while the built-up areas stay dry. A green field in April can also be brown in late May because it was mown or ploughed: the colour of the canal is the least ambiguous signal, and separating mud from agriculture is exactly the first test to pass.
Pairs were not chosen by hand: for each tile the post image with the largest valid area and the most recent pre image among the fullest ones. A pair picked by position alone (Faenza) had 1% valid data.
Detecting the sediment
No model, one rule, calibrated on the Ravenna pair (87% of a known mud block detected, 0% false positives in the urban area), then applied unchanged to all 58 pairs at ~5 m:
The comparison with the pre image removes roofs, bare soils and fields that were already ploughed. What remains undetectable is a field ploughed between the two dates, so the share of mud per municipality is an upper bound.
sediment in Forlì, 8.2% of its surface
of Castrocaro Terme; the Apennines ~0%
Does it hold up?
Pixel by pixel, the mask and the EMS delineation barely overlap: recall 6.2%, precision 11.2%. These numbers are low by construction, not by error. EMS maps the water at the peak (dark, which the mask deliberately excludes); Maxar sees the sediment six days later. They are complementary measures of the same event.
At the municipal level the two converge:
Share of municipal surface: EMS delineation vs sediment mask, municipalities with at least 20% of their surface covered by image pairs.
Landslides in the hills
The same before/after logic, with a different rule: a pixel at ~1.7 m that was vegetated before (excess green ExG ≥ 30) and is bare after (ExG ≤ −5, drop ≥ 50), bright enough to exclude clouds and shadows, grouped into components of 200 m²–10 ha on slopes ≥ 10%. Calibrated on the Modigliana mudflow.
candidate landslides, 47.7 ha, median 406 m²
reliability classes by the date of the pre image
The count is a lower bound with three reliability classes: A, pre image from April 2023 (Modigliana 92, Brisighella 62: the most solid); B, spring pre image from 2020–22; C, blind by construction: an autumn pre image shows bare woods, so the "vegetated before" test fails. The zeros of Tredozio, Marradi and Premilcuore are an artefact of the date, not an absence of landslides.
Who was underneath
Overlaying the EMS footprint on Overture Maps buildings and places gives the exposure: 1,375 places inside the delineation. Transport is over-represented (1.63× its weight in the four provinces; the single most hit category is petrol stations, 50): the water followed the valley floors and the state roads. Hotels (0.52×), schools (0.71×) and culture (0.70×) sit on the rises of the historic centres.
Sensitive services inside the footprint: 22 supermarkets, 12 schools, 11 pharmacies, 10 health facilities, 3 care homes. The sharpest municipal figure: Castel Bolognese, 61.7% of all its places; Conselice 57.2%; Sant’Agata sul Santerno 45.7%.
What annual data do not see
Treatment: 13 municipalities with at least 5% of their surface in the EMS footprint (Conselice 53%, Bagnacavallo and Lugo 29%, Alfonsine 21%). Control: 49 lowland municipalities of the same provinces with no delineation. Parallel pre-trends 2016–2022.
A flood that submerged up to half of a municipality is almost invisible in the annual panels: average taxable income falls 0.4% in tax year 2023 (fragile) and 0.1% in 2024, reabsorbed; population shows no measurable exodus; monthly night lights show no clean signal, the lighting stayed on. Transfers, insurance and reconstruction compensate at the municipality×year scale.
The result is the absence of an effect. To see one, you need a finer grain, or a different outcome.
Where it shows: house prices
Inside the delineated municipalities, the Revenue Agency’s OMI real-estate zones split each town into flooded and dry parts. A zone is "flooded" if at least 10% of it lies in the EMS footprint (zone boundaries of 2023-H1). Comparing flooded against dry zones of the same municipality, ordinary residential property, half-yearly 2021-H1 to 2025-H2:
Flat pre-trends (±0.3 log points), then flooded zones lose 0.6 log points in 2023-H2 and 1.3 in 2024-H1 relative to dry ones, and recover from 2024-H2. A market discount exists, it is about 1%, and it is transitory, consistent with the stickiness of OMI quotations (ranges of ordinary values, not transactions) and with the fiscal invisibility of the previous step.
Caveat: 2024-H1 is the last half-year with a 100% match of zones; the decennial revision from 2024-H2 breaks 12.7% of the links, so later points can move with composition.
Method, data and limits
Imagery. Maxar Open Data, event Emilia-Romagna-Italy-flooding-may23, Cloud-Optimized GeoTIFFs read remotely by range request: nothing downloaded in full. Licence CC BY-NC 4.0. Flood extent. Copernicus EMS rapid mapping, EMSR659 and EMSR664, observed event layers (© European Union, 2023). Exposure. Overture Maps buildings and places. Prices. OMI quotations and zone perimeters, Agenzia delle Entrate (CC BY 4.0). Boundaries. ISTAT 1 January 2021.
Limits. The Maxar strip covers 23% of the second-wave footprint and almost none of the first: Lugo, Bagnacavallo and Sant’Agata are outside it. The post pass is six days after the peak, so the mask measures sediment, not water. Landslide counts in class C are blind. Income and population are annual and municipal; the price effect comes from quotation ranges, not sales.
Reproducibility. Every number on this page comes from scripts in the author’s analysis folder (statistics, pair selection, mask, EMS validation, socio-economic effects, landslides, exposure, OMI zones). The pipeline is available on request; the map publishes only the derived layers shown here.
Citation: Santangelo, A. E. (2026). Reading a flood from space: Emilia-Romagna, May 2023. https://agapitosantangelo.github.io/flood/ · Code: MIT, github.com/agapitosantangelo/flood. Basemap © OpenFreeMap, © OpenMapTiles, data © OpenStreetMap contributors. Built with MapLibre GL JS.