← back to all birds

What is measured, what is modeled, and what is invented

BirdView draws precise-looking figures from data that is, in places, thin. This page says which parts of each figure are observations and which are arithmetic.

In short: of the five segments in each figure, bill and tail are measured; head, neck, and body are modeled. Every total length is modeled, and so are nearly two thirds of the wingspans. The UI labels each value with its status.

For every source, field, join, filter, and model in one place, see how BirdView's data is built.


The provenance system

Every number carries one of three kinds, shown on the detail page for any bird:

Kind Means
measured taken from a source, for this species
near-match measured on a different species this one was recently split from, because no measurement exists for it yet
modeled computed. The model field names the formula and, where one exists, its error

A ~ in the tile footer marks anything not measured. A modeled value is never presented as a measurement; see the decision on length and wingspan sources.

Near-match applies to nine drawn species that are recent splits AVONET does not yet treat separately: American Goshawk, Vega Gull, Hudsonian Whimbrel, Western Cattle-Egret, American Barn Owl, Chihuahuan Meadowlark, Cocos Booby, Siberian Pipit, and Blue-billed White-Tern. Each borrows the parent taxon's measurements, and the detail page names that taxon. For the goshawk, cattle-egret, barn owl, and white-tern, whether AviList's change is a split or only a rename still needs a taxonomic check; near-match is the conservative label until then.

Where the parent is drawn too, the two drawings are identical, and both tiles say Same measurements as the other bird: Siberian and American Pipit, Chihuahuan and Eastern Meadowlark, Cocos and Brown Booby, and Vega and American Herring Gull.


Segment by segment

Part Status Where it comes from
Bill measured — 1073 direct, 9 near-match AVONET Beak.Length_Culmen
Tail measured — 1073 direct, 9 near-match AVONET Tail.Length
Mass measured — 1073 direct, 9 near-match AVONET Mass
Wingspan 391 measured, 6 near-match, 685 modeled Fox et al. 2026 where empirical; otherwise a fitted regression
Head modeled, unvalidated — every species 7.5 × mass^(1/3), × 0.65 for grebes
Body modeled, calibrated — every species 14.0329 × mass^0.37
Neck modeled, invented — every species a per-family (or per-genus) ratio × the trunk
Total length modeled — every species bill + head + neck + body + tail

Counts are over all 1082 species in the dataset. Mass is in grams, lengths in millimeters. The five segments sum to total length; mass and wingspan are not part of the sum.

Why every length is modeled

AVONET, the only openly licensed morphology dataset with this coverage, has no body length and no wingspan. Field-guide lengths are not openly licensed, and Ricklefs (2017), the one comprehensive academic source, prohibits extraction, storage, and redistribution.

So BirdView builds length from parts, two measured and three modeled. The segments sum exactly to the total, so a figure's height is the bird's length and figures at one scale are directly comparable. These lengths will not match your field guide: they are a different quantity, a consistent basis for comparison, not authoritative measurements.

The wingspan model

For the 685 species without an empirical wingspan, BirdView fits its own order-level log-log regression of wingspan on wing chord, trained on the empirical column of the Fox et al. dataset. Five-fold cross-validation, as absolute percentage error:

Model mean median p90
Global log-log on wing chord 9.0% 7.3% 19.0%
Order-level log-log (adopted) 6.0% 4.6% 12.6%

For the 9,194 species Fox et al. predicted, it matches their published predictions to a mean of 1.32%; a test checks that agreement, so a fault in the fitting code shows up against an external reference. Fox et al. give no prediction for species with an empirical value, so there is no held-out comparison: the 1.32% is agreement between two models, and the error figures are our own cross-validation. See the decision record.

Each modeled wingspan states its own error. A passerine reads:

log-log regression of wingspan on wing chord, fitted on 360 empirical species in Passeriformes from the Fox et al. 2026 dataset; cross-validated (five-fold) median error 4.7%

An order group of 5 to 9 empirical species, too few for five folds, is cross-validated by leaving out one species at a time: the woodpeckers read 2.7% from 8 species, and the loons and grebes 3.7% from 9. A small group's error is itself less certain than a large one's.

AVONET's Wing.Length is the wing chord, one folded wing. Doubling it is wrong by a wide, species-dependent margin, and BirdView never does it.

No gamebird (Galliformes) is in the Fox dataset, so every gamebird wingspan is modeled.

The three modeled segments, in descending order of confidence

Body (trunk) — 14.0329 × mass^0.37. Mass is AVONET's only whole-animal measurement, and no dataset of trunk lengths exists to fit against, so the exponent is calibrated against measured skeletons: wingspan 0.368 (n=301), neck length 0.380, and total leg length 0.364 (n=103, Böhmer et al. 2019) agree on how a linear measure grows with mass, and the model sits inside the interval they imply. The coefficient keeps a 63.4 g bird's trunk where an earlier, uncalibrated model put it, so calibration changed proportions across sizes but not overall scale. There is still no measured trunk.

Head (cranium) — 7.5 × mass^(1/3). Isometric scaling, with the coefficient putting a 20 g passerine near 20 mm. A shape function, not a fit, and unvalidated. There is deliberately no minimum: a 2.5 g hummingbird lands near 10 mm, and a floor would only hide a bad coefficient. Grebes, heavy birds with small heads, take 0.65 of it; the bird's page names the factor.

Neck — a per-family ratio of the trunk. This one is invented. No source measures neck length at this coverage. The ratios are a drawing parameter: Ardeidae 0.85, Phoenicopteridae 0.95, Anhingidae 0.90, Gruidae 0.85, down to Strigidae and Tytonidae at 0.08, every other family 0.18. A genus that flies with a different neck from its family has its own: the diving ducks 0.30 to 0.35 against their family's 0.70, which fits the geese and swans, and the grebes 0.45 to 0.80. These overrides, and factors on the modeled head, body, and wingspan, are in data/proportions-review.tsv, each marked proposed or reviewed, and a bird's page says which. They are not checked against photographs: almost none shows a bird as the figure draws it, body and spread wings in one plane, so every other angle would foreshorten some part (decision). They are shown anyway — a heron's page reads "neck modeled as 0.85 × trunk (family-specific ratio for Ardeidae)" — because every figure draws a neck, and hiding the number would hide only the explanation.


Deliberate fictions in the drawing

These are rendering choices; none produces a number.

The figure is a top-down view of a bird in flight, where wingspan and body length lie in one plane, so one scale for both is accurate.

There are no legs. Most birds tuck them in flight. For those that trail them — herons, cranes, storks, ibises, flamingos, stilts, rails, loons, grebes — the projection past the tail depends on the femur, toes, hip, and full tibiotarsus, but AVONET measures only the tarsometatarsus; a leg modeled from it gives a White Ibis −14 mm of projection, though its legs visibly trail. The decision record has the numbers.

Herons and pelicans get an S-shaped neck, because they fly with the head drawn back, the field mark that separates a heron from a crane, stork, or ibis. Two semicircles fit the full neck length into 2/π ≈ 63.7% of that height; the stroke stays to scale. Sorting, sectioning, and the detail view use the full length.

The head is an oval 0.72 as wide as long; tail and body are single strokes, because no source gives their widths.

Colors encode taxonomy only: related hues within an order, distinguishable between families.


BirdView has no photographs, so every tile and species page links to the species on eBird, at ebird.org/species/ and the eBird code shown beside the bird's name. The code comes from AviList with the names, so the link needs no lookup of its own. A short table adds the codes Cornell's current taxonomy gives where they differ: the only code for three species AviList splits more finely than eBird (Ruddy Duck, White-necked Petrel, American Barn Owl, which link to eBird's broader species), and a second code for eight more. Where a species has two, BirdView shows and links with the simpler: fewer digits; if that ties, shorter; so Fork-tailed Swift is fotswi, not fotswi1. Every code finds the species, and /species/<code> redirects to its page. See the decision record.

BirdView links and never copies: no photograph, sound, or map is embedded, hotlinked, or cached. Until 2026-11-01 the link went to iNaturalist, through a table of taxon IDs matched by scientific and common name (the earlier record).

Scope

1082 species: every bird whose records recur in the regions BirdView covers, by BirdView's own rule rather than anyone else's checklist. The regions are the 50 US states and DC, 13 Canadian provinces and territories, 5 US outlying islands, and 5 stretches of open sea. They are not the ABA Area, which includes St-Pierre et Miquelon (no BirdView region) and excludes Johnston, Palmyra, Wake, and Navassa (covered by BirdView). A species is in scope when, in at least one region, its observations fall in repeated episodes across years. Nine are kept by hand because the open record cannot speak for them, each named with its reason.

41 of them are not on the home page, which shows 1041. Recurrence cannot tell a wild population from a steady supply of escapes, so parrots, waxbills, ornamental waterfowl, zoo birds, and a farmed Emu pass the rule too. They are on a declared list with reasons and left off the home page and its controls; the list links to their pages, which stay up. Long-feral populations and wild vagrants are not on it.

Gray-headed Swamphen (Porphyrio poliocephalus) passes the rule but is dropped: eBird splits it and AviList lumps it into Purple Swamphen, and every name BirdView serves comes from AviList.

Out-of-scope birds still have a page. 326 named species are out of scope: 124 vagrant, 38 historical, 164 absent. A URL such as /species/paspig returns a 404 that names the bird and gives the measured reason: Passenger Pigeon recurs across 89 episodes in 52 years in Ontario, the last on 18 May 1902. An absent species shows no counts: no date query was run for it, so its zeros mean not measured, not measured zero. The status is 404, not 410, because scope is re-derived from live records and a bird can clear the bar on a later run.

Names come from AviList v2025b: eBird/Clements (Dovekie) by default, AviList/IOC (Little Auk) on request. The species id in the URL is the eBird code (amerob, not turdus_migratorius), which is assigned rather than derived from the English name, so it does not change with the vocabulary.

Taxonomy is AviList's: names, families, and family order. Where AVONET's BirdLife taxonomy disagrees, AviList wins, joined through explicit synonym tables rather than fuzzy matching: one maps a current name back to the older label AVONET measured under, the other maps BirdView's name forward to a newer label AviList adopted.

Regions

Which regions list a bird is derived: no source states "this species occurs in Pennsylvania" under a license BirdView can serve, so the list is computed from occurrence records.

Regions are GADM level-1 units — US states, Canadian provinces and territories, and the outlying islands — plus five ocean boxes of BirdView's own.

A species is listed in a region when it reaches 1 record per million of that region's bird records, from GBIF: eBird's under CC BY 4.0, and iNaturalist's research-grade observations under CC BY-NC 4.0. The bar decides only the listing: records under it still count in 1 in N. Emu is listed nowhere, yet its 38 records, 19 in Texas, are on its tile.

The bar is relative because eBird's volume doubled in five years, so an absolute bar would admit more species over time for no ecological reason, and because effort differs enormously: California has about 114 million species observations, Hawaii 2.5 million.

eBird first, iNaturalist only where eBird is deliberately silent. eBird's records are reviewed before publication and make up 98.8% of the two sources combined. But eBird withholds sensitive species (listed under When the sources disagree about a name); where its list says a bird is withheld and the records confirm it, BirdView counts iNaturalist instead, and the page says so. eBird alone would omit the birds readers most often look up; iNaturalist everywhere would count escaped pets.

Names are matched verbatim, not through GBIF's taxonomic backbone, which silently drops what it cannot resolve and reports zero US and Canadian records for Cooper's Hawk.

Some species are correctly listed only at sea: Townsend's Storm-Petrel, Bermuda Petrel, Trindade Petrel, Craveri's Murrelet, Guadalupe Murrelet, and Atlantic Yellow-nosed Albatross are seen beyond any land polygon. Six species on the home page, among them Desertas Petrel and Little Curlew, reach the bar nowhere; their pages show no Regions row.

Counties, where BirdView serves them, list a bird under a stricter rule: its state must list it, and it needs at least 5 records in the county as well as 1 per million of the county's records. A bird's page does not name every county. After each state or province it says how many of its counties list the bird out of how many it has: TX 254/254 for American Robin.

The ocean boxes are approximate: disjoint coordinate rectangles, so a record falls in at most one, drawn to separate the birds rather than to follow any published maritime limit.

Known problems with the region rule

The rule counts records, not birds, so a popular rarity outscores a quiet resident: the more surprising a bird, the more people go to see it and file checklists. American Flamingo is listed in Pennsylvania on 931 records of one or two birds at one pond over thirteen days in September 2023, after Hurricane Idalia; Lucy's Warbler, a Sonoran desert bird, is listed in Virginia on 53 records from one site over six days. A region list is where a bird has been reported often enough to clear the bar, not a range map.

No threshold fixes this, because a count measures observer effort: remote and pelagic species are under-reported and vagrants over-reported, errors in opposite directions. Recurrence across years separates them — the flamingo's 931 records span thirteen days; Akikiki, a Hawaiian endemic, has 376 across forty-five years — which is why Scope uses it. Per-region presence still uses the record bar, a known limitation.

Reporting frequency

Each bird shows its reporting frequency: how many of the observations recorded in the selected regions are of that bird, as one in N.

What an observation is

One observation is one species on one checklist, both sources' own unit: eBird's id is OBS1912782147, iNaturalist's is inaturalist.org/observations/333181443. A list of thirty species is thirty observations.

The number of individuals is deliberately ignored: twelve White-throated Sparrows, one Northern Cardinal, and fifty thousand Snow Geese are one observation each. So 1 in N measures how often a bird is logged, not how many there are. It understates flocking birds such as waterfowl and blackbirds and overstates chickadees and cardinals. The denominator also grows with how many kinds of bird people report; dividing by each region's own total cancels that out.

Turning 1 in N into how often you would see it

Divide the average checklist length by N. A bird on a fraction f of checklists of S species holds f/S of the observations, so f = S/N. American Robin is 1 in 32 observations in New York; on two complete New York county-days the average checklist ran 15.6 and 19.1 species, so the robin is on roughly half to three-fifths of outings. Where fewer species are about, lists are shorter and the same 1 in N means fewer outings; the ordering is unaffected.

BirdView does not publish this figure because GBIF's eBird export has no checklist id. Checklists must be reconstructed from observer, date, and rounded coordinates over a complete slice of the data (a partial slice makes almost every checklist look like one species), and doing that for every region would mean downloading all the records.

On the home page the scope is the regions you have selected. A Gyrfalcon is about 1 in 3.6 million species observations across all regions and 1 in 82,000 in Alaska. Selecting a region where a bird is absent lowers its figure, because those records join only the denominator.

On the detail page the scope is every region, and the page names the region where the species makes up the most observations — for a Sharp-shinned Hawk, Top region: Yukon — 1 in 249 species observations there. A bird with no records on screen reads 0 in the full count of more than a billion, because zero in a billion records and zero in nine hundred are different findings.

Seas are measured on open water. The marine boxes include coastline (the Pacific box, 150°W to 115°W, takes in most of coastal California), so a sea counts only records inside it on no land: GADM covers land, so a beach record carries a country and a record on the water none. A sea's denominator is thus comparable with a state's, 44 species are commonest at sea, and selecting both double-counts nothing. See the decision record. Bay water also has no GADM country, so 8.8% of Chesapeake Bay records count as off land, overstating the open-ocean figure by roughly 3%.

When the sources disagree about a name

BirdView serves AviList's names; the records come from eBird and iNaturalist, which do not always agree. Where they disagree the species page says so.

One name, two birds. eBird has not adopted AviList's split of the Yellow Warbler: its more than six million records under Setophaga petechia cover both the northern and the mangrove bird, with none under Setophaga aestiva, and nobody knows how they divide. Both birds show the shared count and name each other. The Elegant and Coppery-tailed Trogons are the same case.

eBird withholds some species to protect them from trapping, disturbance, or hunting, and publishes a list. Ten are withheld everywhere BirdView looks: Spotted Owl, Great Gray Owl, Gyrfalcon, Northern Hawk Owl, Lesser Prairie-Chicken, Gunnison Sage-Grouse, Black Rail, and three Hawaiian endemics. Others only in some places or seasons: Long-eared Owl across the United States, Painted Bunting in Florida from October to mid-April, Short-eared Owl in New York in spring and summer. There the counts are iNaturalist's alone and far below the real record, the month filter does not judge them, and the page says withheld by eBird — a gap in what BirdView can see, not a statement about the bird.

The list is checked, not trusted: it states current policy, while records reflect the policy when published, and for 26 of the 230 places it names eBird's records are present anyway. A bird is marked withheld only where the records show the gap. See the decision record. Because withholding and a different name both return nothing, names were also checked against every plausible alternative.

The same bird under another name. eBird files seven species under names AviList has replaced, and the page names both: five genus moves (Hairy Woodpecker is Leuconotopicus in AviList; eBird files its more than seven million records under Dryobates villosus) and two corrected epithets (Pica nuttallii against eBird's Pica nuttalli). No count is affected, and the page says so. iNaturalist follows AviList on the genus moves and neither uses AviList's epithets, so the page names eBird, the database that holds nothing under the AviList name in every case.

Where the species boundary falls. eBird files three species under a name that AviList gives to a different species, and the page says which case each is. Two are splits eBird has not adopted: North American Hudsonian Whimbrels are under Numenius phaeopus in eBird (556,569 records, none under hudsonicus), while iNaturalist has 17,923 under hudsonicus and nine under phaeopus; Gygis candida against Gygis alba is the same. The third is the only true lump: AviList does not recognize Stejneger's Stonechat, which both databases do, so its 95 eBird and 28 iNaturalist records belong to Saxicola maurus. In each case neither database holds much under the other's name, so it is one set of birds under two names, and BirdView adds the records.

Feral and domestic birds. iNaturalist files feral Rock Pigeons under Columba livia domestica and domestic Muscovy Ducks under Cairina moschata domestica, below species rank. BirdView folds these into the species, and the page says so; no source separates them.

A proportion, not a count

A count says where the birdwatchers are; a proportion says where the bird is. The two pick different top regions for more than half the dataset. California has 35 times Yukon's Sharp-shinned Hawk records but about 117 times its bird observations, so the hawk is about 1 in 250 in Yukon and 1 in 830 in California.

"1 in N", not a percentage, because the figure runs from 1 in 2 to beyond 1 in 100 million: 50% to below 0.000001%. The median bird in its top region is about 0.26%, and many figures round to zero as percentages.

Both numbers are shown: the species' records and the region's total, so a figure resting on nine records says so, instead of an arbitrary minimum-evidence cutoff. A couple of dozen species have a top region resting on fewer than ten records.

What this is not

It is not abundance or a rarity rank, but how often a bird is reported relative to others reported in the same place: the closest open data gets to "how often would I see this", and not the same claim.

Two distortions are uncorrected. Species eBird withholds, such as Spotted Owl, Gyrfalcon, and Northern Hawk Owl, look rarer than they are, because the denominator counts observers whose sightings of them were suppressed. And a bird that draws a crowd is over-reported against a quiet resident, the bias that keeps record counts out of the scope rule; see Scope.

Habitat and diet

Both come from AVONET, with its definitions quoted in the control. AVONET records one primary habitat and one primary diet per species, so a bird that uses two is filed under one, and for a migrant the habitat is in practice where it breeds: Black-bellied Plover is Grassland.

The habitat filter arranges AVONET's eleven habitats in five groups: Shore and sea (Coastal, and Marine as Sea near shore and Open ocean), Fresh water (Wetland, Riverine), Woods (Forest, Woodland), Open country (Grassland, Shrubland, Desert, Rock), and Towns and farms (Human Modified). A group's name is a shortcut that chooses all its habitats; each can also be chosen alone, and a bird is shown when any of its habitats is chosen. A bird's page names its exact habitats. BirdView reviews every bird's habitats, species by species, for where a North American birder sees it; the review is under way.

BirdView adds values where that hides something, and marks every addition: birdview where BirdView decided it, derived where it follows from AVONET's own data by a stated rule. All are listed at /added.

Open ocean or near shore

AVONET's Marine holds both the shearwater seen only from a boat and the eider in the surf, so BirdView refines it, bird by bird, from where its records lie. It counts each bird's eBird records on water: records GADM gives no region code, inside BirdView's sea boxes, in 0.1° cells whose center is not on Natural Earth's land, banded by the center's distance to shore (under 10 km, 10 to 30 km, 30 km or more). Its records on land come from its region counts. Then:

Each bird's page gives its numbers. The habitat review may set either by hand, with a reason. An address that names Open sea (habitat=Marine) now means both.

Known limits. Records logged from the shore often fall just past GADM's coastline, so water within 10 km is mostly the coast itself; that is why this does not separate beach birds from sea ducks, which the review does. Pelagic-trip checklists plotted at the harbor raise a true seabird's share near shore, which is why the second threshold is as high as 80%. Caribbean and Mexican waters fall inside the sea boxes. No natural break separates 77% from 86%, so a few birds near 80% are for review: Sabine's Gull passes on its Alaskan tundra and inland reservoir records, and Masked Booby misses at 79.5%. The decision is openOceanSplit.

The derived rule reads AVONET against itself. Where AVONET's trophic level is Omnivore but its trophic niche names one food, the diet is mixed and the niche names the larger part. For a seed-eater the rest is invertebrates, which follows necessarily; for an insect-eater the plant part is a judgment from the family, so it is tagged birdview.

One category is BirdView's own. AVONET's ten niches have no term for exudates, so the four sapsuckers would be filed under Invertivore or Omnivore without their staple food. Sap fills that gap. The bar for a new term is that no AVONET term fits, not that none fits well: a berry-eater is a Frugivore. Birds that only drink from sapsucker wells — chickadees, nuthatches, kinglets, several warblers — do not get it; sap is only a supplement for them.


Where this is most likely to be wrong

In rough order of how much it would change the picture:

  1. The neck ratios are invented. They are BirdView's largest unforced assumption and set the silhouette of every long-necked family. About a third of families have their own ratio; the rest take 0.18, each non-passerine family individually reviewed and passerines as a group, being compact songbirds with little exposed neck. All are judged from flight silhouettes. Böhmer et al. 2019 measures the full cervical column, including the part inside the body contour, while BirdView models the exposed neck, and converting between them needs an unmeasured per-family fraction (source notes). Vertebra counts are dimensionless: proportions, never millimeters.
  2. The head model is unvalidated, and the body model is calibrated but not fitted. The head coefficient is fitted against nothing; the trunk exponent rests on necks and legs, since no one has published trunk lengths at this coverage. If either is off, every length is off in the same direction.
  3. Lengths are not comparable to published lengths. They are internally consistent, which is what the tool needs, but do not quote them.
  4. Nearly two thirds of wingspans are modeled, with a stated error.
  5. The flight posture is uniform: the same implied wing position and body axis for every bird.
  6. One figure per species. No sex, age, subspecies, or individual variation, though some species differ by 30% between sexes.

If you know a source that fixes any of these, particularly a validation set for head or trunk length or an openly licensed body-length dataset, that is the most useful contribution anyone could make. Adopting a source follows the process in river/requirements/imports/imports.md and the source-adoption path in docs/building.md.

Corrections are welcome, especially to the biology. Report them on the project's issue tracker.

Rendered from BirdView's repository and served here, so citations do not depend on the network.