Cascadia Distribution MapsNorth America Distribution Maps
The Cascadia pilot: modeled distributions of 242 birds at a standard eBird checklist in Oregon, Washington, Idaho and British Columbia, at 1 km, 2002-2018, with population trends, habitat change and checks against Breeding Bird Survey counts.
Modeled distributions across the United States, Canada and Mexico at 3 km: 610 birds from eBird checklists (2002-2018, and 2017-2024 with AlphaEarth satellite embeddings), and 435 mammals, amphibians and reptiles from GBIF records (beta).
A separate map set: Cascadia Distribution Maps (1 km pilot for birds, with deeper analysis) →
Methods
Pilot methods as of November 2026. Every setting was tested on simulated species with known answers before it was used on real species. Where a test failed, this page says so.
What is mapped
The tool covers every bird in the region with at least 1,000 June-July detections, plus the pilot species: 242 in all. Each map shows the modeled probability that one standard checklist reports the species: 60 minutes, 1 km traveled, one observer, 7 am, July 1. It is calculated on a 1 km grid (1.59 million km², Oregon, Washington, Idaho and British Columbia) for each year from 2002 to 2018. It is not occupancy, abundance, or a count of birds.
Data
- Records: eBird complete checklists, so checklists without the species count as non-detections. Stationary or traveling, up to 5 hours, up to 5 km, up to 10 observers, June-July, 2002-2018, with shared checklists collapsed. That leaves 269,172 checklists. Records filed under an eBird form or intergrade of a species count for that species. Until about 2016 most American Coot records were filed as a form, so leaving these out created a false jump, which our quality checks caught. Only American Coot and Red Crossbill lost over 1% of their records, so only those two were refit; for the rest the difference is at most 0.4%.
- Withheld records: eBird withholds some sensitive species, sometimes only by state or date. Treating those gaps as non-detections would create false absences, so affected regions are left unmapped. White-headed Woodpecker and American Barn Owl are fitted and mapped for the US only, because their BC records appear to be withheld (0 detections on 10,483 and 36,756 BC checklists inside their mapped ranges, where 92 and 149 were expected).
Covariates
| Group | Variables | Source |
|---|---|---|
| Land cover | Share of 14 classes, within the 1 km cell and the surrounding 5 km | ESA CCI, annual |
| Vegetation | NDVI and EVI (breeding season, annual), and 5 km breeding NDVI | MODIS, annual |
| Terrain and water | Mean elevation, elevation variation, surface water occurrence | Static |
| Climate | June-July maximum temperature, minimum temperature and precipitation | Daymet, TerraClimate, CHELSA: one product per model |
| Survey effort | Duration, distance, observers, hour, day of year, stationary or traveling | Each checklist. Fixed at the standard checklist when mapping |
Choosing among overlapping climate products by out-of-sample prediction is itself a result of this project. The Show menu above the map displays most of these layers through time.
The models
All models are LightGBM binary classifiers: 400 trees, learning rate 0.05, 31 leaves, at least 50 checklists per leaf, 80% sampling of rows and covariates, L2 penalty 1. These settings are fixed, not tuned per species. 85% of checklists train the model. The other 15%, thinned to one checklist per 3 km cell, week and year, calibrate probabilities by Platt scaling. Apart from the experimental model, no model has a year term, so a map can change between years only through its covariates.
| Model | How it is produced | Best used for |
|---|---|---|
| Average of three climate models | Equal-weight mean of the Daymet, TerraClimate and CHELSA models | Where conditions suit the species. Best on new places |
| Base (land cover and terrain, no climate) | Land cover, vegetation and terrain only. Mean of three fits with different seeds | Change driven by land cover |
| Base + Daymet, + TerraClimate, + CHELSA | Base covariates plus that product's three climate variables. One fit each | Comparing climate products. Single fits are noisier than the average |
| Base + Daymet + Landsat | Base + Daymet, plus 17 Landsat layers (1 km summaries of 62.5 m summer composites, 1984-2025): greenness, burn and moisture indices, tasseled cap, within-cell patchiness, and the share, greenness and regrowth rate of young stands (1-8 years after harvest or fire). One fit | Young-forest habitat and its quality, for example regrowth after harvest. Tests |
| Experimental: location + year | Base covariates plus grid location and year. Mean of three fits. Failed its test with known truth | Broad patterns of spread or decline in reporting, read with the warnings below |
Change views
- Each year: the model applied to that year's covariates.
- 5-year running mean: the mean of a year and the two on each side.
- Change: the 2014-2018 mean minus the 2002-2006 mean.
- Yearly trend: the linear slope across 2002-2018, in probability per year.
- 25 km views: the mean over 25 km blocks.
Habitat area panel. The side panel's Measure menu gives three readings of the same yearly maps, for the region and model chosen: the mean probability; the expected area, the sum of the probabilities over 1 km cells in km², a smooth index in which a broad rise in low probabilities counts as much as a rise in the core; and the area above a threshold, the area of cells at or above a cutoff chosen per species and model to make the sum of the false-negative and false-positive rates smallest on held-out checklists (the maximum of sensitivity + specificity − 1). Areas are counted inside the range clip when the species is clipped. Thresholds are chosen at actual checklist effort while the maps use the standard checklist, and they are low for rare species, so the thresholded area is an index, not a measured occupied area. Neither is a population trend. Tested on 18 simulated species with known habitat change: the expected area followed the true series closely (median correlation 0.94 for the average of three climate models, false change under 0.3% when nothing changed). The thresholded area failed: about 80% of cells sit above the cutoff every year, so it barely moves (median correlation 0.69 for the same model). Use it for the size of the area, not for its change.
Landsat model
Landsat composites (Collection 2, June-September, Landsat 4-9 harmonized to one sensor scale; water, cloud, shadow and snow masked) were checked against MODIS (NDVI r 0.98) and across sensors (Landsat 7 vs 8 in the same year, r 0.99). The harvest detector agrees with Hansen's Global Forest Change on where forest was lost over 2001-2024 (r 0.82 per 1 km cell) but is noisier year by year, so the model uses multi-year young-stand measures, which start in 1994.
- Simulated species whose habitat depends on young-stand quality (share of young stands and how green they regrow): the current models follow their change with correlation 0.08; with Landsat, 0.96. Map pattern 0.60 vs 0.91. Change shown where nothing changed: 1.36 times the current model, within the 1.5 limit. Weather-driven change: 0.86 vs 0.89. Passes all three tests set in advance.
- Real species, held-out checklists (242 species, same rows as the Daymet model): prediction error 0.5% lower for new places (190 of 242 species better; young-forest species 0.7%, 14 of 15), but 0.2% higher for later years (97 of 242 better). So Landsat helps a little with where species are, especially young-forest species, and does not improve prediction of later years.
- The simulated test used the same Landsat layers to build the truth, so it shows the model can use this information when a species depends on it, not that real species do. A Landsat-only model, which could reach back to 1985, created too much false change (1.62 times) and is not shown.
62.5 m detail (2018 map)
At 1 km a cell that is part mature forest and part clearcut gets one value, so a clearcut can look like habitat for a mature-forest bird such as Hermit Thrush. The "62.5 m Landsat detail" model uses, at each checklist's location, the Base + Daymet covariates interpolated smoothly between 1 km cells (so the map has no steps at cell edges) plus the Landsat greenness (NDVI) and burn ratio (NBR) of the 62.5 m pixel, their means within 250 m, and the share of open ground (NBR below 0.1) within 250 m. The map predicts every 62.5 m pixel for 2018 at the standard checklist.
- Test with known truth. 12 simulated species: forest birds that avoid open ground, early-seral birds that prefer it, and birds with no fine-scale preference. Location error was simulated honestly: a travelling checklist's bird could be anywhere along its route. Inside 1 km cells the map ranked sub-pixels as the truth did with a median rank correlation of 0.71 to 0.84 (needed 0.5); for birds with no fine-scale preference it added little invented detail (0.14 of the between-cell spread; limit 0.25); and the 1 km pattern improved (0.92 vs 0.87).
- Real species. On held-out checklists whose location is good to about 1 km (five 50 km folds), the 62.5 m model predicted better than the 1 km Base + Daymet model for 225 of 241 species (median log-loss improvement 2.0%, 95% interval 1.7 to 2.1%), most for shoreline and shrub-steppe birds. A first version that repeated each 1 km value over the cell's pixels drew visible 1 km squares and did less well (1.2%); the interpolated version replaced it.
- How much fine detail to expect. For many forest birds, elevation and climate still shape most of the map, and the contrast between a clearcut and the forest around it is modest (for example Hermit Thrush). It is clearest where habitat edges are sharp: shorelines, shrub-steppe, and young versus mature forest for early-seral birds.
- Limits. eBird records one point per checklist, so the model learns fine-scale habitat from checklists whose location is good to roughly 100 m to 1 km: read stands and landscapes, not single pixels. One year (2018) is shown; no change views at this scale.
Why it's here: what drives each map
The side panel's "Why it's here" card shows what each model relies on. Both measures come from the same five held-out spatial folds used for the model-quality scores: each fold's model is fitted on the other four folds and examined on its own checklists, and the bands and ranges show the spread across folds.
- Share of the map's variation. Each group of covariates (land cover, vegetation greenness, terrain and water, climate, and Landsat stand condition in the Landsat model) is shuffled among held-out checklists, keeping the group's variables together, with survey effort fixed at the standard checklist. The group's importance is how much the prediction moves (mean squared change on the logit scale); shares are importances divided by their total. Groups with many variables are not favored by construction, but a group can borrow credit from a correlated one.
- Survey effort (duration, distance, observers, date, hour) is not habitat and is reported separately: how much held-out prediction error rises when it is shuffled. The maps hold it at the standard checklist.
- Response curves (partial dependence) for the six most important variables: each variable is set to values across its range (2.5th to 97.5th percentile of checklists) on 2,000 held-out checklists at the standard checklist, and the predictions averaged. All curves for a species share one vertical scale, so taller curves matter more. For the average of the three climate models, climate curves average the three products at the same percentile of each, drawn on Daymet's axis.
Test with known truth (Pacific Northwest, 1 km). 18 simulated species with known drivers (climate, needleleaf forest and greenness, or elevation alone) were fitted with the three climate models. Pass rules were set before the results: the top group correct for at least 14 of 18; median share error at most 0.10; curve shapes matching the truth (rank correlation at least 0.8) for at least 80% of true drivers. Results: top group correct for 16 of 18 (both misses were near-ties in the truth too); median share error 0.05; curve shapes matched for 97% of 72 drivers. Known biases: climate's share was understated by about 0.1 on average, mostly credited to land cover instead; elevation-only species kept 73-94% of the share on terrain (truth 100%), so elevation was not wrongly credited to temperature. Curve heights were about half the true effect, partly because the map shows detection at a standard checklist rather than occupancy and partly because tree models flatten curves: read the shapes and the relative heights, not the values. With real species, "drivers" are correlates of where birds are reported, which may stand in for unmeasured causes.
Range clip
The clip uses all extant BirdLife polygons, excluding vagrant ones, plus a 50 km buffer. Cells outside it are drawn faint: flagged, not removed. If under 95% of a species' June-July detections fall inside, the map is not clipped and says that the range map and the data disagree: 189 of 242 species are clipped. Some of the 53 unclipped species look like differences in taxonomy, where BirdLife splits what eBird treats as one species. Range maps do not feed into the models.
Population trends
Population trends come from a separate analysis of the North American Breeding Bird Survey, 1993-2025: a hierarchical first-difference negative binomial model with state by Bird Conservation Region strata, observer-by-route effects and a first-year effect. Definitions follow USGS (Link et al. 2020). Planted-truth recovery and comparison with USGS estimates were run before these fits, and all final fits had zero divergences. These trends measure population change and are a different quantity from the maps.
Validation
Simulated species with known truth. Run on the real checklists and covariates, with effort-dependent detection.
| Test | Result | Reading |
|---|---|---|
| Map pattern, Spearman with truth, species on 5% and 12% of checklists | 0.91-0.95 | Passes the 0.8 gate |
| Map pattern, species on 25% of checklists | Daymet 0.76, TerraClimate 0.88, no-climate truth 0.57 | Fails for two of three: widespread species are less reliable |
| Real vs false change, with the true climate product / the average | 1 km: 2.0 / 2.5 times. 25 km: 5.0 / 6.6 times (5-year view) | Fine detail is often noise. These are best-case figures |
| Weather-driven change, 25 km correlation with truth, average vs base | 0.72 vs 0.12 | Base barely sees weather |
| Land-cover change, 25 km correlation with truth, base vs average | 0.68 vs 0.64. Regional series 0.50-0.75 vs 0.32-0.41 | Base gives the better read |
| Steady shift unrelated to any covariate | Correlation −0.14 to 0.11 | Fails: the maps cannot see this |
Real species, held-out checklists. All 242 species, with full scores on the Model quality tab. New places: five folds of 50 km blocks. Later years: trained on 2002-2014, tested on 2015-2018. American Barn Owl and White-headed Woodpecker are scored for the US only.
| Average of three climate models | New places | Later years |
|---|---|---|
| AUC, median (range) | 0.89 (0.71-0.99) | 0.92 (0.68-0.99) |
| PR-AUC, median (range) | 0.25 (0.02-0.79) | 0.35 (0.01-0.81) |
| Deviance explained, median (range) | 0.25 (0.05-0.62) | 0.32 (0.03-0.64) |
| Mean prediction / observed, median (range) | 0.80 (0.39-1.01) | 0.90 (0.38-1.41) |
| Log loss improvement over one base fit, median across species (95% interval) | 2.5% (2.2 to 2.8) | 1.3% (1.1 to 1.5) |
Part of that improvement comes from averaging three fits rather than from climate. In the first 16 pilot species, compared fairly against three base fits, the gain was 2.2% (1.1 to 3.4) on new places and a tie on later years, 0.2% (−0.6 to 1.1). So climate helps to map new places but has not been shown to help predict later years. Probabilities run low, especially for rare species, so read them as relative. These tests say nothing about whether change over time is right: that is what the tests with known truth and the BBS comparison are for.
Independent check against BBS counts
The North American Breeding Bird Survey is a standardized roadside count that does not feed the maps, so it is an independent test of them. For each species and model, a Bayesian hierarchical model relates BBS counts on each route and year, 2002-2018, to the mean mapped probability within 100 m of the route path. It separates two questions:
- Between routes: do routes with more mapped habitat have more birds?
- Within routes: in years when the map shows more habitat along a route than that route's own average, are counts higher? This is the test of mapped change.
Both are elasticities: 1 means counts rise in proportion to mapped habitat, 0 means no relation. The model also has route effects, each route's own trend, year effects shared by all routes (weather, detectability, regional population swings), observer-by-route effects and a first-year-observer effect, with negative binomial counts. Route paths: USGS National Atlas lines for Oregon, Washington and Idaho, and lines traced from Environment and Climate Change Canada's route maps for British Columbia; 351 of 392 routes.
Tested first with simulated counts on real route designs (50 to 227 routes): effects recovered with bias under 0.15 and 90-100% interval coverage; false positives 0-10% when there was no effect. Two limits: on a species with very sparse counts (about 0.4 birds per route) the sampler got stuck, so fits that do not converge are not shown; and when the yearly maps are noisy versions of the habitat birds respond to, the within-route effect shrinks to about 0.35-0.5 of its true size, so within-route values here are understated.
Loading…
Results, 211 species. Where birds are: every model passes clearly. Routes with more mapped habitat have about proportionally more BBS birds (mean elasticity 0.92 to 0.96), with a clear positive relation for about 90% of species and a negative one for none. Change: year-to-year habitat change along a route predicts BBS change only weakly for the habitat models (about 0.1, clear for 7-14% of species). Because noisy yearly maps understate this, the true link is likely larger, but still small. Read mapped change as change in habitat conditions, not as a forecast of local counts. The experimental location + year model follows BBS change much more closely (0.46), but its year and location terms mean that is not a pure habitat signal.
Known limits
- Why the maps change so little. Each model learns one pooled relationship with habitat from all years. A species that is expanding or declining for reasons the covariates do not capture looks nearly the same every year. Eurasian Collared-Dove stays near 0.05 from 2002 to 2018, while the BBS index rises from 0.002 to 9.4. The experimental model below can show this kind of change, but it has not passed its test.
- The climate and base maps do not track population trends. Across 17 species, 2002-2018 change in the average model correlates 0.20 (−0.42 to 0.74) with BBS trends. A stable map does not mean a stable population.
- eBird yearly index is not shown. It tracks BBS (rank correlation 0.83, interval 0.57 to 0.94), but failed its test with known truth: with no real trend it reported a rise of about 0.6%/yr for species on 2% of checklists, and it reached the wrong direction in BC where trends differed by region.
- Covariates have their own errors. Temperature products disagree on trends, and the 2015-2016 ESA land cover version change was checked and looks acceptable, but cannot be ruled out as a source of small changes. Daymet temperature shows blocky changes of several °C in British Columbia that TerraClimate and CHELSA do not (16% of BC cells change over 2 °C in maximum temperature, against none). This appears to affect the species maps little: the Daymet model's 25 km change in BC correlates about 0.09 with that temperature change (median across species, ignoring sign), and the average's change map correlates 0.94 with an average that leaves Daymet out.
- Experimental location + year model. Shown because it is the only tested model that can show this kind of change, not because it passed. Base covariates plus grid location and year, as a mean of three fits, so year effects can differ by place. It is the only tested model that shows spread: Eurasian Collared-Dove rises from 0.001 in 2002 to 0.064 in 2018. Across all 242 species, a single fit beats one base fit by 3.3% in log loss on new places (2.9 to 3.8) and 2.8% on later years (2.3 to 3.3), and the average of three climate models by 0.8% (0.6 to 1.3) and 1.3% (1.1 to 1.7). Location may carry much of this, and the later-years test holds year at 2014, so it does not test the year effects. For the 17 species with BBS trends so far, its 2002-2018 regional trends correlate 0.73 (0.36 to 0.92) with BBS, against 0.20 for the average. It did not pass its test with known truth: real change was 2.4-2.7 times false change (limit 3), correlation for opposite regional trends was 0.39-0.44 (limit 0.7), and land-cover change was weakly recovered (0.43-0.55, against 0.82-0.85 for base). It runs about 1.8%/yr above BBS and can read changes in eBird reporting as real, so it is neither a habitat map nor a population trend.
- BBS comparison. So far it covers 17 of the 242 species.
- Season and area. The maps cover the June-July breeding season in the pilot region only.
North America methods as of November 2026, beta. The models are the pilot's, rebuilt on a 3 km grid for the continent; the pilot page has the full description of each setting. Every setting was tested on simulated species with known answers; where a test failed, this page says so.
What is mapped
610 birds: every species with at least 1,000 June-July detections on complete eBird checklists in 2002-2020, spread over at least 10 separate 50 km blocks. Each map shows the modeled probability that one standard checklist (60 minutes, 1 km, one observer, 7 am, early July) reports the species, in 3 km cells across the United States (without Hawaii), Canada and Mexico. The western Aleutians, which cross 180 degrees, are left out.
Data and withheld records
- Records: 9.1 million complete June-July checklists, 2002-2025, with the pilot's filters (stationary or traveling, up to 5 hours and 5 km, up to 10 observers; shared checklists counted once). Early years are thin: about 25,000 checklists in 2002 against 1.1 million in 2024.
- Withheld records: eBird hides some sensitive species in some states, so their absence there is not a real absence. A state is left out of a species' training, and drawn faint and marked "not mapped", when the species is on eBird's Sensitive Species list for that state with dates overlapping June-July (list as of May 2026): 92 species-state pairs in 21 species. Check: of 60 listed pairs inside the species' mapped breeding range, 54 have no detections at all and 6 have 1 to 13, where hundreds would be expected. An earlier screen based on zero detections alone removed natural absences too and was replaced.
Models and covariates
| Model | Years | Covariates |
|---|---|---|
| Average of three climate models | 2002-2018 | ESA CCI land cover, MODIS NDVI and EVI, terrain (Copernicus 90 m DEM), surface water, plus Daymet, TerraClimate or CHELSA June-July climate; equal-weight mean of the three |
| Base | 2002-2018 | Land cover, vegetation, terrain and water, no climate; mean of three fits |
| AlphaEarth era | 2017-2024 | AlphaEarth Foundations satellite embeddings (64 annual bands, averaged to 3 km) plus Daymet V4R1 June-July climate; mean of three fits |
The 5 km neighbourhood layers of the pilot become 9 km (3 x 3 cells); cross-validation blocks are 51 km; the coarse change views are 24 km. The single-climate models are shown on the pilot page only. Land cover ends in 2020 and ESA's 2019-2020 forest loss is mostly unconfirmed, so the main models stop in 2018. Range clip: BirdLife extant range plus 50 km, as on the pilot.
Tests with known truth, 3 km
18 simulated species driven by Daymet or TerraClimate climate or by terrain only, at 5%, 12% and 25% of checklists, on the real continental checklists and covariates.
| Test | Result | Reading |
|---|---|---|
| Map pattern, Spearman with truth (Daymet model) | 0.96, 0.98, 0.96 at 5%, 12%, 25% | Passes the 0.8 gate, including widespread species (which failed on the pilot) |
| Real vs false change, 5-year view, about 25 km | Average 4.3 times; Daymet model 4.5; base 2.9 | Average and climate models pass the 3 times gate; base sits just below it |
| Weather-driven change, correlation with truth, 25 km | Average 0.80, base 0.26 | As on the pilot, base barely sees weather |
AlphaEarth era: reliability
AlphaEarth embeddings summarize each year of satellite imagery as 64 numbers per location. They cover 2017 onward, so they are kept in their own model and years rather than mixed into the 2002-2018 series. Two tests with simulated species asked whether they show where species are, whether they follow real change, and whether they invent change.
- Pilot region, 2017-2020 (18 species, truth from land cover, vegetation, climate or terrain): embeddings + Daymet passed: pattern 0.89 (base + Daymet 0.92); false change 1.4 times that of base + Daymet (limit 1.5); real-change correlation 0.76 (0.82). Embeddings alone failed: they barely follow weather-driven change (0.44).
- North America, 2017-2024 (12 species): truth from terrain only, so nothing truly changes, or from Daymet climate plus terrain; embeddings + Daymet compared with a model of climate, terrain and water only.
| Test (median across species) | Embeddings + Daymet | Without embeddings | Gate | Result |
|---|---|---|---|---|
| Drift of the regional mean where nothing changed | -0.19 %/yr (worst 0.5) | 0.01 %/yr | within 1 %/yr | Passes |
| Map pattern, Spearman with truth | 0.94 | 0.99 | 0.8 | Passes |
| Real climate-driven change followed, 27 km | 0.90 | 0.91 | within 0.1 | Passes |
| Change shown where nothing changed, 27 km, 2021-24 vs 2017-20 | 4.2 times the model without embeddings | 1 | at most 1.5 times | Fails |
| Real change relative to that false change | 2.1 times | 9 times | 3 times on the other maps | Below standard |
Why it fails. Each year's embeddings are computed afresh from that year's imagery. Differences in cloud cover, the number and timing of images, the satellites available and the embedding model itself shift the 64 numbers even where the land has not changed. The species model reads those shifts as changes in habitat. Over large regions they cancel, which is why the regional mean does not drift; in any one place they do not.
What to use. The AlphaEarth maps show where species are about as well as the other models, and regional series are sound. The multi-year change views (2017-2020 against 2021-2024) are shown as requested, labeled as failing this test: read only large, consistent patterns, and confirm them with the 2002-2018 models or other data. Year-to-year change in this model is mostly noise. Embeddings cannot say why a species changed. Not yet tested: averaging embeddings over two or three years before fitting, which may reduce the noise.
Held-out checklists
All 610 species, scored on checklists left out of fitting: new places (five folds of 51 km blocks) and later years (2002-2018 models: trained to 2014, tested 2015-2018; AlphaEarth model: trained to 2021, tested 2022-2024). Medians across species, with the middle half in brackets; per-species scores are on the Model quality tab.
| Model | Test | AUC | PR-AUC | Deviance explained | Mean prediction / observed |
|---|---|---|---|---|---|
| Average of three climate models | New places | 0.97 (0.94-0.99) | 0.25 | 0.41 (0.32-0.51) | 0.82 (0.70-0.91) |
| Later years | 0.98 (0.95-0.99) | 0.34 | 0.47 (0.37-0.57) | 0.91 (0.80-0.98) | |
| Base | New places | 0.96 (0.92-0.98) | 0.20 | 0.34 (0.26-0.42) | 0.77 (0.65-0.88) |
| Later years | 0.97 (0.94-0.99) | 0.31 | 0.44 (0.34-0.53) | 0.90 (0.78-0.99) | |
| AlphaEarth era | New places | 0.97 (0.93-0.99) | 0.22 | 0.39 (0.30-0.47) | 0.94 (0.87-0.98) |
| Later years | 0.98 (0.94-0.99) | 0.26 | 0.44 (0.35-0.52) | 0.90 (0.81-0.97) |
- By country (average of three climate models, new places): United States AUC 0.97, deviance explained 0.39; Canada 0.93 and 0.28; Mexico 0.89 and 0.16, where checklists are far fewer.
- Not shown: a model is hidden for a species when it predicts held-out checklists in new places worse than a flat average: seven species-model pairs in six localized or introduced species: Northern Red Bishop (climate average); Rosy-faced Lovebird and Red-masked Parakeet (base); Arizona Woodpecker, King Eider, Rosy-faced Lovebird and Southern House Wren (AlphaEarth). Each species keeps its other models.
- Low predictions: for about one species in six, predictions run below 0.6 of what was observed in one of the tests; their maps say so and should be read as relative.
Why it's here: what drives each map
The side panel's "Why it's here" card shows what each bird model relies on, from the same five held-out spatial folds (51 km blocks) used for the held-out scores below.
- Share of the map's variation. Each group of covariates (land cover, vegetation greenness, terrain and water, climate) is shuffled among held-out checklists, keeping the group's variables together, with survey effort fixed at the standard checklist. The group's importance is how much the prediction moves (mean squared change on the logit scale); shares are importances divided by their total. Groups with many variables are not favored by construction, but a group can borrow credit from a correlated one.
- Survey effort (duration, distance, observers, date, hour) is not habitat and is reported separately: how much held-out prediction error rises when it is shuffled. The maps hold it at the standard checklist.
- Response curves (partial dependence) for the six most important variables: each variable is set to values across its range (2.5th to 97.5th percentile of checklists) on 2,000 held-out checklists at the standard checklist, and the predictions averaged. All curves for a species share one vertical scale, so taller curves matter more. For the average of the three climate models, climate curves average the three products at the same percentile of each, drawn on Daymet's axis.
Test with known truth. The same test was repeated on the 3 km grid with 18 simulated species and all 2.5 million June-July checklists of 2002-2018. It passed: top group correct for 14 of 18 (the minimum allowed), median share error 0.02, curve shapes matched for 93% of 72 true drivers. In two misses the credit moved between land cover and greenness, which are correlated, so read those two groups together. Most curve misses were for the summer maximum-temperature optimum where the true peak was shallow. Shown for the 2002-2018 models (the average of three climate models, and Base); not for the AlphaEarth model, whose embedding bands have no direct meaning.
Mammals, amphibians and reptiles (beta): presence-only maps from GBIF records
These groups have no checklist data like eBird's, so their maps come from presence-only records: GBIF occurrence records from 2002-2020 (GBIF.org 2026, doi:10.15468/dl.encvrb), most from iNaturalist in recent years, kept as one record per species, 3 km cell, year and month. Each map shows where conditions resemble the places a species was recorded, on a relative scale. It is not a probability, an abundance or a population trend, and there are no change views. Every species with at least 300 records is mapped: 155 mammals, 94 amphibians and 186 reptiles.
Maps that meet the standard and lower-performing maps. A map meets the standard when (1) in the simulation test, maps built from that many records matched the true pattern with a rank correlation of at least 0.8, and (2) where an independent survey exists, it scored an AUC of at least 0.7 there. That holds for 62 mammals, 17 amphibians and 38 reptiles. The other 318 maps are lower performing: they are hidden until you tick "Show lower-performing maps", carry a caveat giving the reasons, and their "Model quality explained" link sets out the evidence in plain language.
- Cleaning. Marine species (whales, seals, sea turtles, sea snakes), species of orders with no native North American members (zoo and captive records), non-native primates and domestic animals are left out. Names that differ only because a genus was renamed are merged (for example bullfrog under Lithobates and Aquarana); recently split species are not lumped. For species whose records agree with their range map (at least 95% within the range plus 50 km), isolated records more than 200 km outside it are left out as likely captives or errors.
- Model. Boosted trees comparing a species' records with all other records of its group (a "target-group" background). Because both come from the same observers, this cancels much of the pull of roads, cities and parks. Covariates as for the birds' Base + TerraClimate model, matched to each record's year. The map is the mean of the 2016-2020 predictions, scaled so the species' top 0.5% of land is 1.
- Simulation test. 18 invented species with known habitat, their records drawn the way each group's real GBIF records fall. Rank correlation between map and truth, at 300 / 1,000 / 3,000 records: mammals 0.64 / 0.80 / 0.87; amphibians 0.56 / 0.73 / 0.84; reptiles 0.56 / 0.72 / 0.85. So mammal maps meet the standard from about 1,000 records, amphibian and reptile maps from about 3,000. Change over time failed in every group: where nothing changed, the models still showed about as much change as where it did.
- Independent tests. Mammals: Snapshot USA camera traps (9,626 deployments, 2019-2023). Frogs and toads: the USGS North American Amphibian Monitoring Program calling surveys (NAAMP, 64,081 stop-years, 2002-2015, eastern and central US). Each site was scored by a model that never saw records from its own 51 km block. Boosted trees beat a simpler model in both (median AUC inside the recorded area 0.76 vs 0.72 for mammals, 0.73 vs 0.71 for frogs and toads). Widespread species often scored below 0.7 (for example coyote, beaver, bullfrog, green frog): when a species is found almost everywhere, records alone cannot separate better from poorer habitat. Salamanders, western amphibians and reptiles have no independent test.
- Range maps. Mammals and amphibians: IUCN Red List range maps (IUCN 2026, version 2026-1); reptiles: GARD 1.7 (Roll et al. 2017, Nature Ecology & Evolution 1:1677-1682; Caetano et al. 2022, PLoS Biology). As for the birds, areas more than 50 km outside the range are drawn faint (flagged, not removed) when at least 95% of the species' records fall inside; otherwise the map is not clipped and says the range and the records disagree. Range polygons themselves are not shown. Per-species credits appear with each map.
Known limits
- Uneven data. eBird coverage is far denser in the United States and southern Canada than in the north and in Mexico, and far denser after 2010 than before. Maps are least reliable where checklists are scarce: held-out scores in Mexico are clearly lower (above).
- Habitat change is not population trend. Change views show change in modeled habitat conditions. On the pilot, independent BBS counts confirmed where birds are (elasticity about 0.94) but tracked year-to-year habitat change only weakly.
- Climate after 2024. Daymet ends in 2024, so the AlphaEarth era stops there. MODIS vegetation is missing for 2024 in part of the continent; the AlphaEarth model does not use it.
AlphaEarth Foundations Satellite Embedding dataset produced by Google and Google DeepMind (CC-BY 4.0). Daymet V4R1: ORNL DAAC.
Model quality
How well each model predicted eBird checklists it was not trained on. New places: five folds of 50 km blocks, each predicted by a model fitted to the other four. Later years: fitted to 2002-2014, tested on 2015-2018.
Loading scores…
What the scores mean
- AUC
- The chance a checklist that reported the species is ranked above one that did not. 0.5 is no skill, 1 is perfect. It looks good for rare species almost regardless, so read it with the next two.
- PR-AUC and PR-AUC ÷ prevalence
- How precisely the model finds detections. Its baseline is the share of checklists reporting the species, so the ratio shows skill above that baseline: 1 is no skill.
- Deviance explained
- Improvement in log loss over predicting one constant, the held-out reporting rate. 0 is no better than that constant and below 0 is worse. That constant uses the test data's own rate, so this score is slightly conservative.
- Prediction ÷ observed
- Mean predicted probability over the observed reporting rate. 1 means no overall bias, below 1 means predictions run low.
- Calibration chart
- Checklists grouped by predicted probability, with the share in each group that reported the species. Points on the diagonal mean a predicted 10% happened about 10% of the time. Groups with under 50 checklists are left out.
Scores use actual checklist effort, while maps are drawn at a standard checklist. Base and location + year are scored as single fits, though their maps average three, so they slightly understate the mapped models. Scores are blank when a species has under 10 held-out detections in that region and test. Download all scores (CSV), shared under the same eBird Terms of Use.
