DeepEarth Plant Habitats

US native plant species distributions at 240m scale

A new state of the art: more accurate than SINR, iNaturalist, BIEN and Daru 2024, PNAS on the same independent field plots, with maps 75x finer than Daru's.

16,942
native species
494 from relatives
240
m resolution
75× finer than Daru
0.964
median AUC
vs 0.951 for SINR · higher in 79% of 4,282 tests
440,885
test plots
independent field surveys
Daru (2024) published · ~18 km
Daru's method 240 m
DeepEarth joint model · 240 m
Figure 1. Cercis occidentalis. At 240 m the joint model resolves habitat that 18 km maps cannot.

01Findings

  1. State of the art on independent field plots. Median AUC 0.964 against 0.951 for SINR, the strongest published model, retrained; higher in 79% of 4,282 paired tests.
  2. Ahead of every other published map. Higher in 85–98% of paired tests against SINR's released models, iNaturalist, BIEN and Daru (2024).
  3. Also far from any record. Beyond 10 km from a species' records: 0.939 against 0.924 for SINR, higher in 71% of 4,131 tests.
  4. Each step of the method adds accuracy. On the same 88 tests: Daru's published maps 0.905, his method at 240 m 0.919, trained on real records 0.940, one joint model 0.969.
  5. One model for all 16,942 species. Beats MaxEnt fitted to each species (0.963 vs 0.934, 92% of 2,251 tests) and maps 494 species without records.
  6. Rare species borrow from their relatives. With 5 records, a species scores 0.927 against 0.895 for MaxEnt on the same records (higher for 73% of 725 species).
  7. All maps in 9.8 GB. About 19× smaller than per-species GeoTIFFs, with negligible loss: median AUC changes by 0.0011.
  8. Fast. All 16,448 species fitted in 1848 seconds on one GPU, once the shared representation is learned (55 minutes).

02Model comparison

Every published model, the same plots

Validated on independent field plots. Every comparison is paired: one test is one species on one plot source, scored for both models.

Choose a species

Larrea tridentata Creosote Bushphoto © larivera, some rights reserved (CC BY-NC)

Uniola paniculata Sea Oatsphoto © Alex Abair, some rights reserved (CC BY)

Goodyera tesselata Checkered Rattlesnake Plantainphoto © John Boldt, some rights reserved (CC BY-NC)

Eriophyllum lanatum Common Woolly Sunflowerphoto © p38, some rights reserved (CC BY-NC)

Silene virginica Fire Pinkphoto © ravenclagg5, some rights reserved (CC BY-NC)

a DeepEarth (ours)

AUC VegBank 0.996 · AIM 0.990AUC VegBank 0.997AUC VegBank 0.990AUC VegBank 0.922 · AIM 0.855AUC VegBank 0.960

DeepEarth (ours) map of Larrea tridentataDeepEarth (ours) map of Uniola paniculataDeepEarth (ours) map of Goodyera tesselataDeepEarth (ours) map of Eriophyllum lanatumDeepEarth (ours) map of Silene virginica, open full screen

b SINR (coordinates + environment)

AUC VegBank 0.991 · AIM 0.986AUC VegBank 0.996AUC VegBank 0.991AUC VegBank 0.896 · AIM 0.869AUC VegBank 0.931

SINR (coordinates + environment) map of Larrea tridentataSINR (coordinates + environment) map of Uniola paniculataSINR (coordinates + environment) map of Goodyera tesselataSINR (coordinates + environment) map of Eriophyllum lanatumSINR (coordinates + environment) map of Silene virginica, open full screen

c iNaturalist geomodel

released weights · AUC VegBank 0.990 · AIM 0.989Open Range Map (binary) · AUC VegBank 0.964Open Range Map (binary) · AUC VegBank 0.986Open Range Map (binary) · AUC VegBank 0.678 · AIM 0.672Open Range Map (binary) · AUC VegBank 0.926

iNaturalist geomodel map of Larrea tridentataiNaturalist geomodel map of Uniola paniculataiNaturalist geomodel map of Goodyera tesselataiNaturalist geomodel map of Eriophyllum lanatumiNaturalist geomodel map of Silene virginica, open full screen

d BIEN range

polygon (binary) · AUC VegBank 0.849† · AIM 0.916polygon (binary) · AUC VegBank 0.619†polygon (binary) · AUC VegBank 0.979†polygon (binary) · AUC VegBank 0.700† · AIM 0.739polygon (binary) · AUC VegBank 0.920†

BIEN range map of Larrea tridentataBIEN range map of Uniola paniculataBIEN range map of Goodyera tesselataBIEN range map of Eriophyllum lanatumBIEN range map of Silene virginica, open full screen

e Daru et al. 2024

published MaxEnt · AUC VegBank 0.897 · AIM 0.924published MaxEnt · species not availablepublished MaxEnt · species not availablepublished MaxEnt · species not availablepublished MaxEnt · AUC VegBank 0.962

Not available
no published map in the Dryad extract (557 species)

Not available
no published map in the Dryad extract (557 species)

Not available
no published map in the Dryad extract (557 species)

Daru et al. 2024 map of Larrea tridentataDaru et al. 2024 map of Uniola paniculataDaru et al. 2024 map of Goodyera tesselataDaru et al. 2024 map of Eriophyllum lanatumDaru et al. 2024 map of Silene virginica, open full screen

f Evidence: survey plots

7,508 presences · 117,345 absences (VegBank · AIM)124 presences · 53,673 absences (VegBank)80 presences · 53,717 absences (VegBank)1,529 presences · 123,324 absences (VegBank · AIM)55 presences · 53,742 absences (VegBank)

Evidence: survey plots map of Larrea tridentataEvidence: survey plots map of Uniola paniculataEvidence: survey plots map of Goodyera tesselataEvidence: survey plots map of Eriophyllum lanatumEvidence: survey plots map of Silene virginica, open full screen

Continuous models: score rank within CONUS land (top share)

all landtop 10%top 1%top 0.1%
  • Binary range (inside)
  • Published binary range outline
  • Presence plot
  • Absence plot
Two species chosen in advance and three spanning DeepEarth's margin over SINR (18th–79th percentile of 4,282 tests). Continuous maps are ranked within US land, so models on different scales share one colour scale; † BIEN on VegBank and FIA plots is not independent.
Against every published model, on identical tests
ModelAll plotsVegBankBLM AIMUSFS FIA> 10 kmfrom records
SINR (coordinates + environment)Cole et al. 2023; retrained, strongest configuration0.964 / 0.95179% · 4,2820.969 / 0.95782% · 2,7500.944 / 0.92775% · 1,2810.957 / 0.95071% · 2510.939 / 0.92471% · 4,131
SINR (coordinates)released model0.964 / 0.94385% · 4,2820.969 / 0.95187% · 2,7500.944 / 0.90881% · 1,2810.957 / 0.94380% · 2510.939 / 0.91674% · 4,131
SINR (distilled)released model0.964 / 0.93985% · 4,2820.969 / 0.94488% · 2,7500.944 / 0.91278% · 1,2810.957 / 0.94678% · 2510.939 / 0.91872% · 4,131
iNaturalist geomodelreleased weights; continuous0.933 / 0.87594% · 1250.943 / 0.89196% · 790.922 / 0.81379% · 240.874 / 0.834100% · 220.912 / 0.85093% · 125
iNaturalist range mapsOpen Range Maps; binary0.963 / 0.87498% · 5,3780.970 / 0.89199% · 3,3440.940 / 0.81298% · 1,7750.955 / 0.86499% · 2590.938 / 0.84394% · 5,197
BIEN range mapsbinary; built partly from VegBank and FIA plots0.937 / 0.74398% · 1,8540.969 / 0.841not independent0.937 / 0.74398% · 1,8540.956 / 0.861not independent0.909 / 0.72396% · 1,850
Daru's method at 240 mDaru (2024) MaxEnt recipe re-fitted at 240 m on the same plots0.964 / 0.90396% · 1,8420.968 / 0.90997% · 1,2890.945 / 0.83694% · 3630.946 / 0.90097% · 1900.939 / 0.87987% · 1,769
Daru (2024)published maps, ~18 km0.961 / 0.89795% · 2590.971 / 0.91596% · 1410.938 / 0.84594% · 1040.972 / 0.915100% · 140.931 / 0.87885% · 255

Median AUC, DeepEarth / published model; below, the share of paired tests DeepEarth wins and their number. BIEN on BLM AIM plots only (its ranges used VegBank and FIA plots). Shaded: not significant (Wilcoxon, p ≥ 0.05).

Each step of the method

Daru's published maps, his method rebuilt at 240 m, MaxEnt trained on real records, and the joint model, on the same species and plots.

VegBank, BLM AIM and USFS FIA plots together · 88 species–plot-source pairs

0.500.550.600.650.700.750.800.850.900.951.00AUC0.905Daru (2024)published · ~18 km0.919Daru's method240 m0.940MaxEnt per speciesoccurrences · 240 m0.969DeepEarthjoint model · 240 m0.500.550.600.650.700.750.800.850.900.951.00AUC0.905Daru (2024)published · ~18 km0.919Daru's method240 m0.940MaxEnt per speciesoccurrences · 240 m0.969DeepEarthjoint model · 240 m
Figure 3. AUC per species and median; 1 = perfect, 0.5 = chance; pooled over 88 species–plot pairs.
Median AUC by model and plot source
ModelAll plots88VegBank58BLM AIM19USFS FIA11
Daru (2024)published · ~18 km0.9050.9050.8700.911
Daru's method240 m0.9190.9270.8920.914
MaxEnt per speciesoccurrences · 240 m0.9400.9410.9330.917
DeepEarthjoint model · 240 m0.9690.9640.9720.970

Species with at least 20 presence plots and all four maps; n under each source.

More results

Every paired test against every published model. Medians, the share of tests DeepEarth wins and the Wilcoxon p-value, on all plots and on plots far from records.
ModelPlotsTestsDeepEarthModelHigher forp
SINR (coordinates + environment)All plots4,2820.9640.95179%< 10−300
VegBank2,7500.9690.95782%< 10−242
BLM AIM1,2810.9440.92775%< 10−68
USFS FIA2510.9570.95071%< 10−11
All plots, > 10 km from records4,1310.9390.92471%< 10−163
VegBank, > 10 km2,6050.9460.93072%< 10−114
BLM AIM, > 10 km1,2760.9140.89670%< 10−45
USFS FIA, > 10 km2500.9400.93771%< 10−7
SINR (coordinates)All plots4,2820.9640.94385%< 10−300
VegBank2,7500.9690.95187%< 10−300
BLM AIM1,2810.9440.90881%< 10−107
USFS FIA2510.9570.94380%< 10−19
All plots, > 10 km from records4,1310.9390.91674%< 10−230
VegBank, > 10 km2,6050.9460.92576%< 10−157
BLM AIM, > 10 km1,2760.9140.87972%< 10−67
USFS FIA, > 10 km2500.9400.93274%< 10−11
SINR (distilled)All plots4,2820.9640.93985%< 10−300
VegBank2,7500.9690.94488%< 10−300
BLM AIM1,2810.9440.91278%< 10−88
USFS FIA2510.9570.94678%< 10−15
All plots, > 10 km from records4,1310.9390.91872%< 10−186
VegBank, > 10 km2,6050.9460.92575%< 10−142
BLM AIM, > 10 km1,2760.9140.88868%< 10−46
USFS FIA, > 10 km2500.9400.94062%< 10−4
iNaturalist geomodelAll plots1250.9330.87594%< 10−16
VegBank790.9430.89196%< 10−12
BLM AIM240.9220.81379%0.0039
USFS FIA220.8740.834100%< 10−6
All plots, > 10 km from records1250.9120.85093%< 10−15
VegBank, > 10 km790.9280.87195%< 10−11
BLM AIM, > 10 km240.9030.80979%0.0048
USFS FIA, > 10 km220.8520.815100%< 10−6
iNaturalist range mapsAll plots5,3780.9630.87498%< 10−300
VegBank3,3440.9700.89199%< 10−300
BLM AIM1,7750.9400.81298%< 10−276
USFS FIA2590.9550.86499%< 10−43
All plots, > 10 km from records5,1970.9380.84394%< 10−300
VegBank, > 10 km3,1690.9480.87493%< 10−300
BLM AIM, > 10 km1,7700.9100.77296%< 10−268
USFS FIA, > 10 km2580.9380.84898%< 10−40
BIEN range mapsAll plots1,8540.9370.74398%< 10−294
BLM AIM1,8540.9370.74398%< 10−294
All plots, > 10 km from records1,8500.9090.72396%< 10−283
BLM AIM, > 10 km1,8500.9090.72396%< 10−283
Daru's method at 240 mAll plots1,8420.9640.90396%< 10−284
VegBank1,2890.9680.90997%< 10−204
BLM AIM3630.9450.83694%< 10−52
USFS FIA1900.9460.90097%< 10−30
All plots, > 10 km from records1,7690.9390.87987%< 10−205
VegBank, > 10 km1,2190.9460.89085%< 10−135
BLM AIM, > 10 km3610.9110.80890%< 10−45
USFS FIA, > 10 km1890.9280.87994%< 10−27
Daru (2024)All plots2590.9610.89795%< 10−39
VegBank1410.9710.91596%< 10−21
BLM AIM1040.9380.84594%< 10−16
USFS FIA140.9720.915100%0.00012
All plots, > 10 km from records2550.9310.87885%< 10−29
VegBank, > 10 km1380.9370.90382%< 10−13
BLM AIM, > 10 km1030.9050.82186%< 10−14
USFS FIA, > 10 km140.9420.900100%0.00012

Median AUC; a test is one species on one plot source; p: Wilcoxon signed-rank, paired, two-sided. Distance to records: plots more than 10 km from every training record of the species. BIEN on VegBank and FIA is left out (not independent). Per-test AUCs: work/benchmark/vs_sinr/cards_ringpot_full_shore_maxent240.csv.

Against Daru (2024) on every species with a published map. 0.960 vs 0.896, higher for 96% of 259 tests (190 species).
PlotsTestsDeepEarthDaruHigher forp
All plots2590.9600.89696%< 10−39
VegBank1410.9710.91096%< 10−21
BLM AIM1040.9380.85395%< 10−16
USFS FIA140.9720.916100%0.00012

Median AUC; a test is one species on one plot source; p: Wilcoxon signed-rank, paired.

Against per-species MaxEnt on every species with both maps. 0.963 vs 0.934, higher for 92% of 2,251 tests (1,651 species).
PlotsTestsDeepEarthMaxEntHigher forp
All plots2,2510.9630.93492%< 10−299
VegBank1,5470.9660.93894%< 10−223
BLM AIM4800.9430.90689%< 10−53
USFS FIA2240.9490.93587%< 10−25

Median AUC; a test is one species on one plot source; p: Wilcoxon signed-rank, paired.

Accuracy by number of training records. Median AUC stays at or above 0.92 from fewer than 50 to 5,000 records.
0.800.850.900.951.00< 5050–199200–9991,000–4,9995,000 (cap)training records per species (thinned to ~1 km)221841,0122,32364101716851,0845945206VegBankBLM AIMUSFS FIA0.800.850.900.951.00< 5050–199200–9991k–5k5ktraining records per species (thinned to ~1 km)VegBankBLM AIMUSFS FIA
Figure 4. Median AUC of the deployed maps by number of training records.
Species with few or no records. 5 records: 0.927 vs 0.895 for MaxEnt.
0.40.50.60.70.80.91.0chancemedian AUC on independent VegBank plots5 records, scored within the calibration areaMaxEnt on the same 5 records0.895DeepEarth, 5 records0.927MaxEnt on every record (reference)0.9545 records, the model's score aloneDeepEarth, 5 records0.9350.40.50.60.70.80.91.0chancemedian AUC on independent VegBank plots5 records, scored within the calibration areaMaxEnt on the same 5 records0.895DeepEarth, 5 records0.927MaxEnt on every record (reference)0.9545 records, the model's score aloneDeepEarth, 5 records0.935
Figure 5. Median AUC with 5 records, 725 held-out species: each keeps 5 of its records, the model's stage 2 is refitted, and the species is scored on the VegBank test plots.

Plots over 10 km from any training record (VegBank test half): joint 0.941, MaxEnt 0.917.

Storage. 9.8 GB for all maps; median AUC 0.972 before storage and 0.971 after.
Per-species rasters, uint8, clipped to calibration areas458.5 GBPer-species compressed GeoTIFFs (zstd)186.4 GBDeepEarth: one shared field + 512 numbers per species9.8 GBPer-species rasters (uint8, clipped)458.5 GBPer-species GeoTIFFs (zstd)186.4 GBDeepEarth shared field9.8 GB-0.008-0.0040+0.004+0.0080200400600AUC of the stored map minus AUC of the full model-0.0080+0.0080200400600stored minus full-model AUC
Figure 6. Storage of all maps; change in AUC from storage, per species.
Speed. 16,448 species fitted in 1848 seconds; MaxEnt in PyTorch: 7–8× faster fitting, 263 million cells per second.
Arctostaphylos auriculata162 s21 sFestuca californica213 s30 sVaccinium ovalifolium211 s31 smaxent.jar (CPU)PyTorch (one GPU)Arctostaphylos auriculata162 s21 sFestuca californica213 s30 sVaccinium ovalifolium211 s31 smaxent.jar (CPU)PyTorch (one GPU)
Figure 7. Time to fit one species: maxent.jar against the PyTorch port.
Equations and code. The model, its training objective, the evaluation and the map store, each linked to the code that computes it.
Acer negundo: DeepEarth native range, contiguous United States at 240 m

Figure 8. Acer negundo, joint model, contiguous United States at 240 m. 1 of 16,942 species distributions modeled by DeepEarth.

native range, low to high suitabilitysuitable, outside the rangerest of the 48 states