# Combining datasets in the earth catalog *An analysis of which of the 241 cataloged datasets measure the same thing, which can be aggregated, and where combination is scientifically meaningful — including the specific question of merging land and ocean temperature. July 2026.* --- ## 1. The short answer Yes — large parts of the catalog are families of datasets measuring the same physical quantity by different methods, and combining them is not just possible but standard scientific practice. In fact, several catalog entries *are already combinations* of other catalog entries (GPCP merges gauge and satellite precipitation; IMBIE reconciles three independent ice-sheet mass-balance methods; HadCRUT5 merges land stations with ship/buoy SST). The productive way to think about the catalog is therefore in **families** (many estimates of one quantity), **complements** (different quantities that constrain each other through a physical budget), and **already-merged products** (where the combination work has been done for you, with uncertainty quantified). The single most important caution: **apparent redundancy is not independence.** Most datasets in a family share raw inputs — the same ships, buoys, satellites and stations — so their errors are correlated. Averaging five SST products does not reduce error by √5; their spread is best read as a *lower bound* on structural uncertainty. ## 2. Ocean temperature — the deepest family in the catalog The catalog contains at least nine distinct ocean-temperature datasets, and they form a coherent hierarchy rather than mere duplication. **Surface (SST).** Four gridded analyses overlap in purpose but differ in resolution, record length and method: MUR (1 km daily, 2002–, the layer on the globe), OSTIA (0.05° daily, 1982–), NOAA OISST v2.1 (0.25° daily, 1981–), and the two century-scale reconstructions HadISST (1°, 1870–) and ERSSTv5 (2°, 1854–). ESA's SST CCI adds a climate-quality satellite-only record. These divide naturally by role: **use MUR/OSTIA for spatial detail, OISST for a robust daily workhorse, HadISST/ERSST for trends and long context.** They are calibrated against overlapping in-situ networks (ICOADS ships, drifters, Argo), so treat their agreement as consistency, not confirmation. **Subsurface.** Argo profiles (plus Deep Argo) are the raw truth; EN4 and IAP grid those profiles into monthly T/S fields; the World Ocean Database holds the full historical archive; and the reanalyses (GLORYS12, ORAS5, ECCO, SODA) assimilate all of the above into dynamically consistent states. This is a *vertical* combination chain: profiles → objective analyses → reanalyses, each stage adding coverage and physics at the cost of model dependence. **Heat content.** NCEI/Levitus, IAP/Cheng and EN4-derived OHC are three estimates of the same integral. Their spread (visible in every IPCC figure) is the community's de-facto uncertainty band, and using all three is exactly how the annual "Indicators of Global Climate Change" product does it. **Sensible combinations:** an SST *ensemble layer* (mean and spread of OISST/OSTIA/MUR on a common grid); long-record trend maps from HadISST/ERSST cross-checked against the satellite era; OHC ensembles for the energy side of AMOC analysis; and validation of any gridded product against raw Argo profiles, which are as close to independent ground truth as the ocean offers. ## 3. Land + ocean temperature — sensible, with one physical subtlety Combining land and ocean temperature is not only sensible — it is precisely how the world's headline climate records are built. HadCRUT5, GISTEMP, NOAAGlobalTemp and Berkeley Earth are all *blends*: land air temperature from station networks (GHCN and relatives) merged with SST from ship/buoy archives (ERSST or HadSST). The catalog therefore already contains four independent-ish implementations of the land+ocean merge, and their mutual spread is the standard uncertainty statement for global mean surface temperature (GMST). The subtlety worth understanding before building your own blend: **the two halves are different physical quantities.** Over land these products use air temperature at 2 m; over ocean they use *water* temperature (SST), because SST anomalies are a good and much better-sampled proxy for marine air temperature anomalies. Three practical consequences: 1. **Blend anomalies, never absolutes.** A 15 °C SST and a 15 °C land air temperature are not the same thing, but their *departures from a common 1961–1990 (or similar) baseline* are comparable and can share a grid. Every serious product works in anomaly space. 2. **Sea-ice edges are the awkward seam.** Where ice comes and goes, products must switch between SST and air temperature, and the choices differ between products — a real contributor to their spread in polar regions. 3. **Comparison with models needs care.** Climate models output true global 2 m air temperature; blended observational records run slightly cooler in trend. When the prediction side of this project compares observed GMST to CMIP6, use the models' "blended-masked" diagnostics or apply the known adjustment. For the globe, the practical recipe is: MUR/OISST as the ocean layer, ERA5-Land or Berkeley Earth as the land layer, displayed as anomalies against a shared baseline — or simply render an existing blended product (GISTEMP and Berkeley Earth ship gridded anomaly fields) rather than re-deriving one. ## 4. Other high-value combinations in the catalog **Precipitation** mirrors temperature: gauges (GPCC) are truth-but-sparse, satellites (IMERG, CMORPH, PERSIANN) are complete-but-biased, and the merged products (GPCP for climate quality, CHIRPS for high-resolution land) are the combinations. Building a new merge is rarely worthwhile; comparing the existing ones quantifies uncertainty. **Sea level is the showcase budget.** Altimetry (DUACS/AVISO, total sea level) should equal ocean mass change (GRACE/GRACE-FO) plus steric expansion (Argo/EN4/IAP), with tide gauges (PSMSL/UHSLC) anchoring the coast. Closing this three-way budget is a gold-standard consistency check, and its residuals are scientifically interesting in themselves. A "sea-level budget" dashboard would combine five catalog families in one panel. **The planetary energy budget** links CERES (top-of-atmosphere imbalance, ~1 W/m²) to ocean heat content change, which absorbs about 90 % of it. EEI from CERES and dOHC/dt from IAP/NCEI are independent measurements of nearly the same number — their agreement over multi-year windows is one of the strongest validation checks in climate science. **The carbon budget** is an accounting identity across the catalog's GHG section: emissions (Global Carbon Project, EDGAR, Climate TRACE) minus land and ocean sinks must equal the observed atmospheric growth rate (NOAA GML). The Global Carbon Project performs this reconciliation annually; Climate TRACE's facility-level totals can be validated against national inventories (UNFCCC, PRIMAP) the same way. **The cryosphere already demonstrates the method.** IMBIE exists because three independent techniques — altimetry (ICESat-2), gravimetry (GRACE) and the input–output method (velocity × thickness from ITS_LIVE/BedMachine minus SMB from RACMO/MAR) — gave divergent ice-sheet mass balances until they were formally reconciled. Any combination work in this project should aspire to IMBIE's pattern: independent methods, one reconciled series, explicit uncertainty. **The AMOC stack — most relevant to the prediction goal.** The catalog holds several *semi-independent* estimates of AMOC state: direct transports (RAPID, OSNAP, MOVE, SAMBA, Florida Current cable — unified by AMOCatlas), SST-fingerprint indices (computable from OISST, HadISST *and* ERSST — three versions of the same index whose spread matters), subsurface density/salinity signatures (EN4, Argo), altimetry-based proxies, and reanalysis-derived streamfunctions (GLORYS, ORAS5, ECCO). A combined "AMOC state vector" with these as parallel channels — rather than any single index — is exactly what a credible early-warning analysis should ingest, because the tipping-point statistics (variance, autocorrelation) are sensitive to dataset choice. **Reanalyses as an ensemble.** ERA5, MERRA-2 and JRA-3Q assimilate largely the same observations with different models; their spread is the honest uncertainty of "what the atmosphere did." For any derived index (jet position, blocking, surface winds over the subpolar gyre), computing it in at least two reanalyses is cheap insurance. **Risk stacks multiply rather than average.** Hazard (IBTrACS tracks, FIRMS fires, GloFAS floods, HadEX3 extremes) × exposure (WorldPop, GHSL) × vulnerability (ND-GAIN) is a combination *across* families that yields impact layers — a different logic from aggregating estimates of one quantity, and well suited to the globe. ## 5. Rules of engagement for any combination Work in anomaly space with a common baseline period; regrid conservatively to the coarsest common resolution rather than inventing detail; never treat products sharing inputs as independent; keep reanalyses and observation-only products distinct in your bookkeeping (reanalyses have *assimilated* the observations you might validate them against); check whether the quantities are truly the same (SST vs marine air temperature, skin vs foundation temperature, gauge point vs satellite areal average, calendar-day vs period-mean); and propagate ensemble spread through to every derived number, presenting it as a lower bound on uncertainty. ## 6. What this suggests for the project Three concrete, buildable steps fall out of this analysis. First, an **SST ensemble layer**: OISST + OSTIA + MUR on a common grid with a mean and a spread toggle — the spread map itself is a compelling visualization of where the ocean is well- vs poorly-observed. Second, the **sea-level budget panel**: altimetry vs GRACE + steric, the catalog's most elegant cross-family check. Third, the **multi-channel AMOC state vector** feeding the dashboard: RAPID/OSNAP transports, three SST-fingerprint variants, and an EN4 salinity index side by side, with their disagreement displayed honestly — the right substrate for the tipping-point statistics this project ultimately wants to compute. A supporting improvement to the catalog itself: add a `family` field (e.g. `sst`, `ocean-heat-content`, `gmst-blend`, `precipitation`, `sea-level`, `amoc-index`) so these relationships become machine-readable and the app can show "related datasets" on every entry. --- ## Appendix: why per-pixel differences work for SST but not precipitation The globe's "computed difference" tool inverts a layer's GIBS colormap back to physical units per pixel and subtracts two dates. This is meaningful only for a **smooth, continuous, always-present field** — sea surface temperature is the canonical example: every ocean pixel has a defined value that changes slowly, so "today minus a year ago" is a clean anomaly. Precipitation fails all three conditions. IMERG (the catalog's precipitation layer) is an **instantaneous rate in mm/hr** on a **logarithmic** colour scale, and it is **transparent below 0.1 mm/hr** — most of the map is "no rain right now." Differencing two daily snapshots therefore measures whether it happened to be raining at the satellite overpass, not a climate signal; the result is sparse, noisy salt-and-pepper. The physically correct comparison for precipitation is an **accumulation or mean rate over a window** (e.g. total mm over 30 days) vs the same window in another year — a different operation from per-pixel differencing, and one that needs the gauge-calibrated monthly products (GPCP, GPCC, CHIRPS) rather than instantaneous IMERG. The app reflects this: computed-difference and multi-day aggregation are offered for SST only, and a hint appears if another layer is active in difference mode. Extending honest aggregation to precipitation (accumulated totals from a monthly product) is a sensible future addition. ## Climatology grid layers (GPCP, E-OBS, OISST, MeteoSwiss) Some of the most useful precipitation and ocean products have no global tile service, so the app bakes them into static regular lon/lat grids (`data/.json`) and paints them client-side with a colour ramp. Because these are **long-term means**, they answer the question instantaneous IMERG cannot — "how wet is this place, normally?" — and sidestep the per-pixel differencing problem described above entirely. - **GPCP v2.3** — the gauge-satellite blended, global mean-annual precipitation climatology (2.5°). This is the honest "precipitation on the globe": every cell is a defined multi-decadal average, so the tropical rain band and the subtropical deserts read cleanly. IMERG remains the layer for *where it is raining now*; GPCP is *where it rains on average*. - **E-OBS v31** — the same quantity as GPCP but from dense European rain gauges at 0.25°, land only. It is a **regional** layer (a bounded rectangle over Europe); everything outside stays transparent. Use it to see the orographic detail — wet Alpine and Norwegian coasts, dry Iberian interior — that a 2.5° global product blurs away. - **MeteoSwiss OGD** — the Swiss 1991–2020 precipitation *normal* at ~2 km, the highest-resolution member of the same family, showing the sharp Alpine precipitation gradients within a single country. - **OISST v2.1** — a 1991–2020 SST climatology (0.25°→1°). It complements the live MUR SST tile layer: MUR shows today's field and supports the difference tool; OISST is the long-term baseline those anomalies are measured against. Three of these (GPCP↔E-OBS↔MeteoSwiss) measure **the same physical quantity** at coarsening spatial scales, so they are directly comparable — a nice illustration of how a global product, a regional gridded product, and a national high-resolution product nest together. ## Same quantity, different mission: chlorophyll and salinity Two of the requested ocean datasets have a natural on-globe counterpart already served as NASA GIBS tiles, so the app shows the tile version and treats the climate-record version as its scientific sibling: - **Chlorophyll-a** — the app's live layer is NASA Ocean Color (PACE/OCI). The **ESA OC-CCI** record measures the same variable but merges SeaWiFS, MODIS, MERIS, VIIRS and OLCI into a bias-corrected multi-decadal climate data record; it is the one to use for *trends*, where a single-sensor daily field is not. - **Sea surface salinity** — the app's live layer is SMAP L3 (monthly). **SMOS** (ESA, via CATDS/BEC) retrieves the same L-band surface salinity from a different platform; Aquarius covers the earlier 2011–2015 era. Cross-mission agreement between SMAP and SMOS is itself a validation signal. OC-CCI and SMOS are catalogued (`data/catalog.json`) with their access details; adding them as their own grid layers is a straightforward follow-up once an unauthenticated subset endpoint is wired into `scripts/refresh_data.py`.