Calibration and Validation Data for Agroecosystem Carbon and Greenhouse Gas Models

Curated field observations of soil carbon and greenhouse gas responses to management, with provenance, for calibrating and validating ecosystem and biogeochemical models.

Website: https://ccmmf.github.io/cal-val-data/

This repository holds curated observations, site and treatment metadata, management histories, and bibliographic and methodological context. Values are stored as each source reports them, in reported units, with a trail back to primary archive. It carries no model configuration, no initial conditions, no gap filled or imputed values, and no calibration code. Those depend on a model and belong with whatever consumes data.

Overview

Datasets describe field experiments where a soil carbon or greenhouse gas response was measured against a control, so a model can be tested on two things at once: absolute value, and treatment difference (practice change signal that policy applications care about).

Key features:

• Model agnostic. Observations stay in reported units with no model specific transformation. Mapping to model state variables happens downstream.
• Long format observations, one row per dataset, site, treatment, replicate, time window, and variable, at replicate level where source archive provides it.
• Stable text keys (dataset_id, site, treatment_id, citation) rather than database identifiers.
• Per row provenance through fill_status, recording which source file each value came from.
• Source stated date ranges in min_date and max_date; point dates are never invented.
• Named treatment contrasts in treatment_pairs, so practice change validation (compost vs no compost, cover crop vs none, organic vs conventional) is a first class query.
• Coverage matrix tracking which system, contrast, and target cells are filled, so gaps stay explicit.

Data Flow

Curation happens in a Google Sheet workbook, which is source of truth. CSVs in data/ are generated exports and must not be hand edited.

1. Workbook holds curated tables and is edited and reviewed there.
2. scripts/ingest.R pulls each workbook tab to data/<tab>.csv.
3. scripts/validate.R checks data/ against datapackage.json. 3b. scripts/ingest_benchmarking.R does the same for the synthesis and meta-analysis evidence workbook into data_raw/statewide_benchmarking/.
4. Downstream consumers read committed CSVs.

CSVs are committed so consumers read a stable, documented snapshot. Workbook can keep changing without a commit per edit; CSVs are regenerated and committed when a dataset version is ready.

Datasets

Each dataset pulls replicate level values from a public archive; companion analysis papers supply methods and cross checks. A dataset is curated in its own workbook, and ingest stacks each tab across all workbooks into one CSV per table.

dataset_id	site	lat, lon	period	treatments	variables	citations (DOI)
white_salinas_2020	salinas_socs (USDA ARS, Salinas Valley CA)	36.62, -121.53	2003 to 2011 (8 yr)	8 systems socs_sys1..socs_sys8, 4 blocks (compost with or without cover crop type, frequency, seeding rate)	SOC stock and concentration, bulk density, total N, nitrate, POXC	10.1016/j.dib.2020.106481; 10.1371/journal.pone.0228677
nichols_modesto_2024	modesto_almond_usda (USDA ARS almond fertigation trial, Modesto CA)	37.63, -121.09	2018 to 2019 (2 yr)	compost vs no_compost (annual compost top dress vs control, same drip fertigation)	N2O flux, total C, total N, bulk density	10.1002/saj2.20615; 10.15482/USDA.ADC/26155504
russell_ranch_tautges_2019	russell_ranch_davis (UC Davis Russell Ranch Century Experiment, Davis CA)	38.54, -121.87	1992 to 2014 (study years -1 to 21)	14 systems, 13 with observations (conventional, organic, legume cover, alfalfa phase corn-tomato; irrigated and rainfed wheat; transitional; native grass reference)	total C and N, bulk density to 250 cm, NH4-N, NO3-N, soil organic matter, corn, tomato, wheat, alfalfa and cover crop yields	10.1002/ecy.2105; 10.1111/gcb.14762
raffeld_2024	russell_ranch_davis (UC Davis Russell Ranch Century Experiment, Davis CA)	38.54, -121.87	1993 and 2012	8 treatments, 6 plots each	SOC_stock_Mg_ha at the treatment-specific 0-30 cm reference soil mass, computed per plot with SimpleESM ESM2	10.5061/dryad.p2ngf1w06
anthony_bouldin_2024	US-Bi2 (Bouldin corn) and US-Bi1 (Bouldin alfalfa), Sacramento-San Joaquin Delta CA	38.1091, -121.535; 38.0992, -121.499	2017 to 2021 (4 study years)	single system per site; 9 automated chambers per site as spatial replicates	chamber N2O, CH4 and CO2 flux, annual mean per chamber per year with standard error	10.5061/dryad.qz612jmnx; 10.1007/s10533-023-01095-y

white_salinas_2020

Organic vegetable rotation (romaine lettuce, broccoli) with winter cover crops and compost. Randomized complete block, 4 blocks, so block level is replicate level here. Companion analysis paper focuses on 5 systems (those with optimal cover crop seeding rates); workbook carries all 8 from archive. SOC stock is concentration times bulk density times depth (0 to 30 cm), with bulk density measured only at end of Years 3 and 7 and propagated by equivalent soil mass, so core_sample method row carries that propagation rule. A large Year 0 to Year 1 SOC drop appears across all systems including no compost control, and is documented in source.

nichols_modesto_2024

Long term compost orchard with high frequency low concentration drip fertigation (about 195 kg N ha-1 across 14 events). N2O measured by closed static chamber on 9 dates (Dec 2018 to Aug 2019), 3 replicate arrays per treatment, upscaled to emitter level. Soil total C, total N, and bulk density are single timepoint (August 2019), so this dataset supports N2O magnitude and contrast validation but not a SOC trajectory.

Measured Variables

Core variables, in reported units:

variable	units	description
SOC_stock_Mg_ha	Mg C ha-1	soil organic carbon stock; min_depth / max_depth give the sampled interval for fixed-depth stocks and the reference interval for equivalent-soil-mass stocks
SOC_conc_mg_kg	mg C kg-1 soil	soil organic carbon concentration
bulk_density_g_cm3	g cm-3	soil bulk density
total_N_conc_mg_kg	mg N kg-1 soil	total soil nitrogen concentration
nitrate_N_mg_kg	mg NO3-N kg-1 soil	soil nitrate nitrogen
N2O_flux_g_N_ha_d	g N2O-N ha-1 day-1	nitrous oxide flux
total_C_pct	% dry weight	total soil carbon by dry combustion
total_N_pct	% dry weight	total soil nitrogen by dry combustion

Supporting variables (soil organic matter fractions, texture, pH, biomass, yields, nitrogen inputs) accompany core targets where source reports them. Full list lives in data/variables.csv.

Coverage

data/coverage.csv crosses system and treatment contrast against each target, split into magnitude and treatment contrast, for both SOC change and N2O. Cells marked yes are covered by current datasets; blank cells are tracked gaps.

System	Treatment contrast	ΔSOC magnitude	ΔSOC contrast	N2O magnitude	N2O contrast	source
Annual row / vegetable	cover crop vs no cover crop	yes	yes			white_salinas_2020
Annual row / vegetable	compost / OM vs fertilizer	yes	yes			white_salinas_2020 (N2O not measured)
Perennial woody (orchard / vineyard)	one management contrast with known history			yes	yes	nichols_modesto_2024 (single timepoint SOC)
Annual or perennial	tillage contrast					gap
Annual or perennial	irrigation regime contrast					gap
Rice / flooded	flood vs AWD					gap

The table above predates the AmeriFlux Delta and Bouldin chamber datasets. As of those, N2O is covered at three sites (nichols_modesto_2024 static chamber; anthony_bouldin_2024 automated chamber at US-Bi2 and US-Bi1) and CH4 at three (tower annual budgets for US-Twt, US-Bi1 and US-Bi2, plus chamber CH4 at the two Bouldin sites). Chamber and tower values are different measurement supports and are not interchangeable. These constraints are documented for any consumer.

Repository Structure

.
├── LICENSE                        # BSD-3-Clause (code)
├── README.md
├── 000-config.yml                 # workbook id, data_dir, raw_data_dir (config::get)
├── datapackage.json               # schema: fields, types, keys for each table
├── scripts/
│   ├── ingest.R                   # googlesheets4 pull, one workbook tab to one data/*.csv
│   └── validate.R                 # warn mode integrity checks vs datapackage.json
├── data/                          # generated csv exports land here (see Data Products); do not hand edit
│   └── .gitkeep
├── data_raw/                      # per dataset provenance, one folder per dataset
│   ├── README.md                  # folder per dataset rule
│   └── white_salinas_2020/        # data_entry_report.md lands here
├── docs/                          # data collection protocol lands here
└── tests/
    └── testthat/
        ├── helper.R               # loads validate.R for the tests
        └── test-integrity.R       # datapackage columns present; warn mode gaps surfaced

Data Products

Each workbook tab maps to one CSV in data/. Full field schema, types, and keys live in datapackage.json; two core tables are described below.

Table inventory

CSV	Purpose	Key
citations.csv	bibliographic sources	doi
sites.csv	site name and coordinates	name
treatments.csv	treatment and system definitions, control flag	name
treatment_pairs.csv	named contrasts for treatment difference validation	pair_id
managements.csv	agronomic events (planting, harvest, tillage, fertilization, amendment)	event_id
observations.csv	measured data, long format	(composite)
methods.csv	how each measurement was taken	method name
variables.csv	variable dictionary (name, description, units)	name
crops.csv	crop and cover crop species	scientific name
coverage.csv	coverage planning matrix

observations.csv columns

Column	Type	Description
min_date, max_date	date	source stated window; equal for a point sample, blank when source gives no date
crop, variety	string	crop and variety, where reported
sitename	string	foreign key to sites.name
variable	string	foreign key to variables.name
method	string	measurement method (see methods.csv)
value	number	measured value, in reported_units
n	number	sample size; 1 for replicate level rows
statname, stat	string, number	reported statistic (e.g. SE) and its value; not converted during curation
treatment_id	string	foreign key to treatments.name
replicate_id	number	block, plot, core, or chamber identifier
study_year	number	experiment year or calendar year
min_depth, max_depth	number	cm; sampled interval for depth-based observations and reference interval for equivalent-soil-mass stocks; actual integrated depth is recorded in attributes_json.equivalent_depth_cm
citation	string	DOI, foreign key to citations.doi
fill_status	string	provenance code (see Conventions)
notes	string	free text, including date precision caveats
dataset_id	string	short source packet id, e.g. white_salinas_2020
reported_units	string	units as source reports them
observation_level	string	replicate or zone_mean

managements.csv columns

Column	Type	Description
event_id	string	unique row key, <mgmttype>-NNNN
sites.name	string	foreign key to sites.name
treatments.name	string	foreign key to treatments.name
min_date, max_date	datetime	event window
mgmttype	string	planting, harvest, tillage, fertilization, compost_application
cover_crop	boolean	planting rows: cover crop vs cash crop
crop_name	string	cover crop name on planting rows
level, units	number, string	rate and its unit (e.g. Mg ha-1 (oven-dry), kg N ha-1, cm depth)
notes	string	provenance comment
attributes_keyvalue	string	semicolon delimited key=value extras (material, N content, implement, timing)
citation	string	DOI, foreign key to citations.doi

Smaller tables (citations, sites, treatments, treatment_pairs, methods, variables, crops) are short; see datapackage.json for their fields.

Conventions

• Stable text keys. A site key is physical location only; it does not encode treatment, replicate, or date. Treatment keys (socs_sys1, compost, no_compost) stay stable across a dataset.
• Long format, replicate level. Where a paper has a companion data archive, ingest replicate level values from archive; summary tables in analysis paper are for cross checking, not primary fill source.
• Dates. Record source stated range in min_date and max_date. An exact date sets both equal; a stated month or season bounds interval; an unknown date stays blank with reason in notes. Never invent a day of year.
• Statistics. Store reported statistic (n, statname, stat) as given. Statistic conversion (standard deviation to standard error, and similar) is a downstream step, not curation.
• Basis. Methods state dry vs wet basis (prefer dry) and volumetric vs mass basis. Amendment rates carry their basis in units.
• fill_status. Provenance code for each value. Codes in current data: FILLED_SITE_DATA_XLSX, FILLED_ZOTERO_XLSX, FILLED_DATASET. Additional codes (PAPER_TEXT, NEEDS_FILL, MISSING_IN_SOURCE) are reserved for rows still being curated.

Validation

scripts/validate.R checks every generated data/*.csv against datapackage.json. It runs in warn mode: missing columns, empty keys, broken foreign keys, and odd dates warn but do not stop, so a workbook in progress still passes. It needs only base R plus jsonlite and readr.

Rscript scripts/validate.R

Tests under tests/testthat/ run the same checks via testthat. Tables not yet generated are skipped, so a fresh clone passes.

Rscript -e 'testthat::test_dir("tests/testthat")'

Usage

Load a table

library(readr)
obs <- read_csv("data/observations.csv", show_col_types = FALSE)

# replicate level SOC stock
subset(obs, variable == "SOC_stock_Mg_ha" & observation_level == "replicate")

Pull a treatment contrast

pairs <- read_csv("data/treatment_pairs.csv", show_col_types = FALSE)

# contrasts flagged for validation (e.g. compost vs no compost)
subset(pairs, use_for_validation == TRUE)

Regenerate the snapshot from the workbook

# set workbook id in 000-config.yml first, then authenticate once
# interactively with googlesheets4::gs4_auth()
source("scripts/ingest.R")

Scope

This repository is a data layer: measured values and their provenance. It does not carry model configuration, initial conditions, gap filled or imputed events, day of year imputation, unit or statistic conversion, fold assignment, priors, or calibration code. Those are properties of a particular model and a particular study, so they live downstream with whatever consumes data. Keeping this layer neutral is what lets a second model or analysis reuse the same tables.

License

Code in this repository (scripts/, tests/) is BSD-3-Clause; see LICENSE. Curated data derives from sources (public data archives and journals) that require attribution under CC-BY; cite the original sources listed in data/citations.csv.