Global InVEST

The GEP service library: state and open questions.

Motivation

Nature is invisible in the accounts

Ecosystems supply services the economy depends on, from pollination and soil retention to coastal protection and carbon storage, yet none of them appear in standard national accounts.
What is not measured goes unmanaged: the loss of these services surfaces only later, as lower yields, higher damages and higher prices.
The Gross Ecosystem Product (GEP) account values what ecosystems supply, country by country and service by service, so natural capital can be tracked alongside GDP.
The same service models also drive scenario analysis: how climate and land-use scenarios change the services economies depend on, feeding the macroeconomic pipelines.

Why one reviewed library

Different GEP service models were developed by different contributors in different repositories, with their own layout, path conventions and aggregation choices.
One home with one template makes every service runnable by anyone, verifiable by rerun, and consumable by the GEP account.

Why this matters outside the lab

Central banks and finance ministries are now asked to price nature risk, and the NGFS scenarios they use carry no ecosystem-service feedback at all — the pipeline this library feeds is the first to put one there.
A GEP account changes decisions only if every number can be traced to the model that produced it, so one reviewed library is what makes “verified by rerun” a property of the account rather than a claim about it.
The service models do double duty. The valuation answers what nature is worth in a place; the shock answers what happens to the economy when that changes, and the second is what policy instruments actually act on.
Where a service is not yet trustworthy we say so — livestock’s numbers are known wrong, erosion’s are provisional — because an account used in policy inherits every silent error inside it.

What global_invest is

One module per service: GEP always, a shock where the science allows

The library holds one folder per ecosystem service, all on the same file convention.
Ten of the eleven services carry their GEP valuation in the library; fisheries’ is still external.
Services whose science supports it also produce a scenario shock for the macroeconomic pipeline; not all can be made into shocks.
Where both outputs exist, they are written from the same science, so the valuation and the shock cannot drift apart.
The library is scenario agnostic. Consumers inject their naming through scenario maps and required configuration.

The mental model: a service module reads like a paper

File	In paper terms
functions.py	The equations — pure science; knows nothing about folders or projects.
tasks.py	The estimation steps — each caches its output and skips itself when already done.
initialize.py	The table of contents, plus the plug other pipelines graft in.
run_<service>.py	The main script — opens a workspace and goes. Thin on purpose.
method.qmd	The methods section — rendered from the code, so it cannot silently drift.

The three CSVs are the calibration tables. es_config says what each model reads and how it prices; es_scenarios says which scenarios and maps; es_parameters holds tuning knobs and machine locations.
One rule ties data to code: a CSV cell is a default, and whatever the caller already set wins.

The template every service follows

One runner per task tree, and every run file shares one anatomy, decided with Justin and applied across the library: build_task_tree names the pipeline, and the run is configured where it is constructed.
Configuration is data, in two shared CSVs. es_config.csv holds one row per service in the GEP formula’s own vocabulary (quantity, price, attribution, land base, base year, price convention) plus the crosswalk to the status sheet; es_scenarios_test.csv holds the standard SEALS scenarios and ships the matching test maps, so a standalone shock run works on any machine.
The library fills only what the caller left unset. A pipeline such as ngfs_pnas drives the same tasks from its own scenarios file, and its values always win.
Each task’s first line publishes the inputs it needs (relative to the shared base data; each machine finds its own copy), so any task grafted anywhere is a working piece — rolling out service by service, terrestrial first.
Aggregation is on r250, one row per country. The r264 crosswalk is used only for maps.

The configuration table itself (es_config.csv)

One row per service; the columns are the GEP formula’s roles — Q = quantity input, P = price input, λ = attribution input, land = the LULC base — plus the crosswalk to this group’s status sheet. An empty cell means the role does not apply.

service	sheet crosswalk	base year	price convention	Q	P	λ	land
terrestrial_carbon	global_climate_regulation / terrestrial	2019	rental scc r2%	carbon zones raster	prices xlsx		esa 2019
coastal_carbon	global_climate_regulation / coastal	2019	rental scc r2%		prices xlsx
pollination	pollination	2023		value raster
renewable_energy	wind + solar + geothermal			IRENA production	WB prices	CWON rents
extractive_materials	extractive_materials_provision			WB GDP		WB rents share
coastal_protection	coastal_protection			CWoN mangroves

Five more rows (crop and livestock as the sheet’s commercial subgroups, fish_provision, erosion_control, landslide) carry their crosswalk today and gain cells as their configuration moves in. Every row is under test the moment it is edited.

Where things stand

Services and totals (base year 2019, verified)

Service	Total	Verified against
terrestrial_carbon	$5.365T	the manuscript, to the dollar
coastal_carbon	$22.84bn EEZ-only, $102.18bn all rows	the author’s design, mangrove exact
crop_provision	≈ $678bn (thousand-USD units)	the July reference, to the float, twice
livestock_provision	quarantined	three owner-level defects
coastal_protection	$36.72bn	fresh baseline, rerun identical
extractive_materials	$99.54bn	the author’s reference, exact
renewable_energy	$170.83bn	rerun identical; drive references disagree
pollination	$18.28B raster side (2023)	FAO side $515B, definition open
erosion	$15.54bn PROVISIONAL	internal reconciliation; awaits submission
landslide_mitigation	input stage complete; full run resumes on the cluster	no machine-readable reference yet
fisheries	shock only	valuation external (CWoN)

Verification discipline, and what it caught

Run first, conform second, and verify every conform step against the cached run. An additive gate audits each branch against develop.
Write-time invariants check every shock table before it is written, a canonical-row rule guards every country aggregation, and every row of the configuration table is under test from the moment it is added.
Caught so far: a 23.5 percent double-count in the carbon GEP, a two-spelling baseline bug, and a stale-input trap in the erosion chaining.
It also caught one claim of our own: the coastal report’s numbers are the EEZ-only aggregation by design, matched exactly by our run.
What the next slides ask for is the layer replication cannot supply: it proves our code reproduces your results, not that the results are right. None of it is criticism — most questions exist because verification found exactly where a decision, a source, or an input is missing.

The questions, service by service

The carbon services

terrestrial_carbon: solid

State: GEP replicates the manuscript to the dollar, $5.365T at 2019. The strongest number in the library.
One correction to know about: a double-count in the CODE overstated its total by 23.5 percent (split countries entered once per sub-region — China six times). The published manuscript was right all along; the corrected aggregation reproduces it to the dollar, which is how the fix was validated.
No open questions. This service is the template the others are held to.

coastal_carbon: which side of the shoreline?

State: replicates the author’s design exactly, mangrove component to the cent. One run, two totals: $22.84bn counting ocean-side rows only, $102.18bn counting both sides.
Q1 (design, Yanxu): was EEZ-only chosen to avoid double-counting the land-side pixels with terrestrial carbon? The land side carries 97.5 percent of mangrove and 84 percent of salt marsh area, so the convention moves the total 4.5 times.
Q2 (inputs, Yanxu): which exact seagrass and salt marsh input versions produced the report’s $26.65bn? Our replication gives $22.84bn with mangrove matching exactly, and the report’s own seagrass area (62.21 Mha) exceeds the published range it cites (16 to 26.7 Mha).
Q3 (the room): If both sides should count, who owns the terrestrial-side mask that prevents double-counting?

Marine carbon: unbuilt and unowned

State: the service sheet lists Terrestrial, Coastal, and Marine carbon. We found no marine module in any repository and could not identify a lead — correct us if either exists.
Q1: is marine carbon in scope for the account at all, and on what horizon?
Q2: if yes, who owns it, and which data product drives it?
Q3: most open-ocean carbon sits in the high seas, inside no country’s EEZ. Which attribution convention should a per-country account use?
The sheet lists one more subgroup under the same service, equally without a module or lead that we could find: Other Non-Carbon GHG — the same three questions apply to it. Correct us if either exists.

The biophysical services

pollination: which definition is the GEP?

State: built on the template and running; aggregation verified. Two defensible definitions, 28 times apart.
Q1 (the room): is pollination’s GEP the service nature currently delivers ($18.28B, the sufficiency-weighted raster) or the production value at stake ($515B, dependence-weighted FAO value)? The country ranking is identical; only the definition differs.
Q2: the rasters on hand are 2023 and the account values at 2019. We regenerate at 2019 once the definition is settled. Does anyone object to that ordering?
A consistency note for the discussion: erosion’s GEP records the habitat-supplied share, so the raster-side definition is the internally consistent choice.

erosion: real inputs wanted

State: the full chain runs end to end and reconciles internally at $15.54bn, PROVISIONAL. Sections A and B verified structurally; the valuation runs on interim inputs we built ourselves.
Q1 (the author): Can you share the exact FAO production-value and price series the submitted paper used? Our chain runs on stand-in series we built ourselves; every number moves when yours replace them.
Q2 (the author): Can you share a per-country output table from the submission run? With it we verify by exact replication, the same bar extractive already passed.
Both asks may be one answer: the July thread says the full run lives at /projects/standard/jajohns/shared/sediment_gep on MSI — pointing us there, or copying it to the drive, settles both questions.

landslide_mitigation: a reference to verify against

State: folded from the v0.2.0 release, 50.5 GiB staged. The first run completed the input stage (all twelve global reprojections and the event panels); the compute-heavy remainder resumes on the cluster, where this run belongs. The only machine-readable output predates the method revision.
Q1 (Matt): Can you export one machine-readable per-country output from a v0.2.0 run — a CSV is enough? With it we run, diff, and the service reaches the verified bar.
Q2 (Matt): Which input versions did the release-day run use?

fisheries: one substitution to bless

State: shock-only in the library. The shock file derives from DBEM (Cheung 2021), built by Erwin from CWoN inputs; a fisheries GEP valuation exists outside the library in the group’s own R pipeline, computed from FAO FishStatJ.
Q1 (Erwin): Do you bless the RCP8.5-for-RCP7.0 substitution, or name a replacement? The 8.5 series stands in because no 7.0 series exists — defensible, but currently discoverable rather than stated.
Q2 (Erwin): Which CWoN vintage is behind the HAR file? For the record and the methods writeup.
Q3 (fisheries teams): Should the fisheries GEP come into the library — and from which source? Every service is expected to have a GEP eventually; fisheries’ valuation is still external.
Settled convention (Chiara, 20 Aug): fisheries is an ecosystem service and stays in the nature family — it enters the cc_es pass alongside carbon, pollination and erosion, and is not part of the just_cc runs, even though its shock varies only with the climate pathway. Stated here so the boundary property that makes it unusual (RCP-driven, no land-use map) reads as a characteristic, not an open question.
Found in the project’s own repos: a commercial fisheries GEP exists as an R pipeline (FishStatJ to GEP), and the subsistence repo already carries a per-country output table — so the in-library GEP has source material, and the subsistence scope question has an existing product to evaluate.

The accounting services

crop_provision: right numbers, unlabeled units

State: replicates its reference to the float, twice. The most verified accounting service.
Q1 (the room): Label crop’s thousand-USD column, or convert it to plain USD and regenerate the reference? The column is FAOSTAT-native thousand-USD — the global sum reads 678,413,996 and means $678bn, and nothing says so. Our recommendation is convert: every other service is plain USD, and a cross-service sum is a factor-1000 trap today.
Q2 (the room): Who can confirm whether FAOSTAT’s production values already include imputed subsistence output? If they do, the subsistence question becomes labeling rather than new modeling. We will check the coverage definition; we need the intended scope confirmed.

livestock_provision: three decisions before the numbers can be trusted

State: QUARANTINED numbers, but the owner’s method repo answers more than we hoped — and step 1 of his method is already implemented as a draft: with his 65 FAO item codes replacing the stale 29-name list, India recovers from zero to $136bn, China reads $838bn, and the world total $2.29T, all PROVISIONAL pending his lambda. The old name list matched only 13 of its own 29 items against current FAO files.
Q1 (owner): We found your meat/milk/other item grouping — is it current, and does milk belong as it suggests? The library’s 29-item meats-and-eggs list (which captures 4 percent of India’s FAO livestock value) appears to be a stale fork of your method.
Q2 (owner): Is the intended nature share the GLEAM-3 feed-share lambda — ecosystem feed over total feed, per country — as your lambda.py computes? If so, the answer is neither crop’s coefficients nor pasture rents, and we implement exactly that.
Q3 (owner): Can you share the GLEAM input (gleam3_dmi.xlsx) and one per-country output table? The GLEAM file lives on a personal disk and the repo commits no output; with both, we implement your method in the library and verify by exact replication.

coastal_protection: two candidate totals, one run to settle it

State: verified by identical rerun at $36.72bn (2019, mangrove and coral avoided damage).
Q1 (owner): Which data or code state produced the $73.0bn recorded in the code’s own comment? It is almost exactly twice our verified $36.72bn, and published anchors fit both. If that run’s inputs still exist, one rerun settles it.
Our leading suspect, for what it is worth: coral rows already at year 2019 double-counted by the 2012-to-2019 deflation step.

extractive_materials: what does 0.49 mean?

State: exact per-country replication of the reference, including the rendered tables.
Q1 (owner): What is the source or derivation of the 0.49 factor? The valuation multiplies mineral rents times GDP by it, unsourced in code, submission, and notes; without it the total would be $203bn instead of $99.54bn. The arithmetic is verified; the meaning of the number is not.

renewable_energy: code and references disagree

State: conformed and rerun-identical at $170.83bn (wind 110.31, solar 53.95, geothermal 6.56).
Q1 (owner): Which is blessed — the code’s $170.83bn or the drive CSVs’ $134.6bn? More than 100 countries differ per source, both cannot be right, and the folder’s “gepNEW” method note suggests the CSVs predate a revision.
Q2 (the room): Does the crosswalk satisfy the sheet’s wind/solar/geothermal split, or does anyone want separate modules? The module already computes all three and writes per-source tables.

Cross-cutting decisions

Four conventions the room can settle today

Maps: country values on split-country polygons render partially (China shows one sixth). We propose dissolving to one polygon per country, a one-time cartographic artifact that can never double-count. Any objection?
Base year: ten of eleven services value at 2019; pollination’s raster is 2023 and gets regenerated after its definition is settled. Confirm 2019 as the account’s base year.
Subsistence: crop, livestock, and fisheries carry subsistence subgroups in the service sheet. In scope or not, per service?
The universe crosswalk: the sheet lists roughly 26 services; the library covers ten. We propose a tested crosswalk table (sheet label to module to output) instead of restructuring modules to mirror the sheet. And the sheet’s label typos should be fixed before anything inherits them.

The PM team’s August 31 pass maps onto today

Colleen’s housekeeping email (due August 31) asks every group for consistent parameters and clean write-ups. Today’s answers complete most of it a week early.
2019 dollars: confirms our base year; pollination’s regeneration will follow Marta’s conversion guidance once its definition is settled.
CWON rental rates, not GTAP: renewable and fisheries already use CWON. For extractive, this is context for the 0.49 question — if that factor is a rental-rate choice, the CWON guidance may answer it.
EPA VSL: landslide values mortality, so one more ask for Matt — which VSL does v0.2.0 use?
Zeros vs NAs: our verification hit exactly this — extractive has 9 countries NA for missing WB data (documented), and coastal_protection currently fills unmatched countries with zero in a merge, which needs the true-zero-or-NA review the email asks for. We will do that pass on our modules.
Discount rate 2 percent with 1 and 3 sensitivity: the carbon services’ rental SCC embeds a discounting choice — worth confirming it matches.

One data fix and one naming fix

Kosovo’s row in the shared correspondence carries the country name “Serbia” with the correct label XKX. One cell corrected upstream makes every name-join and map legend right. Who owns the correspondence generation?
The drive folder gdp_inflation_delator is missing an f. We staged under the corrected spelling; the drive folder should be renamed to match. Who owns submissions?

The service sheet itself

Status per the sheet’s own tracking (asks to each team)

Per the sheet’s own status column, three services list their SI table as not filled: Water Supply, Livestock, River Flood Mitigation. Six more are marked Incomplete: Water Quality, Coastal Protection, Pest Control, Timber, Extractive Materials, Fire Protection. (If the sheet is out of date, telling us is the fastest fix.)
Storm Mitigation has no status at all in the sheet, and Fuelwood’s feasibility reads “Unknown” — can their teams say where they stand?
Timber’s draft is In Progress; everyone else reports a shared draft. When does Timber’s land?
Five services are feasible only as case studies, not globally: Water Quality, Pest Control, Health, Peak Flow Mitigation, Moisture Recycling. Is case-study coverage acceptable for the account, or do these wait?

The sheet’s own missing-info flags, by service

Value info missing: Pollination and Fish Provision (the sheet says so itself, which matches our 28-times definitional question).
Quantity and price info missing: Coastal Protection — likely the same gap behind our $73bn-vs-$36.72bn question.
Method and data missing: Health Benefits (value method and data), Fire Protection (price data sources), Air Filtration (value and coverage).
“Relationship to other services” missing: Water Quality, Non-Timber Forest Products, Timber — the same double-counting concern our shoreline question raises for carbon.
Ask: each named team fills its row before the account integrates its service.

Leads, priorities, and hygiene

The sheet itself lists a category “Feasible ES with no assigned lead” — can leads be named today for the unclaimed services?
Which sheet services should enter the library NEXT? Our suggestion for sequencing: the three with complete SI tables and global feasibility that have no module yet (Local Climate Regulation, Recreation, Extractive Energy).
The sheet’s label typos (wind_enery_provision and siblings) now need a coordinated fix: our configuration table’s crosswalk cells reproduce them faithfully, because a crosswalk key must match the sheet as written. Ask: fix the three labels in the sheet, and we update our three cells in the same change — a two-place lockstep edit, since either alone breaks the join.
The CAS and UMN member columns already name people per service — we will read owners from there unless told otherwise.

What we need from you

The take-away list, by person

Yanxu: the shoreline rationale, and the seagrass and marsh input versions behind $26.65bn.
Matt: one machine-readable v0.2.0 output, and its input versions.
Erwin: bless the RCP substitution; name the CWoN vintage.
Erosion author: the submission’s valuation inputs and a reference output.
Livestock owner: the item list, the pasture rents, a blessed reference.
Coastal-protection owner: which run produced $73.0bn.
Extractive owner: the source of 0.49.
Renewable owner: code or CSVs.
The room: pollination’s definition, the shoreline convention, crop’s units, the map dissolve, base year 2019, subsistence scope, marine carbon’s fate, the fisheries GEP’s source.
Every service team: complete your SI-table row (status, value info, relationships to other services); name leads for the unclaimed services; confirm the next-in-line services for the library.

What happens with your answers

Every answer starts an already-scoped step, most of them mechanical: reference outputs trigger exact-replication verification (erosion, landslide, livestock), decisions trigger one-pass conversions (crop’s units, the map dissolve), and the definitional calls trigger the regenerations that are already written up (pollination at 2019, the shoreline diagnostic).
Nothing lands in a notes file. The questions live in a tracked document with an owner per item, and each one closes with a verified number or a recorded decision.
What you saw today — every number reproducible, every gap named — is the state your answers extend to the whole account. The goal for the next session like this one: no questions slide at all.