Philippa Kaur

Field samplingLake Masoko fish were chased into fixed gill nets and SCUBA by a team of professional divers at different target depths determined by diver depth gauge (12× male benthic, 12× male littoral). Riverine fish (11× Mbaka River and 1× Itupi river) were collected by local fishermen. On collection, all fish were euthanized using clove oil. Collection of wild fish was done in accordance with local regulations and permits in 2015, 2016, 2018 and 2019. On collection, fish were immediately photographed with color and metric scales, and tissues were dissected and stored in RNAlater (Sigma-Aldrich); some samples were first stored in ethanol. Only male specimens (showing bright nuptial coloration) were used in this study for the practical reason of avoiding any misassignment of individuals to the wrong population (only male individuals show clear differences in phenotypes and could therefore be reliably assigned to a population). Furthermore, we assumed that any epigenetic divergence relevant to speciation should be contributing to between-population differences in traits possessed by both sexes (habitat occupancy, diet). To investigate the role of epigenetics in phenotypic diversification and adaptation to different diets, homogenized liver tissue – a largely homogenous and key organ involved in dietary metabolism, hormone production and hematopoiesis – was used for all RNA-seq and WGBS experiments.Common-garden experimentCommon-garden fish were bred from wild-caught fish specimens, collected and imported at the same time by a team of professional aquarium fish collectors according to approved veterinary regulations of the University of Bangor, UK. Wild-caught fish were acclimatized to laboratory tanks and reared to produce first-generation (G1) common-garden fish, which were reared under the same controlled laboratory conditions in separate tanks (light–dark cycles, diet: algae flakes daily, 2–3 times weekly frozen diet) for approximately 6 months (post hatching). G1 adult males showing bright nuptial colors were culled at the same biological stages (6 months post hatching) using MS222 in accordance with the veterinary regulations of the University of Bangor, UK. Immediately on culling, fish were photographed and tissues collected and snap-frozen in tubes.Stable isotopesTo assess dietary/nutritional profiles in the three ecomorph populations, carbon (δ13C) and nitrogen (δ15N) isotope analysis of muscle samples (for the same individuals as RRBS; 12, 12 and 9 samples for benthic, littoral and riverine populations, respectively) was undertaken by elemental analyzer isotope ratio mass spectrometry by Iso-Analytical Limited. It is important to note that stable isotope analysis does not depend on the use of the same tissue as the ones used for the RRBS/WGBS samples45. Normality tests (Shapiro–Wilk, using the R package rstatix v.0.7.0), robust for small sample sizes, were performed to assess sample deviation from a Gaussian distribution. Levene’s test for homogeneity of variance was then performed (R package carData v.3.0-5) to test for homogeneity of variance across groups. Finally, Welch’s ANOVA was performed followed by Games–Howell all-pairs comparison tests with adjusted P value using Tukey’s method (rstatix v.0.7.0). Mean differences in isotope measurements and 95% CI mean differences were calculated using Dabestr v.0.3.0 with 5,000 bootstrapped resampling.Throughout this manuscript, all box plots are defined as follows: centre line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; points, outliers.RNA-seqNext-generation sequencing library preparationTotal RNA from liver tissues stored in RNAlater was extracted using a phenol/chloroform approach (TRIzol reagent; Sigma-Aldrich). Of note, when tissues for bisulphite sequencing samples were not available, additional wild-caught samples were used (Supplementary Table 3). The quality and quantity of RNA extraction were assessed using TapeStation (Agilent Technologies), Qubit and NanoDrop (Thermo Fisher Scientific). Next-generation sequencing (NGS) libraries were prepared using poly(A) tail-isolated RNA fraction and sequenced on a NovaSeq system (S4; paired-end 100/150 bp; Supplementary Table 3), yielding on average 32.9 ± 3.9 Mio reads.Read alignment and differential gene expression analysisAdaptor sequence in reads, low-quality bases (Phred score  More

Wolf, S. et al. Nat. Ecol. Evol. https://doi.org/10.1038/s41559-022-01904-x (2022).Article 

Google Scholar 
Kattge, J. et al. Glob. Change Biol. 26, 119–188 (2020).Article 

Google Scholar 
Sabatini, F. M. et al. Glob. Ecol. Biogeogr. 30, 1740–1764 (2021).Article 

Google Scholar 
Łopucki, R., Kiersztyn, A., Pitucha, G. & Kitowski, I. Ecol. Modell. 468, 109964 (2022).Article 

Google Scholar 
Kosmala, M., Wiggins, A., Swanson, A. & Simmons, B. Front. Ecol. Environ. 14, 551–560 (2016).Article 

Google Scholar 
White, C. R. et al. Funct. Ecol. 35, 1572–1578 (2021).Article 

Google Scholar 
Xirocostas, Z. A., Debono, S. A., Slavich, E. & Moles, A. T. Methods Ecol. Evol. 13, 596–602 (2022).Article 

Google Scholar 
Callaghan, C. T. et al. Bioscience 71, 55–63 (2020).
Google Scholar  More

Data collectionIn phase 1 (2017), we first circulated a comprehensive multi-language questionnaire and associated guidelines on protected area personnel numbers to major national protected area agencies, focusing on the 50 countries listed in the WDPA as having the most protected areas. The questionnaire requested information on personnel numbers, type of employers and management levels (from executive to skilled practical workers). Protected area personnel were defined as those spending at least 50% of their work time on protected area-related tasks. The questionnaire also requested information about job titles used for personnel equivalent to rangers. This phase produced usable data for 28 countries/territories.In phase 2 (2018 onwards), we conducted online searches for published data on protected area personnel numbers in the countries/territories not included in the questionnaire survey or where questionnaire responses were incomplete or unclear. The resulting information came from official organizational reports (10 countries/territories), published external studies, project documents and journal papers (35 countries/territories) and websites of protected area organizations or individual sites (9 countries/territories).In phase 3 (2018–2021), we directly requested personal contacts to locate or supply information from official sources both for the remaining countries/territories and to improve or verify data from phases 1 and 2. The minimum data requested were the overall number of protected area personnel, the number of those personnel that could be categorized as rangers, the terrestrial area of protected areas managed by the listed personnel and the source of the information. This phase contributed usable data for 68 countries and territories. Data for a further 17 countries/territories were assembled from multiple sources.The final dataset covered 176 countries/territories: 167 surveyed countries/territories and a further 9 countries/territories that have no WDPA-listed protected areas (Supplementary Table 1), with contributions from more than 150 individuals.Initial data processingTo assess and, where necessary, improve the reliability of data obtained in a wide range of formats and levels of detail and from multiple sources, we scored the data for each country/territory from 0 to 5 for each of four criteria—detail, accuracy, source and age of the data—with a maximum score of 20 (Supplementary Table 1 and Supplementary Fig. 1). For all low-scoring records (a score of less than 15), we sought more-reliable sources in later phases of the study, rejecting any final scores of less than 10.On reviewing the data, we excluded from the analysis protected areas identified in the WDPA as predominantly or entirely marine, Antarctica and countries/territories categorized in the WDPA as polar (Greenland, French Southern Territories, Bouvet Island, Heard Island and McDonald Islands, South Georgia and the South Sandwich Islands). These large, remote and/or largely uninhabited areas are likely to have quite different management models and scales of staffing from terrestrial protected areas (although marine protected areas are also widely understaffed11). For example, in 2012 the 972,000 km2 of Northeast Greenland Protected Area (categorized by the WDPA as polar) was only periodically visited by six two-person teams of naval personnel47, and the 2008 management plan of the 1.51 million km2 Papahānaumokuākea Marine National Monument (Hawai’i, USA) specifies just nine personnel, working in conjunction with several other agencies48. Data for one country were supplied by officials on the agreement that the country was not specifically identified in publications (the country is given the three-letter code ZZZ in relevant tables and figures).Because the format, completeness and level of detail of the data varied widely, from comprehensive personnel lists to single figures, we restricted our raw dataset to six variables that could be consistently extracted from data obtained for each country/territory:

1.

Total number of non-ranger personnel (if known)

2.

Total number of rangers (if known)

3.

Total number of protected area personnel (either the sum of 1 and 2 or provided as an undifferentiated total)

4.

Terrestrial area of protected areas covered by surveyed personnel (km2)

5.

Total terrestrial area of protected areas of the country/territory (km2)

6.

Year of the data

We used the WDPA, official publications and websites to determine (or verify) the area of terrestrial protected areas covered by the personnel listed for each country/territory, using WDPA data if there were discrepancies. Total national terrestrial protected area coverage was taken from the WDPA, with the exception of Turkey, where the area officially reported to the WDPA is significantly less than the nationally published area.The raw data from the survey are shown in Supplementary Table 1.Candidate predictorsTo predict the number of rangers and non-rangers in countries and territories for which we had no data (Statistical analysis), we collected information on the following set of variables, hereafter referred to as candidate predictors:Location dataThe WGS84 latitude and longitude of the centroid of the largest land mass associated with each country/ territory (to obtain the polygons defining the land masses, we used the R package rnaturalearth version 0.1.0; https://github.com/ropensci/rnaturalearth)2020 data from the World Bank (https://data.worldbank.org/indicator)

Area of the country/territory

Population density: the mid-year population divided by land area

Gross domestic product (GDP) in US dollars

GDP per capita in US dollars (GDP divided by mid-year population)

Growth rate of GDP

The proportion of rural inhabitants

The proportion of unemployed inhabitants

The forested proportion of the country/territory

2020 data for each country/territory from the WDPA (https://www.protectedplanet.net/)

The total terrestrial area of WDPA-listed protected areas

The proportion of the terrestrial area of all IUCN-categorized protected areas (Categories I–VI) that falls within protected areas in Category I or II

The proportion of the terrestrial area of all IUCN-categorized protected areas (Categories I–VI) that falls within protected areas in Categories I–IV

2020 data from the Yale Center for Environmental Law and Policy Environmental Performance Index (https://epi.yale.edu/)

Environmental Performance Index (EPI): a composite index using 32 performance indicators across 11 categories

Ecosystem Vitality Index (EVI): an indicator of how well countries preserve, protect and enhance ecosystems and the services they provide

Species Protection Index (SPI): an indicator of the species-level ecological representativeness of each country’s/territory’s protected area network

Not all this information was available for all countries/territories. Most of the missing data were for small territories that account for only a very small proportion of the total area of protected areas worldwide (Supplementary Table 2c).Statistical analysisOur primary objective was to estimate the total number of all personnel engaged in managing all the world’s WDPA-listed terrestrial protected areas and the number categorized as rangers. Our raw data collection yielded full, partial or no information on total personnel and ranger numbers for each country/territory (Supplementary Table 1 shows the completeness of all the data collected). Our first task, therefore, was (1) to impute the information for unsurveyed protected areas on the basis of information from surveyed protected areas within the same countries/territories and (2) to predict those numbers for countries/territories where no information was available on overall personnel numbers and/or ranger numbers on the basis of relationships we could establish between available information and candidate predictors in other countries/territories (Supplementary Table 7). A brief description of these two approaches follows, and full details on the analysis are provided in Supplementary Information.Data imputationFor countries/territories where we had obtained information about numbers of personnel and/or rangers for only some protected areas, our strategy was to populate the unsurveyed protected areas in proportion to the densities of personnel or rangers from the surveyed protected areas of the same countries/territories. For example, for Spain we obtained evidence that there are 619 rangers responsible for protected areas covering 44,328 km2, out of a national total protected area system covering 142,573 km2. To impute the number of rangers for the remaining 98,245 km2, we used the density of rangers in the surveyed area (one ranger per 44,328/619 = 71.6 km2) and applied that to the unsurveyed area, giving a total of 1,991 rangers (619 + (98,245/71.6)). This imputation assumes that unsurveyed areas are staffed at the same density as surveyed areas, whereas in reality the relative densities are likely to vary in unknown ways within different countries/territories. To study the sensitivity of our results to the assumed proportion, we repeated our analysis using the following proportions of the observed densities: 0, 0.25, 0.50, 0.75 and 1.00. This provided a range of personnel numbers from a minimum (based on a proportion of 0) to a presumed maximum (based on a proportion of 1.00). From the data obtained, it was not possible to calculate the actual proportions, but based on the experience of the practitioners in the author team, the unsurveyed areas are highly unlikely to be staffed at higher densities than surveyed areas and, on average, are very likely to be staffed at lower densities. After all, most survey respondents were national or subnational agencies responsible for protected areas subject to stronger formal requirements for protection and management and therefore likely to have larger workforces. Unsurveyed protected areas are more likely to be managed by local entities, with fewer resources, less-stringent management obligations and therefore fewer personnel. The range of proportions we considered to populate unsurveyed areas should therefore yield predictions encompassing the actual (unknown) numbers of rangers and non-rangers with a conservative margin of error. In the main text, we have reported the results of imputation assuming a proportion of 1, which is probably the most optimistic assessment of the current workforce in protected areas within the proportions of the observed densities considered. Results using lower proportions are shown in Extended Data Fig. 2 and Supplementary Tables 4 and 5.Data predictionOur imputation approach was not possible for countries/territories where (1) zero ranger or personnel data had been obtained and (2) specific data had not been obtained that allowed imputation either for rangers or for total personnel (where only total personnel numbers or only ranger numbers had been obtained). To predict the missing information, we used two different statistical approaches: linear mixed models (LMMs)49 and a general implementation of random forests, which we term RF/ETs because it encompasses both random forests sensu stricto (RFs)50 and a variant called extremely randomized trees (ETs)51. LMMs and RFs have been extensively discussed and reviewed in the literature49,52,53. We adopted these approaches because both have proved successful in producing accurate predictions for a wide range of applications and because both are well suited to our data since they both produce predictions from a set of predictors and allow for the consideration of spatial effects54,55. Furthermore, comparing predictions generated through very different methods informs us about the robustness of our results with respect to key statistical assumptions. LMMs come from the ‘data modelling culture’56 and belong to parametric statistics; RF/ETs come from the ‘algorithmic modelling culture’ and belong to non-parametric statistics.We followed the same workflow for both statistical approaches, comprising eight steps: (1) general data preparation; (2) preparation of initial training datasets; (3) selection of predictor variables and of the method used for handling spatial autocorrelation; (4) preparation of final training datasets; (5) fine tuning; (6) final training; (7) preparation of datasets for predictions and simulations; and (8) predictions and simulations (see Supplementary Information for details).Both approaches yielded very similar results with our data. We chose to present the LMM results in the main text, but we provide and compare the results obtained by both approaches in Supplementary Information.SoftwareWe performed all the data analyses using the free open-source statistical software R version 4.157. We used the R package spaMM version 3.9.13 to implement LMMs58 and the R package ranger version 0.13.1 to implement RF/ETs59. To reformat and plot the data, we used the Tidyverse suite of packages60. Details are provided in an R package we specifically developed so that findings presented in this paper can readily be reproduced (see Code availability). Using a workstation with an AMD Ryzen Threadripper 3990 × 64-core processor and 256 GB of RAM, our complete workflow ran in ~3,000 CPU hours.Estimation of required numbers and densities of personnelTo estimate the numbers of personnel and rangers required for effective management of existing protected areas, we referred to ref. 25. This estimates that the minimum budget needed to adequately manage the existing protected area system is US$67.6 billion per year and that current annual expenditure is US$24.3 billion. From these figures, we can calculate that resources invested in the current global system of protected areas are approximately 36% of what is required. We consulted data from https://ourworldindata.org to determine that the proportion of global public expenditure on employee compensation has remained between 21.01% and 23.33% in the years from 2006 to 2019. We obtained these figures from the ‘Government Spending’ section of the site, consulting the chart ‘Share of employee compensation in public spending, 2002 to 2019’ and selecting data for ‘World’. On the basis of this broadly constant proportion and the assumption that total employee compensation is an indicator of total employee numbers, we inferred that current numbers of protected area employees are also around 36% of what is required. We therefore multiplied our estimations of personnel and ranger numbers by 1/0.36 and recalculated the densities on this basis (current requirement = 1/0.36 × current estimate).To estimate staffing requirements for 30% global coverage of protected areas—the global target intended to be reached by 2030—we used the mean personnel and ranger densities calculated as being required at present to ‘populate’ a global area of terrestrial protected areas if increased from the percentage at the time of our study (15.7%) to 30% (current requirement × (0.300/0.157)).Economic calculationsWe based our calculations on published data from 202025, which estimate that expanding the protected areas to 30% would generate higher overall output (revenues) than non-expansion (an extra US$64–454 billion per year by 2050). This figure is only an indicative, partial estimate, generated for the purposes of comparison and to illustrate the substantial return on investment that protected area staff investments imply. Using these figures and our estimates of personnel requirements to ensure effective management of 30% coverage, we calculated the range of sums that each additional protected area staff member has the potential to generate (Supplementary Table 8). For clarity, we rounded these figures to the nearest hundred US dollars in the main text.Our estimates of the gross value added per worker in forestry and agriculture (sectors responsible for similar proportions of the world as protected areas) are included to provide a point of comparison for the figures showing the economic benefit generated per protected area personnel member (see the preceding). The data for the gross annual value of world agricultural production (US$3,550,231,736,000) and the number of workers employed in agriculture (343,527,711) come from the Food and Agriculture Organization of the United Nations30, providing an average gross value of annual agricultural production per worker of US$10,335. We adjusted these 2018 data to 2020 price levels using a deflator based on the US consumer price index (CPI) from the World Economic Outlook database61 (Supplementary Table 9). This ensures that all the economic value data we present are directly comparable for protected area, agricultural and forestry workers. We calculated the gross value of forest production per worker on the basis of direct contribution of forestry of more than US$539 billion to world GDP in 201162 and total forest-sector employment of 11.881 million full-time-equivalent jobs in 201032. These were the most up-to-date global estimates we could locate from credible sources that presented comparable estimates of forest-sector employment and contribution to GDP. This gives an average gross value of forest production per worker of US$45,367 per year. We used the same method as for agriculture to bring these figures to 2020 price levels (Supplementary Table 9). These figures are rounded to the nearest hundred US dollars in the main text. More

Study areaThe present study was conducted in Uttarkashi district, Uttarakhand, located between 38° 28′ to 31°28′ N latitude and 77°49′ to 79°25′ E longitude with an area of about 8016 km2, covering primarily hilly terrain with an altitudinal range of 715–6717 m (Fig. 3). The terrain is mountainous, consisting of undulating hill ranges and narrow valleys with temperate climatic conditions. The district lies in the upper catchment of two major rivers of India, viz., the Ganges (Bhagirathi towards upstream) and the Yamuna. The major vegetation types of the study area are Himalayan moist temperate forest, sub-alpine forest and alpine scrub59. The Uttarkashi district forests are managed under three Forest Divisions viz., (i) Uttarkashi Forest Division (ii) Upper Yamuna Badkot Forest Division and (iii) Tons Forest Division) with two Protected Areas (PAs) (i) Gangotri National Park and (ii) Govind Pashu Vihar National Park. The forested habitats of the study landscape are home to top conservation priority species, including Asiatic Black bear (Ursus thibetanus), Musk deer (Moschus spp.), Common leopard (Panthera pardus), Himalayan brown bear (Ursus arctos isabellinus) and Western Tragopan (Tragopan melanocephalus), Himalayan monal (Lophophorus impejanus). The study was conducted after a study permit issued by the Chief Wildlife Warden, Forest Department, Uttarakhand government, vide letter no. 848/5-6 dated 31/08/2019, we have not handled the species for doing research. Instead, remote camera traps have been used for collecting the data with the permission of the Chief Wildlife Warden, Government of Uttarakhand. Further, informed consent was taken before interviewing the local communities. The data was collected according to the institutional guidelines and approved by the Research Advisory and Monitoring Committee of the Zoological Survey of India.Figure 3Map of the study area Uttarkashi, Uttarakhand. ArcGIS 10.6 (ESRI, Redlands, CA) was used to create the map. (Map created using ArcGIS 10.6; http://www.esri.com).Full size imageSampling protocolThe basic sampling protocol and assumptions for multi-species occupancy modelling are identical to the single-species case7. Briefly, a set of 62 intensive sites, were randomly selected, and each site i was surveyed j times. During each survey, detection/non-detection of S focal species was recorded. Additionally, direct or indirect evidences of species presence from the different areas were also recorded.Data collectionThe complete study area was divided into 10 × 10 km grids, consisting of n = 60 grids. Based on the reconnaissance survey, out of these 60 grids, we selected 25 girds that were accessible to conduct the survey and have the species presence. Further, these grids were divided into 2 × 2 km grids to maximize our effort so that all logistically accessible grids could be covered, and we conducted intensive sampling in N = 62 grids after excluding the grids with human settlements. T The field surveys were conducted during 2018–2019, and a team of researchers systematically visited selected grids to collect data on the detection/non-detection of these ungulates. A total of 62 camera traps were deployed in selected grids, and 650 km were traversed, accounting for N = 54 trails in these sampled grids. These camera traps were visited once in every fifteen days for replacing the batteries as well as documenting the presence of the species through the sign surveys. The ultra-compact SPYPOINT FORCE-11D trail camera (SPYPOINT, GG Telecom, Canada, QC) and Browning trail camera (Defender 850, 20 MP, Prometheus Group, LLC Birmingham, Alabama, https://browningtrailcameras.com) camera traps were used to detect the presence/absence of ungulate species. The cameras were mounted 40–60 cm above ground on natural trails without lures.Data explorationWhile deploying camera traps, we also noted habitat variables through on-site observation such as distance to the village and human disturbance. We tested site covariates for collinearity and discarded one of a pair if the Pearson’s correlation was greater than 0.760. Hence, we assumed each of the site covariates could influence the occupancy and detectability of these ungulates.CovariatesWe hypothesized that habitat variables may influence these ungulates’ occupancy and detection probability. A total of 21 variables were extracted either from the field or using the ArcGIS v. 10.6 software (ESRI, Redlands, CA), and only 14 were retained after collinearity testing60 (Table 3). These covariates were classified into the following categories (Topographic variables, Habitat variables and anthropogenic variables). The topographic variables (elevation, slope and aspect) were generated using 30× resolution SRTM (Shuttle Radar Topography Mission) image downloaded from EarthExplorer (https://earthexplorer.usgs.gov/). The habitat/ land cover classification was carried out using Landsat 8 satellite imagery (Spatial resolution = 30 m) downloaded from Global Land Cover Facility by following the methodology suggested by61 using the ArcGIS v. 10.6 software (ESRI, Redlands, CA). The study area was classified into nine Land use/land cover (LULC) classes viz., West Himalayan Sub-alpine birch/fir Forest (FT 188), West Himalayan upper oak/fir forest (FT 162), West Himalayan Dry juniper forest (FT 180), Ban oak forest (FT 152), Moist Deodar Forest (FT 155), Western mixed coniferous forest (FT 156), Moist temperate Deciduous Forest (FT 157) which were used for further analysis considering their importance to species ecology and behavior60. The values for all the covariates were extracted at 30 m resolution, and a single value per site was obtained by averaging all the pixel values within each sampling site (camera trap locations).Table 3 Habitat variables used for multi species occupancy analysis of three ungulate species in Uttarkashi, Uttarakhand.Full size tableOccupancy modelling frameworkWe used multi-species occupancy modelling62 of barking deer, goral and sambar to estimate the probability of the species (s) occurred within the area (i) sampled during our survey period (j), for accounting the imperfect detection of the species8. Distinguishing the true presence/absence of a species from detection/non-detection (i.e., species present and captured or species present but not captured) requires spatially or temporally replicated data. We used camera stations to record the presence/absence of species along with sign survey in all the studied grids. The camera traps were placed along the trail/transects in the studied grids hence each grid needs to be visited once in every fifteen days to check the camera traps as well as to document the presence of the studied species. Therefore, we treated 15 trap nights as one sampling occasion at a particular camera station resulting in ~ 7 sampling occasions per camera station.Our aim was to record the presence/ absence of the species at a particular gird hence we incorporated sign survey data if the species was not detected in camera station but recorded through sign survey. We pooled the presence/absence data in a single sheet of each species following6 and fitted occupancy and detectability models using programme Mark63,64. We model the species (s) presence (ysij = 1) and absence (ysij = 0) at site i during survey j, and the sampling protocol was identical to single species case65, where the Bernoulli random variable was conditional on the presence of species s (Zs = 1) following6$${text{y}}_{sij} sim {text{ Bernoulli}}left( {{text{p}}_{sij} {text{z}}_{si} } right),$$
where Psij represents the probability of detecting species S during replicate survey j at site i and Zsi = presence or absence of species s at site i.Furthermore, we model the latent occupancy state of species s at site i as a multivariate Bernoulli random variable:$${text{Z}}_{i} sim {text{MVB}}left( {uppsi _{i} } right)$$
where Zi = {Z1i, Z2i….., ZSi} is an S-dimensional vector of 1’s and 0’s denoting the latent occupancy state of all S species and (ψi) is a 2S-dimensional vector denoting the probability of all possible sequences of 1’s and 0’s Zi can attain such that ∑ ψi = 1 with corresponding probability mass function (PMF) adopted from6,64.$$fleft( {{text{Z}}_{i} } right) = {text{ exp}}left( {left( {{text{Z}}_{i} {text{log}}(uppsi_{{text{i}}} {1}/uppsi_{{text{i}}} 0} right) , + {text{ log}}left( {uppsi_{{text{i}}} 0} right)} right).$$The quantity f = log (ψi1/ψi0), is the log odds species S occupies a site often referred to as a ‘natural parameter’.Since we are modeling three ungulate species (S = 3), 2S = 23 the possible encounter histories included in the dataset were eight, if neither of the two species were detected the value of ‘00’ was assigned; similarly ‘01’ indicates detection of species 1; ‘02’ indicates detection of species 2; ‘03’ indicates detection of both the species; ‘04’ indicates detection of species 3; ‘05’ indicates detection of species 1 and species 3; ‘06’ indicates detection of species 2 and species 3 and ‘07’ indicates detection of all the three species. We modelled constant occupancy and detection probability for each of the three species. Hence, we specified 6 f and p parameters, an intercept (β) for each of one-way f parameter and detection parameter p following64.$$f_{{1}}=upbeta_{{{1},}} ;;{text{p}}=upbeta_{{4}}$$$$f_{{2}} = upbeta_{{{2},}} ;{text{p }} = , upbeta 5$$$$f_{{3}} = , upbeta_{{{3},}}; {text{p }} = , upbeta_{{6}}$$We fit a set of models including the detection probability as a constant, p(.), and variable function to occupancy ψ(covariate) for site-specific covariates and models include occupancy as constant ψ(.) and variable function of the detection p(covariates) for the respective site covariates.As we have assumed the independence among all three species, the model shows marginal occupancy probabilities of species 1, species 2 and species 3 varies as a function of environmental variables. We incorporated site-level characteristics affecting species-specific occurrence (f1: occupancy of species 1, f2: occupancy of species 2, & f3: occupancy of species 3) and detection probabilities using a generalized linear modelling approach42. This requires 9 parameters: an intercept (β1, β3, β5) and slope (β2, β4, β6) coefficient for each 1-way f parameter f1, f2, f3 and an intercept parameter for each detection parameter (β7, β8, β9). Below mentioned is the model for 1-way f parameters.$$f_{{1}} = , upbeta_{{{1 } + }} upbeta_{{2}} left( {{text{Covariate}}} right),;;{text{ p }} = , upbeta_{{7}}$$$$f_{{2}} = , upbeta_{{{3 } + }} upbeta_{{4}} left( {{text{Covariate}}} right),;;{text{ p}} = , upbeta_{{8}}$$$$f_{{3}} = , upbeta_{{5}} + , upbeta_{{6}} left( {{text{Covariate}}} right),;;{text{ p }} = , upbeta_{{9}} .$$All covariates were standardized before model fitting. We fitted the most complex model to each species and considered all possible combinations of covariates using the logit link function. Our rationale for including these variables in the occupancy and detectability component of the model was that we expected these variables to influence the occupancy and detectability of the study species.Since multi-species occupancy simultaneously model environmental variables, & interspecific interactions. Further it also allows to understand the influence of environmental variables on one species occupancy, in the presence or absence of other sympatric species64. Hence, we also modeled two species occur together as a function of covariates. We examined how the variables of each camera site influenced the pair-wise interaction of the three ungulate species. This model assumes that the conditional probability of one species varies in the presence or absence of other species. We assumed f123: co-occurrence of species 1, species 2 & species 3 = 0, hence we did not include higher-order interactions in any of our models, we assumed the conditional probability of 3 species occurred together was purely a function of species-specific (f1, f2, f3) and pair-wise interaction (f12: co-occurrence of species1 & species 2, f13: co-occurrence of species 1 & species 3, f23: co-occurrence of species 2 & species 3) parameters. We modeled pair-wise interaction of species varies as a function of environmental variables keeping detection probability constant. Hence, we specified 15 f and p parameters, an intercept and slope coefficient for each of the one-way (f1, f2, f3) and the two-way f parameters (f12, f13, and f23); as well as an intercept parameter for each of the detection models. The model equation below implies for 2-way f parameters:$$f_{{{12}}} = , upbeta_{{{7 } + }} upbeta_{{8}} left( {{text{Covariate}}} right),;;{text{ p }} = , upbeta_{{{13}}}$$$$f_{{{13}}} = , upbeta_{{{9 } + }} upbeta_{{{1}0}} left( {{text{Covariate}}} right),;;{text{ p }} = , upbeta_{{{14}}}$$$$f_{{{23}}} = , upbeta_{{{11 } + }} upbeta_{{{12}}} left( {{text{Covariate}}} right),;;{text{ p }} = , upbeta_{{{15}}} .$$We also fitted models including co-occurrence and detection probability of a species varies as a function of environmental variables. Hence, we specified 18 f and p parameters, an intercept and slope coefficient for each of one-way (f1, f2, f3) and two-way f parameters (f12, f13, f23); and an intercept as well as the slope parameters for each of the detection models. The model equation below implies for 2-way f parameters:$$f_{{{12}}} = , upbeta_{{{7 } + }} upbeta_{{8}} left( {{text{Covariate}}} right),{text{ p }} = , upbeta_{{{13 } + }} upbeta_{{{14}}} left( {{text{covariate}}} right)$$$$f_{{{13}}} = , upbeta_{{{9 } + }} upbeta_{{{1}0}} left( {{text{Covariate}}} right),{text{ p }} = , upbeta_{{{15}}} + , upbeta_{{{16}}} left( {{text{covariate}}} right)$$$$f_{{{23}}} = , upbeta_{{{11 } + }} upbeta_{{{12}}} left( {{text{Covariate}}} right),{text{ p }} = , upbeta_{{{17}}} + , upbeta_{{{18}}} left( {{text{covariate}}} right)$$A total of 38 models were run to test the influence of environmental variables on occupancy and detection probability of species-specific (f1, f2, f3) and pair-wise interaction of the three ungulate species. The best-supported model was identified by selecting the model with the lowest AICc value and highest model weights66, where higher model weights indicate a better fit of the model to the data. Second-Order Information Criterion (AICc)67 values were used to rank the occupancy models, and all the models whose ΔAICc  More

We hypothesized that functionally distinct forest types can be mapped at moderate spatial resolutions, using a combination of canopy foliar traits and canopy structure information. Our analysis of LiDAR and imaging spectroscopy data at spatial resolutions ranging from 4 to 200 m (16 m2–40,000 m2), with an emphasis on the 30 m (900 m2) spaceborne hyperspectral spatial resolution, reveals that few remotely sensed canopy properties are needed to successfully identify ecologically distinct forest types at two diverse tropical forest sites in Malaysian Borneo. In testing our second hypothesis that mapped forest types exhibit distinct ecosystem function, we found that forest types identified using remotely sensed leaf P, LMA, Max H, and canopy cover at 20 m height (Cover20) closely align with forest types defined from field-based floristic surveys29,30,31,32,33 and inventory plot-based measurements of growth and mortality rates (Fig. 4b). Our approach, however, enables mapping of their entire spatial extent (Fig. 1) and reveals important structural and functional variation within areas characterized as a single forest type in previous studies (Fig. 3). Current and forthcoming satellite hyperspectral platforms, including PRISMA (30 m), CHIME (20–30 m), and SBG (30 m), have or will have comparable spectral resolution, higher temporal revisits, and much greater geographic coverage. The ability to conduct this type of analysis using remote sensing measurements at 30 m resolution suggests that our method can be applied to these emerging spaceborne imaging spectroscopy data to reveal important differences in structure and function across the world’s tropical forests.Nested functional forest types revealedTo test our first hypothesis, rather than making an a priori decision about the number of k-means clusters (k), we explored the capacity of remotely sensed data to reveal ecologically relevant variation in forest types. Baldeck and Asner took a similar unsupervised approach to estimating beta diversity in South Africa34. Because the choice of k directly influences analysis outcomes, careful selection of k is required. Different approaches for identifying the number of clusters, using the Gapk and Wk elbow metrics35, yielded varying optimal numbers of clusters for the Sepilok and Danum landscapes (Fig. 1, Supplementary Figs. 4 and 5). However, at both sites, a comparison of results based on different values of k revealed ecologically meaningful structural and functional differences and graduated transitions between forest types (Fig. 2, Supplementary Figs. 7 and 8), indicating that the exploration of traits that aggregate or separate forest types as k changes is a valuable exercise. Overlap between the remotely sensed forest type boundaries and inventory plots within distinct forest types indicate that the series of clustered forests align closely with forest types defined based on in situ data on species composition and ecosystem structure. In part, this type of analysis requires careful selection of the number of clusters. Additionally, however, we gained valuable insights via the exploration of varying numbers of clusters as it relates to biologically meaningful categorization of forest types. Extending this method to other parts of the tropics will require similar decision-making, which will either require user input, or the development of robust automated algorithms for selecting k.Forest types capture differences in ecosystem dynamicsWe further evaluated the canopy traits and structural attributes that were most critical for mapping distinct forest types, hypothesizing that mapped forest types exhibit distinct ecosystem function. Forest types revealed by the cluster analyses were distributed along the leaf economic spectrum, where the leaf economic spectrum characterizes a tradeoff in plant growth strategies36. LMA, which can covary strongly with leaf N and P, is a key indicator of plant growth strategies along the spectrum37. At the slow-return end of the leaf economics spectrum, plants in nutrient-poor conditions with low leaf nutrient concentrations invest in leaf structure and defense, expressed as high LMA, strategizing longer-lived, tougher leaves with slower decomposition rates. This strategy comes at the cost of slower growth. At the quick-return end of the spectrum, plants in nutrient-rich environments with higher leaf nutrient concentrations invest less in structure and defense, enabling faster growth and more rapid leaf turnover, i.e., shorter leaf lifespans. This quick-return growth strategy supports higher photosynthetic rates and more rapid carbon gain36.In this study, the principal components and clustering results yielded forest types that are indicative of community level differences associated with leaf economic spectrum differences. The nutrient rich sites (Danum1 and Danum2, Supplementary Fig. 8) show high canopy N and P and low LMA compared to the nutrient poor and acidic sites (Sandstone and Kerangas), which contributes to lower leaf photosynthetic capacity (Vcmax) and growth (Fig. 4b). Foliar N:P also increased with site fertility, confirming that tropical forests are primarily limited by phosphorus, and not nitrogen38,39, with large implications for carbon sequestration in these forests. Orthogonal differences in canopy structure and architecture between Danum forest types and Sepilok Sandstone and Alluvial forests could be indicative of ecosystem scale differences in the sensitivity of these forests to endogenous disturbance processes40.The significant differences in aboveground carbon stocks and growth and mortality rates between forest types further suggests strong differences in ecosystem dynamics. In general, growth rates varied inversely to aboveground carbon, and higher aboveground carbon corresponded to lower mortality rates. As an example, the Sepilok sandstone forests, which are largely comprised of slow-growing dipterocarp species29,33, had the highest median aboveground carbon (236 Mg C ha−1), with higher canopy P and N, and lower LMA. The taller canopy and low canopy leaf nutrient concentrations are consistent with the low growth and mortality rates found in the sandstone forest, indicating a slow-growth strategy yielding larger trees and higher aboveground carbon stocks. In contrast, alluvial forests exhibit high turnover with mortality and growth rates higher relative to Sandstone forests corresponding to lower aboveground carbon on average. Kerangas forests exhibited low aboveground carbon despite an intermediate plot-level growth rate, and mortality rates that were significantly lower than the Danum or alluvial forest types. Kerangas forests, which were characterized by the highest LMA, lowest foliar P and N (Fig. 2a), and the lowest plot-level aboveground carbon density (186 Mg C ha−1; Fig. 4a), are known to have higher stem densities, lower canopy heights, and long-lived leaves5,32,41, suggesting well-developed strategies for nutrient retention42. Interestingly, despite significantly different aboveground carbon and demography, the kerangas and sandstone forests did not differ in LAI or canopy architecture (P:H); although maximum height, Cover20, and Hpeak LAI were significantly higher in the sandstone forest, highlighting the need to account for differences beyond LAI when scaling processes from leaves to ecosystems.In addition, when three forest types were distinguished at Sepilok, the alluvial inventory plot had significantly higher aboveground carbon than the remote sensing-derived alluvial forest extent (Fig. 4a, p  More

Sakschewski, B. et al. Leaf and stem economics spectra drive diversity of functional plant traits in a dynamic global vegetation model. Glob. Change Biol. 21, 2711–2725 (2015).Article 

Google Scholar 
Berzaghi, F. et al. Towards a new generation of trait-flexible vegetation models. Trends Ecol. Evol. 35, 191–205 (2020).Article 
PubMed 

Google Scholar 
Bruelheide, H. et al. Global trait–environment relationships of plant communities. Nat. Ecol. Evol. 2, 1906–1917 (2018).Article 
PubMed 

Google Scholar 
Joswig, J. S. et al. Climatic and soil factors explain the two-dimensional spectrum of global plant trait variation. Nat. Ecol. Evol. 6, 36–50 (2021).Article 
PubMed 
PubMed Central 

Google Scholar 
van Bodegom, P. M., Douma, J. C. & Verheijen, L. M. A fully traits-based approach to modeling global vegetation distribution. Proc. Natl Acad. Sci. USA 111, 13733–13738 (2014).PubMed Central 

Google Scholar 
Moreno Martínez, A. et al. A methodology to derive global maps of leaf traits using remote sensing and climate data. Remote Sens. Environ. 218, 69–88 (2018).Article 

Google Scholar 
Pérez-Harguindeguy, N. et al. New handbook for standardized measurment of plant functional traits worldwide. Aust. J. Bot. 23, 167–234 (2013).Article 

Google Scholar 
Kattge, J. et al. TRY—a global database of plant traits. Glob. Change Biol. 17, 2905–2935 (2011).Article 

Google Scholar 
Kattge, J. et al. TRY plant trait database-enhanced coverage and open access. Glob. Change Biol. 26, 119–188 (2020).Article 

Google Scholar 
Jetz, W. et al. Monitoring plant functional diversity from space. Nat. Plants 2, 16024 (2016).Article 
PubMed 

Google Scholar 
Butler, E. E. et al. Mapping local and global variability in plant trait distributions. Proc. Natl Acad. Sci. USA 114, E10937–E10946 (2017).Article 
PubMed 
PubMed Central 

Google Scholar 
Boonman, C. C. et al. Assessing the reliability of predicted plant trait distributions at the global scale. Glob. Ecol. Biogeogr. 29, 1034–1051 (2020).Article 
PubMed 
PubMed Central 

Google Scholar 
Madani, N. et al. Future global productivity will be affected by plant trait response to climate. Sci. Rep. 8, 2870 (2018).Vallicrosa, H. et al. Global distribution and drivers of forest biome foliar nitrogen to phosphorus ratios (N:P). Glob. Ecol. Biogeogr. 31, 861–871 (2022).Article 

Google Scholar 
Meyer, H. & Pebesma, E. Predicting into unknown space? Estimating the area of applicability of spatial prediction models. Methods Ecol. Evol. 12, 1620–1633 (2021).Article 

Google Scholar 
Schiller, C. et al. Deep learning and citizen science enable automated plant trait predictions from photographs. Sci. Rep. 11, 16395 (2021).Aguirre-Gutiérrez, J. et al. Pantropical modelling of canopy functional traits using sentinel-2 remote sensing data. Remote Sens. Environ. 252, 112–122 (2021).Article 

Google Scholar 
Homolova, L. et al. Review of optical-based remote sensing for plant trait mapping. Ecol. Complex. 15, 1–16 (2013).Article 

Google Scholar 
Van Cleemput, E. et al. The functional characterization of grass-and-shrubland ecosystems using hyperspectral remote sensing: trends, accuracy and moderating variables. Remote Sens. Environ. 209, 747–763 (2018).Article 

Google Scholar 
Kattenborn, T., Fassnacht, F. E. & Schmidtlein, S. Differentiating plant functional types using reflectance: which traits make the difference? Remote Sens. Ecol. Conserv. 5, 5–19 (2019).Article 

Google Scholar 
Hauser, L. T. et al. Explaining discrepancies between spectral and in-situ plant diversity in multispectral satellite earth observation. Remote Sens. Environ. 265, 112684 (2021).Article 

Google Scholar 
Wäldchen, J. & Mäder, P. Plant species identification using computer vision techniques: a systematic literature review. Arch. Comput. Methods Eng. 25, 507–543 (2018).Article 
PubMed 

Google Scholar 
Jones, H. G. What plant is that? Tests of automated image recognition apps for plant identification on plants from the British flora. AoB Plants 12, plaa052 (2020).Article 
PubMed 
PubMed Central 

Google Scholar 
Hampton, S. E. et al. Big data and the future of ecology. Front. Ecol. Environ. 11, 156–162 (2013).Article 

Google Scholar 
WÜest, R. O. et al. Macroecology in the age of big data—where to go from here? J. Biogeogr. 47, 1–12 (2020).Article 

Google Scholar 
Mäder, P. et al. The Flora Incognita app—interactive plant species identification. Methods Ecol. Evol. 12, 1335–1342 (2021).Article 

Google Scholar 
Di Cecco, G. J. et al. Observing the observers: how participants contribute data to iNaturalist and implications for biodiversity science. BioScience 71, 1179–1188 (2021).Article 

Google Scholar 
Mahecha, M. D. et al. Crowd-sourced plant occurrence data provide a reliable description of macroecological gradients. Ecography 44, 1131–1142 (2021).Article 

Google Scholar 
Botella, C. et al. Jointly estimating spatial sampling effort and habitat suitability for multiple species from opportunistic presence-only data. Methods Ecol. Evol. 12, 933–945 (2021).Article 

Google Scholar 
iNaturalist Research-Grade Observations (GBIF, accessed 5 January 2022); https://www.gbif.org/dataset/50c9509d-22c7-4a22-a47d-8c48425ef4a7Callaghan, C. T. et al. Three frontiers for the future of biodiversity research using citizen science data. BioScience 71, 55–63 (2020).
Google Scholar 
Dickinson, J. L., Zuckerberg, B. & Bonter, D. N. Citizen science as an ecological research tool: challenges and benefits. Ann. Rev. Ecol. Evol. Syst. 41, 149–172 (2010).Article 

Google Scholar 
Kosmala, M. et al. Assessing data quality in citizen science. Front. Ecol. Environ. 14, 551–560 (2016).Article 

Google Scholar 
Boakes, E. H. et al. Patterns of contribution to citizen science biodiversity projects increase understanding of volunteers’ recording behaviour. Sci. Rep. 6, 33051 (2016).Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Bowler, D.E. et al. Temporal trends in the spatial bias of species occurrence records. Ecography 2022, e06219 (2022). https://doi.org/10.1111/ecog.06219GBIF Occurrence Download (GBIF, 4 January 2022); https://doi.org/10.15468/dl.34tjreBruelheide, H. et al. sPlot—a new tool for global vegetation analyses. journal of vegetation science. J. Veg. Sci. 30, 161–186 (2019).Article 

Google Scholar 
Sabatini, F. et al. sPlotOpen—an environmentally balanced, open access, global dataset of vegetation plots. Glob. Ecol. Biogeogr. 30, 1740–1764 (2021).Article 

Google Scholar 
Whittaker, R.H. et al. Communities and Ecosystems (Macmillan/Collier Macmillan, 1970).Olson, D. M. et al. Terrestrial ecoregions of the world: a new map of life on Earth: a new global map of terrestrial ecoregions provides an innovative tool for conserving biodiversity. BioScience 51, 933–938 (2001).Article 

Google Scholar 
Joswig, J., Wirth, C. & Schuman, M. Climatic and soil factors explain the two-dimensional spectrum of global plant trait variation. Nat. Ecol. Evol. 6, 36–50 (2022).Article 
PubMed 

Google Scholar 
Díaz, S. et al. The global spectrum of plant form and function. Nature 529, 167–171 (2016).Article 
PubMed 

Google Scholar 
Ploton, P. et al. Spatial validation reveals poor predictive performance of large-scale ecological mapping models. Nat. Commun. 11, 4540 (2020).Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Meyer, H. & Pebesma, E. Spatial validation reveals poor predictive performance of large-scale ecological mapping models. Methods Ecol. Evol. 12, 1620–1633 (2021).Article 

Google Scholar 
Schrodt, F. et al. Bhpmf—a hierarchical Bayesian approach to gap filling and trait prediction for macroecology and functional biogeography. Glob. Ecol. Biogeogr. 24, 1510–1521 (2015).Article 

Google Scholar 
Kuppler, J. et al. Global gradients in intraspecific variation in vegetative and floral traits are partially associated with climate and species richness. Glob. Ecol. Biogeogr. 29, 992–1007 (2020).Article 

Google Scholar 
Scheiter, S., Langan, L. & Higgins, S. I. Next-generation dynamic global vegetation models: learning from community ecology. New Phytol. 198, 957–969 (2013).Article 
PubMed 

Google Scholar 
Taubert, F. et al. Confronting an individual-based simulation model with empirical community patterns of grasslands. PLoS ONE 15, e0236546 (2020).Roger, E. & Klistorner, S. (2016) Bioblitzes help science communicators engage local communities in environmental research. J. Sci. Commun. https://doi.org/10.22323/2.15030206 (2016).Legendre, P. & Legendre, L. Numerical Ecology 3rd edn (Elsevier, 2012).Warton, D. I. et al. Smatr 3—an R package for estimation and inference about allometric lines. Methods Ecol Evol 3, 257–259 (2012).Article 

Google Scholar 
Wolf, S. et al. iNaturalist_traits: iNaturalist trait maps version 1 (January 5, 2022) Zenodo https://doi.org/10.5281/zenodo.6671891 (2022). More

Bates, H. W. XXXII. Contributions to an insect fauna of the Amazon Valley. Lepidoptera: Heliconidæ. Trans. Linn. Soc. Lond 23, 495–566 (1862).Article 

Google Scholar 
Müller, F. Ituna and thyridia: A remarkable case of mimicry in butterflies. Trans. Entomol. Soc. Lond. 1879, 20–29 (1879).
Google Scholar 
Baxter, S. W. et al. Convergent evolution in the genetic basis of Müllerian mimicry in Heliconius butterflies. Genetics 180, 1567–1577 (2008).Article 
CAS 

Google Scholar 
Sheppard, P. M., Turner, J. R. G., Brown, K., Benson, W. & Singer, M. Genetics and the evolution of Muellerian mimicry in Heliconius butterflies. Philos. Trans R. Soc. Lond. B, Biol. Sci. 308, 433–610 (1985).Article 
ADS 

Google Scholar 
Mallet, J. & Gilbert, L. E. Jr. Why are there so many mimicry rings? Correlations between habitat, behaviour and mimicry in Heliconius butterflies. Biol. J. Lin. Soc. 55, 159–180 (1995).
Google Scholar 
Brower, A. V. Parallel race formation and the evolution of mimicry in Heliconius butterflies: A phylogenetic hypothesis from mitochondrial DNA sequences. Evolution 50, 195–221 (1996).Article 
CAS 

Google Scholar 
Wilson, J. S. et al. North American velvet ants form one of the world’s largest known Müllerian mimicry complexes. Curr. Biol. 25, R704–R706. https://doi.org/10.1016/j.cub.2015.06.053 (2015).Article 
CAS 
PubMed 

Google Scholar 
Wilson, J. S., Williams, K. A., Forister, M. L., Von Dohlen, C. D. & Pitts, J. P. Repeated evolution in overlapping mimicry rings among North American velvet ants. Nat. Commun. 3, 1272 (2012).Article 
ADS 

Google Scholar 
Wilson, J. S., Pan, A. D., Limb, E. S. & Williams, K. A. Comparison of African and North American velvet ant mimicry complexes: Another example of Africa as the ‘odd man out’. PLoS ONE 13, e0189482. https://doi.org/10.1371/journal.pone.0189482 (2018).Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Plowright, R. & Owen, R. E. The evolutionary significance of bumble bee color patterns: A mimetic interpretation. Evolution 34, 622–637 (1980).Article 
CAS 

Google Scholar 
Williams, P. The distribution of bumblebee colour patterns worldwide: Possible significance for thermoregulation, crypsis, and warning mimicry. Biol. J. Lin. Soc. 92, 97–118 (2007).Article 

Google Scholar 
Hines, H. M. & Williams, P. H. Mimetic colour pattern evolution in the highly polymorphic Bombus trifasciatus (Hymenoptera: Apidae) species complex and its comimics. Zool. J. Linn. Soc. 166, 805–826 (2012).Article 

Google Scholar 
Koch, J. B., Rodriguez, J., Pitts, J. P. & Strange, J. P. Phylogeny and population genetic analyses reveals cryptic speciation in the Bombus fervidus species complex (Hymenoptera: Apidae). PLoS ONE 13, e0207080 (2018).Article 

Google Scholar 
Ezray, B. D., Wham, D. C., Hill, C. E. & Hines, H. M. Unsupervised machine learning reveals mimicry complexes in bumblebees occur along a perceptual continuum. Proc. R. Soc. B 286, 20191501 (2019).Article 

Google Scholar 
Bateson, W. The alleged “Aggressive Mimicry” of volucellæ. Nature 46, 585 (1892).Article 
ADS 

Google Scholar 
Poulton, E. B. The volucellœ as alleged examples of variation “almost unique among animals”. Nature 47, 126 (1892).Article 
ADS 

Google Scholar 
Cockerell, T. D. New social bees. Psyche A J. Entomol. 24, 120–128 (1917).Article 

Google Scholar 
Koch, J., Strange, J. & Williams, P. In: Bumble bees of the western United States (US Forest Service, San Francisco California, 2012).
Google Scholar 
Williams, P. H., Thorp, R. W., Richardson, L. L. & Colla, S. R. In: Bumble bees of North America: An identification guide Vol. 87 (Princeton University Press, Princeton, 2014).
Google Scholar 
Ruxton, G. D., Franks, D. W., Balogh, A. C. & Leimar, O. Evolutionary implications of the form of predator generalization for aposematic signals and mimicry in prey. Evol Int. J. Org. Evol. 62, 2913–2921 (2008).Article 

Google Scholar 
Rowe, C., Lindström, L. & Lyytinen, A. The importance of pattern similarity between Müllerian mimics in predator avoidance learning. Proc. R. Soc. Lond. Ser. B Biol. Sci. 271, 407–413 (2004).Article 

Google Scholar 
Beatty, C. D., Beirinckx, K. & Sherratt, T. N. The evolution of Müllerian mimicry in multispecies communities. Nature 431, 63 (2004).Article 
ADS 
CAS 

Google Scholar 
Chittka, L. & Osorio, D. Cognitive dimensions of predator responses to imperfect mimicry. PLoS Biol. 5, e339 (2007).Article 

Google Scholar 
Dittrigh, W., Gilbert, F., Green, P., McGregor, P. & Grewcock, D. Imperfect mimicry: A pigeon’s perspective. Proc. R. Soc. Lond. Ser. B Biol. Sci. 251, 195–200 (1993).Article 
ADS 

Google Scholar 
Sherratt, T. N., Whissell, E., Webster, R. & Kikuchi, D. W. Hierarchical overshadowing of stimuli and its role in mimicry evolution. Anim. Behav. 108, 73–79 (2015).Article 

Google Scholar 
Beatty, C. D., Bain, R. S. & Sherratt, T. N. The evolution of aggregation in profitable and unprofitable prey. Anim. Behav. 70, 199–208 (2005).Article 

Google Scholar 
Kazemi, B., Gamberale-Stille, G., Tullberg, B. S. & Leimar, O. Stimulus salience as an explanation for imperfect mimicry. Curr. Biol. 24, 965–969 (2014).Article 
CAS 

Google Scholar 
Kikuchi, D. W., Dornhaus, A., Gopeechund, V. & Sherratt, T. N. Signal categorization by foraging animals depends on ecological diversity. Elife. 8, e43965 (2019).Article 

Google Scholar 
Rapti, Z., Duennes, M. A. & Cameron, S. A. Defining the colour pattern phenotype in bumble bees (Bombus): A new model for evo devo. Biol. J. Lin. Soc. 113, 384–404 (2014).Article 

Google Scholar 
Wilson, J. S., Sidwell, J. S., Forister, M. L., Williams, K. A. & Pitts, J. P. Thistledown velvet ants in the desert mimicry ring and the evolution of white coloration: Müllerian mimicry, camouflage and thermal ecology. Biol. Lett. 16, 20200242 (2020).Article 

Google Scholar 
Ascher, J. & Pickering, J. Discover Life bee species guide and world checklist (Hymenoptera: Apoidea: Anthophila) (2019).iNaturalist. Available from https://www.inaturalist.org. Accessed [2022].Bombus Latreille, 1802 in GBIF Secretariat (2021). GBIF Backbone Taxonomy. Checklist dataset https://doi.org/10.15468/39omei accessed via GBIF.org on 2021-12-03. More

Asian elephants spend most of their time outside protected areas because they prefer the food they find there, an international team of scientists reports. But this behaviour is putting the animals and people in harm’s way, say researchers.The finding has important implications for the long-term survival of the animals because protected areas are a cornerstone of global conservation strategies to protect threatened species, say researchers.If protected areas do not contain animals’ preferred habitats, they will wander out, says Ahimsa Campos-Arceiz, who studies Asian elephants (Elephas maximus) at the Chinese Academy of Sciences’ Xishuangbanna Tropical Botanical Garden in Menglun, China. “It’s a good intention, but doesn’t always work out that way.”Human–elephant conflict is the biggest threat for Asian elephants. Over the past few decades, animals in protected areas have increasingly wandered into villages. They often cause destruction, damaging crops and infrastructure and injuring and even killing people.Wandering elephantsTo understand how effective protected areas are for conserving Asian elephants, Campos-Arceiz and his colleagues set out to get a precise picture of Asian-elephant movements. They collared 102 individuals in Peninsular Malaysia and Borneo, recording 600,000 GPS locations over a decade. They found that most elephants spent most of their time in habitats outside the protected areas, at the forest edge and in areas of regrowth. The findings were published in the Journal of Applied Ecology1 on 18 October.The researchers suspect that the elephants venture out because they like to eat grasses, bamboo, palms and fast-growing trees, which are common in disturbed forests and relatively scarce under the canopy of old-growth forests.Philip Nyhus, a conservation biologist who specializes in human–wildlife conflict at Colby College in Waterville, Maine, says Asian elephants live deep in dense forest and so are much more difficult to study than African elephants, which roam open savannahs. “The sample size is impressive,” he says.The finding is not unexpected given past anecdotal observations of elephant behaviour, says Nyhus. But now the data show that this is a common strategy for the survival of these animals, and not just something seen in a subset of the population. The research provides strong evidence for how to set up suitable protected areas that reduce the risk of elephants wandering out, he says.‘There will be conflict’The results do not diminish the importance of protected areas, which provide long-term safety for the animals, says Campos-Arceiz, who did the field work while at the University of Nottingham Malaysia in Selangor. “But they are clearly not enough.”The study suggests that “there will be conflict between humans and elephants”, says Guo Xianming, director of the Research Institute of Xishuangbanna National Nature Reserve in Jinghong.Asian elephants wander into villages owing to a combination of reasons: an increase in elephant populations, forests in many reserves have grown denser and have become unsuitable for the animals, and increasing habitat loss and degradation outside.Last year, two herds of elephants made global headlines as they wandered out of the Xishuangbanna National Nature Reserve and travelled for hundreds of kilometers, wreaking havoc along the way. One herd spent five weeks at the botanical garden where Campos-Arceiz works. “It was intense,” he says.There is an urgent need to understand how people and elephants can better share the landscape, says Guo. And the first step is by better protecting people’s lives and livelihoods. “It’s the only way of peaceful co-existence.”
The reporting of the story was supported by International Women’s Media Foundation’s Howard G. Buffett Fund for Women Journalists. More