Skip to content

Observations and datasets

working pillar 8 · Data management, statistics and UQ

In one minute

  • Reanalysis is not observation. ERA5 is a model constrained by observations — excellent, and wrong in specific, documented ways.
  • Radar measures a beam volume that grows with range and rises above the surface. Most radar artefacts follow from that geometry.
  • Satellites measure radiances, not geophysical variables. Everything else is a retrieval with assumptions baked in.
  • GRIB2 is lossy by design. Packing precision is a per-field choice, and it can be coarser than the signal you are looking for.
  • Every dataset has a representativeness scale. Comparing a point gauge to a 13 km grid cell is comparing different quantities.

Key results

Observation types and what they actually measure:

Platform Measures Representativeness Main limitation
Surface station point, 2 m / 10 m metres siting, exposure, urban heat
Radiosonde a drifting profile, ~2/day a slanted curtain sparse in time and over ocean
Radar backscattered power in a beam volume 0.5–4 km, range-dependent geometry, clutter, attenuation
Geostationary satellite radiance, 0.5–2 km, 1–15 min column or cloud top retrievals, no vertical resolution
Polar orbiter radiance, higher spatial res swath twice-daily revisit
Aircraft (AMDAR) along flight path point concentrated near airports and cruise levels
Reanalysis model state grid cell not an observation

Radar. Reflectivity factor in dBZ, related to rain rate by an empirical Z–R relation:

\[ Z = a R^{b},\qquad a \approx 200,\ b \approx 1.6 \ \text{(Marshall–Palmer)} \]

That relation is a fit, varies by precipitation type, and is a leading source of QPE error. Dual polarisation (\(Z_{DR}\), \(K_{DP}\), \(\rho_{HV}\)) constrains drop shape and improves both classification and rainfall estimation substantially.

Beam geometry sets most of the artefacts: the beam widens with range (about 1° beamwidth, so ~1.7 km at 100 km) and rises with Earth curvature plus refraction, roughly

\[ h \approx r\sin\phi + \frac{r^2}{2\cdot(4/3)R_\oplus} \]

Satellite channels (GOES-R ABI as the reference case): visible (0.64 µm) for cloud and albedo; near-IR (1.6, 2.2 µm) for phase and particle size; water vapour (6.2, 6.9, 7.3 µm) for mid/upper tropospheric moisture and flow; window IR (10.3, 11.2 µm) for cloud-top temperature.

Reanalysis. ERA5: ~31 km global, hourly, 1940–present, 137 levels. MERRA-2 and JRA-3Q are the main alternatives; regional reanalyses exist for some domains. Key properties to keep in mind:

  • It is a model forecast constrained by observations, so in data-sparse regions and eras it is mostly model.
  • The observing system changes over time — satellites arrive in 1979 and again in the 2000s — so trends across those boundaries can be artefacts.
  • Different variables have very different reliability: upper-air winds and temperature are well constrained; precipitation and surface fluxes are model products.

Data formats:

Format Used for Watch out for
GRIB2 operational NWP output lossy packing; template complexity
NetCDF4/HDF5 research, reanalysis, satellite chunking affects read speed by orders of magnitude
BUFR observation exchange tabular, needs decoding tables
Zarr cloud-native array storage increasingly standard for large datasets

Mental model

Every dataset is an instrument plus a processing chain, and both leave fingerprints. The useful habit is to ask, for any field: what was physically measured, and what was inferred? For a radiosonde temperature, almost all measured. For satellite-derived precipitation, almost all inferred. For reanalysis precipitation, entirely model.

That question also settles most comparison disputes. A model and an observation disagreeing may simply be reporting different quantities over different volumes.

Numerics / practice

  • Read the dataset documentation on known issues before using it. ERA5, HRRR and GOES all publish them, and they are specific and useful.
  • Chunk NetCDF/Zarr to match the access pattern. A time-series read from space-chunked data can be 100× slower than necessary.
  • Check GRIB2 packing precision when differencing fields or computing gradients — the quantisation step can exceed the signal.
  • Use xarray + dask for anything larger than memory, and cfgrib/pygrib for GRIB. MetPy for unit-aware meteorological calculations.
Failure modes

Reanalysis used as ground truth. Verifying a model against ERA5 partly verifies it against another model — and if they share physics or assimilate the same observations, the agreement is not independent. For precipitation and surface fluxes especially, ERA5 is a model product.

Radar bright band. Melting snow produces a layer of strongly enhanced reflectivity near the freezing level. A radar beam intersecting it reports rain rates several times too high in a characteristic ring around the site. Well known, still routinely forgotten in QPE.

Beam overshoot at long range. At 150 km the beam centre is above 2 km, so shallow precipitation is missed entirely. Apparent precipitation gradients in a radar climatology often trace radar geometry rather than weather.

Trends across observing-system changes. A "trend" in reanalysis that begins in 1979 is very likely the arrival of satellite data. Any long-term analysis must check whether the change coincides with an observing-system transition.

Satellite retrieval assumptions ignored. Cloud-top height from IR brightness temperature assumes an opaque cloud and a known temperature profile; for thin cirrus both fail, and heights are biased low by kilometres.

Point-to-grid comparison. A rain gauge measures ~0.02 m²; a 3 km grid cell is \(9\times10^6\) m². Their variances differ by construction, so RMSE against gauges includes a representativeness component that no model improvement can remove.

Station metadata changes. Instrument swaps, relocations and changes in observation time produce step changes that look meteorological. Homogenisation exists for this, and raw records need it.

Worked example

Radar beam height, and why distant coverage is not coverage:

import numpy as np

Re_eff = (4.0/3.0) * 6371e3            # effective Earth radius for standard refraction
beamwidth = np.deg2rad(0.95)           # WSR-88D

for r_km in (25.0, 60.0, 120.0, 200.0):
    r = r_km * 1e3
    for elev_deg in (0.5,):
        h = r*np.sin(np.deg2rad(elev_deg)) + r**2/(2*Re_eff)
        width = r * beamwidth
        print(f"range {r_km:5.0f} km   beam centre {h:6.0f} m AGL   "
              f"beam width {width/1e3:5.2f} km")

range    25 km   beam centre    255 m AGL   beam width  0.41 km
range    60 km   beam centre    735 m AGL   beam width  0.99 km
range   120 km   beam centre   1895 m AGL   beam width  1.99 km
range   200 km   beam centre   4100 m AGL   beam width  3.32 km

At 200 km the lowest beam is 4 km up and 3.3 km wide — it cannot see shallow precipitation at all, and what it does see is smeared over a volume larger than a convection-permitting grid cell.

Connections

Sources

  • Rinehart, Radar for Meteorologists.
  • Hersbach et al. (2020), The ERA5 global reanalysis.
  • ECMWF and NOAA dataset documentation — the known-issues pages are worth reading in full.