Meteorologist, oceanographer, Python solvem probler, experimenter, Raspberry Pi tinkerer, geek dad, manager, space enthusiast, ham radio lover, science columnist, ADHDer, synthetic intellect curious, linux admin cosplayer.

Post 48 — “The Hemisphere That Caught Up”: Scientific & Technical Research

Satellite RADIANCE data assimilation — the observation side of the forecast revolution. Remit: scientific/technical substance. Status: COMPLETE (first pass).

PRIMARY SOURCE NOTE: The single richest source is J. R. Eyre’s own review “Progress achieved on assimilation of satellite data in NWP over the last 30 years” (ECMWF Seminar, 3–7 Sept 2007). Eyre is a central human character in this post, and he narrates the whole retrieval-to-direct-radiance arc first-hand, including the cost-function equation and the retrieval-error analysis. Full text extracted and read. Page/line references below are to that PDF unless noted. Cross-checked against ECMWF’s “Fifty years of data assimilation at ECMWF” (2025) and the Simmons & Hollingsworth 2002 QJRMS paper.

1. What a satellite sounder measures — radiances, brightness temperatures, weighting functions

A passive sounder does NOT measure temperature. It measures radiance — the intensity of upwelling radiation reaching the satellite at a set of wavelengths (channels) — which is conventionally re-expressed as a brightness temperature (the temperature a blackbody would need to emit that radiance). The atmosphere is semi-transparent, so radiation reaching the sensor in any one channel originates from a range of altitudes.

Weighting function. Each channel’s radiance is a vertically-weighted average of the Planck emission along the viewing path; the vertical profile of that weighting is the channel’s weighting function (the derivative of transmittance with respect to height/log- pressure). A channel placed on the strongly-absorbing centre of a gas absorption band is opaque and its signal comes from high in the atmosphere; a channel on the weakly-absorbing wing sees deep toward the surface. By choosing channels across an absorption band you build a “stack” of weighting functions peaking at different levels, and that is how a sounder resolves a vertical temperature profile.

Which gas. Temperature sounding uses channels in a band of a gas of known, ~uniform concentration — the CO2 bands near 15 µm and 4.3 µm in the infrared, and the O2 complex near 50–60 GHz in the microwave. Because the mixing ratio is known, radiance variations map to temperature. Humidity sounding uses water-vapour bands — the 6.3 µm H2O band in the IR, and ~183 GHz in the microwave — where the absorber amount is itself the unknown.

TOVS specifics (Eyre §3.1, from Smith et al. 1979):

  • HIRS CO2 channels (15 and 4.3 µm): tropospheric + lower-stratospheric temperature, ~40 km spatial sampling.
  • HIRS 6 µm H2O channels: tropospheric humidity.
  • MSU: tropospheric/lower-stratospheric temperature, fewer channels → lower vertical resolution, ~160 km sampling, BUT works in cloud.
  • SSU (pressure-modulator radiometer): extends temperature into the upper stratosphere, ~200 km sampling.
  • Individual weighting functions are 5–10 km wide; even combining overlapping channels the effective system vertical resolution of TOVS is only ~3 km — its central limitation. (Advanced IR sounders AIRS/IASI later reached ~1 km; see §8.)

Infrared vs microwave — why microwave sees through cloud. This is the pivotal physical point of the post. Cloud droplets are comparable in size to infrared wavelengths (~10 µm), so they strongly absorb/scatter IR — an IR sounder effectively sees only the cloud top and is blind below it. Microwave wavelengths (a few mm) are much larger than cloud droplets, so non-precipitating cloud is nearly transparent to microwave; the microwave sounder retrieves temperature in cloudy air. Since meteorologically active regions (fronts, cyclones) are precisely the cloudy ones, the microwave channels’ all-but-all-weather capability is what made satellite sounding decisive for NWP. Eyre §5.1: the introduction of AMSU “provided sounding quality in cloudy areas that had only been available hitherto in cloud-free areas … crucial to the performance of NWP systems.”

2. The RETRIEVAL problem — ill-posed inversion, McMillin, split-window

The retrieval-then-assimilate paradigm (1970s–1980s). The instinct was to make satellite data “look like a radiosonde”: invert the measured radiances into a vertical temperature/ humidity profile (a “retrieval”/”sounding”), then feed that profile to the analysis as if it were a sonde. Eyre frames this with pioneer Verner Suomi’s “11th commandment”: “Thou shalt not worship the radiosonde.” Treating satellite soundings as “poor-quality sondes” was, in Eyre’s words, “an ultimately flawed approach.”

Why the inversion is ill-posed. Recovering a continuous profile from a handful of broad, overlapping weighting functions is a classic ill-posed inverse problem: many different profiles produce nearly identical radiances, so the solution is non-unique and unstable without extra information (a first guess / prior). The measured radiances are accurate with simple error characteristics, but the retrieved profile at high vertical resolution has low accuracy and complicated, correlated errors.

McMillin, radiative transfer and the split-window. Larry (Lawrence) McMillin at NOAA/NESDIS was central to operational retrieval and to correcting radiances for atmospheric attenuation. The split-window technique (McMillin, ~1975; theory & validation McMillin & Crosby / McMillin 1984) exploits the differential absorption in two adjacent IR window channels: because water-vapour absorption differs between the two nearby channels, the difference in their brightness temperatures estimates — and removes — the atmospheric correction, recovering surface (sea-surface) temperature. It is the archetypal example of using channel differences to cancel an unknown atmospheric effect, and the intellectual ancestor of using multiple channels to separate signals.

Why retrieval-then-assimilate discards information and injects correlated error. Eyre’s error analysis (§3.5.3) is the technical heart. Write the retrieval as a linear analysis:

x_a − x_b = K · (y_o − H[x])            (retrieval)
y_o − H[x] = H · (x − x_b) + ε          (forward/RT model, ε = measurement error)

Combining gives the retrieval error:

x_a − x_t = (I − K·H)·(x_b − x_t) + K·ε          (Eyre eq. 11)

where x_t is the true profile. The second term (K·ε) maps measurement error into the retrieval — expected and benign. The first term maps the background/first-guess error into the retrieval. In a statistical-regression retrieval the effective background is a climatological mean, so adjacent soundings share nearly the same background error → the retrievals inherit systematic, spatially- and synoptically-correlated errors that no local observation-error covariance can represent. Andersson et al. (1991) documented exactly these “synoptically correlated biases” in TOVS layer-mean temperature increments. This is why, theoretically, assimilating retrieved profiles is fundamentally more problematic than assimilating radiances directly — the correlated retrieval error is invisible to an assimilation system that assumes uncorrelated observation errors.

The late-1980s crisis of confidence. As short-range forecasts improved, TOVS-retrieval impact in the Northern Hemisphere went from positive (Uppala et al. 1984) to neutral (1987 system) to negative (1988 system; Andersson et al. 1991). Kelly & Pailleux (1988) tried to fix it by coarsening the retrieved profile (14 thickness layers in the early 1980s → 11 in 1985 → 7 in 1987) to represent only vertical scales TOVS could actually resolve — SH impact stayed positive but NH stayed mixed. This crisis forced the re-think that produced direct radiance assimilation.

3. The DIRECT RADIANCE turn — H as a radiative-transfer forward model inside the cost function

The idea. Instead of inverting radiances into a profile and assimilating the profile, put the radiative-transfer forward model directly inside the assimilation as the observation operator H, and assimilate the raw brightness temperatures. The analysis then adjusts the model state so that radiances simulated from the model match the measured radiances. No inversion is performed; the ill-posedness is handled implicitly and optimally by the background constraint already present in the analysis. Crucially, when simulated radiances equal measured radiances the scheme produces zero analysis increment — it never fabricates structure the data don’t support (Eyre §3.6 on 1D-Var).

Why variational DA made this natural (the link to Post 47). Eyre §3.5.2 gives two reasons direct radiance assimilation rode in on the back of variational methods (OI → 3D/4D-Var):

  1. Nonlinearity. The RT operator H[x] is nonlinear (Planck function, transmittances). Optimal Interpolation is built on linear theory and can’t easily handle it; variational minimisation of a cost function can.
  2. Dimensionality. OI’s weight matrix (eq. K = B·H^T·(H·B·H^T + E + F)^-1) requires inverting a matrix the size of the number of observations — infeasible for many thousands of radiances. Variational methods minimise the cost function iteratively and sidestep the explicit inversion, so millions of radiances become tractable. So the arrival of 3D-Var/4D-Var and the arrival of direct radiance assimilation are two faces of the same event. Direct radiance assimilation is the first great customer of the variational machinery Post 47 describes.

1D-Var — the bridge (Eyre 1989; Eyre et al. 1993). The one-dimensional (vertical-only) form of the variational equations, applied at a single sounding location using the forecast profile and its error covariance as the constraint. J. R. Eyre developed it — theory in Eyre (1989), “Inversion of cloudy satellite sounding radiances by nonlinear optimal estimation. I: Theory and simulation for TOVS” (QJRMS 115, 1001–1026) — and it became operational on TOVS at ECMWF in June 1992 (Eyre, Kelly, McNally, Andersson & Persson, “Assimilation of TOVS radiance information through one-dimensional variational analysis,” QJRMS 1993). In the early ECMWF system 1D-Var sat between the raw radiances and the OI analysis: it converted radiances into a forecast-consistent retrieval that OI could then use, without the climatological-background pathology of statistical retrievals.

Into 3D-Var / 4D-Var. The seminal methodological paper is Andersson, Pailleux, Thépaut, Eyre, McNally, Kelly & Courtier (1994), “Use of cloud-cleared radiances in three/four- dimensional variational data assimilation” (QJRMS). Operational milestones:

  • NCEP: 3D-Var direct radiance assimilation operational October 1995 (Derber & Wu; SSI analysis).
  • ECMWF: 3D-Var operational January 1996 (Andersson et al. 1994).
  • ECMWF: first direct radiance assimilation in 4D-Var, November 1997 — the operational switch that Post 47 is built around. (Eyre §3.6.) ECMWF’s own 50-year retrospective: “the transition to variational assimilation and direct radiance assimilation resulted in the largest changes to operational forecast scores at the end of the 1990s.”

Tangent-linear and adjoint of the RT model. Because H now lives inside a gradient-based minimisation, the analysis needs the derivative of H — its Jacobian H = ∇_x H[x] (how each brightness temperature responds to a change in temperature/humidity at each level) — and, for efficient cost-function-gradient computation, the tangent-linear and adjoint versions of the radiative-transfer code. This requirement is exactly why the fast RT model (§4) had to be differentiable, not just fast.

4. RTTOV — the fast radiative-transfer model (the “enabling machine” is SOFTWARE)

What it is. RTTOV (originally “Radiative Transfer for TOVS”) is a fast radiative- transfer model: given a model profile of temperature, humidity and trace gases it computes top-of-atmosphere radiances/brightness temperatures for a sensor’s channels — and provides tangent-linear, adjoint and Jacobian versions for assimilation. It is the concrete implementation of the observation operator H for radiance assimilation.

Why speed is the binding constraint. A variational analysis evaluates H (and its adjoint) for every radiance at every iteration of the minimisation — millions of radiative-transfer evaluations per analysis cycle, within a fixed operational time window. A line-by-line RT computation (summing absorption over thousands of spectral lines) is far too slow. RTTOV replaces it with regression predictors: layer optical depths/transmittances are parameterised as fast polynomial functions of profile variables, with regression coefficients pre-trained against accurate line-by-line models. This trades a little accuracy for orders-of- magnitude speed — and that trade is what makes operational radiance assimilation possible at all. Hence the post’s thesis: the enabling “machine” here is software, not hardware.

Lineage (from the NWP SAF / Saunders & Matricardi history):

  • Eyre & Woolf (1988) — original fast-transmittance basis.
  • Eyre (1991) — first version of RTTOV, maintained at ECMWF in the early 1990s.
  • Successive modifications: Rayer (1995), Rizzi & Matricardi (1998), Saunders & Matricardi (1999) — “An improved fast radiative transfer model for assimilation of satellite radiance observations,” QJRMS (ECMWF Tech Memo 282).
  • Mid-1990s: EUMETSAT set up its Satellite Application Facilities; the NWP SAF, led by the Met Office (UK), adopted RTTOV as a core package and has developed it ever since. Roger Saunders (Met Office) is the long-time lead. RTTOV now has 1000+ registered users worldwide and versions through RTTOV-13/14.

Institutional point. Eyre §6 notes that maintaining expertise on every satellite observation operator exceeded any single NWP centre’s means. This recognition in the mid-1990s spawned collaborative structures — the EUMETSAT NWP SAF in Europe and the Joint Center for Satellite Data Assimilation (JCSDA) in the USA (whose analogous fast model is CRTM). The fast RT model was thus not just code but shared scientific infrastructure.

5. Variational Bias Correction (VarBC) — Dee 2004; Auligné, McNally & Dee 2007

The problem. Direct radiance assimilation compares measured brightness temperatures to model-simulated ones. But both carry systematic errors: instrument calibration drift, imperfect spectroscopy in the fast RT model, air-mass-dependent biases. These biases are typically larger than the atmospheric signal the analysis is trying to extract, and vary from channel to channel, scan position to scan position, air mass to air mass. Uncorrected, they would poison the analysis.

The solution — VarBC. Model the bias of each channel as a small linear combination of bias predictors (e.g. constant offset, scan angle, layer-thickness / air-mass predictors, surface skin temperature), and estimate the predictor coefficients inside the variational cost function, simultaneously with the atmospheric state, by augmenting the control vector. The bias correction thus adapts automatically and continuously as instruments drift and as new sensors are added — no manual offline retuning. Key references:

  • Dee (2004), “Variational bias correction of radiance data in the ECMWF system” (ECMWF workshop/proceedings; foundational statement).
  • Auligné, McNally & Dee (2007), “Adaptive bias correction for satellite data in a numerical weather prediction system,” QJRMS 133, 631–642 — the definitive VarBC paper; reviews/updates coefficients at all analysis cycles adaptively.
  • VarBC became operational in the ECMWF IFS in 2006. (FLAG: exact month commonly cited as September 2006 — confirm before asserting a month.)
  • Dee & Uppala (2009), “Variational bias correction of satellite radiance data in the ERA-Interim reanalysis,” QJRMS — VarBC in reanalysis.

Why an unbiased anchor is needed. VarBC has a subtle failure mode: if every observation type is bias-corrected against the model background, the whole system can drift together toward a biased state (nothing pins the absolute reference). The analysis needs some observations that are NOT bias-corrected — trusted “anchor” data that hold the system to the true climate. Radiosondes and, especially, GPS radio occultation serve as these anchors: RO bending angle is traceable to the fundamental measurement of time/frequency (an atomic-clock-referenced Doppler/phase measurement) and has systematic errors < 0.2 K, so it is assimilated without bias correction and effectively anchors VarBC (see §8). This is the deep reason RO is prized out of proportion to its data volume.

6. Channel selection & information content for hyperspectral sounders

Hyperspectral IR sounders (AIRS 2378 channels, IASI 8461 channels; §8) produce far more channels than can be assimilated operationally, and many are redundant or contaminated (cloud, surface, trace gases). One selects an informative subset using information-content theory (Clive Rodgers, “Inverse Methods for Atmospheric Sounding,” 2000):

  • Degrees of Freedom for Signal (DFS) = trace of the averaging-kernel matrix A = ∂x_a/∂x_t; it counts how many independent pieces of information the observations add beyond the background. For a hyperspectral IR sounder the effective DFS for temperature+humidity is only of order ~10–15 even from thousands of channels — the weighting functions overlap heavily.
  • Shannon information content H_S = ½ ln|B / A_analysis| (entropy reduction) is the other standard metric.
  • Channel selection (e.g. Rodgers’ iterative maximum-information method; Collard’s operational IASI selection) picks the channels that maximise cumulative DFS/entropy subject to avoiding cloud- and surface-sensitive channels and accounting for inter-channel error correlation. The pitch for the post: thousands of channels do NOT mean thousands of independent facts about the atmosphere; the system vertical resolution improves to ~1 km precisely because thousands of overlapping narrow weighting functions are combined — but the true information content is a modest number of degrees of freedom, and knowing which channels carry it is its own science.

7. The accessible equation (variational observation term with H = radiative-transfer operator)

Use Eyre’s own cost function (Eyre eq. 4), which is exactly the Post-47 variational cost function with the observation operator explicitly a radiative-transfer model:

J[x] = ½ (x − x_b)ᵀ B⁻¹ (x − x_b)  +  ½ (y_o − H[x])ᵀ (E + F)⁻¹ (y_o − H[x])
  • x = atmospheric state being solved for (temperature, humidity, … profiles)
  • x_b = background / short-range forecast first guess
  • B = background-error covariance
  • y_o = the observations = measured radiances (brightness temperatures)
  • H[x] = the observation operator = the radiative-transfer forward model (RTTOV): it turns a model atmospheric state into simulated radiances
  • E, F = error covariances of the observations (E) and of the observation operator (F)

The whole conceptual move of the post lives in one substitution: in classical DA, y_o is a temperature and H just interpolates the model to the observation location. In direct radiance assimilation, y_o is a raw brightness temperature and H is a physics model of how photons leave the atmosphere. The analysis minimises J — pulling the model state toward the forecast (first term) and toward agreement with what the satellite actually measured (second term) — and, because H is the radiative-transfer model itself, no ill-posed inversion is ever performed. (For the writer: the minimisation needs ∇_x J, hence the adjoint of H — tie back to §3/§4 and Post 47’s adjoint.)

Companion equation for GPS-RO (Eyre eq. 12) — refractivity as the observation quantity:

N = κ₁ p/T + κ₂ e/T² + κ₃ n_e/f² + κ₄ W

N = refractivity = (n−1)×10⁶; p pressure, T temperature, e water-vapour pressure, n_e electron density, f frequency, W liquid-water density. Term 3 (ionosphere) removed using the two GPS frequencies 1.575 & 1.227 GHz; term 4 negligible at GPS frequencies; term 2 (water vapour) dominates the lower troposphere, term 1 (density) dominates aloft. Shows why RO gives clean temperature aloft and humidity below.

8. Instrument-chronology scaffold (dates verified)

  • TIROS-1, 1960 — first satellite images of Earth’s weather.
  • Nimbus-3, April 1969 — first satellite temperature sounders (SIRS, IRIS). SIRS data “bogussed” into NMC analyses by Smith et al. (1970) — the very first satellite-sounding assimilation.
  • NOAA-2, 1972 — first operational temperature sounder; carried VTPR (Vertical Temperature Profile Radiometer), the IR filter radiometer flown on NOAA 2–5 (1972–79).
  • TIROS-N, launched 13 October 1978 (Vandenberg, Atlas E/F, ~850 km sun-synchronous) — new generation of operational polar orbiters; introduced TOVS = HIRS/2 + MSU + SSU. TOVS flew from TIROS-N through NOAA-14, providing operational soundings for 20+ years.
  • FGGE / First GARP Global Experiment, Dec 1978 – Nov 1979 — first global test of the space-based Global Observing System; basis for many OSEs and later reanalyses.
  • 1D-Var operational on TOVS at ECMWF, June 1992 (Eyre et al. 1993).
  • Direct radiance assimilation: NCEP 3D-Var Oct 1995; ECMWF 3D-Var Jan 1996; ECMWF 4D-Var Nov 1997.
  • NOAA-15 (NOAA-K), launched 13 May 1998 (Vandenberg, Titan-II) — introduced ATOVS = AMSU-A + AMSU-B + HIRS/3. AMSU-A: 15 channels, 23.8–89 GHz, O2-band temperature sounding surface-to-~3 hPa, with HIRS-comparable horizontal/vertical resolution in cloud. The single biggest step-change in sounder impact on NWP. (AMSU-B → later “look-alike” MHS.)
  • AIRS on Aqua, launched 4 May 2002 — first advanced/hyperspectral IR sounder; grating spectrometer, 2378 channels, 3.7–15.4 µm; research/demo instrument but heavily exploited by NWP. Assimilated at ECMWF from 2003–04.
  • IASI on MetOp-A, launched 19 October 2006 — Fourier-transform interferometer, 8461 channels, 3.6–15.5 µm; first hyperspectral interferometer on an operational meteorological mission. System vertical resolution ~1 km (vs ~3 km for HIRS/TOVS).
  • GPS radio occultation: GPS/MET (Microlab-1) demonstrated the technique 1995–97 (Kursinski et al. 1996; Rocken et al. 1997); CHAMP from 2000 (continuous from 2001, near-real-time from 2006); GRACE-A, SAC-C; COSMIC/FORMOSAT-3, launched 15 April 2006, 6-satellite constellation (~3000 occultations/day; Anthes et al. 2000); MetOp/GRAS operational from 2007. Assimilation options (Eyre 1994): retrieved T/q (rejected — can’t separate T and humidity effects on refractivity), retrieved refractivity, or (preferred) direct bending-angle assimilation via a 1-D bending-angle observation operator (Healy & Thépaut 2006, integrated into ECMWF 4D-Var). RO: high vertical resolution 0.5–1 km, low horizontal resolution ~200 km, ~1 K random error, < 0.2 K systematic error, all-weather — hence the VarBC anchor role (§5). Key people: Sean Healy (ECMWF), Christian Rocken, Richard (Rick) Anthes (UCAR/COSMIC).

9. The payoff — SH↔NH skill convergence (Simmons & Hollingsworth 2002)

Canonical paper: A. J. Simmons & A. Hollingsworth, “Some aspects of the improvement in skill of numerical weather prediction,” QJRMS 128, 647–677 (2002) — the SAME verification paper anchoring Post 47. DOI 10.1256/003590002321042135.

The headline finding (verbatim from the abstract): across three global NWP systems there had been a gain of “about a 1-day gain in predictability of mean-sea-level pressure and 500 hPa height over the last decade in the northern hemisphere, with a similar gain over the last 3 years in the southern hemisphere.” The SH — historically trailing the NH by roughly a full day of skill because it is an observational desert (few radiosondes, ships, aircraft) — closed most of that gap in a few years, and the agent of convergence was improved use of satellite data (advanced sounders + direct radiance assimilation + variational methods). By the mid-2000s the two 500 hPa geopotential-height anomaly-correlation curves had essentially converged.

Later corroboration. ECMWF’s own long-term skill charts (e.g. “Fifty years of data assimilation at ECMWF,” 2025) plot the lead time at which 500 hPa height anomaly correlation falls below the skill threshold and show the NH/SH gap essentially closing in the 2000s; Bauer, Thorpe & Brunet (2015), “The quiet revolution of numerical weather prediction,” Nature 525, 47–55, made the converged NH/SH skill curves iconic and attributed the change substantially to satellite data + 4D-Var.

Data-denial / OSE evidence. OSEs consistently show the SH forecast collapses toward ~1980s skill without satellites. Even the 1982 Gilchrist FGGE OSE report already showed ECMWF “useful predictability reduced from 5.5 to 4.5 days in NH and from 5 to 3 days in SH” when satellite data were withheld — the SH always loses far more. Later FSOI work (below) quantified which satellite system matters most.

Observation impact / FSOI. Adjoint-based Forecast Sensitivity to Observations Impact (Cardinali 2009, “Monitoring the observation impact on the short-range forecast,” QJRMS; operational FSOI at ECMWF since ~2009/2012) and the analogous adjoint method of Langland & Baker (2004) (Tellus; NRL) rank observing systems by their measured contribution to reducing short-range (24–48 h) forecast error. Result repeatedly reported: AMSU-A is the single largest-impact observing system, typically followed by IASI, radiosondes (TEMP), aircraft and AMVs. (Also Joo et al. at the Met Office.) FSOI is limited to short range because it relies on a simplified/linearised adjoint of the assimilation system.

Human centres — quick attribution notes for the writer

  • Lawrence “Larry” McMillin (NOAA/NESDIS) — operational retrieval, radiative transfer (McMillin–Fleming), the split-window technique (~1975; McMillin 1984 SST validation).
  • John R. Eyre (Met Office → ECMWF → Met Office) — 1D-Var (1989, 1993), the retrieval-error theory, first RTTOV (1991), RO assimilation options (1994); author of the definitive 30-year review used here. Central protagonist.
  • Erik Andersson, Anthony “Tony” McNally, Jean-Noël Thépaut, Graeme Kelly, Jean Pailleux, Philippe Courtier — the ECMWF team that carried direct/cloud-cleared radiances into 3D/4D-Var (Andersson et al. 1994). (Courtier links to Post 47.)
  • Roger Saunders (Met Office) — long-time RTTOV lead; NWP SAF.
  • Sean Healy (ECMWF), Christian Rocken, Richard “Rick” Anthes (UCAR) — GPS RO.
  • Carla Cardinali (ECMWF), Rolf Langland & Nancy Baker (NRL), Sanghyun Joo (Met Office/KMA) — observation impact / FSOI.
  • Dick (D. P.) Dee (ECMWF) and Thomas Auligné — variational bias correction.

Cross-links to already-published posts (exact URLs — provided in brief)

  • Post 41 Lorenz/chaos: /weather/hpc/history/2026/05/15/The-coyote-who-found-chaos.html
  • Post 44 optimal interpolation / Lorenc: /weather/hpc/history/2026/05/17/The-book-that-crossed-the-curtain.html
  • Post 46 Kalnay / reanalysis: /weather/hpc/history/2026/05/25/A-textbook-for-the-grandmothers.html
  • Post 47 4D-Var at ECMWF: /weather/hpc/history/2026/05/30/Eleven-years-from-the-adjoint.html

FLAGS / NOT-VERIFIED items

  • VarBC exact operational month at ECMWF — “2006” is well supported; “September 2006” is commonly cited but I did not confirm the month from a primary ECMWF source. FLAG.
  • McMillin split-window original year (1975) and the precise “McMillin 1975” citation come from secondary/derivative literature (search snippets), not a primary McMillin 1975 PDF. The primary I can vouch for is McMillin & … 1984 JGR (“multiple/split window SST”). Verify the 1975 origin against a primary source before asserting a year. FLAG.
  • Split-window attribution as the McMillin technique is standard in the LST/SST remote- sensing literature; the NWP relevance is via radiative-transfer/retrieval, not directly the split-window per se — present carefully.
  • AIRS assimilation start at ECMWF (2003 vs 2004) — Eyre implies “very encouraging results” pre-2007; ECMWF operational AIRS often dated to 2003. Minor; verify exact date if used.
  • COSMIC launch date 15 April 2006 — from the brief; Eyre’s table lists “2006” without day. Consistent, day not independently re-verified here (widely reported as 15 April 2006 UTC).
  • Cardinali FSOI operational-since year — sources say “since 2009” (paper) and “routinely since 2012” (implementation); use “from around 2009” to be safe.
  • DFS numeric magnitude (~10–15 for hyperspectral IR) — order-of-magnitude, illustrative; actual value depends on instrument/config. Present as “of order ten,” not a hard number.

Sources (URLs actually fetched or directly cited from fetched search results)

  • Simmons & Hollingsworth 2002, QJRMS (abstract + DOI): https://rmets.onlinelibrary.wiley.com/doi/10.1256/003590002321042135
  • Eyre, “Progress achieved on assimilation of satellite data in NWP over the last 30 years,” ECMWF Seminar 2007 (FULL TEXT extracted & read — primary source for §1–3,8,9): https://www.ecmwf.int/sites/default/files/elibrary/2008/9341-progress-achieved-assimilation-satellite-data-numerical-weather-prediction-over-last-30-years.pdf
  • “Fifty years of data assimilation at ECMWF,” 2025 (FULL TEXT extracted & read — §3,9): https://www.ecmwf.int/sites/default/files/elibrary/81650-fifty-years-of-data-assimilation-at-ecmwf.pdf
  • Eyre et al. 1993, “Assimilation of TOVS radiance information through one-dimensional variational analysis,” QJRMS: https://rmets.onlinelibrary.wiley.com/doi/abs/10.1002/qj.49711951411
  • Saunders & Matricardi 1999 (RTTOV), ECMWF Tech Memo 282 (RTTOV lineage): https://nwpsaf.eu/oldsite/deliverables/rtm/papers/tm282.pdf
  • RTTOV overview / NWP SAF (Saunders lead, lineage Eyre-Woolf 1988 → Eyre 1991 → Rayer 1995 → Rizzi & Matricardi 1998): https://nwp-saf.eumetsat.int/site/software/rttov/
  • Dee 2004, “Variational bias correction of radiance data in the ECMWF system” (ECMWF): https://www.ecmwf.int/sites/default/files/elibrary/2004/8930-variational-bias-correction-radiance-data-ecmwf-system.pdf
  • Auligné, “Bias correction of satellite data at ECMWF,” 2005 (ECMWF; VarBC development): https://www.ecmwf.int/sites/default/files/elibrary/2005/15847-bias-correction-satellite-data-ecmwf.pdf
  • Dee & Uppala 2008/2009, “Variational bias correction in ERA-Interim” (ECMWF): https://www.ecmwf.int/sites/default/files/elibrary/2008/8936-variational-bias-correction-era-interim.pdf
  • “Bias correction of satellite radiance observations at ECMWF,” 2020 (NWP SAF overview PDF): https://nwp-saf.eumetsat.int/site/download/ECMWF-VarBC.pdf
  • Healy, “An introduction to GPS radio occultation and its use in NWP,” ECMWF 2008 (fetched): https://www.ecmwf.int/sites/default/files/elibrary/2008/9342-introduction-gps-radio-occultation-and-its-use-numerical-weather-prediction.pdf
  • Healy & Thépaut 2006, “Assimilation experiments with CHAMP GPS radio occultation measurements,” QJRMS: https://rmets.onlinelibrary.wiley.com/doi/abs/10.1256/qj.04.182
  • Cardinali 2009, “Monitoring the observation impact on the short-range forecast,” QJRMS (ResearchGate record): https://www.researchgate.net/publication/229466083_Monitoring_the_observation_impact_on_the_short-range_forecast
  • ECMWF, “Forecast sensitivity to observations (FSO) as a diagnostic tool,” 2009: https://www.ecmwf.int/sites/default/files/elibrary/2009/8578-forecast-sensitivity-observations-fso-diagnostic-tool-monitoring-impact-observations-short.pdf
  • McMillin 1984, “Theory and validation of the multiple window sea surface temperature technique,” J. Geophys. Res. Oceans (abstract): https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/JC089iC03p03655
  • NOAA-15 launch (13 May 1998) / ATOVS era, ECMWF news: https://www.ecmwf.int/en/about/media-centre/news/2018/noaa-satellite-launch-20-years-ago-marked-start-new-era
  • TIROS-N (launched 13 Oct 1978), Wikipedia: https://en.wikipedia.org/wiki/TIROS-N
  • AMSU (channels/Aqua), NASA Aqua project: https://aqua.nasa.gov/content/amsu
  • Collard 2011, “From Observations to Forecasts – Part 8: The use of satellite observations in NWP,” Weather (RMetS): https://rmets.onlinelibrary.wiley.com/doi/10.1002/wea.736
  • Bauer, Thorpe & Brunet 2015, “The quiet revolution of numerical weather prediction,” Nature: https://www.nature.com/articles/nature14956
  • eoPortal, “Satellite Inputs to Numerical Weather Prediction (NWP)”: https://www.eoportal.org/other-space-activities/nwp