An interactive web application demonstrating a Glacial Lake Outburst Flood (GLOF) hazard-screening workflow for the Nepal Himalaya. It covers 25 high-elevation glacial lakes from 2000 to 2024 and applies multi-factor hazard scoring, machine learning, climate projections, and population exposure analysis.
Lake names, coordinates, basins, districts and elevations are real. Lake areas, growth rates, dam types, downstream slopes and settlement distances are simulated by
data/generate_data.py, so every hazard score, risk class, ML probability, change-detection alert and climate projection derived from them illustrates the method rather than describing these lakes.Population and building counts are real (WorldPop 2020 + OpenStreetMap), but they are summed over flood corridors drawn from the simulated inventory.
To run the app on real data, load a real inventory with
data/fetch_icimod.pyand real imagery withdata/fetch_sentinel.py. Do not use the shipped figures for planning, early warning or risk communication.
Live site: himalayaglof.bikal3.com.np | GitHub: bikal3/himalaya-glof
| Page | Description |
|---|---|
| 🏔️ Home | Overview metrics and map of all 25 lakes |
| 🗺️ Risk Map | Interactive Leaflet map with basin filters, risk class filters, and area threshold |
| 📈 Lake Trends | Area time-series, basin totals, and risk score distribution charts |
| 🌡️ Climate Projections | Lake area forecasts to 2100 under RCP 4.5 and RCP 8.5 scenarios |
| 🤖 ML Risk Scoring | Random Forest classifier trained on the GLOF event catalogue vs formula score |
| 🛰️ Change Detection | Baseline vs latest cached area comparison with 15% alert threshold |
| 👥 Population Exposure | WorldPop + OSM building counts within each lake's downstream flood corridor |
| 📋 Methodology | Spectral indices, hazard scoring table, data sources, provenance table, and GEE script |
| ⬇️ Downloads | GeoJSON, CSV, JSON, area cache, ML model, and PDF report |
git clone https://github.com/bikal3/himalaya-glof.git
cd himalaya-glof
pip install -r requirements.txtpython site/build.py --serve # build, then preview on http://localhost:8000
python -m pytest tests/ -q # tests--serve mirrors Cloudflare Pages' clean-URL behaviour (/map, not /map.html), so local
preview matches production exactly. Python is only needed to build the site; the published
result is static files and runs no Python at all.
requirements.txt holds only what the deployed app needs. The offline scripts have their own dependencies:
pip install -r requirements-offline.txt
python data/compute_exposure.py # WorldPop + OSM exposure (~100 MB raster download)
python data/fetch_sentinel.py # real Sentinel-2 areas (needs Sentinel Hub credentials)
python data/fetch_icimod.py --shapefile /path/to/icimod_nepal_lakes.shp # real inventory
python data/train_model.py # retrain the Random Forest after changing either inputsite/
build.py # Generates dist/ — the published site
assets/ # style.css and one JS file per interactive page
vendor/ # Leaflet, Plotly and Source Sans 3 (no CDN at runtime)
utils/ # Shared logic, imported by the build and the tests
risk_score.py # Hazard formula
climate_projections.py # RCP projection model
ml_model.py # Random Forest train / infer
change_detection.py # Area cache diff logic
exposure.py # Population exposure loaders
provenance_text.py # "Which numbers are simulated" notice wording
data/
generate_data.py # Generates the simulated inventory and time-series
lakes_risk.geojson # 25 lake Point features with hazard scores (simulated)
lakes_timeseries.csv # Annual area values 2000–2024 (simulated)
flood_corridors.geojson # 8 downstream LineString corridors
flood_corridors_buffered.geojson # 25 ±2 km Polygon corridors
population_exposure.json # Pre-computed population + building counts (real)
sentinel_cache/ # Per-lake area cache; each file records its own `source`
glof_events.csv # GLOF event catalogue (HKH-wide, unverified attributes)
compute_exposure.py # Offline: WorldPop + OSM exposure script
fetch_sentinel.py # Offline: Sentinel Hub API fetch
fetch_icimod.py # Offline: normalise a real ICIMOD inventory
train_model.py # Offline: train and save the Random Forest
models/
glof_risk_model.pkl # Trained Random Forest (joblib)
gee_scripts/
lake_detection.js # Google Earth Engine lake delineation script
| Dataset | Provider | Resolution | Use | In shipped data? |
|---|---|---|---|---|
| Landsat 5/7/8/9 Surface Reflectance | USGS / NASA | 30 m | Lake delineation (MNDWI) | No — pipeline only |
| Sentinel-2 MSI | ESA | 10 m | Recent area measurements | No — needs fetch_sentinel.py |
| Copernicus DEM GLO-30 | ESA / Copernicus | 30 m | Downstream slope | No — slopes are simulated |
| ICIMOD GLOF Database | ICIMOD | — | Event catalogue for ML training | Events yes, attributes unverified |
| WorldPop Nepal 2020 | WorldPop / Univ. of Southampton | 100 m | Population exposure | Yes |
| OpenStreetMap | OSM contributors | — | Building footprints | Yes |
A 2000–2024 Landsat record spans three sensors: Landsat 5 TM and 7 ETM+ for 2000–2012, Landsat 8 OLI from 2013, and Landsat 9 from late 2021.
- Open Google Earth Engine Code Editor.
- Paste the contents of
gee_scripts/lake_detection.js. - Click Run, then export results to Google Drive via the Tasks panel.
site/build.py loads the datasets through the utils/ modules, precomputes every hazard
score, projection, ML probability and change figure, and writes dist/ as plain HTML + JSON.
The browser only renders; it never recomputes a number, so the page cannot drift from what
the Python produced. The published site runs no Python at all.
Assets are content-fingerprinted (style.585bec62.css), so a redeploy always publishes a new
URL and a cached stylesheet can never shadow new markup.
python site/build.py # build into dist/
python site/build.py --serve # build, then preview on http://localhost:8000
wrangler pages deploy dist --project-name himalaya-glof --branch maindist/ is generated and gitignored; rebuild it after changing any data file or util.
Cloudflare Pages is not connected to this repository, so pushing to main does not update
the live site; run the two commands above.
The custom domain himalayaglof.bikal3.com.np is a proxied CNAME to himalaya-glof.pages.dev.
Leaflet, Plotly and Source Sans 3 are committed under site/vendor/ deliberately: the page
then loads nothing from a third-party CDN, which is what makes the strict
Content-Security-Policy in site/build.py workable.
MIT