Methodology
What the Observatory measures, what it does not, and exactly how each number on the map is calculated.
What the Observatory measures, what it does not, and exactly how each number on the map is calculated.
Local units, districts and provinces are the Survey Department of Nepal's official local-unit boundaries (the 2020 map, which includes Limpiyadhura, Lipulekh and Kalapani in Byas Rural Municipality, Darchula). The source is in the Everest 1830 datum; it is transformed to WGS84 with the Nagarkot parameters (+towgs84 = 293.17, 726.18, 245.36), which lines it up with international open data to within a metre. Each unit carries the OCHA COD-AB P-code, so the same code works in this app, in the history database and in other humanitarian datasets. Twenty-two protected-area cores are separate polygons in the source and are kept as their own units.
| Number | How it's computed |
|---|---|
| Tree cover in 2000 | Hansen/UMD Global Forest Change v1.13 treecover2000, pixels with ≥30% canopy, each counted at its true ground area and summed inside the official polygon (scripts/build-place-stats.py). Districts and provinces are sums of their units. |
| Loss by year | The same pixels with a lossyear of 2001–2025, summed per year. "Share lost" is total loss ÷ tree cover in 2000. |
| People | WorldPop Global2 2025 (100 m, constrained, UN-adjusted) summed inside the polygon — a modelled estimate, not the census. |
| Climate normals | CHELSA V2.1 monthly temperature and rainfall, 1981–2010, averaged over every ~1 km cell inside each municipality's official boundary (until September 2026 it was sampled at the municipality's centre, which read warm and wet in large mountain municipalities); districts, provinces and Nepal are area-weighted means of their municipalities. CHELSA models mountain rainfall from reanalysis and usually gives more rain on high ground than CHIRPS, which the Rainfall trend analysis uses; the two are never mixed in one calculation. |
| Warming | ERA5-Land annual mean temperature at the same point, 1950–2024, as anomalies from that point's own 1991–2020 mean. The trend is a least-squares fit from 1979 (the back-extension before 1979 is the weakest part of the record over the Himalaya). |
| Projection | Seven CMIP6 HighResMIP models via Open-Meteo, high-emissions forcing, 20–50 km grid, at the place's centre. The app reports the change between the 2000s and the 2040s, not the absolute temperature, because coarse models run warmer or cooler than a mountain valley. |
| Ranks | "#12 of 77" is the place's position among all places of the same level, highest first. The map's colours are five classes with equal numbers of places; the key prints each class's real range. |
Tree-cover loss is the Hansen/UMD Global Forest Change dataset (v1.13, 2001–2025). A pixel (~30 m) counts when it is classified as stand-replacement disturbance in that year and its year-2000 canopy density was at least 30%.
"Loss" is not synonymous with "deforestation." The method flags any stand-replacement disturbance — logging, fire, landslides, river erosion, disease or clearing — without attributing a cause or saying whether the land later regrew. GeoSutra follows the dataset's own definition and always says "tree-cover loss." "Tree cover" is any vegetation taller than 5 m, so it includes plantations and orchards and is not the legal forest area in Nepal's Forest Resource Assessment.
| Field | How it's computed |
|---|---|
| Area and perimeter | Geodesic (spherical-excess) area and haversine perimeter over the exact vertices. |
| Tree cover and loss | Your browser downloads Global Forest Watch's public tiles of the same Hansen data (the ≥30% tree-cover mask and the encoded loss-year tiles), draws your shape onto each tile, and adds up the ground area of every pixel inside it. Zoom 12 (about 34 m pixels in Nepal) is used when the shape needs up to 90 tiles, coarser for very large shapes; the pixel size used is printed with the result. Expect small differences from the per-place figures, which use the full-resolution rasters. |
| Fires, incidents, earthquakes | The live points already on the map, filtered with a point-in-polygon test, for the time window selected. |
| People | For each municipality the shape overlaps, its population density × the part of the shape inside it — an estimate that assumes people are spread evenly within each municipality. |
| Climate and warming | The normals and trends of the overlapped municipalities, weighted by overlap. |
| Elevation | Copernicus GLO-90 at up to 100 evenly spaced points inside the shape, via Open-Meteo. |
Fire points come from NASA FIRMS' VIIRS S-NPP active-fire product, fetched through GeoSutra's /api/firms proxy and tagged with the municipality they fall in. Each point is a satellite-detected thermal anomaly, not a confirmed wildfire — crop burning and brick kilns also trigger detections, and cloud or overpass timing can miss real fires.
Disaster incidents come from the BIPAD portal (Ministry of Home Affairs) as reported, with deaths, injuries, missing people and houses destroyed; a handful of records dated in the future (data-entry errors) are left out. Earthquakes are USGS ComCat events of magnitude 2.5 and above.
History since 2012. A GeoSutra job (observatory-snapshot, a Supabase Edge Function) counts FIRMS detections and BIPAD incidents per district per day and stores the totals. It uses the FIRMS standard (science-quality) product where it exists and the near-real-time product after that, runs every three hours, and re-counts the last few days each time so late detections and late reports land. The backfill covers February 2012 to today. A district-day with nothing recorded simply has no row — nothing is estimated.
The canopy-height layer (Meta/WRI Canopy Height Map v2) is a machine-learning estimate trained on satellite and lidar reference data, not a direct measurement. It is shown on a fixed 0–40 m colour scale, and it is a single-epoch snapshot.
MODIS and VIIRS true colour, vegetation greenness (MODIS NDVI, 8-day), snow cover and land-surface temperature come from NASA GIBS for the date you pick. These products exist only down to about 250 m (zoom 9) or coarser; when you zoom in further the map enlarges the finest real image rather than asking for tiles that do not exist.
Everything above is a third-party, openly licensed dataset, used as published or aggregated. The exception is the Shree Manakamana Community Forest pilot: a DJI Mini 3 flew 4,269 RGB images at 4.4 cm resolution, and a DeepForest (RetinaNet) model fine-tuned on hand-annotated local training tiles detected individual tree crowns, validated against field plots. This is the site's only VALIDATED layer, and none of the others should be read with the same confidence as a field-checked survey.
Every dataset above has real, documented limitations — cloud contamination, coarse resolution relative to an individual tree, detection thresholds that miss small or gradual change, and definitions (like "loss") that do not map cleanly onto plain-language terms like "deforestation." GeoSutra's role is to make those limitations visible next to every number, not to paper over them with a confident-looking dashboard. Where a claim matters — a legal boundary, a carbon baseline, a community forest's actual condition — it belongs in a field survey, not a satellite pixel count.
| Label | Meaning |
|---|---|
| OBSERVED | Measured by a satellite or sensor (a fire point, a tree-cover-loss pixel, an elevation). |
| DERIVED | Calculated from observations (a trend, an area, a hectare count). |
| MODELLED | From a model or reanalysis (population, climate normals, projections). |
| ESTIMATED | GeoSutra's own approximation (people inside a drawn shape). |
| REPORTED | Taken from an agency as published (BIPAD incidents). |
| VALIDATED | A GeoSutra field or UAV observation, checked against ground data. |
Questions typed into the search box are answered only by the recipes below. A keyword parser (and, where enabled, a small open model) reads the question and picks a recipe. It never calculates anything. The recipe itself is fixed code over the datasets listed. "Partial" recipes return only what they can, and "planned" ones return "This analysis is not currently available" along with what it would take to build them. No recipe yet has site-specific accuracy for Nepal, and every answer says so.
| Recipe | Status | Method |
|---|---|---|
| Loading… | ||
Questions about a specific number on the map? Contact GeoSutra, or open the full source list.