Forest conservation
Canopy from satellite is free and continuous; ground methods carry the species claims. Bat identification stays at genus.
Every environmental report compares your site to a regional average. Naturecode compares it to its own history — and states what change is large enough to be real here, on the methods you actually used.
01
Mean average precision on clean single-species recordings, against F0.5 on 286 hours of real soundscape. The marketed figure does not apply to passive monitoring.
Accuracy at locations the model was trained on, against locations it was not. Species assignment does not travel between places.
Datasheet against field RMSE under generic factory calibration in clay-rich and organic soils. Calibration accounts for 42% of total variation; the sensor model only 29%.
Which is why we do not make it. Detection is what eDNA gives you.
Standard report · Regional baseline
An expected range is an average across places that are not yours. It cannot tell you whether this reef lost something last season, because it was never about this reef.
Naturecode · This site's own threshold
A threshold learned from this site's own repeat sampling, on the protocol you used — and widened when the protocol slips. We say which species, with what confidence, and when we cannot say.
02
Canopy from satellite is free and continuous; ground methods carry the species claims. Bat identification stays at genus.
eDNA carries the community; satellite is useful for turbidity and surface temperature, not for reef health directly.
Water extent from satellite, community from eDNA, birds from acoustics.
Soil calibration dominates here — calibrate to this soil or the readings are about the probe rather than the ground.
Siting class matters more than sensor choice in a built environment — an exposure class can add 1 to 5 °C before the instrument is considered.
Flowing water carries DNA from upstream, so a detection places a species in the catchment rather than at the sampling point. We report it that way.
Teams
A project outlives the people monitoring it. Access is granted through a team and can be withdrawn without the record becoming unreadable — which is why the key belongs to the project rather than to one person's device.
Borrowing
One site's scatter can sharpen what a project expects. A project average may never become a site's threshold — that is how a regional baseline gets in through the back door.
03
Free, full and open, commercial use explicit — the only modality here that costs nothing per site.
One PCR replicate recovers 70.9% of taxa; one field replicate 55.4%. Eight or more PCR replicates are the working recommendation.
0.414 F0.5 on 286 hours of real soundscapes, against 0.791 mean average precision on clean single-species recordings. The confidence score is a sigmoid output with a user-tunable sensitivity — it is not a calibrated probability, and it must be thresholded per site.
Animal-versus-empty detection genuinely works: 99.2% recall at 97.3% precision. Species-level identification is the weak link — accuracy falls from 95.6% at trained locations to 68.7% at new ones.
Unusable at species level for the cases that matter: one widely used classifier returns a positive predictive value of 0.05 for a federally endangered bat — five correct calls in a hundred. Genus level is sound (0.92–0.99). Class imbalance makes overall accuracy read above 0.80, so quoting accuracy here is itself the overclaim.
Factory calibration gives about ±0.053 m³/m³ in clay-rich and organic soils against a ±0.03 datasheet figure. Calibration choice accounts for 42% of total variation and the sensor model only 29% — how you calibrate matters more than what you buy.
0.2 K achievable against 0.1 K required for air temperature. Siting dominates the sensor: an exposure class adds 1, 2 or 5 °C of uncertainty, and a passive shield can contribute +2 °C at solar noon — larger than the sensor's own spec.
04
Refuted, and mechanistically rather than just correlationally: at equimolar input, read counts span three to four orders of magnitude, and the allometric exponent of about 0.67 means eDNA tracks neither abundance nor biomass. Detection is what eDNA gives you. Quantity is not.
Refuted as a proxy — correlations with actual species richness sit around r = 0.33 and fall under scrutiny. A single number cannot carry a community. We report named detections with their own confidence instead.
Measured against us: for one species (Eastern Black Rail), the 99%-precision cutoff was 0.60 in Colorado, 0.74 in Texas, 0.83 in Florida and 0.91 in South Carolina — and precision at fixed settings ran from 0.87 down to 0.46. A global threshold gives F1 below 0.5 across every dataset tested.
Richness inflates roughly twofold for every tenfold increase in sequencing depth. A deeper run looks like a richer site. We report inverse Simpson, which is robust to depth, and refuse Chao1 and ACE on ASV data entirely.
The hardest part of this problem
Without one fixed protocol, about 90% of the variation between eDNA results is the laboratory rather than the site. A change measured by one lab and confirmed by another is confounded with the difference between them.
Ecology is not noise
Roughly one taxon in five appears at a single time point and never again. A site's own variation is part ecology and part instrument, and a threshold has to clear both before a change means anything.