Lakes that flooded more than once
Three ice-dammed lakes account for a fifth of the whole database. They fill and drain almost every summer, so each row of the file is a drainage cycle, not a distinct hazard site.
Where the floods start
Counts follow the record, not the risk. Kyrgyzstan and Kazakhstan rank high partly because Soviet-era field observers logged small annual drainages that go unrecorded elsewhere.
Timing
618 of 766 events carry a usable year. The other 148 are satellite detections with no documentary date and are left out of every chart on this page.
Events per decade since 1900
The rise after 1950 tracks observation capacity as much as hazard. Satellite coverage from the 1970s, then high-cadence imagery after 2000, made small and remote events visible for the first time. The database cannot separate a real increase in flood frequency from an increase in detection, and it does not claim to.
Month of occurrence
Of the 390 events with a known month, most fall in the summer melt and monsoon window.
How precisely each event is dated
Year_approx carries text forms such as "Before 1966", "2002-2004" and "1960s". The cleaned file resolves each to a single year and records the assumption alongside it, so approximate dates never masquerade as exact ones.
Consequences
The database records 8,996 deaths. Two events supply 89% of that figure, and 96% of records carry no casualty entry at all — absent, not zero.
Deaths by event
A log scale, because the range runs from 1 to 6,000. Kedarnath (2013) and the Booni Gol water-pocket burst (2010) dominate any regional total that includes them.
Every event with a recorded death toll
What triggered the flood
Counts are out of the 97 events with an identified trigger, not out of 766. The remaining 87% are recorded as unknown.
How the dam failed
What to check before you analyse this
Seventeen findings from the cleaning run, ordered by how badly each one could distort a result. Nothing was deleted: every repair is reversible and every original string is preserved.
How much of each field is actually filled
Transformation log — every change made to the file
Four releases, 2022 to 2025
The record grew by 69 events. Matching on GF_ID would give a wrong answer, so these comparisons match on lake coordinate plus date instead.
Growth of the record
Release by release
Edits to events that were already in v1.0
602 events appear in both v1.0 and v4.0. These fields changed underneath them.
Records in an old release with no exact match in the next one
Nine of the ten resolve to the same lake with a corrected date, a corrected coordinate or a new name. Only one — Bhairav Taal, Nepal, dated 5 July 2024 in v3.0 — has no record within 2 km in v4.0. Verify that one against the source before concluding it was withdrawn.