Every number on the site is auditable: each traces back to a published satellite or reanalysis dataset, a park boundary polygon from USGS, and a specific statistical test run on the aggregated daily series.
Temperature, precipitation, snowpack and radiation come from two reanalysis-grade grids; park shapes come from USGS. Every source is free and publicly re-downloadable.
NASA/ORNL, 1980–present. Daily values are interpolated from
ground-station observations onto a 1 km grid — not a
reanalysis or model output. Covers CONUS, Hawaii, Puerto
Rico. Bands: prcp (mm), tmax/tmin (°C),
srad (W/m²), swe (kg/m²), vp (Pa).
ECMWF/Copernicus, 1950–present. Fills gaps DAYMET can't: parks in Alaska, American Samoa, and Virgin Islands, plus evaporation, wind, and snow depth.
Matched by Unit_Nm with alias handling for
Alaskan "National Park and Preserve" units, dissolved per
park, simplified to 50 m in Albers, then reprojected to WGS84.
For each park we load the boundary, split disjoint sub-units
(Saguaro's Rincon and Tucson districts; the five islands of
Channel Islands), and reduce each dataset over the geometry
with ee.Reducer.mean() at native grid scale —
1 km DAYMET, 11 km ERA5-Land — using bestEffort=True
and maxPixels=1e13.
Exports run as serverless Earth Engine batch tasks
(ee.batch.Export.table.toDrive) with an explicit
selectors list so the CSV writer emits one column
per band. End dates are passed to filterDate
exclusively, the EE convention.
_sum (precipitation, snowfall, snowmelt,
evapotranspiration) are preserved as daily totals per
ECMWF/Copernicus convention; everything else is a daily mean.
All conversions happen in the display layer — raw pipeline outputs stay strictly in SI units (the international scientific system: Kelvin for temperature, metres for depth, kg/m² for mass) so Parquet/CSV consumers get unambiguous values.
The methodology favours tests that don't assume normality or homoscedastic residuals — a conservative choice for short-to-medium climate series, which often have skewed distributions (precipitation, snow) and mild autocorrelation.
Non-parametric rank test for monotonic trend. Uses tie correction and the normal-approximation p-value, which is standard for n ≥ 10. Applied to the annual series. Does not correct for serial autocorrelation; for series longer than a few decades, the Hamed–Rao variant7 is the recommended substitute (see Limitations below).
Slope estimator: the median of all pairwise year-over-year slopes. Robust to outliers and to non-normal residuals. We prefer it to OLS for skewed climate variables (precip, snow). The reported slope tends to be more conservative than an OLS fit on the same series.
Calendar year. Sums for flow variables (precipitation, snow, ET); means for state variables (temperature, radiation, snow depth, vapour pressure). DAYMET v4 and ERA5-Land are gap-free historical records, so this aggregation drops only the year currently in progress at export time — the < 330-day cutoff prevents a partial calendar year from contributing a biased annual aggregate.
Standard meteorological seasons: DJF / MAM / JJA / SON. December is assigned to the following year's DJF so that each winter is contiguous in calendar time.
Reference-period mean and σ computed over 1981–2010 (WMO 30-year normal). z-scores drive the climate stripes on the home and park pages.
Classical additive decomposition at monthly resolution: 12-month centred rolling mean as the long-term trend, plus per-period month-of-year climatologies (1981–1995, 1996– 2010, 2011–2025) with a 10th–90th percentile inter-annual envelope. Classical (rather than STL or X-13) keeps the method dependency-free and reproducible from pandas alone; the trade-off is that the seasonal component is fixed within each period rather than allowed to evolve smoothly.
A few parks span geographically separate polygons. For these
we compute one series per polygon (using
ee.Geometry.geometries()) and a union-level
park-wide series. Both are exported; per-part series live in
the downloadable summary JSON.
Pipeline outputs are audited against four classes of check — unit / range bands per variable, annual-aggregation arithmetic recompute, Mann–Kendall ↔ Theil–Sen consistency, and an external warming benchmark for 20 parks across climate zones drawn from Gonzalez et al. (2018), the Greater Yellowstone Climate Assessment, NPS climate-change pages, and the NOAA / NCEI State Climate Summaries. Each finding is tagged PASS / WARN / FAIL with a documented rubric.
The audit lives at
docs/DATA_QC.md
on GitHub and is regenerated by
scripts/qc_pass.py against
site/public/data/parks/.
Tagged by category — dataset (upstream DAYMET / ERA5-Land properties), statistics (trend-test caveats), method (choices in this pipeline), and coverage (where the grid doesn't reach).
DAYMET v4 has cold-bias at high elevation; ERA5-Land's ~11 km grid smooths elevation-driven temperature gradients. The trend signal is more robust than the absolute level.
Polygon means are reanalysis / interpolated-grid values reduced over the entire park boundary. They will not match a specific weather station inside the park.
The dataset's bare-soil assumption inflates
pet_mm by ~1.5–2× vs FAO Penman–Monteith.
Trend direction is informative; absolute level is biased.
aet_mm is unaffected.
Lag-1 autocorrelation depresses MK p-values. The deployed normal-approximation is fine for n ≈ 45 with mild dependence; longer-horizon work needs the Hamed–Rao variant7.
~14 variables × 63 parks ≈ 880 tests; ~5% will be false-positive at p < 0.05 under the null. Apply Benjamini–Hochberg FDR10 or Bonferroni for inference.
The union series for disjoint-geometry parks (Saguaro, Channel Islands, Kings Canyon, Redwood, American Samoa) averages per-polygon values without area weighting. Per-part series are exposed in the summary JSON.
WMO-standard, but back-loaded inside a 1980–present record, which compresses early-period anomalies. The 1991–2020 normal would shorten the comparison window further.
The EE export uses bestEffort=True with
maxPixels=1e13. For very large polygons
(e.g. Wrangell-St. Elias), it may sub-sample rather than
aggregate every pixel; the spatial mean still converges.
Park-wide means smooth within-park heterogeneity (coastal vs interior, north vs south slope). For pixel-level series, use Earth Engine directly with the package's helpers.
Three small island parks (American Samoa, Dry Tortugas, Virgin Islands) fall inside sea-masked ERA5-Land pixels and aren't covered by DAYMET. The site labels them "outside land-grid coverage" and renders no charts.