Zhenyi Tsai

This project was designed to create an interactive, web-based situational map of the 26 August 2026 Bhote Koshi flash flood, documenting where the event originated, how it propagated through the Gyirong–Nepal corridor, and which communities were affected as conditions evolved. Rather than presenting a static map alone, the interface combines terrain context with live-style incident updates, making it easier to understand both the physical hazard and the human impact in a single, navigable environment.

The system is built as a lightweight GitHub Pages deployment: a single HTML page loads Three.js from a CDN and renders a height-mapped terrain model generated from a Copernicus DEM encoded as a Terrain-RGB PNG. Flood extent, damage buffers, the flood-origin marker, and settlement-based updates are structured as client-side geographic and narrative data, allowing the map to function as a disaster-tracking and reporting tool without requiring a backend or external map tiles. The repo documents the encoding workflow, data sources, and visual architecture so the project can be reused for other rapid-response or event-tracking scenarios.

A glacial lake outburst flood begins where ice, meltwater, and a fragile natural dam meet. Across the Koshi Basin, roughly two thousand lakes sit in that arrangement — and only a handful will ever fail. The problem is not a shortage of lakes to worry about; it is knowing which ones, and who sits below them.

This project builds a lightweight, repeatable way to triage that question. Rather than model every lake in exhaustive detail, it scores all of them on the three variables that most reliably separate the dangerous from the benign, then carries the survivors through validation, proximity exposure, and downstream flood-path analysis — the same path a hazard analyst would walk, made legible on a map.

A screening model is only as trustworthy as its agreement with slower, field-based work. Validated against ICIMOD's authoritative inventory of 42 potentially dangerous glacial lakes, the three-factor model recovered 29 of them — including the two Rank-I benchmarks, Imja and Tsho Rolpa. For a method this lightweight, a 69% catch is a strong result, and the disagreements are useful in their own right: they mark exactly where rapid screening and detailed assessment diverge.

Risk is only meaningful once people enter the frame. Drawing a 5 km danger zone around the 79 high-risk lakes captures roughly 10,808 residents — a deliberately conservative baseline, since outburst floods travel far beyond five kilometres down-valley. Exposure concentrates in the thin, high-mountain valleys threaded between the glaciers.

Proximity is a blunt instrument; a flood follows the valley, not a circle. Tracing the actual flow paths of the two marquee lakes turns the abstraction concrete — and reveals how much the downstream geography matters. Tsho Rolpa's corridor down the Rolwaling Khola threatens 9 settlements and about 1,134 residents. Imja's corridor crosses more villages — 20 — yet only ~224 modelled residents, because the upper Khumbu is sparsely and seasonally peopled.

Two Rank-I lakes, comparable hazards, and an order-of-magnitude gap in human consequence. That contrast is the argument for doing flow-path analysis at all.

The point was not just the storm data. It was getting comfortable with the modern geospatial stack — CLI, GDAL, GeoParquet, and GitHub — by using each piece in a practical workflow. The script finds NOAA's current publication of the selected year's details file, downloads and extracts it, creates point geometry from latitude and longitude, and writes a GeoParquet dataset in WGS 84.

The raw download and extracted CSV stay out of version control, while the processed output is ready to query in DuckDB, inspect in GeoPandas, or open directly in QGIS.