Climate Home Atlas

Engineering case study

Finding places across scales

A climate-aware relocation or site-selection search starts with ordinary questions. How warm is the region? How long is the dry season? Is the terrain level enough for the intended use? Each answer comes from a different map.

Try a regional search

A worked example

The Atherton Tablelands

The one-degree square containing Atherton averages 21.61 °C over 1991–2020. It clears the tropical preset’s rainfall thresholds and misses its 22 °C temperature minimum. Its median sampled elevation is 630 m across 100 representative land points.

21.61 °C1991–2020 historical mean
24.15 °C2071–2099 · SSP3–7.0

Under this late-century pathway the same cell passes the preset. The terrain samples stay fixed. The result describes the whole climate cell, which includes land well beyond the town.

Inspect the Atherton cell

System architecture

From source to screen

The public explorer computes every search on the visitor’s device. The private research application runs independently.

Public data path
  1. Published observations & modelsClimate grids, elevation, land masks and burned area
  2. Python preparationVerify hashes, read raster windows and preserve missing values
  3. Reviewed releaseStatic datasets, cartography, browser code and a checksum manifest
  4. Cloudflare PagesStatic HTTPS delivery from the isolated release repository
  5. Visitor’s browserJavaScript applies hard filters. MapLibre renders evidence and comparisons.

Private Atlas 1.8

FastAPI validates requests from the manual UI, MCP tools and optional AI interfaces. Bounded raster adapters support finer climate profiles, experimental terrain, hazard layers and stored studies with per-metric provenance.

Possible future public extension

An isolated read-only API could serve a bounded high-resolution analysis or a constrained AI-assisted query. It would need its own data, resource limits and access controls. This is future work.

Methods

What counts as a match?

Climate. Python calculates annual temperature, rainfall, monthly temperature range, the longest run of dry months below 60 mm, and rainfall variability from complete sets of twelve monthly values. The search compares full-precision numbers. Rounding happens on screen.

Projections. The historical period is 1991–2020. The two future periods are 2041–2070 and 2071–2099 for SSP1–2.6, SSP3–7.0 and SSP5–8.5. The published monthly ensemble means and classification product have different resolutions. The authors screened 67 climate models and retained up to 42 depending on availability. A pathway represents a conditional future. These snapshots provide no probability or local confidence interval.

Terrain. Each dot is one representative land elevation from the 30-arcsecond surface grid within a 0.1° square. Each relief value is the 90th minus 10th percentile land height within 10 km, with at least eight valid source samples. A climate cell reports the median across its indexed land points, requiring complete values for that measure. The colour scale changes only the drawing.

Fire history. The quarter-degree layer gives estimated mean annual burned area as a percentage of burnable area during 2001–2020. It includes agricultural and managed burns. For a one-degree search, all sixteen subcells must have values before Atlas reports their maximum. Estimating a property’s fire probability or losses would need separate hazard, exposure and vulnerability evidence.

Decisions. A known failed condition excludes a cell. If the known conditions pass and required evidence is missing, the result is unknown. The five warmest passing cells form a shortlist. Visitors can inspect the comparisons that produced each result, then pin two cells side by side.

Implementation & verification

Keeping the evidence traceable

The preparation step reads published raster inputs with Python, NumPy and geospatial raster tools, checks their hashes, and exports just the three regional windows. Each release records its Git source revision and SHA-256 artifact checksums. GitHub holds the development history and a separate release repository with the reviewed output.

Independent tests compare all 560 cell-period records to the original monthly arrays and georeferenced class rasters. Other checks decode terrain sample positions, recalculate relief and burn summaries, exercise missing-data rules, and reproduce the packaged files. Chromium tests exercise WebGL, touch-sized layouts, projection changes, comparisons and failed requests. Cold-load and warm-search benchmarks use the original demo as a baseline.

The recurring engineering lesson was that resolution is part of the measurement. Replacing a missing coastal fire value with zero would admit an unsupported match. Painting a hillshade at a finer display resolution would leave the original elevation measurements exactly as coarse as before.

The project owner defined the geographic problem, source constraints and acceptance criteria. Development used AI coding assistance for implementation and testing. In the private research application, optional AI tools translate requests into validated application operations. Visitors use explicit controls in this public explorer.

Provenance & licences

Source register

Climate and Köppen–Geiger classes

Beck et al. (2023), High-resolution (1 km) Köppen–Geiger maps for 1901–2099 based on constrained CMIP6 projections. Atlas subsets the V3 1° monthly means and counts the dominant class among valid 0.1° class pixels in each degree square. Ties use the lower published class ID. The regional outputs are adapted under CC BY 4.0.

Elevation and relief

NOAA NCEI ETOPO 2022 v1, MacFerrin et al. (2025). Heights use metres relative to EGM2008. Numerical samples use the 30-arcsecond source. The backdrop uses 15-arcsecond source elevations, reprojected into Web Mercator and shaded with threefold vertical exaggeration. Coastlines and chart colours supply orientation.

Land mask and cartography

Terrain selection uses the 2020 GLC_FCS30D land-cover overview by Liu, Zhang & Zhao, distributed via the Natural Capital Alliance, intersected with Natural Earth land polygons. Made with Natural Earth’s public-domain country outline at 1:50m and land, towns, generalized roads and rivers at 1:10m. Road lines give geographic context only.

Historical burned area

Chen et al. (2023), Global Fire Emissions Database GFED5 Burned Area v0.1. The preparation sums the “Total” area across 240 months, divides by 20 and by burnable area, and exports regional subcells with their masks. Repeated burning contributes to the estimate.

Software and delivery

MapLibre GL JS 6.18.0 ships locally under its BSD licence and bundled third-party notices. DejaVu glyphs are rasterized into the chart images. The browser loads scripts, images and data from the demo origin. Preparation uses fixed source downloads, and the reviewed derivatives can be cached and redistributed under the terms above.