The Brazil Nut Research Atlas

The family is a Neotropical giant with a paleotropical tail: Bertholletia excelsa, the Brazil nut; Couroupita, the cannonball tree; Eschweilera and the genera Kew has recently split from it, the subtly flowered “matchbox” trees that dominate Amazonian forests; the timber genera Couratari and Cariniana; and the Old World Barringtonia of tropical coasts. COMPASS links each of the 401 accepted species to OpenAlex, tags every study, and geocodes each to its study site. What emerges is a portrait of a family the world eats, trades and logs, and has barely begun to catalogue.

About a third of studies (1,208 of 3,588) touch some economic dimension, and the pattern is unmistakably one of a standing-forest economy: edible product, commercial processing, non-timber forest-product harvest and trade, rural livelihoods, and sustainable-use management. The seed, the nut, is the single most-studied plant part. The Brazil nut is its emblem: a globally traded commodity gathered from wild trees that resist cultivation, depending on intact forest, its orchid-bee pollinators and the agoutis that crack and disperse it.

Research concentrates on a familiar few, and most of the family has little or no indexed research. The largest genus on the WCVP census is the Old World coastal genus Barringtonia, 73 species, 68% of them with no indexed paper. Eschweilera keeps 66 accepted species, 53% of them with no indexed paper, after Kew's recent revision moved many former Eschweilera names into Scottmoria, Chytroma, Imbiriba, Guaiania and Pachylecythis; together these matchbox trees remain a structural backbone of Amazonian forest, and the two largest segregates fare no better: Scottmoria 57% of 35 species, Chytroma 61% of 18. Gustavia is 55% without an indexed paper (26 of 47), Foetidia 72% (13 of 18). Many are trees with edible fruits, useful timbers and little-studied chemistry, potential products of a bioeconomy that still rests on a single nut.

Among least-concern species, 43% have no indexed research; among the vulnerable, 83%; among the data-deficient, 86%. Some of this is structural: a species is often listed as threatened because it is rare or narrowly ranged, the same traits that depress study, so it is an association rather than simple neglect. But the consequence is concrete: the very species a diversified forest economy might one day rest on are the ones being lost before they are documented. Conservation framing appears in just 6% of studies.

For where research sits against species ranges, coverage cells and access distances, see section 6; for the same sites set against environmental change and split by decade, section 7.

Brazil accounts for more of the located studies than every other country combined: 563 of the 928 studies with a locatable study site have one in Brazil, at 354 distinct sites, against 56 studies in Indonesia, 53 in Colombia and 35 in Peru. The Brazil-nut economy has made the family a Brazilian research subject. But the family, and its potential, are pan-Amazonian and pan-tropical.

Where the species with no indexed research are

Map instead the species with no indexed research, using each species' native range countries from the WCVP distribution table, and the frontier moves off Brazil. They concentrate across the northern Amazon and Guiana Shield, Colombia (43 species with no indexed research, against 42 in Brazil), Venezuela (21), Panama (21), Guyana (16), Ecuador (15) and French Guiana (15), and into the paleotropics: Malaysia (25), Indonesia (24), Gabon (19) and Cameroon (18). Colombia now holds more species with no indexed research than Brazil, on a fraction of the research. A species counts in every country of its native range.

1 · Species ranges against study sites

For each of the 401 species we compared two country lists: where it is known to occur, and where its indexed papers were carried out. Known range countries come from the Kew World Checklist of Vascular Plants distribution table (TDWG level-3 areas mapped to countries, native and introduced), so a country counts as range wherever the checklist records the species; in-range is a country-level test, generous within a large country and conservative for a species studied only just across a border. Of the 400 species with range data, 286 (71.5%) have no locatable study site inside any range country. That figure mixes two things that only a paper-level check can separate: species with no research (223 of 401 have no indexed research) and species whose papers name no place a geoparser can locate (63 of the 178 studied species). A further 1 species were studied only outside their known range countries.

Range shortfall by country

Each country is shaded by the share of its range species that have no locatable study site inside it: darker means a larger share of the species known to occur there have not been studied in the field there. Unshaded land has no census species in range. The four countries with the most range species lacking a site are labelled (Colombia, Brazil, Venezuela, Indonesia); they also head the table below.

2 · Where the study sites are

Each cell is one hexagon of the H3 grid, drawn on its true boundary so that the cells tile without overlap; darker cells hold more distinct study sites. Most cells hold one site; the fieldwork of the whole corpus concentrates in a small number of them.

The densest window, enlarged

The dashed rectangle on the map above marks the window that holds the most study sites (86°W to 32°W, 24°S to 4°N). Here it is at a larger scale and at the grid's own resolution 3 (about 12,400 km² per cell), with its own log ramp.

Many species, few studies

The ten range countries with the most census species and two or fewer locatable study sites from this corpus. Because the range lists include cultivated and museum records, a country can appear here through garden and herbarium holdings rather than wild populations; read the list as a prompt to check, not a verdict.

3 · How far the sites are from a city

Distance from each study site to the nearest populated place of 100,000 people or more is a rough measure of how reachable the fieldwork was. The busiest access points, counted greedily so that no site is counted twice, are Castanhal (Brazil, 21 sites within 100 km), Rio Claro (Brazil, 19 sites within 100 km), Cabo de Santo Agostinho (Brazil, 18 sites within 100 km); 17.7% of all sites lie within 100 km of one of the ten busiest.

These are straight-line distances, not travel times; roads, terrain, sea ice and protected-area access are not modelled. Study-site coordinates are geoparsed from abstracts, and region-level places sit at their centroid.

4 · Method note

Sources. known range countries = Kew WCVP distribution table (TDWG level-3 areas mapped to countries, native and introduced; entity_places role=range, source wcvp / wcvp_introduced; no polygons); access points = Natural Earth ne_10m_populated_places_simple, GitHub master 5.2.0-pre (nvkelso/natural-earth-vector, downloaded 2026-09-29) (populated places with 100,000 people or more), public domain; grid = H3 resolution 3 (h3-js 4.5.0), drawn on the true cell boundaries from h3-js and aggregated one level (h3 cellToParent) on the main maps. A study site is a place a paper's abstract names as the location of the work, resolved to coordinates by a gazetteer and kept only when it came from one of the six trusted geocode sources; 928 of 3,588 studies (25.9%) have one. Density and distance use precise sites only (not a whole country, not a rounded centroid): 660 distinct places.

Caveat. Study-site coordinates come from automated geoparsing; a 40-site manual check found 31 plausible; a 100 to 200 pin human check is planned. Range coverage is a country-level test, so a study in the right country but far from the species' actual distribution still counts as in range.

Provenance. Generated 2026-10-14 from the published database. Figures in this section are recomputed by phase1/spatial_shortfall.js and checked by the figures audit.

In this corpus the rate of environmental change and the density of located research are essentially unrelated: Spearman's rank correlation between the two, unit by unit, is near zero. For forest loss it is 0.074 across 431 ecoregions and 0.059 across 1,143 admin1 regions; for drying it is −0.262 across 167 ecoregions and −0.149 across 215 admin1 regions. These are descriptive figures: they do not show that change causes or deters research, and the units differ in area, accessibility and richness. What they do show is where fast change and little located research coincide, which is the list each block below ranks.

1 · Forest loss 2001 to 2024

Change here is the share of the area that was forest in 2000 (tree cover of 30 percent or more) with a loss event between 2001 and 2024, from Hansen Global Forest Change v1.12 at 30 m. Each ecoregion and each first-level administrative region is placed in one of four quadrants: the top third of the universe by forest loss is high change, the bottom third by located research density (papers of this corpus with a precise study site inside the unit, per 10,000 km²) is low research. The high-change, low-research quadrant is the candidate gap list.

Loss is any stand-replacing event, including fire and the rotational harvest of plantations, and regrowth is not subtracted, so a high value can be a working plantation landscape as easily as a cleared forest; the 30 percent canopy threshold makes open woodland "not forest".

Forest loss 2001 to 2024, cell by cell

2 · Aridity trend 1990 to 2024

Change here is the linear trend of the annual aridity index (precipitation over potential evapotranspiration, P/PET) from 1990 to 2024, as percent change per decade of the 1990 to 1999 mean, from TerraClimate at about 4.6 km; a negative number means drying. Each ecoregion and each first-level administrative region is placed in one of four quadrants: the top third of the universe by drying is high change, the bottom third by located research density (papers of this corpus with a precise study site inside the unit, per 10,000 km²) is low research. The high-change, low-research quadrant is the candidate gap list.

TerraClimate is an interpolated climate surface, so a 35-year trend at 4 km carries station-density and interpolation error, and a relative trend is unstable where the baseline is near zero; the ranking is restricted to drylands (mean P/PET under 0.65) and excludes hyper-arid units (mean under 0.05) for that reason.

Aridity trend 1990 to 2024, cell by cell

3 · When the research happened

The same study sites split by decade of publication. Of the 853 located papers, 75.7% were published in the 2010s or 2020s and 50.5% since 2016. The whole corpus of 3,588 papers, located or not, is 76.1% from those two decades, so located research is about as recent as the corpus. Of the 140 ecoregions with a located paper, 16 have none published since 2010; of the 165 admin1 regions, 26. 2 located papers have no publication year.

The decade with the most located papers, cell by cell

4 · Method note

Change layers. Forest loss 2001 to 2024 = UMD/hansen/global_forest_change_2024_v1_12 (v1.12 (2024 update)), 2001-2024, sampled at 1,000 m (native 30 m), CC BY 4.0; Aridity trend 1990 to 2024 = IDAHO_EPSCOR/TERRACLIMATE (TerraClimate monthly, EE catalogue copy (1958-01 to 2024-12 at run time)), 1990-2024, sampled at 4,638 m (native 4,638 m), TerraClimate: CC0 1.0 (public domain). MODIS: NASA, no restrictions; dryland classes after UNEP 1992. Values are area means of a systematic sample over each H3 cell of the grid (the maps above draw them one resolution coarser, as valid-area-weighted means of the child cells) and each unit outline (RESOLVE Ecoregions 2017 and GADM 4.1, simplified for display), computed in Google Earth Engine; boundary pixels are weighted fractionally. Layers are remote-sensing and reanalysis products, not field measurements.

Research. 853 of 3,588 papers in this corpus have a precise trusted study site (the same selection as the density map above); a paper counts once per cell and once per unit it has a site in, and a unit's density is papers per 10,000 km² of its area. Only units of 1,000 km² or more are ranked, and only units in a country where a census species is known to occur, so that a city-state's density is not set against a province's. 5 precise places (open sea, Southern Ocean, atoll centroids) fall in no ecoregion and 1 in no admin1 region; they are in the cell maps but not in the unit tables. Range species without a site is a country-level test (every unit in a country is credited with the country's whole range list), so it is an upper bound.

Reading it. A near-zero correlation says that, across these units, more change does not go with more or with less located research. It does not say that change deters research or that research prevents change, and units differ in area, accessibility, richness and national research capacity. A unit in the high-change, low-research quadrant is a candidate gap, not a proven one: papers whose text names no locatable site are invisible here, and the geoparsing caveat of the section above applies.

Provenance. Generated 2026-10-14 from the published database and the Earth Engine cache. Figures in this section are recomputed by phase1/research_vs_change.js, drawn by phase1/gen_atlas_change_section.js and checked by the figures audit.

Beta This section fits a model of research attention: the number of papers of this corpus with a precise study site inside each ecoregion and each first-level administrative region, predicted from the unit's area, the number of census species whose known range countries include the unit (country level, so it separates countries rather than places within one), its distance to the nearest city of 100,000 people, forest loss and the aridity trend, its biome, realm and latitude band, and how much located research the rest of its country carries. The model predicts where research tends to go. It does not predict ecological importance, threat or need, and a unit with less research than expected is a candidate gap: a place where the record is thinner than in comparable places. It is never proof of neglect. Fit quality is measured out of sample by 5-fold spatial cross-validation (whole realms or whole countries held out): pseudo-R2 0.67 for ecoregions and 0.44 for administrative regions (admin1). Each unit carries a 90 percent prediction interval from the negative-binomial count spread about its expectation. Because located research is strongly clustered (theta under 1), that interval is wide: a unit with no located paper falls below it only when the model expects more than 190.8 papers there (ecoregions) or 1786.9 papers there (administrative regions (admin1)), and in this run no unit of either type does. The rankings therefore order units by the tail probability, the chance of a count at or under the observed one given the expectation, and carry the interval label alongside.

1 · Ecoregions: expected against observed Beta

Fit. 578 ecoregions in the universe, 135 with a located paper; negative binomial, dispersion theta 0.50, mean by Poisson quasi-likelihood, ridge lambda 30. 5-fold spatial cross-validation blocked by realm: pseudo-R2 0.67 (0.59 without the national-effort covariate), Spearman rho 0.55 out of fold. Labels against the 90 percent interval: 0 below, 560 within, 18 above.

The "more than expected" list doubles as a geocoding audit: a small unit with tens of papers against an expectation under one is as likely a place-name collision, a herbarium or a field station as a genuine concentration of fieldwork.

2 · Administrative regions (admin1): expected against observed Beta

Fit. 1,356 administrative regions (admin1) in the universe, 159 with a located paper; negative binomial, dispersion theta 0.35, mean by Poisson quasi-likelihood, ridge lambda 30. 5-fold spatial cross-validation blocked by country: pseudo-R2 0.44 (0.44 without the national-effort covariate), Spearman rho 0.40 out of fold. Labels against the 90 percent interval: 0 below, 1,315 within, 41 above.

The "more than expected" list doubles as a geocoding audit: a small unit with tens of papers against an expectation under one is as likely a place-name collision, a herbarium or a field station as a genuine concentration of fieldwork.

3 · Calibration

Out of fold, by decile of expected count (observed / expected): ecoregions: deciles 3 to 8 between 0.44 and 1.70, top decile 2.39; administrative regions (admin1): deciles 3 to 8 between 0.52 and 1.88, top decile 1.50. The middle deciles are the best calibrated; the top decile is under-predicted in every model, because the busiest places (long-term field stations, provinces with a herbarium or a university) exceed what any smooth covariate model gives. That is one more reason the residuals are candidate gaps and candidate artefacts rather than findings.

4 · Method note Beta

Response and exposure. 853 of 3,588 papers in this corpus have a precise trusted study site (23.8 percent); the response per unit is the number of those papers with a site inside it (point in polygon, or the nearest unit within 25 km, as the place explorer counts), so a paper with sites in two units counts once in each: 1,010 unit-papers over 578 ecoregions and 936 over 1,356 administrative regions (of 847 and 3,686 units in all). The exposure is log(area), with log(area) also entering as a slope so the data set the area elasticity.

Covariates. log_area_km2, log_richness, log_access_km, forest_loss_pct, forest_loss_missing, aridity_trend_pct_per_decade, aridity_trend_missing, log_national_loo_density, national_loo_undefined, biome, realm, lat_band; continuous covariates standardised, categorical levels with fewer than 10 units pooled. Protected-area share is not available (no protected-area table in the database) and is stated as the natural next covariate. Richness is a country-level proxy: every unit of a country is credited with the national list.

Model and validation. Negative binomial (NB2) GLM with a log link and a ridge penalty on the standardised slopes (ecoregions lambda 30, administrative regions (admin1) lambda 30); the mean estimator and lambda are chosen on out-of-fold Poisson deviance over the same spatial folds; the dispersion is estimated by maximum likelihood. 5-fold spatial cross-validation blocked by realm (ecoregions) or country (admin1), never random rows; a held-out realm's own dummy is unseen in training. The 90 percent interval is the 5th and 95th percentile of NB(expected, theta) and carries no parameter uncertainty, so it is if anything too narrow. Zero rule: a unit with no located paper and an expectation under 0.5 is never labelled below and never ranked.

Limits. Located papers only: papers whose text names no locatable site are invisible here, and the geoparsing error rate of the sections above applies. The national-effort covariate carries most of the between-country signal; the fit without it is printed above. Terrestrial units only. This is a Beta layer: the numbers are honest as stated, and they will move when the geocoding, the corpus or the covariates change.

Provenance. Generated 2026-10-14 from the published database (SELECT only) and the change layers' cache. Refresh with node phase1/expected_research.js brazil-nut, redraw with phase1/gen_atlas_expected_section.js; every figure in this section is checked by the figures audit.

This atlas is one view of a larger queryable corpus. A single genus’s evidence profile, the wild-harvest-versus-cultivation split for any species, or a shortlist of economically promising trees with no indexed research, ready for survey are each a query away, on the same reproducible foundation.