Instrument-Criticality: what climate modelling would lose if an Earth-observation instrument went dark

Framing (non-negotiable). The unit of analysis is a GCMD instrument label carried by a spaceborne platform in nasa-gesdisc-kg (288 labels on 254 platforms), scored against the climate-modelling literature in climatemodelskg (2,000 papers). Every number below is a bibliometric or catalogue-structural association — a measure of how visibly the published modelling record leans on an instrument, and of how the catalogue is wired. It is not a measurement of scientific irreplaceability, and not an engineering or programmatic risk assessment. A low score is evidence of low observed dependency in this evidence base, never evidence that an instrument does not matter — §8 Claims 3 and 7 show exactly that failure mode. Keep this caveat attached to every downstream claim.

Abbreviations. GCMD = Global Change Master Directory (NASA's controlled instrument/platform vocabulary); DAAC = Distributed Active Archive Center; CMR = Common Metadata Repository; DOI = Digital Object Identifier; ORCID = Open Researcher and Contributor ID; ROR = Research Organization Registry; CMIP = Coupled Model Intercomparison Project; ECV = Essential Climate Variable; GCOS = Global Climate Observing System; ERB = Earth Radiation Budget; TWSA = terrestrial water storage anomaly; NLP = natural-language processing; KG = knowledge graph; DB = dependency breadth; IR = irreplaceability; R1–R5 = the five dependency routes defined in §3; ρ = Spearman rank-correlation coefficient; L2 = satellite Level-2 (retrieved geophysical) product.

1. Executive summary

The OKN federation describes Earth-observation infrastructure at real scale but not at the resolution the retirement question demands. nasa-gesdisc-kg catalogues 921 instruments on 455 platforms across 8,058 datasets and 457,085 publications; restricting to spacecraft gives 254 platforms carrying 288 instrument labels and 4,931 datasets. Of those labels only 243 are actual science instruments — 25 are generic GCMD class names (RADIOMETERS, SAR, GPS, NOT APPLICABLE) and 20 are spacecraft-bus subsystems (star trackers, gyros, laser retroreflectors). Establishing that before ranking anything matters, because the generic labels sit near the top of every raw volume count.

Dependency was established along five independent routes — two textual (an instrument or its platform named in a climate-modelling paper), two structural (NASA's own record that a DOI-matched modelling paper used that instrument's data; NASA-side publications with a modelling term in the title citing its datasets), and one capability route (the instrument is the only measurer of a variable that model components produce). The routes reach 73, 118, 44, 76 and 162 instruments respectively, and they do not agree with each other: the two paper-level textual routes correlate at ρ = 0.79 and the two structural routes at ρ = 0.74, but across the two families agreement ranges only ρ = 0.32–0.55. Only 24 of 243 instruments show dependency on all five routes; 49 show it on none. Route agreement, not the magnitude of any single route, is what carries weight here.

The ranking is led by MODIS (score 100.0/100, 5/5 routes, 29.5% of all instrument mentions in the matched corpus), followed by AMSR-E, MISR, MOPITT, AMSU-A, VIIRS, AIRS, SSMIS, AVHRR and ASTER. Three risk classes were kept separate rather than collapsed: 58 instruments broadly relied on, 2 narrow-but-irreplaceable, and 9 with a substantial data footprint and no modelling uptake at all.

On the asymmetry — the finding matters, but not in the form it was posed. At population level, data footprint and criticality are strongly correlated (ρ = 0.727, n = 243, p < 1e-40; ρ = 0.656 using fractional attribution). The blanket claim that the least survivable losses are not the largest archives is not supported. What is supported is sharper and more useful: the correlation breaks down exactly where decisions get made. Only 6 of the top ten by criticality are also in the top ten by volume, and individual rank gaps reach +133/−150 places. PALSAR (13 datasets) outranks 133 higher-volume instruments; ACE-FTS (10 datasets) outranks 120; CERES-FM5 (527 datasets) ranks 112th. The 49 instruments with no dependency signal on any route hold only 2.7% of the spaceborne dataset attributions. Volume is a decent prior and a bad decision rule.

2. Sources used

Both joins are the federation's own precomputed, hand-verified crosswalks (EO1, EO2, PB1, PB2 in the crosswalk catalogue) and both were re-established by logged query rather than taken on trust. No other federation graph was queried, and none is credited.

3. Design & rules

What counts as observing infrastructure. nasa-gesdisc-kg types every platform, so "flying on spacecraft" is a filter, not a judgement: platforms typed Earth Observation Satellites, Space Stations/Crewed Spacecraft, Solar/Space Observation Satellites, Navigation Satellites, Spacecraft or Space-based Platforms. That yields 254 platforms, 288 distinct instrument labels and 4,931 datasets. Everything airborne, shipborne, balloon-borne or ground-based — the great majority of the catalogue's 921 instruments — is out of scope and reported as such in Figure 1A.

What the labels actually are. The 288 labels were classified by hand into 243 science instruments, 25 generic GCMD class names and 20 platform/bus subsystems. Only the science instruments are scored. This is not tidying: NOT APPLICABLE alone carries 1,855 attributed datasets and would otherwise rank second in the catalogue by volume.

How a dataset is attributed to an instrument. It is not, directly. nasa-gesdisc-kg has no Dataset→Instrument edge: datasets attach to platforms (HAS_PLATFORM) and platforms carry instruments (HAS_INSTRUMENT), so every instrument on a platform inherits all of that platform's datasets. Terra's 793 datasets are credited identically to MODIS, MISR, MOPITT, ASTER and CERES-FM1. Both a raw count and a fractional count (a platform's datasets divided evenly among its instruments) are carried through the analysis, and the asymmetry test is run on both.

The five dependency routes, each with an explicit meaning and an explicit evidence type:

A sixth axis, R5 (capability), is not a dependency count but a substitutability measure: a variable that is both measured by an instrument and produced by a model component is sole-measured when exactly one distinct instrument name in the corpus measures it.

Scoring. Dependency breadth (DB) is the unweighted mean of the five log-normalised route scores; irreplaceability (IR) is the log-normalised count of sole-measured variables plus sole-source GCMD keywords; corroboration is the count of non-zero routes. Criticality = 100 × (0.55·DB + 0.30·IR + 0.15·(corroboration/5)), rescaled to a 100-point maximum. The weights are a stated judgement, not a fitted quantity — no ground truth exists to fit them against. Full replicator specification in Instrument-Criticality_reproducibility.md.

Inventory, rebuilt live.

Panel B is the reason the attribution rule had to be stated before any ranking: 447 instruments appear to hold 200+ datasets each, but that is co-flight inheritance, not instrument-specific volume. Panel D shows the routes are not interchangeable in reach — R4 sees 162 instruments, R2 only 44 — which is the first sign that no single route can carry the ranking.

4. Confidence tiers

Tier is deliberately a function of corroboration, not of score magnitude, so that a high score resting on one route cannot be mistaken for a robust one. The distribution is bottom-heavy — 112 of 243 instruments sit in tier C, and 48 score zero outright. Median criticality across the scored set is 10.7.

5. Findings by axis

5.1 Independent routes agree only within their evidence family

The five routes were built to be independent in construction, and they turn out to be substantially independent in result. The two paper-level textual routes (an instrument named in a paper; the same restricted to papers that also use a model) correlate at ρ = 0.79. The two structural routes (DOI-matched NASA usage records; NASA-side modelling-title publications) correlate at ρ = 0.74. Across families, agreement spans only ρ = 0.32–0.55.

This is the methodological result of the study. Two routes that both claim to measure "climate modelling depends on this instrument" rank the catalogue differently enough that picking one would produce a materially different priority list. Only 24 instruments are visible on all five; 49 are visible on none and 63 on exactly one, so 112 of 243 — 46% of the catalogue — rest on evidence too thin to act on.

The matrix also refuses the tidy story its own panel boxes suggest. R1b — a textual route — sits closer to the structural family (ρ = 0.53 with R3, 0.55 with R4) than to its textual siblings (ρ = 0.3 with R1). Naming a platform behaves more like citing a dataset than like naming an instrument, which is what one would expect if authors reach for the mission name ("Aqua", "Terra") when describing data provenance and for the instrument name when describing a retrieval. The families are a useful device for reading the figure, not a property of the data, and the honest summary is that all five routes are only loosely concordant.

5.2 The ranking, and what it rests on

The head of the ranking is dominated by broad-swath imagers and sounders with long records — MODIS, AMSR-E, MISR, MOPITT, AMSU-A, VIIRS, AIRS, AVHRR, CrIS — which is unsurprising and is partly a reassurance that the score is measuring something real. The informative entries are the ones that do not fit that pattern: SSMIS at 8th on 170 datasets, GOME-2 at 35th on 72, WINDSAT at 38th on 21. Those are instruments whose standing comes from corroboration rather than from volume, and §8 Claim 2 confirms the passive-microwave cluster independently.

5.3 The asymmetry: strong at population level, broken at the top

The population-level correlation is high and the fractional-attribution version is barely lower (ρ = 0.656), so the correlation is not an artefact of co-flight inheritance. Total publication count correlates even more strongly (ρ = 0.863), which is close to tautological and is reported for that reason. The instinct that big archives and critical instruments are different populations is, at this scale, wrong.

What survives, and matters more for a review board, is the local breakdown. Only 6 of the top ten and 19 of the top 25 are shared between the two rankings. PALSAR (+133 places), WINDSAT (+127), ACE-FTS (+120), GLAS (+107), SRTM (+100), GFO Altimeter (+97) and GOME (+94) are far more depended-upon than their volume suggests; AQUARIUS_RADIOMETER and AQUARIUS_SCATTEROMETER (-150), the three PACE instruments (−117) and CERES-FM5 (−92) run the other way. A decision rule based on archive size would protect the wrong half of that list.

5.4 Substitutability is concentrated — and mostly invisible to the join

climatemodelskg carries 3,144 Variable nodes resolving to 237 distinct measured variable names and 2,947 model-produced ones; 184 variable names are both measured by an instrument and produced by a model component. 90 of those (48.9%) are measured by exactly one distinct instrument name in the extracted literature.

Three results sit in this panel, and the second and third matter more than the first.

First, the shape is extreme but the counts are small. Under the strict GCMD-label join that feeds the score, only 3 instruments hold any sole-measured variable at all — MODIS (4: albsn, nppLut, tsSprd, vegFrac), GOME-2 (1: total_solar_irradiance) and GEDI (1: TOTVEGC_ABG) — 6 variables in total. That is the whole irreplaceability signal the analysis is entitled to score.

Second, alias fragmentation hides most of it. Resolving the free-text mention names to instrument families lifts the count to 32 variables across 15 families — and the joint largest is CERES, with 7 sole-measured variables and a strict-join score of zero, because papers write "CERES" or "clouds and earth's radiant energy system" while the catalogue writes CERES-FM1…CERES SCANNER. MODIS gains three more under its spelled-out name. TMI gains 3, SeaWiFS and AVHRR and CloudSat-CPR 2 each. The alias map is a declared, hand-built supplement and deliberately does not feed the score; its purpose is to quantify how much the strict join misses, which is roughly a factor of five.

Third — and this is the most consequential number in the section — most irreplaceable measurement in this corpus is invisible to a spaceborne, GCMD-labelled analysis. 58 of the 90 sole-measured variables (64%) resolve to no scored instrument. The large majority of those are genuinely non-satellite: automatic weather stations, Argo floats, eddy-covariance towers, radiosondes, ceilometers, Winkler titration, the Mauna Loa in-situ record. A minority are spaceborne but still unscorable — generic category names (satellite instruments, passive microwave satellite sensors, satellite altimeter), the two generic GCMD class labels excluded in §3 (SAR → ws60m, SCATTEROMETERS → WIND), merged products rather than instruments (ISCCP, HadCRUT5, COBE-SST2), a simulator (modis simulator), and two airborne surveys. Either way the conclusion holds: any inference about "irreplaceable observing infrastructure" that stops at named spaceborne instruments is missing most of the sole-source measurements the modelling literature actually depends on.

The attempt to measure substitutability structurally rather than textually failed, and the failure is itself a finding. GCMD science keywords should support it: they are a controlled vocabulary describing what a dataset measures. But only 122 of 1,609 keywords — 7.6% of the vocabulary — are attached to any dataset at all, and at that granularity only five keywords have five or fewer spaceborne instruments, three of them with exactly one (BATHYMETRY/SEAFLOOR TOPOGRAPHY → ATLAS; TERRESTRIAL ECOSYSTEMS → the generic SAR class; WATER QUALITY → DDMI). A vocabulary that coarse cannot distinguish substitutable from irreplaceable, so the structural substitutability test is reported as run and uninformative, not omitted.

6. Domain analyses

6.1 Three kinds of risk, kept apart

Collapsing these into one ranking is exactly the error the analysis is meant to avoid, so they are defined separately and allowed to overlap.

Class A (58 instruments) is the conventional answer and the least interesting: MODIS, AMSR-E, MISR, MOPITT, AMSU-A, VIIRS, AIRS, SSMIS, AVHRR, ASTER, CrIS, OMI, CERES SCANNER, TOVS, CERES-FM1/FM2, MLS, TES, SCIAMACHY, SMMR and 38 others. These are protected by their own visibility — a review board is unlikely to retire MODIS unaware that modelling depends on it.

Class B is the class the question was really about, and it has 2 members. Only ATLAS (the only spaceborne source of bathymetry/seafloor topography in the catalogue) and DDMI (the only source of the water-quality keyword) satisfy "irreplaceable but not broadly relied on". GEDI and GOME-2, the other two strict sole-measurers, score highly enough on the dependency routes to fall outside the class. Two instruments is not a credible estimate of how much narrow-but-irreplaceable infrastructure exists; it is a measure of how little of that structure the federation encodes. §5.4 gives the reason — the class is thin because substitutability is almost entirely unrepresented, not because the risk is rare, and the alias-resolved supplement suggests the true figure is several times larger.

Class C (9 instruments) — HARP2, OCI, SPEXone, TIRS-PREFIRE, TEMPO, TMS, TMWS, DDMI and SIRS — reads at first as pure waste: substantial archives nobody models with. It is not. Eight of the nine have their most recent dataset start in 2021–2025, and PACE's three instruments (OCI, HARP2, SPEXone) only reached stable calibration in 2024–25 (§8 Claim 6). Class C is dominated by publication latency, and for seven members the correct interpretation is "not yet evaluated". Two are different cases, and the distinction turns on first-light rather than latest dataset start: SIRS, a 1960s Nimbus instrument, is the member for which "no modelling uptake" is a settled fact; and DDMI (CYGNSS), whose data reach back to 2017 and whose datasets carry 120 publications overall, is a genuine uptake gap rather than a latency artefact — it is also the one instrument that satisfies both class B and class C.

6.2 The boundary-spanning community, and where the literature looks

Who. The people who work across both sides are themselves infrastructure, and the federation lets that be counted without relying on name matching. The 651 papers that appear in both graphs by DOI (34.1% of climatemodelskg's DOI-bearing papers) carry 4,397 distinct author names, 3,169 of them resolving to a distinct ORCID. That is the identifier-anchored boundary-spanning cohort, and it is the more defensible of the two available measures. The cohort's size is the finding: a few thousand people, spread across 121 countries, connect the world's climate models to the observations they are checked against.

The weaker measure is the name join: 8,391 of climatemodelskg's 10,029 distinct author names (83.7%) also appear as a nasa-gesdisc-kg author label. That figure must not be read as overlap — NASA's author pool is 905,086 nodes, large enough that common names collide by chance, and the federation's own crosswalk documentation records that 11.3% of matched names resolve to more than one ORCID. It is reported only to show that the DOI-anchored number is the conservative one.

The cohort is heavily concentrated: the top five countries (US, GB, DE, FR, CN) hold 53.5% of those author–country records, and 16 of 121 countries hold 80%. Whatever resilience this community provides is not evenly distributed, and a funding change in one or two countries would move a large fraction of it.

Where. The same corpus resolves study locations to GeoNames, giving 159 named regions with coordinates and 215 countries.

Research attention is oceanic and polar before it is territorial: Southern Ocean (231 papers), Pacific Ocean (219), Arctic (170), Mediterranean Sea (141) and Atlantic Ocean (117) lead, with the Sahel (45), Sahara (49), Middle East (50) and South Eastern Asia (50) forming a second tier. At country level China leads with 540 papers and the top ten countries account for 37.6% of all country mentions.

The thin-evidence finding is the important half. 92 of 215 countries are named by ten or fewer papers in the entire corpus — including most of Central and West Africa, the Pacific island states, and much of Central America and the Caribbean. These are places where climate projections carry high decision stakes and where the modelling literature that would justify any particular observing system is close to absent. Note the confound: a PAPER_MENTIONS Country edge records a mention, not a study focus, so these counts are an upper bound on attention, which makes the sparse tail worse rather than better.

Analysis families run and skipped. Run: all five dependency routes (R1, R1b, R2, R3, R4); the capability/substitutability axis (R5) in three forms — strict GCMD join, alias-resolved supplement, and the structural GCMD-keyword test; the footprint asymmetry test on both raw and fractional attribution; the community cohort by both DOI anchor and name join; geography at region and country level. Skipped, with reasons: temporal-coverage analysis of instrument lifetimes — schema:startDate and endDate exist and are populated on 7,959 and 4,837 datasets, but they describe a dataset's temporal coverage, and platform-mediated attribution pollutes them badly (MODIS inherits a 1950 start from a Terra-hosted reanalysis product), so no instrument-lifetime measure can be derived; data volume in bytes — no such field exists anywhere in the graph; mission status, launch or decommission date, and successor-instrument relations — none of these are represented, which is why this study cannot answer "when" for any instrument, only "how much would be lost"; citation-graph centrality of instruments — CITES exists between publications but resolving it to instruments would compound the platform-attribution error across two hops.

7. Discussion

Three things follow from the analysis, in decreasing order of confidence.

First, and most solid: the evidence route determines the answer, so no single route should be trusted. Cross-family agreement spans only ρ = 0.32–0.55 (§5.1), and the one route that bridges the two families is R1b — a platform-level textual route, not an instrument-level one. A review that asked "which instruments do modelling papers cite?" and a review that asked "which instruments' data do NASA's records show modelling papers using?" would produce materially different priority lists, and neither would be wrong. The practical recommendation is procedural: rank by corroboration count first and by score second. On that basis the 24 five-route instruments are the defensible core, and the 112 tier-C instruments should be treated as unassessed rather than as low priority.

Second: the asymmetry is real but local. Population-level correlation between volume and criticality is high (§5.3), so "big archive" is a reasonable prior. It fails precisely at the top of the distribution and for specific instruments, which is where retirement decisions are actually made. The operational form of this finding is the rank gap, not the correlation: an instrument whose criticality rank is 100+ places ahead of its volume rank (PALSAR, WINDSAT, ACE-FTS, GLAS, SRTM) is one whose loss would be under-weighted by any volume-based triage. §8 Claim 4 confirms this independently for ACE-FTS — the published literature already describes an "imminent data desert" for exactly the Aura/SCISAT limb-sounding group this analysis places in tier A, with ACE-FTS itself the one member whose criticality far outruns its volume.

Third, and most tentative: the classes that matter most are the ones the data can barely see. Class B contains 2 instruments not because narrow-but-irreplaceable infrastructure is rare, but because the federation encodes almost nothing about substitutability — no ECV mapping, no successor relations, a science-keyword vocabulary 92% of which is unused, and an instrument vocabulary so alias-fragmented that the largest single holder of irreplaceable measurement capability in the corpus (CERES, 7 sole-measured variables) scores zero on the axis that is supposed to detect it. The same gap explains the study's clearest failure: GRACE and GRACE-FO, whose eleven-month gap generated an entire reconstruction literature, rank in the bottom half here (§8 Claim 7). That is not a marginal miss; it is a demonstration that a low score in this framework means "not visible in this evidence base", full stop.

Testable predictions. (1) If the SSMIS→AMSR2 transition degrades sea-ice concentration consistency, models evaluated against sic and siextentn will show discontinuity at the 2025 handover — the corpus records SMMR, SSMIS and AMSR-E among the few instruments measuring those variables at all. (2) The PACE instruments should move from class C to class A within roughly three years of their 2024 data release; if they do not, that is a genuine uptake failure rather than latency. (3) An expert-elicited criticality ranking would agree with this one at the head (MODIS, AMSR-E, VIIRS, AIRS, CERES) and disagree in the tail, with GRACE and CERES-FM5 moving up sharply.

What would fix the analysis. Four fields would change what is answerable: a Dataset→Instrument edge (removing the co-flight attribution error entirely), an ECV or measurement-type mapping per instrument (making substitutability computable), mission status with launch/decommission dates (making "when" answerable at all), and an instrument-alias table linking free-text mention strings to GCMD labels (removing the CERES-class blind spot). None requires new science — all four exist in NASA CMR, the GCMD keyword service and the CEOS/CGMS ECV Inventory, and are simply not carried into this graph.

8. Comparison with prior work

Claims were checked against the primary Earth-observation and climate literature — agency documentation (NASA Earthdata, NSIDC, NOAA NESDIS, GCOS/WMO, ESA CCI), peer-reviewed journals (BAMS, Geoscientific Model Development, Remote Sensing of Environment, Atmospheric Measurement Techniques) and the National Academies Decadal Survey, retrieved by web search. The per-claim record with full citations is in Instrument-Criticality_literature_comparison.md.

Full-text verification. No claim above rests on a full-text read. All are abstract-, documentation- or landing-page-level checks, and are labelled accordingly rather than marked verified. The PubMed and Paperclip connectors were available but are not evidence sources here: PubMed indexes only biomedical and life-sciences literature and returns essentially no coverage of satellite Earth observation, climate-model evaluation or mission continuity. That is a tool-scope limitation, recorded rather than presented as a null result.

Where the KG evidence diverges from the literature. Three divergences are errors in the graphs: CERES flight-model labels that no paper uses, which zero out both its mention route and its irreplaceability score (Claim 3); GRACE payload fragmentation across engineering labels combined with GES DISC's partial gravimetry coverage (Claim 7); and a science-keyword vocabulary 92.4% of which is attached to no dataset (Claim 9). Two are differences of scope rather than error: reanalysis dominance reflects genuine community practice (Claim 5), and class C measures the corpus's time window rather than the instruments (Claim 6).

9. Full ranked results

The complete table of 243 scored science instruments — with all five route counts, both footprint measures, the rank gap, tier, risk class, and the sole-measured variable list — is in Instrument-Criticality_results.xlsx, sheet Ranked Results, and as data/instrument_criticality.csv. The workbook's other sheets carry the full 288-label catalogue inventory, the three risk classes, the route-agreement matrix, the sole-measured variables and their alias resolution, the study regions and country mentions, the boundary cohort, every verified quantity, and a Methods & Rules sheet with the complete rule set and abbreviations.

Tip: click a column header to sort, type in the box to filter, and use the drop-downs to select a tier, risk class or corroboration level. The sources (n) column counts the federation KGs behind each row — nasa-gesdisc-kg supplies the catalogue, footprint and the structural routes; climatemodelskg supplies the modelling literature, the textual routes and the capability axis.

A representative slice of the head and the divergent tail:

The rank column is sequential; the rank gap uses tie-aware ranks, so for the 48 instruments tied at zero criticality (HARP2 among them) the two are not arithmetically consistent.

The ranking's centre of gravity is where it should be, and its interest is in the divergent rows. A 2,070-dataset instrument at rank 13 with only 3/5 routes (CERES SCANNER), a 13-dataset instrument at rank 59 with 4/5 (PALSAR), a sole-source instrument at rank 78 with a single route (ATLAS), and a 36-dataset instrument at rank 143 that the literature treats as a landmark continuity failure (GRACE's KBR) between them show why the score should be read together with its corroboration count and its known coverage gaps, never alone.

10. Summary of findings & limitations

Findings. The OKN federation describes 921 instruments on 455 platforms, of which 288 labels are spaceborne and 243 are genuine science instruments. Dependency on climate modelling was established along five independent routes whose cross-family agreement spans only ρ = 0.32–0.55, so corroboration across routes — not any single route — carries the claim. 24 instruments are visible on all five routes; 49 on none.

The ranking is led by MODIS (100.0/100), AMSR-E, MISR, MOPITT, AMSU-A, VIIRS, AIRS, SSMIS, AVHRR and ASTER. Three risk classes were kept separate: 58 instruments broadly relied on, 2 narrow-but-irreplaceable (ATLAS, DDMI), and 9 with a large footprint and no modelling uptake — the latter dominated by instruments whose data begin in 2021–2025, where the absence of uptake is latency rather than redundancy. 90 of 184 model-relevant variables have a single measurer, but only 6 of those attach to a GCMD-labelled spaceborne instrument under the strict join; alias resolution lifts it to 32 across 15 instrument families, and 58 of the 90 are measured by in-situ instruments with no spacecraft at all.

On the asymmetry: at population level footprint and criticality correlate strongly (ρ = 0.727), so the general claim is not supported. It holds locally and sharply — 6/10 top-ten overlap, rank gaps to +133/−150 — with PALSAR, WINDSAT, ACE-FTS, GLAS and SRTM more critical than their volume implies. The boundary-spanning community numbers 3,169 ORCID-identified researchers across 121 countries, 53.5% of the author–country records in five countries, and 92 of 215 countries are named by ten or fewer papers in the entire modelling corpus.

Limitations.

1. No Dataset→Instrument edge. Datasets attach to platforms only, so every instrument on a

platform inherits all of that platform's datasets. All raw footprint figures over-count for co-flying instruments; the fractional measure mitigates but does not fix this, since it assumes equal division. This is the single largest structural weakness in the analysis.

1. No data volume in bytes, anywhere in the graph. "Data footprint" is a dataset count, which

is a poor proxy — a 25-year daily global L3 product and a single-campaign file count the same. The asymmetry question as posed ("largest data volumes") therefore cannot be answered directly.

1. No mission status, launch date, decommission date, or successor relation. The question that

motivates the study — when would this go dark, and what replaces it — is unanswerable from this federation. Dataset temporal coverage exists but describes products, not instruments, and is polluted by co-flight attribution (MODIS shows a 1950 earliest start).

1. climatemodelskg's instrument vocabulary is NLP-extracted and alias-fragmented.

1,490 instrument nodes include "MODIS", "moderate resolution imaging spectroradiometer (modis)", "Aqua MODIS", "Terra-MODIS" and "MODIS simulator" as separate entities. Only 115 match a GCMD label at all. §5.4 quantifies the cost: alias resolution multiplies the visible irreplaceability signal roughly fivefold, and MODIS's apparent dominance is inflated by having the most aliases.

1. The literature routes miss instruments the literature names differently from the catalogue.

CERES is the demonstrated case: zero mentions and zero irreplaceability across all its GCMD flight-model labels despite 210 NASA-side modelling citations and 7 alias-resolved sole-measured variables (§8 Claim 3). Any instrument whose common name differs from its GCMD label is systematically under-scored.

1. Coverage is GES DISC-centred. nasa-gesdisc-kg is built around NASA's GES DISC holdings.

Instruments archived primarily elsewhere are under-represented — GRACE gravimetry is the clearest casualty (§8 Claim 7), and non-NASA missions generally are visible only where GES DISC holds derived products.

1. The GCMD science-keyword vocabulary is 92.4% unused (122 of

1,609 attached to any dataset), so the structural substitutability test is uninformative and class B is almost certainly a severe undercount.

1. Most sole-source measurement in this corpus is not spaceborne. 58 of the 90 sole-measured

model-relevant variables are measured by weather stations, Argo floats, flux towers, radiosondes and similar. A spaceborne-scoped study structurally cannot see them, so nothing here should be read as a ranking of all irreplaceable observing infrastructure.

1. Model evaluation runs mostly against reanalyses. The dominant observational references in

climatemodelskg are ERA5, CRU, GPCC and similar. Because reanalyses assimilate satellite radiances, an instrument can be load-bearing without appearing in any evaluation's dataset list. This indirect dependency is named but not measured, and it biases the analysis against sounders whose contribution flows through assimilation.

1. The scoring weights (0.55 / 0.30 / 0.15) are a judgement, not a fit. No ground truth exists

to calibrate against. The rank ordering is fairly robust to reweighting; the absolute scores are not, and they should not be compared across any future version of this analysis.

1. The author-name join is not an identity join. The 8,391-name overlap is

reported for context only; 11.3% of matched names resolve to more than one ORCID, with severe fan-out for common Chinese-origin names. Only the DOI-anchored cohort (3,169 ORCIDs) supports person-level claims, and even that is limited to authors with an ORCID on record (60% of NASA-side authors). The per-country figures are author–country records, not people.

1. PAPER_MENTIONS Country records a mention, not a study focus. Geographic counts are an upper

bound on research attention, which strengthens rather than weakens the thin-evidence finding but means no country count should be read as a study count.

1. 32.0% of datasets have any recorded publication use

(2,581 of 8,058). The citation graph is built by crawling, so absence of a link is weak evidence of absence of use, particularly for recent and for non-NASA-archived data.

1. **The 2.7% figure for zero-signal instruments is measured against the

co-flight-inflated attribution total**, not against the 4,931 distinct spaceborne datasets. Against the latter the same instruments account for roughly 18%. The inflated denominator is used consistently across the footprint analysis, but the two are not interchangeable.

1. Only NASA-catalogued infrastructure is in scope. ESA, EUMETSAT, JAXA, NOAA and commercial

instruments appear only where GES DISC holds their data. Nothing here should be read as a complete picture of global observing-system dependence.

11. Reproducibility

Everything needed to replicate this analysis — the originating prompt verbatim, the full replicator specification (scoping rules, attribution rule, route definitions, thresholds, join recipes, scoring formulas, verified quantities and limitations), every supporting SPARQL query verbatim with its row count, the pinned KG versions and the timing — is in Instrument-Criticality_reproducibility.md, with the scripts in scripts/ and the intermediate extracts in data/.

12. References

Retrieved by web search against the primary Earth-observation and climate literature and agency documentation. The PubMed and Paperclip connectors were available but are not evidence sources here — PubMed indexes only biomedical and life-sciences literature and does not cover this domain. Full per-claim citations are in Instrument-Criticality_literature_comparison.md.

1. NASA LAADS DAAC / Earthdata. MODIS to VIIRS Transition — mission end dates, orbital drift, GCOS climate-data-record continuity requirement. 2025. https://ladsweb.modaps.eosdis.nasa.gov/learn/modis-to-viirs-transition/
2. Calibration of the SNPP and NOAA-20 VIIRS sensors for continuity of the MODIS climate data records. Remote Sensing of Environment. 2023. doi:10.1016/j.rse.2023.113716
3. Continuity between NASA MODIS Collection 6.1 and VIIRS Collection 2 land products. Remote Sensing of Environment. 2024. https://www.sciencedirect.com/science/article/pii/S0034425723005151
4. National Snow and Ice Data Center. SSMIS sunsets, AMSR2 rises. NSIDC Sea Ice Today. 2025. https://nsidc.org/sea-ice-today/analyses/ssmis-sunsets-amsr2-rises
5. National Snow and Ice Data Center. SMMR and SSM/I-SSMIS and AMSR2. https://nsidc.org/data/smmr_ssmi
6. NOAA/NSIDC. Climate Data Record of Passive Microwave Sea Ice Concentration, Version 6. https://nsidc.org/data/g02202/versions/6
7. Ageing Satellites Put Crucial Sea Ice Climate Record at Risk. Scientific American. https://www.scientificamerican.com/article/ageing-satellites-put-crucial-sea-ice-climate-record-at-risk/
8. Decades of science results and new technologies related to measurements of Earth's Radiation Budget from space and a pathway for continuity of observations. Science of Remote Sensing. 2026. https://www.sciencedirect.com/science/article/pii/S2950630126000086
9. Loeb N, et al. Risk and Impact of a Data Gap in the Earth Radiation Budget Satellite Record. AGU. 2023. NASA NTRS. https://ntrs.nasa.gov/api/citations/20230017173/downloads/LOEB_AGU_2023.pdf
10. NOAA NESDIS. Libera — CERES follow-on mission. https://www.nesdis.noaa.gov/our-satellites/currently-flying/joint-polar-satellite-system/libera
11. The Imminent Data Desert: The Future of Stratospheric Monitoring in a Rapidly Changing World. Bulletin of the American Meteorological Society 106(3). 2025. https://journals.ametsoc.org/view/journals/bams/106/3/BAMS-D-23-0281.1.xml
12. UNEP Ozone Secretariat. The Future of Stratospheric Monitoring in a Rapidly Changing World. 2025. https://ozone.unep.org/sites/default/files/2025-04/The%20Future%20of%20Stratospheric%20Monitoring%20in%20a%20Rapidly%20Changing%20World.pdf
13. Livesey N, et al. The Continuity Microwave Limb Sounder (C-MLS). AGU Fall Meeting. 2022. https://ui.adsabs.harvard.edu/abs/2022AGUFM.A52Q1224L/abstract
14. Continuing the MLS water vapor record with OMPS LP using neural networks. Atmospheric Measurement Techniques 19. 2026. https://amt.copernicus.org/articles/19/3601/2026/
15. Eyring V, et al. Earth System Model Evaluation Tool (ESMValTool) v2.0 — an extended set of large-scale diagnostics. Geoscientific Model Development 13. 2020. https://gmd.copernicus.org/articles/13/3383/2020/
16. Evaluating simulated climate patterns from the CMIP archives using satellite and reanalysis datasets (CMATv1). Geoscientific Model Development 13. 2020. https://gmd.copernicus.org/articles/13/3627/2020/
17. Waliser D, et al. Observations for Model Intercomparison Project (Obs4MIPs): status for CMIP6. Geoscientific Model Development 13. 2020. https://gmd.copernicus.org/articles/13/2945/2020/
18. Teixeira J, et al. Evolving Obs4MIPs to Support Phase 6 of the Coupled Model Intercomparison Project (CMIP6). Bulletin of the American Meteorological Society 96(8). 2015. https://journals.ametsoc.org/bams/article/96/8/ES131/69444/Evolving-Obs4MIPs-to-Support-Phase-6-of-the
19. ESA eoPortal. PACE (Plankton, Aerosol, Cloud, ocean Ecosystem) Mission. https://www.eoportal.org/satellite-missions/pace-mission
20. NASA Earthdata. PACE HARP2, SPEXone, OCI products released. 2024. https://www.earthdata.nasa.gov/data/alerts-outages/pace-harp2-spexone-oci-products-released
21. NASA Earthdata. PACE OCI V3.1 Reprocessing Completed. 2025. https://www.earthdata.nasa.gov/data/alerts-outages/pace-oci-v3-1-reprocessing-completed
22. Bridging the gap between GRACE and GRACE-FO using a hydrological model. Science of the Total Environment. 2022. https://www.sciencedirect.com/science/article/abs/pii/S0048969722007513
23. Bridging the gap between GRACE and GRACE-FO missions with deep learning aided water storage simulations. Science of the Total Environment. 2022. https://www.sciencedirect.com/science/article/abs/pii/S0048969722017946
24. Improving prediction of terrestrial water storage anomalies during the GRACE and GRACE-FO gap with Bayesian convolutional neural networks. arXiv (preprint — not peer-reviewed). 2021. https://arxiv.org/pdf/2101.09361
25. Yang X, et al. A Two-Step Linear Model to Fill the Data Gap Between GRACE and GRACE-FO Terrestrial Water Storage Anomalies. Water Resources Research 59. 2023. doi:10.1029/2022WR034139
26. Bridging Terrestrial Water Storage Anomaly During GRACE/GRACE-FO Gap Using SSA Method: A Case Study in China. Sensors. 2019. PMC6806599. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6806599/
27. Multidecadal reconstruction of terrestrial water storage changes by combining pre-GRACE satellite observations and climate data. Earth System Science Data 18. 2026. https://essd.copernicus.org/articles/18/1747/2026/
28. Joint CEOS/CGMS Working Group on Climate. ECV Inventory. https://climatemonitoring.info/ecvinventory/
29. GCOS / WMO. About Essential Climate Variables. https://gcos.wmo.int/site/global-climate-observing-system-gcos/essential-climate-variables/about-essential-climate-variables
30. ESA Climate Change Initiative. What is an Essential Climate Variable? https://climate.esa.int/en/about-us-new/climate-change-initiative/what-are-ecvs/
31. On the Determination of GCOS ECV Product Requirements for Climate Applications. Bulletin of the American Meteorological Society 106(5). 2025. https://journals.ametsoc.org/view/journals/bams/106/5/BAMS-D-24-0123.1.xml
32. Observational Data for Next-Generation Climate Model Evaluation. Bulletin of the American Meteorological Society 107(4). 2026. https://journals.ametsoc.org/view/journals/bams/107/4/BAMS-D-25-0079.1.pdf
33. National Academies of Sciences, Engineering, and Medicine. Thriving on Our Changing Planet: A Decadal Strategy for Earth Observation from Space. National Academies Press. 2018. https://www.nationalacademies.org/read/24938/chapter/13
34. Proto-OKN federated SPARQL endpoint (nasa-gesdisc-kg v0.0.6, climatemodelskg v0.0.15), queried via the mcp-okn MCP server. Query log and versions in Instrument-Criticality_reproducibility.md.