Is public data enough to really test this?#

Researched 2026-10-07, after testing the APIs rather than reading about them.

Short answer: it validates about half of layer zero, and the half it cannot touch is the half we claim is novel.

The data model and the quality checks can be scored against real sensors with real faults, for free, over decades. The transport contract — store-and-forward, the duty cycle, batching, eviction, deduplication, cumulative acknowledgement — cannot be tested by any public dataset, because every archive has already thrown that layer away. A harmonised hourly series is what survives after the problems §5 exists to solve.

What is usable, in order of value#

1. ISMN — the one that matters#

The International Soil Moisture Network hosts 71 networks and 2,842 stations with data from 1952 to the present, the longest continuous records running about forty years. SCAN, SNOTEL, USCRN, ARM and COSMOS all contribute, so US and Tennessee coverage comes with it. Free with registration, CC-BY-4.0, with an MIT-licensed Python reader (ismn) that exposes measurements and their flags as pandas columns.

Four properties make it the right validation set, and the first is decisive:

• Flagged values are retained, not deleted. "Measurements that are detected as erroneous by the quality control procedures are not deleted from the database but tagged as such." That is the difference between a labelled fault dataset and a list of holes.
• The flag vocabulary maps onto §4.2. C flags are observations exceeding a plausible geophysical range — C01 soil moisture below 0 m³/m³, C02 above 0.6, C03 above the saturation point — which is precisely our out_of_range. D flags are questionable or dubious observations from consistency and spectrum-based checks, the family our flatline and rate checks belong to. M is missing, G is good.
• Units already match. Everything is converted to volumetric soil moisture in m³/m³, which is our cm3/cm3 exactly — the same quantity, the same UCUM code, no conversion and therefore no conversion error.
• Harmonised to hourly UTC, which §2 wants anyway.

What this buys, concretely:

• Score the range check on real data. ISMN's C flags are an independent judgement about the same readings. Precision, recall and false-alarm rate against them is a real measurement, not a measurement against our own fixtures.
• Test the quality model's expressiveness. Mapping C01–C03 and D01–D10 into §3.3's flag set is a concrete interop exercise, and if the mapping is lossy that is a finding about our model rather than about theirs.
• Set thresholds from forty years instead of from guesswork. The canonical farm had to invent min, max and max_rate_per_s; §4.2 now says a rate threshold derives from the fastest legitimate process, and this is where that number comes from.
• Real drift. Forty-year records at some stations contain the slow failures the drills showed a single probe cannot see.

2. SCAN directly, via the NRCS AWDB REST API#

Tested on 2026-10-07, no token: https://wcc.sc.egov.usda.gov/awdbRestApi/services/v1/, OpenAPI documented.

Three Tennessee stations — Allen Farms (Giles), McAllister Farm (Lincoln), Eastview Farm (Franklin) — soil moisture and temperature at 2, 4, 8, 20 and 40 inches, hourly through monthly, since June 2003, live to the current day, with QC and QA flags. The element code carries its depth: elements=SMS:-2, not elements=SMS.

But SCAN deletes rather than tags. Measured over 22 years of daily 8-inch readings across the three stations: 23,240 readings, of which 686 carry qcFlag=S for suspect — and the suspect count equals the missing-value count exactly. Asking for returnOriginalValues=true returns neither a value nor an original for those rows. Thirteen readings are E for estimated and do carry values.

So SCAN gives labelled gaps, not labelled bad values: useful for §5.4's gap handling and §4.2's staleness, useless for scoring a value-based check. It is still the best source of regional threshold evidence, and its live endpoint is the simplest thing to point an ingest adapter at.

3. Small datasets with real, documented faults#

• AquaIntel (IEEE DataPort) — 3,886 multi-sensor readings from an IoT edge node built for irrigation and anomaly detection, with per-timestamp ground-truth labels, a real sensor-displacement fault where a level probe lost contact, and a runaway pump event ending in total saturation. Small, but the faults actually happened.
• A published fault-injected soil moisture dataset covering drift, bias, spike and malfunction — independently the same taxonomy as sim/faults.py, which is reassuring about the fault library and means the two are comparable.
• myClim's mc_prep_TMSoffsoil — a published R detector for out-of-soil measurements from TMS loggers. A probe lifted out of the ground is the exact case our person-override example uses, and here is somebody else's detector for it to compare against.

4. The Tennessee Mesonet, when it exists#

95 stations, one per county, the same five depths, five-minute cadence, real-time and public, half due operational by the end of 2026. Unlike SCAN's hourly, five minutes is inside §4.2's cross-check window, so a station in the pilot's county is a genuine peer.

What no public data can test#

Everything the transport contract claims, which is most of what is distinctive here:

This is the honest shape of it: public data validates the half of layer zero that looks like everyone else's, and cannot touch the half that is the reason the project exists. The simulated farm covers the transport half, and §9.2 is right that a simulator is not a season.

What to do about it#

1. A scoring run against ISMN. Done 2026-10-08 — see ismn-scoring.md: 2.8 million Tennessee readings, the checks quiet in the root zone, and two real defects found (Q23, Q24). tools/ismn_score.py, the ag counterpart of the HVAC fault-detection scoring in campus_twin: load an ISMN download, map its flags into §3.3's, run sdk/checks.py over the real series, and report precision, recall, false-alarm rate on G rows, and per-episode detection. Skips cleanly when the download is absent, like every other optional stage. This is §9.2's "replay of real recorded field data", and it is the strongest validation available without hardware.
2. An ISMN flag mapping, written down. If §3.3 cannot express C03 (above saturation, which needs a per-soil saturation figure the datastream does not carry) then that is a gap worth knowing about.
3. Derive the thresholds the canonical farm invented from ISMN statistics, and say in the spec where each number came from.
4. Leave the transport half to the sim and the pilot. Do not pretend an archive tested it.

Caveats#

• ISMN was not downloaded (resolved 2026-10-08: downloaded, and the flag vocabulary verified against the archive — it matched). Originally: ismn.earth refused the connection from here, so the flag semantics above come from the HESS paper and the ISMN documentation rather than from the data. Verify the flag vocabulary against an actual download before building the scorer, and check how many C and D rows the Tennessee SCAN stations actually carry — a flag family that is rare in practice scores nothing.
• CC-BY-4.0 requires attribution, and ISMN asks for citation of the network papers.
• The AquaIntel and fault-injected datasets were found by search and not inspected.
• Scoring our checks against another system's QC flags measures agreement, not correctness. Where we disagree with ISMN, either could be right, and the honest report says so rather than treating their flags as truth.

Sources#

• ISMN: serving Earth system science for over a decade (HESS 2021) · ISMN quality flags · ISMN flag overview · ISMN data · TUW-GEO/ismn Python reader (MIT)
• NRCS AWDB REST API · R {awdb} · Python soildb
• AquaIntel multi-sensor IoT telemetry dataset · Fault detection in smart farming, drift/bias/spike/malfunction · myClim mc_prep_TMSoffsoil · Self-supervised anomaly detection of rogue soil moisture sensors
• NCSMMN partner networks