A decade of European flood recurrence, in one query¶

A worked example with the euroflood package, reproducing the paper's headline application (main text Section 7.1, Fig. 5). It answers a continental question, how often has each place in Europe been detected as flooded over 2015 to 2024?, from a single discovery query. Answering it from the raw archive would mean downloading, aligning, and stacking thousands of event rasters; with the index it is a windowed read, no rasters downloaded.

The archive and the index¶

The observations come from the CEMS-EFAS satellite-derived flood-depth maps (reconstructed from Sentinel-1), served through euroflood's inverted raster index: each approximately 90 m cell stores the identifiers of every event that inundated it, so a per-cell recurrence count is len(events at that cell), read straight from the index. Discovery returns a FloodFrame (a geopandas.GeoDataFrame); no source rasters are downloaded.

Runs against euroflood >= 0.2.0 with the viz extra, querying the published index (needs network on first use, or a locally mirrored index).

In [1]:

import json

import matplotlib.pyplot as plt
import pandas as pd

import euroflood as ef

import _support as S

import json import matplotlib.pyplot as plt import pandas as pd import euroflood as ef import _support as S

1. One query, a continent¶

A single query over the European window returns every archived event of the decade. This is the largest query in the series and the slowest cell (about two minutes to build the geometry for ~2,900 events), but it still downloads no rasters, only a windowed read of the index.

In [2]:

cat = ef.floods(bbox=(-25.0, 34.0, 45.0, 72.0), start="2015-01-01")
print(f"{len(cat):,} archived events, {str(cat['date'].min())[:10]} to {str(cat['date'].max())[:10]}")
cat.head()

cat = ef.floods(bbox=(-25.0, 34.0, 45.0, 72.0), start="2015-01-01") print(f"{len(cat):,} archived events, {str(cat['date'].min())[:10]} to {str(cat['date'].max())[:10]}") cat.head()

2,942 archived events, 2015-01-05 to 2024-12-30

Out[2]:

EuroFlood catalogue: 5 flood events · 2015-02-02 … 2015-04-06 · 5,649.1 km² total

	collection	event_id	date	year	end_date	cluster_id	filename	download_url	area_km2	geometry
0	historic	3004549128	2015-02-02	2015	2015-02-09	028	WD_MERGE_2015-02-02---2015-02-09_duration_7_da...	https://jeodpp.jrc.ec.europa.eu/ftp/jrc-openda...	431.572	POLYGON ((45 34, 45 72, -25 72, -25 34, 45 34))
1	historic	451134548	2015-02-16	2015	2015-02-23	038	WD_MERGE_2015-02-16---2015-02-23_duration_7_da...	https://jeodpp.jrc.ec.europa.eu/ftp/jrc-openda...	22.125	POLYGON ((45 34, 45 72, -25 72, -25 34, 45 34))
2	historic	3265023219	2015-03-30	2015	2015-04-20	067	WD_MERGE_2015-03-30---2015-04-20_duration_21_d...	https://jeodpp.jrc.ec.europa.eu/ftp/jrc-openda...	3.838	POLYGON ((45 34, 45 72, -25 72, -25 34, 45 34))
3	historic	229685871	2015-03-30	2015	2015-05-18	068	WD_MERGE_2015-03-30---2015-05-18_duration_49_d...	https://jeodpp.jrc.ec.europa.eu/ftp/jrc-openda...	5163.998	POLYGON ((45 34, 45 72, -25 72, -25 34, 45 34))
4	historic	2498803972	2015-04-06	2015	2015-04-13	075	WD_MERGE_2015-04-06---2015-04-13_duration_7_da...	https://jeodpp.jrc.ec.europa.eu/ftp/jrc-openda...	27.532	POLYGON ((45 34, 45 72, -25 72, -25 34, 45 34))

2. The continental recurrence surface¶

Turning those events into a per-cell recurrence surface for the whole continent is best left to the paper's figure pipeline: at full resolution the European grid is billions of cells, so the published map streams the index and max-pools for display. It is shown at the end of this notebook (the paper's Fig. 5). First, a zoom we can build live.

3. The most-flooded place in Europe, computed live¶

At a zoom the recurrence surface is cheap to build from the public API. The single most repeatedly flooded location in the archive is the Vouga floodplain around the Aveiro lagoon in Portugal. We query it, build the per-cell recurrence with recurrence_grid (the index-only helper behind the recurrence map), and read the count back.

In [3]:

from euroflood.viz._raster import recurrence_grid  # noqa: E402

aveiro = ef.floods(bbox=(-8.80, 40.60, -8.45, 40.95))
counts = recurrence_grid(aveiro)[0]
wet = counts[counts > 0]
print(f"{len(aveiro)} events at this reach · up to {int(counts.max())} detections in one cell "
      f"· {(counts >= 2).sum() / wet.size:.0%} of wet cells flooded more than once")

from euroflood.viz._raster import recurrence_grid # noqa: E402 aveiro = ef.floods(bbox=(-8.80, 40.60, -8.45, 40.95)) counts = recurrence_grid(aveiro)[0] wet = counts[counts > 0] print(f"{len(aveiro)} events at this reach · up to {int(counts.max())} detections in one cell " f"· {(counts >= 2).sum() / wet.size:.0%} of wet cells flooded more than once")

89 events at this reach · up to 87 detections in one cell · 83% of wet cells flooded more than once

The maximum, 87, matches the continental headline exactly: this really is the most repeatedly detected cell in Europe. Explore the archived events at this reach on an interactive grey basemap (explore(footprints=True); pan and zoom the Vouga floodplain):

In [4]:

aveiro.explore(footprints=True, tiles="grayscale")

aveiro.explore(footprints=True, tiles="grayscale")

Out[4]:

Make this Notebook Trusted to load map: File -> Trust Notebook

4. The distribution of recurrence¶

Across Europe the per-cell recurrence is strongly right-skewed: most detected ground floods once or twice, a long tail floods many times. (Read from the paper's frozen continental result, since a full-resolution continental recurrence raster does not fit in memory.)

In [5]:

hist = pd.read_csv(S.RESULTS / "recurrence_hist.csv")
fig, ax = plt.subplots(figsize=(6.6, 3.4))
ax.bar(hist["recurrence"], hist["cells"], width=1.0, color="#1f5c8b")
ax.set_yscale("log")
ax.set_xlabel("times a cell was detected flooded (2015 to 2024)")
ax.set_ylabel("number of cells (log scale)")
ax.set_title("Europe: distribution of per-cell flood recurrence")
ax.margins(x=0.01)
plt.tight_layout()
plt.show()

hist = pd.read_csv(S.RESULTS / "recurrence_hist.csv") fig, ax = plt.subplots(figsize=(6.6, 3.4)) ax.bar(hist["recurrence"], hist["cells"], width=1.0, color="#1f5c8b") ax.set_yscale("log") ax.set_xlabel("times a cell was detected flooded (2015 to 2024)") ax.set_ylabel("number of cells (log scale)") ax.set_title("Europe: distribution of per-cell flood recurrence") ax.margins(x=0.01) plt.tight_layout() plt.show()

In [6]:

# Key figures. The continental aggregates are the paper's frozen result (a full-continent
# recurrence raster does not fit in memory); the max is also reproduced live above at Aveiro.
stats = json.loads((S.RESULTS / "fig_a_stats.json").read_text())
from IPython.display import Markdown  # noqa: E402

Markdown(
    f"### Key figures\n"
    f"| | |\n|---|---|\n"
    f"| **Archived events (2015 to 2024)** | **{len(cat):,}** |\n"
    f"| **Area flooded at least once** | **{stats['area_flooded_at_least_once_km2']:,.0f} km²** |\n"
    f"| **Cells flooded more than once** | **{stats['cells_recurrence_ge2_fraction']:.1%}** |\n"
    f"| **Most-flooded cell (Vouga / Aveiro, PT)** | **{max(int(k) for k in stats['cells_by_recurrence'])} times** |\n"
)

# Key figures. The continental aggregates are the paper's frozen result (a full-continent # recurrence raster does not fit in memory); the max is also reproduced live above at Aveiro. stats = json.loads((S.RESULTS / "fig_a_stats.json").read_text()) from IPython.display import Markdown # noqa: E402 Markdown( f"### Key figures\n" f"| | |\n|---|---|\n" f"| **Archived events (2015 to 2024)** | **{len(cat):,}** |\n" f"| **Area flooded at least once** | **{stats['area_flooded_at_least_once_km2']:,.0f} km²** |\n" f"| **Cells flooded more than once** | **{stats['cells_recurrence_ge2_fraction']:.1%}** |\n" f"| **Most-flooded cell (Vouga / Aveiro, PT)** | **{max(int(k) for k in stats['cells_by_recurrence'])} times** |\n" )

Out[6]:

Key figures¶

Archived events (2015 to 2024)	2,942
Area flooded at least once	376,857 km²
Cells flooded more than once	51.3%
Most-flooded cell (Vouga / Aveiro, PT)	87 times

The published figure¶

The paper's Fig. 5: the full-resolution per-cell recurrence surface for the whole continent (streamed from the index and max-pooled for display), shown here for comparison.

In [7]:

S.published_figure("recurrence")

S.published_figure("recurrence")

Out[7]:

Interpretation¶

One discovery query, read from the index alone, yields a decade-long recurrence surface for a continent, which is the point: no event rasters are downloaded, aligned, or stacked. Two cautions on reading it:

1. Recurrence is a detection count, not a return period. It counts how many archived events were detected at a cell, not statistically independent floods, and cannot be converted into an annual exceedance probability or a return period.
2. An event is an archive cluster. Each archived event is a spatio-temporal cluster of satellite observations, not necessarily one hydrological flood, and detection reflects Sentinel-1 acquisition timing and detectability as much as flood occurrence. The continental map is max-pooled for display only.

References and reuse¶

This notebook reproduces the continental recurrence application (main text Section 7.1, Fig. 5) of:

J. Hackl (2026). EuroFlood: a Python library and queryable index for the CEMS satellite-derived flood-depth archive of Europe.