Transboundary flood discovery: Storm Boris, September 2024¶

A worked example with the euroflood package, reproducing Section S1.1 (Fig. S1) of the EuroFlood paper. Storm Boris flooded much of Central and Eastern Europe in September 2024. Here, one index query recovers its archive footprint across nine countries, attributes the detected area to each, and then pulls native-resolution depth and modelled hazard only where we need them, without ever collecting or harmonising national flood maps.

The event¶

Between 9 and 16 September 2024, Storm Boris brought persistent frontal rainfall over the Odra, Morava, and upper Danube basins, causing severe flooding across Poland, Czechia, Austria, Slovakia, and the downstream Danube states.

It is a textbook transboundary episode, and that is what makes it hard to study: one meteorological event affects many countries, yet flood records, impact data, and mapping products are each compiled within national systems using different formats and spatial reference systems. The question this notebook answers is whether a single euroflood query can recover the geographically distributed archive footprint and attribute detected flood area by country, without first harmonising national flood maps.

The archive and the index¶

The observations come from the CEMS-EFAS satellite-derived flood-depth maps, the only open, continental record of observed flood depth: roughly 3,280 dated events (2015 to 2024) at 20 m, reconstructed from Sentinel-1 (Copernicus GFM extent, spatio-temporal clustering, then FLEXTH depth). The archive is open but hard to query, several thousand GeoTIFFs in a flat HTTP directory geolocated only by a single centroid in each filename, even though one file can bundle flooded areas up to 2,870 km apart.

euroflood adds an inverted raster index: each approximately 90 m cell stores the identifiers of every event that inundated it, so footprints and recurrence come from a windowed read with no rasters downloaded. Access is two-stage:

• Discover with floods(...), returning a FloodFrame (a geopandas.GeoDataFrame).
• Extract native 20 m depth with download(), and modelled return-period hazard with hazard(...), only for what you select.

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 geopandas as gpd
import matplotlib.pyplot as plt
from IPython.display import Markdown

import euroflood as ef

import _support as S

import geopandas as gpd import matplotlib.pyplot as plt from IPython.display import Markdown import euroflood as ef import _support as S

1. Discover¶

A single query over a box around Central and Eastern Europe returns every archived cluster that started in September 2024 and whose footprint intersects the window. It reads a small window of the index and downloads nothing.

In [2]:

cat = ef.floods(bbox=(8.0, 43.5, 24.0, 55.0), start="2024-09-01", end="2024-09-30")
cat

cat = ef.floods(bbox=(8.0, 43.5, 24.0, 55.0), start="2024-09-01", end="2024-09-30") cat

Out[2]:

EuroFlood catalogue: 3 flood events · 2024-09-09 … 2024-09-23 · 1,512.2 km² total

	collection	event_id	date	year	end_date	cluster_id	filename	download_url	area_km2	geometry
70	historic	1506465524	2024-09-23	2024	2024-09-30	210	WD_MERGE_2024-09-23---2024-09-30_duration_7_da...	https://jeodpp.jrc.ec.europa.eu/ftp/jrc-openda...	14.454	POLYGON ((24 43.5, 24 55, 8 55, 8 43.5, 24 43.5))
71	historic	2681670024	2024-09-23	2024	2024-11-11	367	WD_MERGE_2024-09-23---2024-11-11_duration_35_d...	https://jeodpp.jrc.ec.europa.eu/ftp/jrc-openda...	217.568	POLYGON ((24 43.5, 24 55, 8 55, 8 43.5, 24 43.5))
140	historic	772644797	2024-09-09	2024	2024-11-18	366	WD_MERGE_2024-09-09---2024-11-18_duration_49_d...	https://jeodpp.jrc.ec.europa.eu/ftp/jrc-openda...	1280.155	POLYGON ((24 43.5, 24 55, 8 55, 8 43.5, 24 43.5))

2. Pick the Storm Boris cluster¶

The query returns three clusters. Storm Boris is simply the largest, so we take it by mapped area rather than hard-coding an identifier, and reconstruct its approximately 90 m footprint with .footprints() (still index-only, no download).

In [3]:

boris = cat.nlargest(1, "area_km2").footprints()
boris[["event_id", "date", "end_date", "duration_days", "extent_km2"]]

boris = cat.nlargest(1, "area_km2").footprints() boris[["event_id", "date", "end_date", "duration_days", "extent_km2"]]

Out[3]:

	event_id	date	end_date	duration_days	extent_km2
0	772644797	2024-09-09	2024-11-18	70	1253.306

The cluster covers roughly 1,253 km² over a 70-day window (9 September to 18 November 2024). An archived event is a spatio-temporal cluster of observations, not necessarily one hydrological flood, so this record combines the initial Boris flooding with later autumn inundation along the wider Danube system. That temporal spread is exactly why we inspect date and end_date before drawing conclusions.

3. See it: an interactive map, straight from the index¶

With the viz extra, a FloodFrame renders without any raster download. cat.explore(footprints=True, tiles="grayscale") returns an interactive folium map on a grey basemap, each cluster on its own toggleable layer. Pan and zoom along the Odra and the Danube to follow the cross-border cascade from Poland down to the Balkans.

In [4]:

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

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

Out[4]:

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

4. Attribute the detected area by country¶

Because a FloodFrame is a GeoDataFrame, country attribution is a single spatial overlay: reproject the footprint to the equal-area ETRS89-LAEA system (EPSG:3035), intersect it with GISCO NUTS level 0 boundaries (2021, 1:10 million), and sum area per country.

In [5]:

countries = (
    gpd.read_file(S.DATA / "gisco" / "nuts0_10m_2021.geojson")
    [["NUTS_ID", "NAME_LATN", "geometry"]]
    .rename(columns={"NUTS_ID": "cc", "NAME_LATN": "country"})
    .to_crs("EPSG:3035")
)
inter = gpd.overlay(boris.to_crs("EPSG:3035"), countries, how="intersection")
inter["area_km2"] = inter.geometry.area / 1e6
per_country = (
    inter.groupby("cc")["area_km2"].sum()
    .round(1).sort_values(ascending=False)
)
per_country = per_country[per_country > 0.5]
per_country.to_frame("area_km2")

countries = ( gpd.read_file(S.DATA / "gisco" / "nuts0_10m_2021.geojson") [["NUTS_ID", "NAME_LATN", "geometry"]] .rename(columns={"NUTS_ID": "cc", "NAME_LATN": "country"}) .to_crs("EPSG:3035") ) inter = gpd.overlay(boris.to_crs("EPSG:3035"), countries, how="intersection") inter["area_km2"] = inter.geometry.area / 1e6 per_country = ( inter.groupby("cc")["area_km2"].sum() .round(1).sort_values(ascending=False) ) per_country = per_country[per_country > 0.5] per_country.to_frame("area_km2")

Out[5]:

	area_km2
cc
PL	497.7
DE	297.5
CZ	125.3
HR	121.1
HU	95.1
AT	50.8
SK	27.1
RO	18.7
RS	17.3

In [6]:

# The country breakdown as a bar chart (Table S2 as a figure).
ordered = per_country.sort_values()
fig, ax = plt.subplots(figsize=(6.4, 3.2))
ordered.plot.barh(ax=ax, color="#1f5c8b")
ax.set_xlabel("detected area (km$^2$)")
ax.set_title("Storm Boris cluster: detected flood area by country")
ax.margins(x=0.12)
for i, (cc, v) in enumerate(ordered.items()):
    ax.text(v + 6, i, f"{v:,.0f}", va="center", fontsize=8, color="#555555")
plt.tight_layout()
plt.show()

# The country breakdown as a bar chart (Table S2 as a figure). ordered = per_country.sort_values() fig, ax = plt.subplots(figsize=(6.4, 3.2)) ordered.plot.barh(ax=ax, color="#1f5c8b") ax.set_xlabel("detected area (km$^2$)") ax.set_title("Storm Boris cluster: detected flood area by country") ax.margins(x=0.12) for i, (cc, v) in enumerate(ordered.items()): ax.text(v + 6, i, f"{v:,.0f}", va="center", fontsize=8, color="#555555") plt.tight_layout() plt.show()

The result reproduces Table S2. The largest shares fall in Poland (along the Odra and Nysa systems) and Germany (along the Elbe and lower Oder). The remaining area extends through Czechia, Austria, Slovakia, Hungary, Croatia, Romania, and Serbia, forming a geographically coherent sequence down the Danube. Once the footprint has been retrieved, this whole attribution is a single spatial overlay against harmonised administrative boundaries.

In [7]:

# Key figures, computed from the results above.
area = float(boris["extent_km2"].iloc[0])
duration = int(boris["duration_days"].iloc[0])
top = per_country.head(2)
Markdown(
    f"### Key figures\n"
    f"| | |\n|---|---|\n"
    f"| **Detected flood area** | **{area:,.0f} km²** |\n"
    f"| **Countries** | **{len(per_country)}** |\n"
    f"| **Archive cluster** | **{duration} days** (9 Sep to 18 Nov 2024) |\n"
    f"| **Largest shares** | **{top.index[0]} {top.iloc[0]:,.0f} km²**, "
    f"**{top.index[1]} {top.iloc[1]:,.0f} km²** |\n"
)

# Key figures, computed from the results above. area = float(boris["extent_km2"].iloc[0]) duration = int(boris["duration_days"].iloc[0]) top = per_country.head(2) Markdown( f"### Key figures\n" f"| | |\n|---|---|\n" f"| **Detected flood area** | **{area:,.0f} km²** |\n" f"| **Countries** | **{len(per_country)}** |\n" f"| **Archive cluster** | **{duration} days** (9 Sep to 18 Nov 2024) |\n" f"| **Largest shares** | **{top.index[0]} {top.iloc[0]:,.0f} km²**, " f"**{top.index[1]} {top.iloc[1]:,.0f} km²** |\n" )

Out[7]:

Key figures¶

Detected flood area	1,253 km²
Countries	9
Archive cluster	70 days (9 Sep to 18 Nov 2024)
Largest shares	PL 498 km², DE 298 km²

5. Zoom in: native-resolution depth¶

Discovery is index-only; the 20 m source depth is fetched only where wanted. Two reaches of the cluster tell the transboundary story, upstream at the border and downstream weeks later.

The Oder at Schwedt (the German-Polish border)¶

Here the river itself forms the border, and the archive records inundation on both banks. We take the same largest-area event, download() the ROI-cropped raster, and show it on the grey basemap with DepthRaster.explore().

In [8]:

sel = ef.floods(bbox=(14.10, 52.90, 14.50, 53.15), start="2024-09-01", end="2024-09-30")
schwedt = ef.open_depth(ef.download(sel.nlargest(1, "area_km2"), str(S.ROOT / "out" / "boris_notebook"))[0])
schwedt.explore(tiles="grayscale", vmax=2.0)

sel = ef.floods(bbox=(14.10, 52.90, 14.50, 53.15), start="2024-09-01", end="2024-09-30") schwedt = ef.open_depth(ef.download(sel.nlargest(1, "area_km2"), str(S.ROOT / "out" / "boris_notebook"))[0]) schwedt.explore(tiles="grayscale", vmax=2.0)

Out[8]:

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

The Drava at Kopacki Rit, near Osijek (downstream, weeks later)¶

Part of the same archive cluster, this is the later downstream signal along the Danube system in Croatia, hundreds of kilometres from Schwedt and separated in time. It is a concrete reminder that one indexed event can span both the September peak and later autumn inundation.

In [9]:

sel = ef.floods(bbox=(18.62, 45.50, 18.98, 45.72), start="2024-09-01", end="2024-09-30")
osijek = ef.open_depth(ef.download(sel.nlargest(1, "area_km2"), str(S.ROOT / "out" / "boris_notebook"))[0])
osijek.explore(tiles="grayscale", vmax=2.0)

sel = ef.floods(bbox=(18.62, 45.50, 18.98, 45.72), start="2024-09-01", end="2024-09-30") osijek = ef.open_depth(ef.download(sel.nlargest(1, "area_km2"), str(S.ROOT / "out" / "boris_notebook"))[0]) osijek.explore(tiles="grayscale", vmax=2.0)

Out[9]:

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

6. Compare with modelled hazard¶

EuroFlood serves modelled river-flood hazard too (CEMS-GLOFAS return-period scenarios), so observed and modelled extents live in one workflow. We retrieve the 10-year scenario for the Schwedt reach and summarise its depth with .stats().

In [10]:

haz = ef.hazard(bbox=(14.10, 52.90, 14.50, 53.15), return_period=10)
haz.download(str(S.ROOT / "out" / "boris_notebook"))
haz.stats()

haz = ef.hazard(bbox=(14.10, 52.90, 14.50, 53.15), return_period=10) haz.download(str(S.ROOT / "out" / "boris_notebook")) haz.stats()

Out[10]:

	return_period	wet_pixels	max_depth_m	mean_depth_m	p95_depth_m	flooded_area_km2	volume_m3	volume_Mm3
0	10	21427	4.9	2.67	4.543	124.6478	332791423.2	332.7914

The modelled 10-year scenario is markedly deeper than the observed detected water above. That is expected: the archive records what Sentinel-1 actually saw, often delayed, residual, and shallow, so the observed depth is a lower bound rather than the modelled peak. Comparing the two envelopes systematically is the subject of the Zutphen case study in the same paper.

The published figure¶

The full multi-panel Fig. S1 (the country cascade, the Schwedt and Osijek depth chips, and the summary) is assembled by the paper's figure script from the same query above, shown here for comparison.

In [11]:

S.published_figure("boris")

S.published_figure("boris")

Out[11]:

Interpretation¶

One index query retrieves a geographically distributed archive cluster and produces an auditable country-level attribution, without collecting or harmonising national flood datasets first. Three qualifications matter for reading the result:

1. A cluster is not one flood. The record is an archive-defined spatio-temporal cluster spanning 9 September to 18 November 2024 (70 days), so it combines the September peak with later autumn inundation along the wider Danube. Detected areas in downstream countries should not be read as the September peak alone; for physical event attribution, inspect the date/end_date fields and re-segment or aggregate accordingly.
2. Date filters act on the cluster start date. A narrow window such as 10 to 30 September would miss this record, whose start date is 9 September. Robust discovery uses a padded window and inspects the returned dates.
3. Detected area is a lower bound. The country areas inherit opposing biases: radar under-detection omits inundation that was not observed, while conservative resampling to the approximately 90 m grid enlarges the footprint (median inflation factor 1.49 relative to the native 20 m wet area). Detected area and recurrence are counts of archived detections, never a return period.

References and reuse¶

This notebook reproduces the Storm Boris case study (Supplementary Material, Section S1.1, Fig. S1) of:

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