Methodology

Evidence before interpretation.

Every published dataset carries its source, version, dates, license, processing method, coverage, limitations, and uncertainty. Missing measurements remain UNKNOWN.

Overall data pipeline, CRS and evidence semantics

Source acquisition happens offline. Pipelines pin exact source objects or API responses, validate identity and required fields, preserve source licences/dates, transform only what the documented feature requires, then publish immutable browser artifacts with SHA-256 checksums. The public site reads those local artifacts; it does not silently refresh third-party APIs at visitor time. The Data Catalog exposes the release-level lineage and Sources exposes originating providers.

Vector interchange uses OGC:CRS84 longitude/latitude unless a feature explicitly documents an analysis CRS. Equal-area calculations such as exposure area weighting use EPSG:6933. Native raster CRSs/grids are retained for numerical analysis when required; display reprojection is a separate presentation step. Web Mercator or RGB terrain products are never substituted for the native numerical source merely because they are convenient to render.

Geometry simplification and tiling are presentation operations unless a feature explicitly defines otherwise. Administrative display geometry is topology-preserving, major-road display geometry is simplified within the documented tolerance, and terrain/satellite display tiles have their own resolutions. Analysis features use the source/native or explicitly prepared analysis geometry named in their methodology rather than inferring precision from pixels on screen.

Observed

Measurements or records supplied by a documented source.

Reported

A source assertion or event report; not automatically an independently observed measurement.

Historical

Source-reported past events or archival observations; historical does not mean current.

Derived

Deterministic output calculated from source data using a documented method.

Estimated

A quantified estimate whose assumptions/coverage remain explicit; not an observed total.

Modelled

Output produced by a model or hypothetical spatial assumption.

Simulated

A modelled scenario run with explicit parameters; never an official forecast unless a future validated forecasting product explicitly says so.

Hypothetical

An explicitly assumed scenario or input, not an observed event.

Unknown

Reliable information is unavailable or the concept is not applicable.

Uncertainty is not reduced to one generic score. Each feature preserves the uncertainty/limitations its source or method can support. UNKNOWN, partial coverage, unavailable and incompatible states are intentionally different from zero.

Population for public exposure (HRSL)

Feature 48 adds Meta/CIESIN High Resolution Settlement Layer v1.5 (CC BY 4.0) as a separately registered public population source; WorldPop remains excluded from the public build. The two 1 arc-second source tiles intersecting Nepal are pinned by SHA-256. People in 1″ cells whose centres fall inside the COD-AB v02 Nepal boundary are summed exactly into a 3″ analysis raster (offline exposure) and a 30″ browser grid (scenario screening). HRSL allocates census counts to building-detected cells, so inside Nepal a missing cell contributes zero people; outside Nepal the grid is UNKNOWN. The sum is preserved (29.16 million) and no resampling or smoothing is applied. HRSL is a modelled allocation, not a census; no per-cell confidence interval exists.

Flood and GLOF corridor scenarios

Feature 49 registers 18 hypothetical release points: the ten largest glacier-fed Nepal lakes in the Glacial Lake Observatory, each linked to the nearest HydroRIVERS reach within 2 km (declared assumption — the real outlet is UNKNOWN), and the main stems of the eight largest river outlets traced upstream to where they enter the mapped network. Paths follow NEXT_DOWN until the network edge. Offline, corridors of 250, 500 and 1,000 m each side of reach-complete path prefixes (10, 25, 50, 100, 200 km and the full path) are intersected with HRSL population (exact equal-area cell fractions) and mapped OSM assets (each element once). In the simulator a user-declared triangular or rectangular release hydrograph conserves volume exactly and is translated without attenuation at a declared celerity ensemble. Arrival ranges are sensitivity to that ensemble; exposure ranges are sensitivity to corridor width. Neither is a confidence interval, a flood extent or a forecast. Depth, inundation extent, damage and losses remain UNKNOWN.

Earthquake shaking scenarios

Feature 51 implements the BSSA14 ground-motion model (Boore, Stewart, Seyhan & Atkinson 2014, NGA-West2, doi:10.1193/070113EQS184M) with global coefficients, independently verified against 1,320 pygmm reference cases to floating-point precision. Scenarios replay USGS ComCat epicentres and magnitudes or use a custom epicentre; mechanism, a uniform V_S30 site condition and an optional declared line rupture are explicit assumptions. Population (HRSL 30″) and mapped assets are counted by median PGA band and with ln PGA shifted by ±1 total sigma — a single-site variability bound, not a correlated field. Inputs outside the model domain (M 3–8.5, normal M ≤ 7, R_JB ≤ 400 km, V_S30 150–1500 m/s) are refused. Basin effects are not modelled. Recorded Gorkha 2015 ground motion is UNKNOWN in this build. Building damage, casualties and losses are UNKNOWN.

Rainfall, snow, drought and heat context

Features 54 and 55 use NASA POWER v10 monthly MERRA-2 reanalysis (CC BY 4.0) on its native 0.5° × 0.625° cells that intersect Nepal; only the source chunks covering Nepal are pinned by SHA-256 and nothing is interpolated. For the latest 12 complete months the Hazards page shows precipitation, snowfall, snow depth, snow-cover fraction and the month's highest hourly 2 m temperature against 1991–2020 normals. SPI-3 (McKee et al. 1993) fits a gamma distribution with a zero-probability term to 1991–2020 three-month totals for each calendar month and cell, then converts the latest total to a standard normal deviate; the fit is re-derived by the release verifier from the published baseline totals. National shares are weighted by each cell's Nepal area. This is reanalysis context, not station observation, a drought declaration, a heatwave classification or a forecast; in the dry season small totals can give large SPI values. The snapshot is labelled STALE two months after its latest month.

Terrain steepness for landslide awareness

Feature 53 mosaics the 45 pinned Copernicus GLO-90 tiles over Nepal on their native 3″ grid, computes Horn (1981) slope with latitude-dependent spacing, and reports for each COD-AB district the area share in declared slope classes (0–5°, 5–15°, 15–30°, 30–45°, ≥ 45°), mean and median slope and elevation, weighted by cell area. About 30% of Nepal is steeper than 30° at this resolution. Steepness is context only: landslide susceptibility, probability and hazard are UNKNOWN because no reviewed, inventory-calibrated model with out-of-sample validation is available. GLO-90 is a surface model whose 90 m cells smooth local slopes.

Hazards hub and gated features

Feature 56 brings these contexts, the scenario simulator and three gated features together. Gated features — DHM/NDRRMA-BIPAD feeds (44), verified physical flood cases (50) and earthquake damage (52) — are built in a disabled state; their unlock conditions are listed on the page and in docs/hazards-hub.md, and nothing is estimated before they are met.

Evidence analyst

Feature 58 retrieves verbatim paragraphs from the public feature documents and the method, uncertainty and limitation fields of public model releases (atlas-public-evidence). Ranking is the documented lexical score; the analyst quotes passages with citations and never generates or paraphrases factual text, calls a language model or contacts an external service. An empty result is reported as insufficient evidence, never as safety or zero impact.

Live conditions page and bilingual bulletin

The Live conditions page presents only verified Feature 43 snapshots: the same-origin index and each snapshot are size-, SHA-256-, schema- and source-review-checked before any value is shown. Bulletin sentences are generated from those records with fixed English and Nepali templates; unverified, corrupt or absent data produces an UNAVAILABLE sentence rather than a reading. Observations, reported events, model forecasts and official warnings carry distinct glyph-and-text labels, never colour alone. Every record shows its source, observation or issue time, fetched_at and a FRESH / STALE / UNAVAILABLE state recomputed on the viewer's clock from source, fetch and workflow timestamps. Workflow health and last successful fetch are shown separately. GFS grid squares outline native 0.25° cells for display only; values are not interpolated and legend bins are not hazard thresholds. Official warnings are not ingested, and their absence never means no warnings: DHM and NDRRMA/BIPAD remain the authorities. Air quality stays OFF with AQI UNKNOWN. Damage, loss, inundation and casualties remain UNKNOWN. Nepali copy is project-authored and requires native-speaker review.

Offline shell and last-known copies

A service worker written for this atlas keeps a small app shell (home, Live conditions, Methodology, Sources and an offline page) on the device. Every shell file is pinned at build time by byte size and SHA-256 and verified before installation; a mismatch keeps the previous version. Live data is network first: any HTTP answer is treated as current, and only a network failure or timeout falls back to a stored copy. A copy is stored only after it matches its index reference, is re-verified before every offline use, and is deleted after seven days. Offline copies are labelled LAST KNOWN with the time they were saved on this device and are never shown as freshly checked; source freshness is still recomputed from the original timestamps, so an expired copy reads LAST KNOWN · STALE. Caches are versioned, and an update activates only when the reader chooses to reload. The map library is not precached; offline, records remain in the tables.

Air quality: disabled until provider review

OpenAQ PM2.5 acquisition is off by default and prohibited for public builds. Its provider allowlist is empty. A research-only adapter validates stationary Nepal station/provider/sensor identity, exact licences, original units, averaging periods, timestamps and quality flags. Flagged or unknown-period/quality concentrations remain UNKNOWN; source values are retained separately for review. AQI remains UNKNOWN because no calculation standard or sufficient averaging/completeness inputs have been reviewed. No browser API key or request is used.

Manual open feeds: USGS and NOAA GFS

Bounded offline downloads preserve source revisions and raw SHA-256. USGS retains preferred USGS-network earthquake summaries for seven days in the regional box; other networks are withheld. NOAA GFS native 0.25° APCP gives one six-hour interval at cycle +18 to +24 hours, with its interval decoded from GRIB2 template 4.8. It is a model forecast, not observed rainfall, a daily total or an official Nepal warning. No interpolation or hazard threshold is applied.

Immutable snapshots and source-review manifests are committed only to a dedicated live-data branch by a manual workflow. Static builds import a full pinned data commit and verify every referenced byte and source policy. A small latest index is promoted only after staging succeeds. Failed attempts retain the last verified snapshot as STALE; a valid empty collection is separate from failure. Source freshness and workflow health use their original timestamps. The default deployment is unconfigured until a reviewed data commit is pinned.

Live contracts: evidence, freshness and workflow health

Feature 42 defines validated static JSON contracts for periodically updated conditions. It acquires no feeds and adds no forecast or warning service. The Data Catalog exposes synthetic contract cases containing no measurements or locations, with fixed test timestamps and explicit SYNTHETIC FIXTURE labels.

Evidence (OBSERVED, REPORTED, DERIVED, MODELLED, HYPOTHETICAL or UNKNOWN), product purpose (observation, reported event, model forecast, official warning or scenario), source freshness and workflow health are separate fields. Forecasts require modelled evidence and issue/validity times. Unknown values remain null/UNKNOWN, while a source zero remains zero. Physical inundation, destroyed buildings, casualties, repair costs, hydropower downtime and economic loss must all remain UNKNOWN in these contracts.

Freshness uses source issue time or the oldest known observation in a snapshot, never a renewed fetch time. Explicit expiry and warning validity can shorten that deadline. Unknown warning validity remains STALE. A failed fetch retains the last verified snapshot as STALE; a missing snapshot is UNAVAILABLE. A successful empty response is EMPTY and cannot mean all-clear. Workflow health has its own heartbeat deadline, last attempt and last successful fetch. The browser rechecks deadlines as time passes and when a tab resumes.

Only bounded, same-origin JSON is read. Schema, semantic identity, byte size and SHA-256 checks must pass before display. Checksums detect changed bytes; they do not establish scientific accuracy, authenticity or redistribution permission. Source-specific review and immutable provenance remain required before acquiring or publishing real feeds.

Official warning authorities: DHM · NDRRMA · BIPAD. The Atlas is not a real-time warning service.

Search

The global Search index is derived offline from the exact immutable atlas releases and split into eight gzip shards, each capped at 2 MiB compressed and 16 MiB decoded. Every indexed record keeps its source dataset/version, stable feature/source ID, entity type, representative position, source date and disambiguating context. The browser verifies each shard checksum before use and processes shards sequentially on explicit search submission.

Unicode NFKD normalization supports diacritic-insensitive matching without inventing different-script transliterations. Ranking is canonical exact, alias/stable-ID exact, canonical prefix, alias prefix, canonical substring, alias substring, then explicit type terms with deterministic tie-breaking. Duplicate canonical names remain separate results. Current HydroRIVERS reaches and GLO records have no verified common-name linkage in these releases, so unsupported river/lake names are not manufactured.

Compare Mode

Compare Mode uses Search only to resolve stable identities, then reloads the two exact versioned source datasets. The current contract supports district-to-district, mountain-to-mountain, river-reach-to-river-reach, glacier-to-glacier, glacial-lake-to-glacial-lake, BIPAD-event-to-BIPAD-event and USGS-earthquake-to-USGS-earthquake comparisons. Arbitrary Location Explorer coordinates are excluded because no versioned numerical location snapshot is published yet.

Compatibility depends on entity family, metric definition and unit; equal units alone do not make two measurements comparable. Missing values remain UNKNOWN, source zero remains zero, and incompatible metrics are blocked. BIPAD nominal estimated loss is excluded because values are not normalized across years. The UI never ranks entities and always shows source, dataset version, dates, temporal coverage, spatial resolution and evidence type for both sides.

Simulation UI

The Feature 29 workbench is a user interface over Feature 28’s registered network approximations, not a new physical hazard model. Before execution it verifies both immutable scenario examples and requires the same source HYRIV reach, input dataset versions/hashes and ordered pathway. Level 1 uses network-path@1.0.0 with no physical parameters. Level 2 uses constant-celerity-pulse@1.0.0 with explicit celerity (0.1–10 m/s), hypothetical release volume (0–10,000,000 m³) and duration (60–86,400 s). Invalid or nonfinite values are blocked before calculation, and Level 3+ remains unavailable because no physical adapter is registered.

Level 2 applies only the validated translation equations already documented by the Scenario Engine: source cumulative reach length divided by user-assumed signal celerity gives relative entry/exit delays; duration extends the rectangular pulse; volume divided by duration gives signal discharge. Celerity is not measured water velocity and relative delays are not predicted arrival times. Default values independently reproduce the published Feature 28 reference calculation; custom values have no separate real-event or published-reference validation.

Every result is labelled MODELLED SCENARIO — NOT AN OFFICIAL FORECAST, retains model/version, exact input source/version/hash metadata, assumptions, partial network coverage, validation status and limitations, and keeps source-derived geometry separate from modelled quantities. Current models return no inundation footprint, depth, hydraulic velocity or confidence interval, so the UI reports those as UNKNOWN and does not run the Exposure Engine. Side-by-side run comparison is allowed only for the same level, model/version and verified source pathway. Flat and perspective map views share exactly one simulation output; camera perspective does not change the model.

Time Machine

One UTC observation date synchronizes water imagery/comparisons, satellite context, monthly climate and reported event days. Acquisition, source publication and retrieval remain separate. Missing dates stay unavailable; no nearest date is substituted. Monthly data remain whole monthly aggregates. Events use occurrence dates, not report dates. Before/after water imagery shares a grid and chronological dates, while its quality gate still governs measured change. Different satellite footprints cannot form a change comparison. Other inventories and scenario layers retain their own dates and are not reconstructed historically.

Water Change

Three April Sentinel-2 observations around Phewa Lake share a native 20 m UTM grid. Calibrated green/NIR reflectance, NDWI above 0.05 and SCL water must agree. Land requires NDWI below −0.05 and valid non-water SCL. Clouds, shadows, snow, a 20 m exclusion buffer, uncertain pixels and water patches below 0.0036 km² remain UNKNOWN. Only jointly valid pixels contribute to gain, loss and persistence. Both full-window and mapped-water-union coverage must reach 80%; failing comparisons have UNKNOWN metrics. Display masks are reprojected separately from native area counts.

These dates do not establish seasonal behaviour, trends, flood extent or causes. Mapped area is partial coverage, not total lake area. Cross-sensor, atmospheric and shoreline uncertainty remain; no independent accuracy estimate is available. Coverage QA · Source windows and calibration.

Population preparation

The Atlas pins the WorldPop Nepal 2025 R2025A v1 constrained Float32 GeoTIFF by SHA-256 and validates its EPSG:4326 CRS, 9,773 × 4,921 dimensions, 3 arc-second cell spacing, -99999 NoData value, finite non-negative valid cells, and source raster bounds. Valid zero population cells remain distinct from NoData.

The native source raster remains the numerical analysis input. A separate EPSG:3857 zoom 5–10 PNG pyramid is generated only for visualization using bilinear reprojection and a log-scaled intensity transform. Display pixels therefore cannot be sampled or summed as population counts. The national source-cell sum is retained as an integrity check, not presented as a census total or an administrative validation target.

Future exposure calculations must use the pinned native people-per-grid-cell raster and explicitly define raster/vector alignment, NoData handling and partial-cell weighting. The current layer performs no hazard exposure estimate.

Satellite observation preparation

The Atlas uses three fixed Sentinel-2 Collection-1 Level-2A observations selected from 1 April through 15 June 2026 as separate western, central and eastern context windows. Candidate scenes must have source cloud cover at or below 5% and zero Scene Classification Layer NoData. The selected 10 m true-colour COG and 20 m SCL objects are pinned by exact URL, ETag and byte size before they are read; a changed upstream object fails the build instead of silently replacing a release input.

The SCL is independently counted for cloud classes 8, 9 and 10, NoData class 0 and snow/ice class 11. A 768 × 768 RGB preview is then produced from each full-resolution TCI by bilinear resampling and checksum-verified in the browser before display. The preview does not retain 10 m analytical resolution. Clouds, shadows, snow and visual differences between observation dates remain image content only: this feature performs no water classification, land classification or change detection.

The three scenes have different acquisition timestamps and disjoint/overlapping footprints, so they are deliberately presented as observation windows rather than a seamless or single-date Nepal mosaic. Areas outside those footprints are unavailable, not filled from another source.

Climate preparation

The Atlas acquires NASA POWER monthly T2M and PRECTOTCORR for 1991–2020 from the source-native MERRA-2 grid. The API reports 0.5° latitude × 0.625° longitude spacing, UTC monthly periods, temperature in degrees Celsius and corrected precipitation as a mean daily rate in mm/day. POWER annual fields are deliberately excluded; no daily or station-level precision is reconstructed from monthly data.

For each of the 360 months, every POWER grid box intersecting Nepal is overlaid with the pinned, unsimplified COD-AB v02 national boundary. Intersection pieces are transformed to equal-area EPSG:6933, and valid grid values are weighted by the Nepal area inside each cell. The release uses 66 contributing grid cells and requires complete valid-area coverage; every published month has 100% coverage. This produces a Nepal-wide national summary, not a local interpolation or a new climate raster.

The 12 displayed 1991–2020 monthly normals are arithmetic means of the 30 area-weighted values for each calendar month. The accompanying minimum and maximum are the range of those 30 monthly reanalysis values and show historical reanalysis variability only; they are not confidence intervals or measurement uncertainty. MERRA-2 assimilation, model physics, topographic representation and POWER processing limitations remain applicable. The baseline is historical context, not a forecast or current-condition estimate.

Exposure engine

Exposure means a spatial intersection with an explicit footprint. It does not establish vulnerability, risk, confirmed damage or losses. The prepared corridors are hypothetical areas of interest around one retained downstream trace, with assumed widths of 250 m and 1,000 m each side. They are not flood simulations.

The offline engine accepts versioned, valid Polygon or MultiPolygon footprints in OGC:CRS84 within the Nepal region. Event points and bare river lines are not exposure areas. It overlays the pinned native WorldPop 2025 GeoTIFF, never the population display tiles. For each valid source cell, its population is multiplied by the intersected fraction of cell area in equal-area EPSG:6933, then summed. This assumes uniform population within a cell. Native resolution is preserved and results are rounded to people for display.

NoData and out-of-grid areas remain UNKNOWN. A known-cell subtotal is shown separately, and any unknown area makes the whole-footprint population estimate UNKNOWN. No uncertainty interval is invented. OSM geometry intersection counts mapped points, lines and polygons, including boundary contact; polygons count as whole assets. Shared OSM IDs are deduplicated, while road and bridge representations are tested separately. Buildings and dams have no inventory in this calculation.

District area pieces are made disjoint in P-code order, with unassigned coverage retained. Each matched asset is assigned to the greatest intersected district area or length, with lowest P-code breaking ties. Category counts can overlap and must not be added as a unique total. Source geometry simplification, representative facility centres, mapping completeness and differing source dates limit interpretation.

Calculations run offline to avoid national raster processing in a visitor’s browser. Downloaded result JSON, footprint requests and spatial layers record source versions/hashes, calculation timestamp, method version, CRS, assumptions and limitations. A new footprint is computed through the offline pipeline and published as a new immutable result before it is registered in the viewer.

Downstream trace

The Atlas combines both verified Nepal river partitions and follows HYRIV_ID / NEXT_DOWN pointers from a selected whole reach. Duplicate IDs, inconsistent internal links, invalid lengths and cycles prevent tracing. Each retained reach is counted once, and LENGTH_KM values are summed in kilometres. The first downstream ID outside this release ends the trace with a coverage-boundary notice; an absent source connection is labelled a source outlet.

Results are ATLAS DERIVED network analyses. Gold lines and playback show reach order, not a flood footprint, flow velocity or travel time. The selected reach is included in full, irrespective of where it was clicked. Source geometry is unchanged, with no inferred connections across gaps. The downloaded JSON records method version, ordered reach IDs, termination, source versions, hashes and limitations.

Lake/glacier outlet links and settlement, infrastructure and population intersections remain UNKNOWN. A shared basin or nearby river is insufficient to establish an outlet connection. No terrain-flow or hydraulic model is used. See the HydroRIVERS attribute definitions.

Glacial-lake preparation

The Atlas pins the GLO v1.02 Sentinel-2 unique-lake centroid and polygon-attribute GeoPackages by SHA-256 and joins them by stable GLO_ID. The EPSG:4326 source centroids are used directly for browser geometry, while source equal-area polygon area/perimeter attributes are retained without reprojecting those polygons.

AREA_DISSOLVED is the source dissolved maximum mapped extent across 2017–2024, not a current lake-area measurement. Expansion rate, uncertainty and significance are source time-series/statistical fields and are never converted into a hazard class. Specific glacier and river links remain UNKNOWN because this source layer does not identify them.

Glacier inventory preparation

The Atlas pins RGI 7.0 South Asia West and South Asia East WFS subsets by SHA-256, selects complete glacier features intersecting the validated Nepal country polygon, and retains stable RGI/GLIMS IDs, source area, source outline date and RGI elevation attributes.

Some GLIMS WFS display geometries are invalid after its coordinate transformation. Only the browser geometry is repaired with make_valid and topology-preserving simplification; source RGI area remains the measurement. QA records every repair, and the UI labels every outline as a dated inventory record rather than a current margin.

River network preparation

The Atlas selects every FAO Rivers 2026 reach whose complete source geometry intersects Nepal. It does not clip or simplify the network. HYRIV_ID and NEXT_DOWN are retained so later downstream analysis can follow source topology, including links crossing the two browser-delivery partitions.

The partitions exist only to keep decoded GeoJSON under 8 MiB. Average discharge and flow-regime attributes are long-term/source descriptors, not live hydrological observations.

Mountain catalogue preparation

The atlas pins the GeoNames Nepal country dump by SHA-256, selects only terrain feature codes PK and MT, retains stable GeoNames IDs, source coordinates, explicit source elevations, names and bounded aliases, and rejects duplicate IDs or implausible coordinates/elevations. The GeoNames DEM fallback is deliberately not substituted for a missing summit elevation.

Infrastructure preparation

The Atlas pins one build-time OpenStreetMap snapshot for bridge/facility/settlement positions and five smaller major-road geometry slices by SHA-256. The road slices are merged by stable OSM way ID; repeated slice occurrences are removed without geographic/name heuristics. Every retained point is inside the unsimplified Nepal COD-AB v02 country polygon, and road ways are intersected with that polygon before browser simplification.

Major roads retain source way geometry but use a topology-preserving 0.0002° display simplification. School, health, emergency and bridge ways/relations are represented by source-provided Overpass centres for browser exploration; those points do not replace source OSM geometry for future precise analysis. Buildings are deliberately omitted because uneven national OSM completeness would make exposure counts misleading. Inventory presence is not an exposure, vulnerability, damage or risk result.

Administrative boundary preparation

The atlas pins the Nepal COD-AB v02 source archive by SHA-256. The offline pipeline checks the expected 1 country, 7 provinces, 77 districts, and 775 level-3 pieces; unique P-codes; closed and valid Polygon or MultiPolygon geometry; parent containment; neighboring units; and complete Nepal coverage without unexplained gaps or overlaps.

Level 3 includes 753 local-government units and 22 protected or special-area pieces. The atlas retains both categories because together they cover Nepal. Ward boundaries are unavailable in this release.

Display and analysis geometry

The source uses OGC:CRS84 longitude and latitude. Administrative web geometry is simplified as a shared coverage so neighboring edges remain aligned, then compressed for delivery. It supports identification and display only. Pinned unsimplified/raw sources remain outside the browser for future analysis where licensing permits.

Reported events and hazard records

BIPAD disaster events, USGS earthquakes, reported floods and reported landslides remain source-reported records. Event points are locations supplied by their source; they are not automatically affected-area polygons. Reported casualty/loss fields remain source values, missing values stay UNKNOWN, and event-page narrative is limited to supported source fields. Proximity to a reported event does not establish exposure, damage, cause or future hazard.

Terrain preparation

Copernicus GLO-90 native Float32 rasters are retained at their original 3 arc-second sample spacing and aligned in an offline analysis mosaic without resampling. The source is a surface model, including buildings and vegetation, with EGM2008 heights in metres.

Web display uses bilinear reprojection to aligned Web Mercator grids and Terrain RGB tiles at zooms 5–9. Each tile is checked against its recorded SHA-256 before rendering. Finest display spacing is roughly 270 m near Nepal; encoding in 0.1 m increments is not an accuracy claim. Elevation inspection always samples that finest display grid, independently of map zoom and visual exaggeration. Hillshade is a derived lighting effect, not an observation or slope analysis.

A full rectangular tile domain surrounding Nepal prevents artificial zero-height seams at national borders. Regional plausibility probes are checked near Everest, Kathmandu and Biratnagar; they do not substitute for survey validation. No hydrological conditioning or bare-earth correction is performed.

Regional overview

A pinned Mapzen terrain pyramid fills the Asia overview. Its source heights are not normalized to the Copernicus vertical datum and are not used for measurement. Negative heights are displayed as sea surface. In the context buffer outside Nepal, display heights transition over one degree between the detailed tile domain and the coarse backdrop; interpolation operates on decoded heights, never on RGB colour channels. Copernicus elevation inspection reads the original verified tiles.

Research boundary

No operational forecast, hazard classification, confirmed-damage estimate or risk score is included in this release. The exposure engine reports explicitly hypothetical spatial intersections, while source-reported areas, satellite imagery and change statistics remain descriptive evidence rather than predictions or impact claims.

Hazard Graph

Directed edges carry source records, methods, evidence class and assumptions. River NEXT_DOWN links are derived drainage topology. Corridor exposure links are modelled and conditional on hypothetical footprints. Dashed inferred/modelled links are never presented as observed causes. Missing links and unsupported confidence remain UNKNOWN. Diagram positions are schematic; traversal is bounded and cycle-safe.

Scenario Engine

Level 1 traces source NEXT_DOWN pointers. Level 2 translates a hypothetical rectangular release at one explicitly assumed celerity: delay equals source distance divided by celerity and discharge equals volume divided by duration. No attenuation, terrain hydraulics, depth, flood footprint or water velocity is calculated. Celerity is not water velocity. These are educational network approximations with synthetic analytical validation only, not forecasts or emergency predictions. Coverage exits and unsupported quantities remain explicit.

Definitions, assumptions, model versions, units, source hashes and processing versions accompany every result. Invalid inputs fail offline before publication; the browser only inspects immutable, verified runs. USACE lag-model concept provides context for translation without attenuation; the Atlas does not execute or claim calibration from HEC-HMS.

Evidence retrieval

The evidence browser retrieves exact excerpts from eight versioned Atlas methodology documents. It preserves document hashes, sections, dates and linked dataset manifests. These are project descriptions, not independent observations or a complete report archive. Keyword matches do not establish factual support; missing evidence and UNKNOWN freshness remain explicit. Corpus excluded from the public release pending redistribution review..

AI Analyst

The local analyst answers supported questions using exact project-document excerpts, verified downstream topology and registered hypothetical scenarios. Citations retain source versions, dates and limitations. Calculations and modelled results remain distinct from observations; unsupported measurements and interpretations stay UNKNOWN. It introduces no new upstream source or runtime language-model service.

Physical GLOF modelling project

Explore Nepal focuses on glacial lakes, river connectivity, terrain and optional infrastructure. The research atlas retains the wider scientific tools and historical layers.

The physical project covers the 4,150 lake records in the Nepal/transboundary GLO inventory as candidates for input assessment. A transboundary location alone does not establish downstream drainage into Nepal. Every candidate currently requires a verified outlet, terrain, bathymetry or a surveyed volume relation, lake stage, breach assumptions, roughness and boundary conditions.

The offline ANUGA 4.0.1 adapter solves shallow-water flow on a supplied triangular mesh for an explicitly assumed instantaneous breach. Dry dam-break analytical checks, mesh refinement and lake-at-rest checks verify numerical behaviour on synthetic cases. These checks are not validation of a Nepal GLOF. No real-lake inundation result has been published.

Publishing a result requires independent input review, mass balance, domain and mesh convergence, sensitivity analysis and relevant observational or laboratory validation. Water-only shallow-water flow does not model sediment-rich debris flow, evolving breach erosion, structural failure or destruction. A validated flood footprint could support a subsequent asset intersection; damage would additionally require appropriate vulnerability functions. Until then, flood extent, affected assets and damage remain UNAVAILABLE.

ANUGA validation methods. The separate offline model is not run in the browser.

Geographic names

District display names follow the referenced district list; Eastern Rukum and Western Rukum retain the source variants Rukum East and Rukum West. The 753 local-government names were compared with the government directory; Manang Ngisyang uses its municipality’s full name, while Dodhara Chandani follows its district administration. Local governments are not an exhaustive inventory of cities or settlements.

Named river locations come from the GeoNames Nepal snapshot of 22 September 2026. Original names and alternate spellings remain visible. These points are not a name crosswalk to HydroRIVERS: nearby reaches and entire downstream chains do not inherit a name. Missing reach identities remain UNKNOWN.

Government local-level directory · GeoNames source and attribution.