The Refugia habitability index is a composite raster across seven environmental factors. Each factor is scored on a continuous gradient between a "fully habitable" anchor (score 1.0) and an established uninhabitability threshold (a hard cliff at score 0). Within the habitable band, scoring is graduated rather than binned — a boreal forest does not score the same as a Mediterranean climate the way a 3-class scheme would force. The composite at each cell is the geometric mean of the seven factor scores, but if any factor hits its cliff the cell is marked uninhabitable (composite = 0).
Each factor scores cells on a 0 – 1 scale. The ramps below show the colour of a cell's score from cliff (red, score 0) to fully habitable (green, score 1). Tick marks call out the values at which a cliff or anchor sits, and any intermediate breakpoint that's methodologically meaningful.
The climate-driven factors (wet-bulb temperature, drought, sea-level rise) are computed from a multi-model ensemble rather than any single climate model's projection. Above: each line is one CMIP6 SSP5-8.5 long-extension model's annual global cos(lat)-weighted mean tasmax anomaly relative to the 2015–2034 reference period, drawn from the cached monthly NetCDF on this machine. Hover a line to identify the model. The fan grows from a tight cluster near 2020 (all models still close to the historical baseline) to a 9–15 °C spread by 2300 — that spread is irreducible structural uncertainty about how sensitive the climate system is to forcing, and it is the central reason the habitability surfaces should be read as ranges, not point estimates.
The production central estimate for wet-bulb & humidity is the ECS-screened 6-model mean: CanESM5 (ECS 5.62 K) and UKESM1-0-LL (ECS 5.36 K) sit above AR6 WG1's "very likely" 5.0 K upper bound and inflate the multi-model mean by ≈ 0.9 °C at 2300 (Hausfather et al. 2022, Nature 605:26). Both hot models are still shown here (dashed red) for spread visibility but excluded from the central estimate. CESM2-WACCM is in the per-epoch ensemble means but drops from this trajectory plot because its run begins 2101, below the 2015–2034 baseline window. The broader CMIP6 SSP5-8.5 ensemble contains roughly 30 additional models that all reach 2100 but lack the long extension.
Slope limits where people can build and farm. Steep ground drives up construction and transport costs and rules out most mechanised agriculture, so population density falls off as terrain steepens. People settle a wide range of slopes, but above a certain grade sustained settlement and farming stop being practical.
Slope is computed from the GMTED2010 1 km global mean elevation product (Danielson & Gesch 2011, USGS OFR 2011–1073) using Horn's algorithm (Horn 1981, Proc. IEEE 69:14) and aggregated as cell-mean to the 30 km analysis grid. The resulting continuous slope-in-degrees raster is scored as a piecewise-linear suitability ramp: ≤ 4° → 1.0 (flat to gently undulating, fully habitable for any settlement or agriculture); ≥ 30° → 0 (steeply mountainous, hard cliff — structurally infeasible for sustained settlement); linear interpolation in between. An earlier 3-class split used roughly 8%/16% slope-class boundaries (≈ 8° / 16°), which over-penalised mid-slopes where settlement remains common; the single 30° cliff is the more defensible engineering limit for sustained settlement. Topography is time-invariant across all epochs.
Soil sets how much food a place can grow locally. Global trade moves calories around, but local productivity still matters for sustained habitation, and it matters most when long supply chains are disrupted. Permafrost is a special case, and warming cuts both ways on it: frozen, it is unbuildable and biologically inert; thawing, it becomes waterlogged and unstable and starts venting methane. Neither state supports settlement or agriculture.
Soil-quality classes from the FAO/IIASA Harmonized World Soil Database (HWSD v1.2, 1 km native, "soil quality 1 — nutrient availability" derivative) are scored on a continuous gradient: class 1 (no/slight constraints, best) → 1.0; class 6 (very severe constraints) → 0.15; linear interpolation in between. Cliffs: class 7 (FAO's own label "unsuitable, ineffective" — desertic or glacial soils with no agricultural potential) and class 0 (non-soil / inland water) both → 0. For 2100 and 2300 epochs, permafrost zones projected to thaw under CMIP6 SSP5-8.5 mean annual temperature are reclassified to a marginal-equivalent score of 0.5 — thawed permafrost is thermokarst-prone, of variable fertility, and foundation-unstable. Chadburn et al. (2017) constrains where permafrost thaw occurs; the 0.5 score is our judgement, not their result. HWSD SQ1 measures nutrient chemistry rather than aridity, so sand-desert cells (e.g. central Sahara) are often not class 7 — they're constrained by water rather than nutrient deficiency, and the composite captures that via the CDD and water factors.
No community lasts without renewable water. Water stress — withdrawals as a fraction of renewable supply — indicates whether a catchment can keep meeting demand. Severely stressed catchments lose supply reliability and degrade their groundwater and ecosystems; very dry catchments reach the same point on low supply alone, whatever their demand. Irrigation, desalination, and inter-basin transfer can postpone the limit but not remove it.
Water availability is encoded as the WRI Aqueduct Baseline Water Stress score (BWSs, 0–5), the ratio of total annual withdrawals to available renewable supply per catchment. Scoring is continuous and has no cliff: BWSs ≤ 1 (Low) → 1.0; ≥ 5 (Extremely High) → 0.05 (steep penalty but not zero); linear ramp in between. We deliberately omit the cliff because Aqueduct's "Extremely High" tier conflates two physical cases — engineered settlements drawing on long-distance transfers or aquifer mining (Phoenix, Cairo, Riyadh) and arid no-water-use cells — and we don't want to falsely flag the engineered-habitable cases as uninhabitable. The 40% withdrawal-to-availability boundary between "stressed but managed" and "severely stressed" is the long-standing Falkenmark / UN SDG 6.4.2 convention that Aqueduct's own tiering adopts; Wada et al. 2011 (Hydrol. Earth Syst. Sci. 15:3785) maps global water stress consistent with it. The baseline raster (Aqueduct Global Maps 2.1, 2010-vintage climatology) is held constant across all four epochs — this isolates climate-driven changes in other factors against a fixed hydrological baseline.
The body sheds heat by sweating, and sweat only cools if the air can absorb the moisture. Wet-bulb temperature (Tw) folds heat and humidity into one number for how much cooling capacity is left, which dry-bulb temperature alone misses. Near 35 °C Tw a healthy adult can no longer shed metabolic heat even resting in shade with water at hand, so it is a hard physiological ceiling rather than a comfort threshold. An area whose annual peak Tw crosses it is unsurvivable outdoors, and indoors too whenever the power fails.
Tw is computed from CMIP6 monthly maximum temperature (tasmax,
Amon) and monthly relative humidity via the Stull (2011) approximation; the raster
reports the annual maximum of the monthly-mean Tw at each
cell — the hottest month's value, not a true daily/sub-daily peak. This monthly
construction understates the real annual-max Tw: we measured the gap
against daily tasmax (MIROC-ES2L, 2081–2085) at a tropical-cliff-zone median of
~0.9 °C (larger, 2–6 °C, in dry mid-latitude interiors
where Tw sits far below the cliff and the gap doesn't matter). The reported
cliff-exceedance fractions are therefore conservative lower bounds;
and because the present-day baseline uses daily tasmax while the future epochs use
monthly, the warming signal is, if anything, slightly understated. A daily-resolved
future ensemble is not cleanly possible here — daily humidity is unpublished for
two of the six production models — so the monthly construction is retained and
disclosed rather than half-corrected. Future epochs draw from the SSP5-8.5 long-term-extension
subset (8 models reaching 2300). Production central estimate uses an
ECS-screened 6-model subset, dropping CanESM5 (ECS ~5.6 K) and UKESM1-0-LL
(~5.3 K; Meehl et al. 2020) — both above AR6 WG1's "very likely" 5.0 K upper
bound (Hausfather et al. 2022). Humidity is per-epoch from each model's
own hurs (5-model ECS-screened RH ensemble at 2100 / 2200 / 2300; the
2014 baseline still uses ERA5). Scoring is a continuous ramp: Tw ≤
24 °C → 1.0; ≥ 35 °C → 0 (cliff); linear between. The
2300 ensemble puts ~37 % of land past the cliff under ECS-screened.
Two-limit framing. Two distinct wet-bulb limits matter and are easily conflated. The 35 °C cliff is the thermodynamic survivability ceiling (Sherwood & Huber 2010). A second, lower limit from Vecellio et al. 2022 (PSU HEAT; J. Appl. Physiol. 133:340) — the empirical compensability ceiling — sits near 31 °C Tw in humid conditions for healthy young adults (and lower, ~25–28 °C, in hot-dry conditions); sustained exertion is thermally uncompensable above it, and the limit is lower for vulnerable populations. Refugia cliffs at 35 °C because it is the harder physical limit; the real habitability boundary for vulnerable populations is several degrees lower.
Uncertainty framing. Per-model spread is real (4.7 °C
median 17-83% inter-model width over land at 2300; 32 % of land has the
35 °C cliff sitting inside the 17-83% range), widest at high latitudes and
narrowest in the tropical cliff zone. Full quantification is documented in
research/white-paper/wetbulb-uncertainty-quantification.md; read
the production map as the centre of a wide envelope, not a point forecast.
Cold is the one factor that eases under warming. Far-northern Russia, interior Greenland, the Canadian Arctic Archipelago, and interior Antarctica are uninhabitable today because their mean annual temperature sits below the cold end of the human climate niche (Xu et al. 2020): the growing season is short, the ground stays frozen, and heating is costly. Warming under SSP5-8.5 raises mean annual temperature in those regions and lifts the cold limit. A composite with no cold-side term would see only the habitable mid-latitudes shrinking and would miss the high-latitude land that conditionally opens up.
tas (Amon) from ESGF for the 8 long-extension models; 6-model ECS-screened production ensemble (ACCESS-ESM1-5, CESM2-WACCM, EC-Earth3-Veg, GISS-E2-1-G, GISS-E2-1-H, MIROC-ES2L).
The factor is computed from the CMIP6 SSP5-8.5 long-extension monthly
tas (mean surface T) for the 6-model ECS-screened production ensemble
(same ensemble as wet-bulb). MAT per cell per epoch = annual mean of monthly tas,
20-year window centred on each epoch. Anchor (score 1.0) at
MAT ≥ 5 °C — a calibration set on the cold side of the
~11–15 °C peak of the Xu et al. 2020 human climate niche; linear ramp to a
cliff at -5 °C, a proxy for the boreal/tundra ecotone informed by
the Beck et al. 2018 Köppen mapping (Köppen ET is a warmest-month rule, so -5 °C
MAT is our proxy, not a Köppen-defined threshold); soft floor of
0.05 between -5 and -15 °C (Yakutia, Nunavut, northern
Fennoscandia all have demonstrably small but non-zero population — a hard
zero here would be empirically wrong); hard zero below -15 °C (continuous
tundra, ice cap).
Anticipated pushback. A critic might claim this "makes the polar regions look too habitable under SSP5-8.5". The cold factor only removes a cliff — it does not add positive habitability. A formerly-frozen cell that warms past -5 °C still has to pass soil (HWSD permafrost-thaw cliff persists for decades after MAT crosses zero; IPCC AR6 WG1 Ch.9), water stress, flooding, and wet-bulb. The boreal wildfire pressure that follows warming is a known gap.
Current-epoch baseline note. Local cache lacks CMIP6 historical
1995–2014 tas; we use the SSP5-8.5 2015–2034 mean as the
"current" baseline. ~+0.3 °C warm bias vs the historical baseline; doesn't
change which cells fall in the cliff or soft-floor bands.
How long the dry season runs unbroken decides whether rainfed farming, livestock, and native vegetation survive it. A long enough dry spell drains soil moisture and kills perennials before the rains return. Irrigation offsets this where it exists, but the lengthening of dry spells under warming is a recognised driver of agricultural failure and out-migration, most acutely in continental interiors and the subtropics.
The maximum number of consecutive dry days (CDD, days with precipitation < 1 mm) per year is sourced from the IPCC AR6 Interactive Atlas (Santander Meteorology Group) ensemble export of CMIP6 projections under SSP5-8.5: 1995–2014 historical mean for the current epoch, 2081–2100 mean for the 2100 epoch, with 2300 reusing the 2100 raster since no published Atlas projection extends past 2100. Scoring is a continuous ramp: ≤ 30 days → 1.0 (no meaningful dry-season constraint); ≥ 180 days → 0 (cliff); linear between. The 180-day cliff is an internal calibration — roughly six unbroken dry months, the point at which rainfed agriculture becomes infeasible, and broadly the dry-spell length of arid/hyper-arid zones (FAO/UNESCO aridity is itself defined on the Aridity Index and growing-period length, not on consecutive dry days, so this is an approximate correspondence, not an FAO-defined CDD threshold). The published ETCCDI annual-max-CDD climatology smooths extreme single-year stretches and caps around 260 days even for Atacama-class deserts, so the 180-day cliff (rather than 200+) is calibrated to the actual data envelope while remaining physically defensible.
Flooding displaces people and destroys infrastructure outright. Sea-level rise takes low-lying coast permanently; riverine flooding hits floodplains episodically, including otherwise-habitable ones. Together they set the lower-elevation limit on where dense settlement and farming can hold, and they account for much of this century's flood-driven displacement to date.
Two flood hazards are combined into a single 3-class suitability mask, taking the worst classification at each cell. Coastal sea-level rise is keyed to the GMTED2010 DEM. For the 2100 epoch: cells with elevation ≤ 1 m are unsuitable (deep inundation under the IPCC AR6 likely range of ~0.6–1.0 m by 2100 under SSP5-8.5), 1–5 m are marginal (high-tide and storm-surge exposure), and > 5 m are suitable. For the 2300 epoch we use a wider 65 m commitment marginal band: ≤ 5 m are unsuitable (within IPCC AR6 likely 1.7–6.8 m at 2300 under SSP5-8.5; Fox-Kemper et al. 2021), 5–65 m are marginal (encompassing the ~65 m total ice-volume sea-level equivalent of Antarctica + Greenland; the multi-millennial total-melt-commitment framing follows Clark et al. 2016 and Van Breedam et al. 2020), and > 65 m are suitable. The 65 m number is a commitment ceiling, not a 2300 prediction — it bounds where coasts are guaranteed safe over the long run, not where the 2300 shoreline actually sits. Riverine flooding uses the WRI Aqueduct 100-year return-period inundation depth: depth ≥ 1 m is unsuitable, 0.1–1 m is marginal, and < 0.1 m is suitable. The current epoch has no SLR component and applies only the riverine layer. Riverine depth at the future horizon is the per-pixel ensemble maximum across four CMIP5 GCMs (NorESM1-M, GFDL-ESM2M, HadGEM2-ES, IPSL-CM5A-LR) for the RCP 8.5 / 2080 horizon; the historical (WATCH) baseline is used for the current epoch, and the 2080 horizon is reused as a proxy for 2300 since no published Aqueduct projections run past 2080.
Biodiversity does not constrain where people can live; humans settle biodiverse landscapes readily. It enters the model the other way around, as an asset that climate-driven displacement puts at risk. As habitable land contracts, the intact biomes that remain draw pressure from migration, development, and farm expansion. The regions that look like 21st-century climate refugia — boreal forest, montane terrain, high-latitude wetland — are largely the same places that still hold intact ecosystems, and mapping that overlap is what makes the conflict visible.
Until 2026-05 the biodiversity asset was held at the static 2020 PREDICTS BII baseline across all four epochs because no public gridded future-scenario BII raster exists (Newbold/De Palma & Purvis 2018 publish only country aggregates, not rasters). The asset has now been swapped to a per-epoch BII proxy derived from LUH2 v2.1 land-state grids (Hurtt et al. 2020). Per cell: natural-land fraction = primf + primn + secdf + secdn (primary forested, primary non-forested, secondary forested, secondary non-forested), 20-year mean centred on each epoch from the SSP5-8.5 v2.1f file (2015–2100) and the v2.1e extension (2100–2300). LUH2 is the only public gridded land-state product that natively reaches 2300, matching the wet-bulb long-extension ensemble. The proxy is narrower in concept than species-abundance intactness but moves in the same direction — Newbold et al. 2015 (Nature 520:45) attribute the dominant BII signal to land-use change, the variable LUH2 carries. Cross-validation: GLOBIO 4 MSA SSP5-RCP8.5 2050 (Schipper et al. 2020) is conceptually closer to BII (true species abundance) but is published only for one future epoch; a future iteration may bias-correct the LUH2 proxy against GLOBIO at 2050. Superseded 2026-05-29. The LUH2 natural-land-fraction proxy described above is no longer the biodiversity stakes layer: the redesigned biodiversity-risk methodology (see Biodiversity Risk — methodology below) now uses a dedicated bio_value term — canonical NHM BII v2.1.1 (species-abundance intactness) × endemism-weighted IUCN/BirdLife richness — and the encroachment term dropped its LUH2 Ncurrent mask entirely. The per-epoch LUH2 panels remain visible here only as transparency on the SSP5-8.5 land-use trajectory itself, not as an input to any index.
The seven factors are scored continuously on [0, 1]. Each factor has two anchors: a "fully habitable" lower bound (below which → 1.0) and a hard cliff (above which → 0, where the factor crosses an established uninhabitability threshold). Linear ramp in between gives the gradient within habitability. Cliffs: slope ≥ 30° (alpine, structurally unbuildable), soil HWSD SQ1 class 7 (FAO "unsuitable, ineffective" — desertic/glacial soils with no agricultural potential) or class 0 (no soil), annual-max wet-bulb Tw ≥ 35 °C (Sherwood & Huber survivability ceiling), mean annual T ≤ -15 °C (continuous tundra / ice cap), consecutive dry days ≥ 180 (six unbroken dry months — rainfed agriculture infeasible, desert-margin classification), flood depth ≥ 1 m (RP100 riverine, or 5 m / 65 m SLR commitment for 2100 / 2300). Water stress is gradient-only — no cliff, since BWSs=5 conflates engineered settlements (Phoenix, Cairo) with arid no-use cells, and we don't want to falsely flag the former as uninhabitable. The composite is the geometric mean of factor scores unless any factor hits its cliff, in which case the cell is marked uninhabitable (0). Ocean and inland water bodies are masked out using the HWSD soil-class land mask (class 0 = ocean / no soil, class 7 = inland water), matching the basemap water mask.
Scale of valid inference. The analysis grid is 30 km, with cell-mean slope and a monthly climate ensemble regridded from ~0.5° CMIP6. At that resolution a single value stands for a 30 km cell, so a mountainous cell averages away the inhabited valley floors within it, and sub-grid refugia are invisible. Read the output as continental-to-regional contrasts and relative change between epochs — a screening tool, not a cell-level prediction of where individual settlements can persist.
What is actually 2300. The four epoch panels advance the temperature-driven factors with real long-extension data, but several inputs are frozen or reused at the later horizons — there is no defensible projection for them past 2100. The table makes the vintage explicit:
| Input at the 2300 panel | Source used | Genuinely 2300? |
|---|---|---|
| Wet-bulb Tw, mean annual T (cold) | 2281–2300 long-extension ensemble | Yes |
| Tair anomaly, climate velocity (bio) | 2300 ensemble | Yes |
| Consecutive dry days (CDD) | reuses the 2100 raster (Atlas stops at 2100) | No — frozen at 2100 |
| Water stress (BWSs) | 2010-vintage Aqueduct, static all epochs | No — frozen |
| Riverine flood | 2080-horizon Aqueduct reused; SLR band is 2300 commitment | Partial |
| bio_value (biodiversity stakes) | observed-2020, held static | No — frozen 2020 |
| Population (human exposure) | 2100 SSP5 grid reused | No — copy of 2100 |
| Encroachment (urban) | linear extrapolation of the 2015–2100 Chen trend | Extrapolated |
One consequence is directional and worth stating plainly: freezing the moisture-side factors (water stress, dry-day length) at their 2100-or-earlier vintage while heat advances to 2300 means the composite sees a hotter world with unchanged aridity. Because those factors can only pull habitability down, the frozen inputs make the 2300 habitability panel read optimistically — a self-consistent 2300, with drying advanced to match the heat, would be somewhat harsher.
The biodiversity-risk raster is a priority score — hazard × stakes — bio_risk = joint_pressure × bio_value, structured after the IPCC AR6 WGII risk decomposition (hazard × exposure × vulnerability). It is a designed composite, not an expected loss in calibrated units: joint_pressure is the hazard term — how hard a cell is being hit; bio_value is the stakes term — how much irreplaceable, intact biodiversity is there to lose (defined in the Biodiversity Value block below). A cell scores high only when a real pressure meets real stakes: Saharan interior and high-Arctic cells take heavy climate pressure but carry little stake, so their score stays low; the Andes, Madagascar, New Guinea and Sundaland carry enormous stake, so even moderate pressure registers.
The joint_pressure term itself combines two physically distinct
pressures — human-encroachment pressure that "places becoming habitable
for humans" exert on intact ecosystems, and the direct climate damage
biodiversity itself takes (heat anomalies, drought, climate velocity, polar
warming) — via probabilistic OR with an overlap discount on shared biota:
joint_pressure = 1 − (1 − encroachment) × (1 − climate_threat × (1 − α × encroachment)),
with α = 0.3. Each term is in [0, 1]; either pressure alone
can saturate it. The (1 − α × encroachment)
factor discounts the climate-threat term where encroachment is already
converting the land. A pure probabilistic OR assumes the two pressures hit
different biota, but they target the same intact-natural-land
cells in practice, so independence-OR over-counts joint loss; we apply an
overlap discount (α = 0.3) for this, a tuning choice we estimate
removes a single-digit-to-low-teens percent over-count at typical operating
points (the choice of aggregation rule matters: Tonmoy et al. 2014,
WIREs Clim Change 5:775). Until 2026-05-11 the raster was instead a single
habit_suit × BII_proxy product, which collapsed these
distinct pressures into one term; the redesign separated them, and the
2026-05-29 revision added the bio_value stakes multiplier above.
The four epoch panels share a single cross-epoch colour scale
(p99-normalized) so they are directly comparable.
2026-05-20 revision. A second-pass expert review
(research/expert-review/05-biodiversity-review.md) found
that the original cliff-aware geometric-mean aggregator over
six floored factors forced climate_threat → 1.0
across most of the SSP5-8.5 2300 land surface as soon as 2–3
partially-correlated factors hit cliff — over-aggressive relative
to the underlying literature, which presents most of these stressors
as graded rather than binary. The aggregator was replaced by
a weighted arithmetic mean on threat (1 − score)
with per-cell active-weight renormalisation; the Tair-anomaly,
climate-velocity, and water-stress factors were demoted from cliffs to
continuous gradient scores; the cold-loss gate was softened from a
hard MAT ≤ 0 °C cutoff to a linear taper out to
MAT ≤ 3 °C (capturing the cool-temperate vulnerability
the prior gate exempted; Brandt 2013, Gauthier 2015); the CDD floor
was raised from 0.05 to 0.15 (McDowell 2020 is probabilistic, not
binary); the flood factor was made biome-aware (riverine-driven
drowning gets a softer 0.30 floor reflecting floodplain adaptation,
while SLR-driven drowning keeps the 0.05 Schuerch coastal-squeeze
floor); and the encroachment ⊕ climate-threat OR combination
gained an α = 0.3 overlap discount on shared biota.
encroachment(epoch) = nuf(epoch), the cell-aggregated fraction of 1 km cells flipping from non-urban to urban between Chen et al. (2020)'s 2015 baseline and the target year, under SSP5 (matched to our SSP5-8.5 climate trajectory). This is now a pure pressure signal. Until 2026-05-29 it was multiplied by an LUH2 natural-land-fraction mask (Ncurrent) to weight pressure by what was intact; with the new bio_value stakes term carrying intactness (and irreplaceability) for the whole asset, that mask became redundant and was actively harmful — LUH2 classes biodiverse pasture and rangeland as managed, so multiplying by it zeroed urban encroachment onto African savanna grazing land and Central-Asian steppe. Dropping it makes encroachment a clean hazard signal; stakes are applied once, downstream, via bio_value.
Chen 2020 publishes 1 km decadal urban-land projections through 2100. For 2200 and 2300 epochs we linearly extrapolate the cumulative cell-level newly-urbanized fraction (multiply by (year − 2015) / 85, where 2015 is the Chen-published baseline year), clipped to [0, 1]. Our pipeline's "current" epoch nominally anchors at 2014, a 1-year offset that is negligible (~1.2% of the 85-year extrapolation interval).
climate_threat = 1 − weighted arithmetic mean of six biodiversity-specific per-factor suitability scores, each in [0, 1]. Threats are aggregated on the (1 − score) side so the result reads as a damage signal that increases monotonically as factors stack. Per-cell weights renormalise across active factors only — e.g. the cold-loss factor's weight is tapered out in cells where the historical MAT is too warm for cold-loss to be physically meaningful, and the remaining factors share that weight pro-rata. The factor set differs from the human-habitability composite — not all human-habitability factors apply to ecosystems, and the ones that do apply at different thresholds:
Combination weights: Tair anomaly 0.25, climate velocity 0.20, cold-loss 0.20, water stress 0.15, CDD 0.10, flood 0.10 (sum 1.0). Weights are elicited from how IPCC AR6 WGII Chapter 2 frames multiple-stressor risk; they are documented judgement, not derivable from a single paper.
research/expert-review/05-biodiversity-review.md (2026-05-20).
Biodiversity Value — the stakes term.
joint_pressure above measures how hard a place is hit;
bio_value measures how much is at stake there, so that
bio_risk = joint_pressure × bio_value reads as a
hazard × stakes priority score rather than raw pressure. It replaces
the earlier LUH2 natural-land-fraction "BII proxy" with two canonical,
complementary axes combined as a weighted geometric mean
bio_value = BII0.4 × WE0.6:
Intactness (BII, weight 0.4). The Natural History Museum's
Biodiversity Intactness Index v2.1.1 — the modelled abundance of
originally-present species relative to an intact baseline, fit on the
PREDICTS database of >3.2 million records (Newbold et al. 2016 Science
353:288; Hudson 2017). Intact wilderness ≈ 1; degraded cropland and
cities ≈ 0. This is true species-abundance intactness, not the
land-cover proxy it supersedes.
Irreplaceability (WE, weight 0.6). An endemism-weighted
species-richness surface built from ~33,000 expert range maps (IUCN Red List
mammals, amphibians, reptiles; BirdLife BOTW 2025 birds), where each species
contributes 1 / range-area — so a Madagascan frog
confined to one massif counts for thousands of times more per cell than a
continent-spanning generalist. This is the Jenkins 2013 (PNAS
110:E2602) endemism-weighting, extended with reptiles (Cox 2022 Nature
605:285). Irreplaceability gets the larger exponent because, in systematic
conservation planning, what cannot be substituted dominates what is merely
intact (Brooks 2006 Science 313:58; Margules & Pressey 2000
Nature 405:243).
Held at observed-2020 across all four epochs (v1). There is
no published, defensible global projection of species ranges or BII to 2100,
let alone 2300, so we freeze bio_value at today's observed stakes rather than
fabricate one. The four panels therefore read as future climate damage
scaled by today's biodiversity stakes. This is a deliberate, one-sided
under-count: SSP5-8.5 cropland and pasture expansion through 2100 would
further erode BII (most in the Cerrado, Madagascar dry forest,
Sundaland transition zones, and the Sub-Saharan savanna belt), so those
regions' future bio_risk here under-states land-use attrition. Direct urban
encroachment, which is projected (Chen 2020 SSP5), still enters
through the separate encroachment term. A v2 path would unfreeze 2100 stakes
once a projected-BII raster is obtainable.
Data Sources
Adversarial pushback (pre-empted).
(1) Aren't you double-counting drought? No — the human-hab CDD cliff
(180 d) and the bio CDD score (gradient, no cliff) live in two independent composites and
are combined via probabilistic OR, which by construction does not double-count.
(2) The BII proxy isn't BII. Fixed in the 2026-05-29 revision —
the LUH2 land-cover proxy was replaced by the canonical NHM BII v2.1.1
(species-abundance intactness) inside the new bio_value term.
(3) Climate velocity is scale-dependent. True (Burrows 2014, Methods); we
publish at the native 30 km analysis resolution.
(4) Encroachment assumes humans actually move there. True — this is a
pressure surface, not a realised land-use trajectory. Chen 2020 SSP5 is itself
a socioeconomic-driven projection, not observed migration. The companion realised-LUC
layer is LUH2 v2.1f/v2.1e shown above.
(5) Cold-loss as a threat is counter-intuitive. Boreal/Arctic biomes
depend on cold; warming above +6 °C breaks the regime those ecosystems
are adapted to. The factor is fully active in cells with historical MAT ≤ 0 °C,
tapers linearly to inactive at historical MAT = +3 °C (catching the
southern-boreal / cool-temperate band the prior hard MAT ≤ 0 gate exempted —
Brandt 2013, Gauthier 2015), and is zero in warmer biomes.
(6) Why drop topography? Slope is a human-build constraint; biomes have
already colonised steep terrain.
(7) Why graded scores instead of hard cliffs? The 2026-05-20 expert review
found that a cliff-aware geometric mean over six floored factors forced
climate_threat to ≈ 1.0 across most of SSP5-8.5 2300 whenever
two or three (often-correlated) factors crossed cliff — over-aggressive relative
to the underlying literature (Trisos 2020 is a timing metric not a magnitude cliff;
Burrows 2014 / Schloss 2012 report ~91% of mammals tracking projected climate velocity;
Vörösmarty 2010 is a 23-driver index of which BWS is one input). Replacing the cliffs
with graded ramps and switching to a weighted arithmetic mean preserves the
"stacking is worse" property without forcing the composite to floor.
(8) Why endemism-weighted richness instead of raw species count?
Raw richness peaks in the lowland Amazon, where most species are wide-ranging;
it also tracks sampling effort. Weighting each species by 1/range-area surfaces
irreplaceability — the small-range endemics (Andes flank, Madagascar,
New Guinea, Cape, Mesoamerica, Western Ghats) whose loss is permanent — and
damps the rich-but-replaceable interior. In our 30 km surface the western
Andean flank scores ~5× the adjacent Amazon lowland for exactly this reason.
(9) Why hold bio_value static at 2020 when climate panels run to 2300?
Because no defensible global range-shift or BII projection exists past 2100. A
fabricated projection would add false precision; freezing the stakes and disclosing
the resulting one-sided land-use under-count (see the Biodiversity Value block) is
the more honest choice. The climate hazard is fully time-varying; only the
stakes are frozen.
(10) Why did encroachment drop its intactness mask? The old
nuf × Ncurrent double-counted intactness (now in
bio_value) and zeroed urban pressure on biodiverse pasture/rangeland that LUH2
classes as managed. Encroachment is now pure hazard; stakes are applied once.
What the redesign does NOT capture (yet).
A Fire Weather Index factor (Van Wagner 1987; Abatzoglou 2019) is defined in
composite_utils.py but not yet wired in — it requires a
precipitation-on-grid preparation step deferred to a follow-on iteration. A
permafrost frost-index factor (Nelson & Outcalt 1987) is similarly framed
but deferred. Both are pure derivations from cached CMIP6 inputs — no external
downloads — and can be added to the climate_threat composite later without
changing the combination formula. ISIMIP3b burnt-area products would supply a
validation cross-check on the derived FWI but are not required.
The composite habitability raster describes where the planet remains liveable,
but not what stands to lose by its decline. The two paired surfaces here cross
habitability with the human population on one side, and combine projected urban
expansion with direct climate threat to intact ecosystems on the other.
Human Exposure = (1 − habitability) × population: a static-population
cartogram surfacing where today's population is concentrated in cells projected to lose
habitability.
Biodiversity Risk is a hazard × stakes priority score documented in the
methodology section above:
bio_risk = joint_pressure × bio_value, hazard × stakes.
joint_pressure combines pure-pressure encroachment (Chen et al. 2020 SSP5
urban-land projection) with a weighted-arithmetic climate-threat composite of six
biodiversity-specific factors (Tair anomaly, cold-loss, climate velocity,
water stress, drought, flood) via probabilistic OR. bio_value scales that
by what is at stake — NHM BII intactness × endemism-weighted IUCN/BirdLife
irreplaceability, held at observed-2020 — so the map reads as where biodiversity
loss matters, not merely where climate changes. Panels share a single
cross-epoch (p99-normalized) colour scale.
Exposure is not a migration projection. The climate-mobility literature (Black et al. 2013; Cattaneo & Peri 2016; Cundill et al. 2021) is unambiguous that the most climate-vulnerable populations are often least able to relocate — migration is expensive, and the same conditions that erode habitability also erode the resources needed to leave. A high-population × low-habitability cell may therefore contain trapped populations facing maximum exposure precisely because they cannot move, not displacement-prone populations on the verge of moving; this surface cannot distinguish the two. Population is also held at the 2020 WorldPop baseline, so the 2100 and 2300 panels read present-day settlement against future climate fields and should be interpreted as exposure cartograms, not forecasts of where people will actually live. No peer-reviewed migration projection exists at the 2300 horizon.