Inventory of glacial lake outburst floods (GLOFs) in High Mountain Asia 2022-2025 | HMAGLOFDB review

An interactive review of the HMAGLOFDB, ICIMOD's inventory of glacial lake outburst floods (GLOFs) in High Mountain Asia: across the four annual releases from 2022 to 2025.

Dashboard: https://prazg.github.io/GLOF_dashboard/

---

Why this exists

HMAGLOFDB is the most complete GLOF inventory available for the region and is published openly under CC BY 4.0. Using it well takes some care: several fields documented as integers contain prose, the event identifier is reassigned at every release, and roughly a fifth of events carry no date. This repository does not correct the database — that is ICIMOD's to do — it makes those properties visible before they reach an analysis, and preserves every original value.

What the review found

Finding	Why it matters
GF_ID is reassigned at every release	Of the 697 IDs shared by v1.0 and v4.0, only 16% point to a record with the same lake coordinate. A stored reference to a GF_ID silently points at a different flood after an update.
One record genuinely withdrawn in four releases	Nine of the ten records without an exact match resolve to the same lake with a corrected date, coordinate or name. The exception is Bhairav Taal, Nepal, dated 5 July 2024 in v3.0, with no record within 2 km in v4.0.
Free text in integer fields	Hydropower holds "1200 Teesta III, 510 Teesta V, 500 Teesta VI"; Residential_damaged holds "13 houses fully damaged; 10 partially damaged". Summing these naively either errors or drops them.
Non-breaking spaces as empty-cell placeholders	U+00A0 appears in 17 fields. Those cells look populated to a parser.
766 records cover 385 distinct lakes	450 records are flagged as repeat events; Merzbacher alone contributes 85. Treating rows as independent events roughly doubles the apparent number of hazard sites.
19% of events have no year, 87% no identified trigger, 96% no casualty figure	Any rate or proportion has to state which denominator it used.

The fatality total in the cleaned file sums to 8,996, matching the figure in the v4.0 metadata abstract. Two events supply 89% of it.

Caveats

• The "date or name refined" verdicts in the version diff are inferred here from a 2 km spatial match plus a year comparison. They are not published by ICIMOD and are leads to check, not conclusions.
• No changelog accompanies the releases, so there is no independent confirmation of what ICIMOD intended to change between versions.
• The rise in events per decade tracks observation capacity as much as hazard. The database cannot separate a real increase in flood frequency from an increase in detection, and does not claim to.
• Casualty counts are lower bounds. Entries carrying + mean "at least this many"; the cleaned file moves that marker into its own _is_minimum column so it survives aggregation.

Repository layout

index.html                 generated dashboard, served by GitHub Pages
src/build_glofdb.py        the whole pipeline: clean, check, diff, render
src/template.html          dashboard markup with a /*__DATA__*/ injection point
data/raw/                  the four HMAGLOFDB releases exactly as published (cp1252)
data/processed/            everything the pipeline produces
  HMAGLOFDB_v4_clean.csv     766 rows x 81 columns, UTF-8
  qc_report.json             17 findings, completeness table, transformation log
  version_diff.json          per-release added, unmatched and edited records
  glof_payload.json          aggregates inlined into index.html

Rebuilding

pip install -r requirements.txt
python src/build_glofdb.py

Roughly two seconds. It rewrites data/processed/ and index.html. Nothing under data/raw/ is ever modified.

When v5.0 is released

1. Drop the new CSV into data/raw/.
2. Add one line to the VERSIONS dict at the top of src/build_glofdb.py:

   "v5.0": ("HMAGLOFDB_v5_0_DDMMYYYY.csv", "2026-12-13"),

3. Rerun. The diff chain, the growth chart and the release-by-release tab extend on their own; the cleaning step always targets the last entry in the dict.

If the schema changes, the QC step will flag new field names as newly empty rather than failing silently — check the completeness table after any update.

What the cleaning does

Reversible repairs only. No row is dropped and no original string is lost.

• Non-breaking and zero-width spaces removed, whitespace collapsed, blanks set to NA
• Degree symbols stripped from coordinates so all 766 events parse as floats
• Free text in numeric fields moved to a parallel <field>_note column, numeric set to NA
• + split into a boolean <field>_is_minimum column
• year_best derived from Year_exact first, then parsed from text forms in Year_approx ("Before 1966", "2002-2004", "1960s"), with the assumption recorded in year_precision
• event_date built where day, month and year are all known
• Driver_GLOF and Mechanism split on both ; and , into normalised _terms lists
• "Unnamed" and "Unknown" flagged in <field>_is_placeholder, original text kept

The full list, with row counts, is in data/processed/qc_report.json and on the dashboard's data quality tab.

Data source and attribution

ICIMOD (2025). GLOF database of High Mountain Asia [Data set]. International Centre for Integrated Mountain Development. https://doi.org/10.26066/RDS.1973283 — licensed CC BY 4.0.

The files in data/raw/ are redistributed unchanged under that licence. data/processed/ contains derived works, also CC BY 4.0. The code is MIT. See LICENSE and LICENSE-DATA.md.

This repository is an independent review and is not affiliated with or endorsed by ICIMOD.