DRT · Extreme value
Drought
L-moment GEV, region-of-influence RFA (~400-cell climate-similarity pools) · v2.1 · 2026-07
SPEI-12 rp100 drought severity · California + SW USA, 90 m
Drought-severity (SPEI-12) susceptibility and return levels from 85 years of ERA5 climate reanalysis, 1940-2024. Polar and hyperarid cells are left unscored.
What ships
Surfaces + Returns — 0–1 drought susceptibility (N2-normalized rp100 |SPEI-12|, v2.1) plus per-return-period |SPEI-12| magnitudes (rp5/10/25/50/100, σ), region-of-influence regional fit.
Every release: 0–1 score · 5-class band · per-pixel uncertainty · input-coverage mask · STAC item · methodology doc.
Inputs
ERA5 monthly precipitation + 2m temperature 1940-2024, Beck 2018 Köppen-Geiger 1km, Thornthwaite 1948 PET, climate_indices Pearson III SPEI-12
Version history
What changed between releases — the same notes that ship with the data. Newest first.
Surfaces
v2.1
2026-07
current
- First Drought Surfaces release. The companion susceptibility set to the live Drought Returns v2.1 — it reduces the 100-year drought extreme (|SPEI-12| at rp100) to a single, globally-comparable susceptibility score so drought sits on the same 0–1 scale as every other Petrel hazard Surface. Four layers: susceptibility, a 5-class classification, uncertainty, and input coverage.
- N2 global normalization. Susceptibility is the global empirical-CDF percentile of the rp100 |SPEI-12| field — 0.90 means a location's 1-in-100-year drought is more severe than ~90 % of scored land worldwide.
- Real fit uncertainty. The uncertainty layer is the *relative standard error* of the 100-year return level from the region-of-influence RFA fit (SE(rp100)/rp100), Monte-Carlo-calibrated — a genuine, tail-driven fit uncertainty, not the earlier placeholder.
- Region-of-influence regionalization (v2.1). Same science and masks as Returns v2.1: ERA5 1940–2024 → Thornthwaite PET → SPEI-12 (Pearson III) → L-moment region-of-influence RFA — each cell pooled with its ~400 most climatically similar cells, so the Köppen class-boundary seams of the earlier pooling are gone. Polar (permanent ice) and hyper-arid cores stay NoData, never scored low: a standardized anomaly is undefined where the baseline water balance is ~0.
Returns
v2.1
2026-07
current
- Continuous growth curves — the Köppen class-boundary artifact is gone. v2.0 pooled cells for the regional frequency analysis by discrete Köppen-Geiger class, so the growth factor was constant inside a class and stepped at every class edge. Those steps were visible in the return surfaces as hard, angular seams — a diagnostic over Nevada found 75% of the sharpest rp100 jumps sitting exactly on a Köppen boundary, with full data coverage (not a data gap). Köppen is drawn on temperature and precipitation thresholds, which need not mark any change in the drought-extreme distribution, so the steps were partly spurious.
- Region-of-influence pooling. Each cell now forms its own pool of the ~400 most climatically similar cells — nearest in a space of great-circle geography plus log-aridity — and fits its own L-moment GEV growth curve. Neighbouring cells share nearly all of their pool, so the growth factor varies smoothly with no boundaries anywhere, at the same statistical strength (tens of thousands of pooled samples per fit). Geography is embedded on the unit sphere, so there is no seam at the antimeridian or the poles.
- What changed in the numbers. Return levels shift by roughly ±0.2σ against v2.0, in both directions, as the pooling is redrawn — not a systematic re-levelling. Monotonicity across return periods is preserved everywhere.
- Sharper 90 m detail. The downscale from the 0.25° analysis grid to 90 m is now bicubic rather than bilinear, removing the faint kinks bilinear left at coarse-cell edges.
- Read the scale from the band. Each COG now carries its scale factor as a GDAL band scale, so raw * src.scales[0] yields physical σ directly. It is also on the band as the SCALE tag. Previously the factor was only in a dataset tag and the documented snippet returned raw integers.
v2.0
2026-06
- Rebuilt on 85 years of ERA5, 1940–2024. The record more than doubles from 30 years to 85, computed through Thornthwaite PET → SPEI-12 (annual water balance). Thornthwaite is used because ERA5 wind and humidity only reach back to 1991; a Penman-Monteith PET would have truncated the record and defeated the point of the long baseline.
- Regional frequency analysis instead of per-cell fitting. Return levels come from an L-moment GEV fit over Hosking-Wallis homogeneous regions defined by the Beck 2018 Köppen-Geiger map, replacing the per-cell GEV of v1.1. This removes the pixel-level splotchiness the old method produced and yields spatially coherent return surfaces.
- Polar and hyper-arid cells are masked, not stretched. Beck Köppen EF (polar ice cap) and UNEP aridity below 0.05 (Sahara core, Atacama) are excluded, along with a goodness-of-fit gate, so the product no longer emits degenerate values where a drought index is undefined.
- Trend-adjusted rp100 companion. A parallel 100-year layer from a non-stationary GEV with a time-varying location parameter (Cheng & AghaKouchak 2014), evaluated at 2024.
v1.1
2026-04
- Initial release. Per-return-period |SPEI-12| drought magnitudes from per-cell GEV fits on a 30-year ERA5 record. This vintage is now deprecated: the per-cell fitting produced visually splotchy return maps, and the fit degenerated over polar and hyper-arid regions where the drought index is not well defined — both fixed in v2.0.
Other hazards
LandslideAvalancheWildfireEarthquakeWindstormExtreme heatExtreme coldFrost / freezeLightningSoil erosionHurricaneVolcanic EruptionTsunamiStorm Surge