Data of the Insect Biome Atlas: a metabarcoding survey of the terrestrial arthropods of Sweden and Madagascar

Site selection

In Sweden, sampling sites were selected following a stratified design based on the major landscape types as identified by the National Inventory of Landscapes in Sweden (NILS)²⁷, for which extensive data on land cover, land use and landscape structure spanning 18 years (2003–2020) are available (https://landskap.slu.se/nils/dv). Malaise traps were set up within a radius of 5 km from the closest NILS sampling plot. A total of 203 Malaise traps were deployed within six major habitats across 100 sites: 102 traps in forests, 33 in croplands, 26 in wetlands,16 in grasslands, 14 in alpine and 12 in urban areas (Fig. 1). The proportion of Malaise traps by habitat is representative of the contribution of each habitat type to the overall Swedish landscape (Table 1), with down-weighting of the most common habitat (forests, which account for 60% of the landscape according to NILS) and some up-weighing of the remaining, sparser habitats (croplands and wetlands, which account for 8% of the landscape each; and urban sites and grasslands, which each account for 3% of the landscape). At 21 of the 100 sampling sites, we implemented a hierarchical design by deploying up to 6 Malaise traps instead of one. In Sweden there were 12 multi-trap sites with between 2 and 6 traps, and in Madagascar there were 9 multi-trap sites with between 2–4 traps. Distances between Malaise traps at these multitrap sites ranged from 90 to 1200 meters, sampling two different habitats when possible (three Malaise traps per habitat). In Madagascar, sampling sites were associated with forested habitats located within protected areas only. In total we deployed 50 Malaise traps across 33 sites: 27 Malaise traps in rainforests (16 sites) and 23 Malaise traps in dry forests (17 sites) (Fig. 1). At eight of the 33 sites, we implemented a hierarchical design by deploying 3 traps within a single habitat. Distances between Malaise traps at these multitrap sites ranged from 100 to 4700 meters. Malaise traps were set up by a team of fieldworkers together with the trap managers at each site: 150 volunteers in Sweden and 107 local community members and park rangers in Madagascar. At the time of setting up the Malaise traps, trap managers were trained on how to set up and maintain the trap, retrieve the samples from the trap and collect metadata associated with the sampling event.

Standing characteristics and other abiotic data

At each Malaise trap location in Madagascar, we measured a set of standing characteristics. To assess tree density, we set up four 10 m long transects, one on each side of the trap, and measured the diameter at breast height (DBH) of all trees with its roots or stem within 10 cm of the transect. For multi-stemmed trees, we measured only the DBH of the largest stem. We included measurements for trees, lianas (L), palms (P) and ferns (F). We also measured the circumference of all trees with a DBH larger than 30 cm within a 10 m radius of the Malaise trap. Finally, we measured canopy cover at each Malaise trap location by taking five photographs of the canopy at 2 meters height straight up to the canopy. One photograph was taken at the center of the Malaise trap and the other four photographs were taken five meters away from the trap, one on each side of the trap. Each photograph was analysed with ImageJ software to get a percentage of canopy cover.

At each Malaise trap location for both Sweden and Madagascar we also measured soil nutrients and soil pH. Topsoil (0–20 cm) was sampled at 5 sites around each Malaise trap: one soil core (6 cm diameter) at the center of the trap and one soil core on each of the four “sides” of the trap, five meters away from the trap. Soil samples were taken as composite samples from the five locations. At each site, prior to taking the soil cores, we measured leaf litter depth using a ruler and soil humidity with the SM150 soil moisture kit, Delta-T, United Kingdom. Soil samples from Sweden were sent to Eurofins to measure the following macronutrients: AL-extractable P, K, Mg and Ca; NH₄; NO₃; K/Mg ratio and mineral N (Nmin). The concentrations of soil nutrients are presented as mg·100 g⁻¹ air-dry sieved soil (<2mm). Soil samples from Madagascar were sent to the Laboratoire des Radioisotopes, Madagascar, to measure the following macronutrients: organic C (g/Kg), total N (g/Kg), total P (g/Kg), exchangeable K (cmol/Kg), exchangeable Ca (cmol/Kg) and exchangeable Mg (cmol/Kg).

Sampling for biotic characterization

Malaise trap samples

Arthropods were collected in individually barcoded bottles pre-filled with 400 mL of 95% ethanol attached to each Malaise trap. To facilitate the recording of metadata associated with each sampling event, trap managers used a mobile application specifically designed for the project. The app allowed scanning the individually barcoded Malaise trap and sample bottle at the time of collection so that each sampling event was automatically associated with a specific trap. When the barcode of each sample was scanned, the app automatically registered the GPS coordinates and the date and time of the collection event in addition to any other metadata manually inserted by the user, such as the trap condition at collection. In Sweden, Malaise traps were active between January and December 2019. Samples were collected every week during spring to autumn (March/April to September/October depending on latitude) and monthly or bi-weekly in the winter (October/November to March/April, depending on latitude). In the northern part of the country, we did not collect at all during the winter months, when snow and strong winds prevented proper operation of the traps. The sampling strategy was based on previous experiences from the Swedish Malaise Trap Project²², and was deemed to result in minimal loss of sampled specimens and an acceptable loss of phenological resolution. The effort resulted in 4,753 insect community samples from Sweden. In Madagascar, Malaise traps were active between August 2019 and July 2020. Samples were collected every week throughout that period, resulting in 2,566 insect community samples. Every fourth sample from each trap location (638 samples in total, roughly one sample per month), was left at the Madagascar Biodiversity Center (https://www.madagascarbio.org/) to help build a natural history collection of Malagasy insects in the country of origin. The remaining samples were shipped to Sweden for DNA extraction and metabarcoding as described below.

Soil and litter samples

Soil and leaf litter arthropod communities were sampled during the growing season (26th of June 2019 to 27th of July 2019) at each Malaise trap location in Sweden. Leaf litter arthropod communities were sampled by collecting a total volume of 1 L of leaf litter from five different sites around each Malaise trap, 2.5 meters away from the nearest extremity of the trap. Soil arthropod communities were sampled using soil cores (5.5 cm diameter) with a depth of 10 cm at exactly the same five sites where litter samples were taken. The five soil core samples from each Malaise trap location were pooled and mixed by hand immediately after collection, then stored in a cool place. Within 48 h after collection, living arthropods were extracted from soil and litter samples over a period of 96 hours using Berlese funnel³⁵. In brief, arthropods were extracted by placing the soil samples in metal net baskets (1.5 mm mesh) in the top of stainless metal funnels (diam. 18 cm). Twenty funnels were placed together on a wooden board with holes for the funnels. Approximately, 17 cm above the funnels, a heating plate (max 950 watt, starting at 25% and gradually increasing to 75%) were switched on to create a heat gradient and LED-lamps were switched on to provide light. To cause the arthropods to exit the substrate before being trapped in the dried-out structure, we gradually increased the temperature over the first 24 hours to a temperature of 52 °C. This temperature was then held constant for the remaining 72 hours. Arthropods were collected directly into 100 mL plastic bottles filled with 95% ethanol. Leaf litter arthropod communities were also sampled at each Malaise trap location in Madagascar. Here we collected four samples in each direction of the Malaise trap (back, front, left and right), five meters away from it. Leaf litter sampling involved sifting and concentrating 2 L of leaf litter from a 1 m² square using a Winkler-sifter. Before sifting, the leaf litter was minced with a machete to dislodge any insects hiding in the twigs or decayed logs. After sifting, the leaf litter was stored in a cloth bag before extracting the arthropods. Within 12 hours of sampling, living arthropods were extracted at ambient temperature overnight into a 100 mL bottle filled with 95% ethanol using a mini-Winkler extractor for a total of 12 hours³⁶.

DNA extraction

Malaise trap samples

DNA was extracted from Malaise trap samples using both non-destructive (mild lysis and preservative ethanol) and destructive methods (homogenization) (Fig. 2).

1. (1)

   Mild lysis: DNA was extracted from 6,483 Malaise trap samples (4,560 from Sweden and 1,923 from Madagascar) using a non-destructive mild-lysis protocol (FAVIS protocol, steps 1-17²⁹). In brief, ethanol was first decanted from each sample and insect wet biomass was measured. Lysis buffer, proteinase K and biological spike-ins (specimens from foreign species, i.e., species that don’t occur in the respective countries) were added and samples were incubated for 2h45m at 56 °C in a dry shaking incubator. After the incubation period the lysate was drained and each insect community was remixed with the previously-decanted ethanol for long-term storage. DNA was purified from 225 μL of lysate using silica-coated magnetic beads with the KingFisher Cell and Tissue DNA kit on a KingFisher Flex 96 robot (both Thermo Scientific, ThermoFisher Inc, United States of America) according to manufacturer instructions. After DNA purification, DNA extracts from 12 samples in each 96-well plate were quantified using Qubit™ dsDNA HS Assay Kit on a Qubit Fluorometer (Invitrogen™, Thermo Scientific, ThermoFisher Inc, United States of America). DNA concentration of those samples is available as Supplementary Table S1. The list and number of biological spike-ins added in each country (SE and MG) can be retrieved at from Figshare³⁷.

2. (2)

   Homogenization: Approximately every fourth sample collected at each sampling site in Sweden (n = 873) was further processed using a destructive homogenization protocol³⁰ after mild lysis. In brief, after decanting preservative ethanol we homogenized each bulk sample into an “arthropod soup” using the ULTRA-TURRAX® Tube Drive P, IKA®-Werke GmBH & Co. KG, Germany. For the homogenization we used single use DT-50 Dispersing tubes. After homogenization, the entire “arthropod soup” was digested with lysis buffer and proteinase K for 2h45min at 56 °C in a dry shaking incubator. To make sure that we used all available DNA from each sample, we combined the homogenate with the respective lysate obtained during mild lysis before proceeding with DNA purification. At this time we also added a standardized amount (5 million copies) of two synthetic oligonucleotide sequences (synthetic spike-ins) to each homogenate aliquot. Synthetic spike-ins were produced as described in Iwaszkiewicz-Eggebrecht et al.³⁸ and their sequences can be found at Figshare³⁷. DNA was purified from a 225 µL subsample of homogenate using silica-coated magnetic beads with the KingFisher Cell and Tissue DNA kit on a KingFisher Flex 96 robot (both Thermo Scientific, ThermoFisher Inc, United States of America) according to manufacturer instructions. After DNA purification, DNA extracts from 12 samples (including positive and negative controls) in each 96-well plate were quantified using Qubit™ dsDNA HS Assay Kit on a Qubit Fluorometer (Invitrogen™, Thermo Scientific, ThermoFisher Inc, United States of America). DNA concentration of those samples is available as supplementary material (Dataset 1).

3. (3)

   Preservative ethanol: For 15 Malaise trap samples we also extracted DNA from the preservative ethanol (prior to mild lysis and homogenization). Ethanol was first decanted from each sample as described in step 7 of the FAVIS protocol²⁹. The decanted ethanol was then manually filtered using a Millipore® Sterivex™ filter unit (pore size of 0.22 μm) (Merck KGaA, Germany). After filtration, DNA was lysed inside the Sterivex™ unit by adding 540 µl of lysis buffer (ATL, Qiagen, Germany) and 60 µl proteinase K (Qiagen, Germany) and incubating the unit at 56 °C overnight. DNA was transferred from the Sterivex™ unit into a DNeasy® Blood and Tissue Kit column (Qiagen, Germany) for purification, following manufacturer instructions.

Schematic representation of sample processing and bioinformatic pipeline. Malaise trap sample processing in the lab (left panel). DNA was extracted from (A) the ethanol filtered from malaise trap samples; (B) the lysates using the FAVIS mild lysis protocol²⁹; and (C) the homogenates³⁰. Before lysis, biological spike-ins and synthetic spikes were added to the samples in B and C, respectively. After lysis, DNA was purified using silica-coated magnetic beads with the KingFisher Cell and Tissue DNA kit on a KingFisher Flex 96 robot (1). After DNA purification, a 418 bp fragment of the mtDNA cytochrome c oxidase subunit 1 (CO1) gene was amplified using a 2-step PCR approach. In the first step (2), the target region was amplified using broad-spectrum primers BF3 CCHGAYATRGCHTTYCCHCG⁴² and BR2 CDGGRTGNCCRAARAAYCA⁴³. In the second step (3), indexes were added and Illumina adapters completed. Samples were pooled and library pools were sequenced on a NovaSeq 6000 instrument using the ‘NovaSeqXp’ workflow in ‘SP 500-cycle’ flow cells, with 384 double-uniquely indexed samples per lane. Schematic representation of processing sequence data from raw reads to ASV clusters (right panel). Briefly, paired end reads were trimmed in a series of steps using the program cutadapt⁴⁷ and filtered to retain only sequences that were between 403 and 418 nt in length (in incremental steps of 3 nt) and that did not contain any in-frame stop codons. Preprocessed reads were then denoised using DADA2⁴⁹ to infer amplicon sequence variants (ASVs). The ASVs were then processed using the HAPP pipeline³¹ to taxonomically annotate the ASVs (using SINTAX classifier and a purposely built CO1 database). Taxonomic assignments obtained from SINTAX were refined using phylogenetic methods with EPA-NG⁵⁵ into an insect phylogeny⁵⁶, followed by taxonomic assignment with gappa⁵⁷. ASVs were then clustered into OTUs using Swarm⁵⁹. Finally, the clustered data was cleaned from NUMTs and other types of noise using the NEEAT algorithm³¹.

Soil and litter samples

Arthropod litter samples from Madagascar were processed using the same mild lysis protocol used for the Malaise samples described above, with the exception of not adding biological spike-ins to the samples. For Sweden, we used a 1 mm wire-sieve to separate each sample (from soil or litter) into two components based on size: the macrofauna subsample, composed mainly of adult and larval Coleoptera, and the mesofauna subsample, dominated by mites and springtails. To remove debris and dirt accumulated in the mesofauna subsamples, we processed them further using a combination of flotation in distilled water, adapted from the Flotation-Berlese-Flotation protocol in^39,40, followed by vacuum pump filtration (41 µm nylon filter). DNA was extracted separately from the macrofauna and the mesofauna subsamples using the Thermo Scientific KingFisher Cell and Tissue DNA Kit, ThermoFisher Inc, United States of America. Ethanol was first dried from each subsample in a dry incubator at 40 °C for 4-5 hours. After drying, specimens were manually homogenized with a pestle in a 50 mL single use falcon tube and lysed overnight at 56 °C by adding 800 µL of lysis buffer and 100 µL of proteinase K (provided in the Kingfisher Cell and Tissue DNA kit). After each use pestels were sterilized by washing in a 10% bleach bath for 15 minutes, followed by rinsing with water. DNA was then purified from a 225 µL aliquot of homogenate using silica-coated magnetic beads with the KingFisher Cell and Tissue DNA kit on a KingFisher Flex 96 robot (both Thermo Scientific, ThermoFisher Inc, United States of America) following manufacturer instructions. DNA extracts of all samples were quantified using Qubit dsDNA HS Assay Kit on a Qubit Fluorometer (Invitrogen™, Thermo Scientific, ThermoFisher Inc, United States of America), and the original sample was reassembled by combining DNA extracts from each subsample pair in a ratio of 1:10 (amount of DNA of macrofauna: amount of DNA of mesofauna). This minimizes the bias in the sequencing depth due to biomass differences between the two components of each sample^39,40.

Library preparation and sequencing

To characterize arthropod communities, we amplified a 418 bp fragment of the mtDNA cytochrome c oxidase subunit 1 (CO1) gene, using the purified DNA extracted from the Malaise trap samples (lysates, homogenates and preservative ethanol) and soil and litter samples. Each bulk sample was metabarcoded using a two-step PCR approach for library preparation (method 4 of⁴¹). In the first step (PCR 1), the target region was amplified using broad-spectrum primers BF3 CCHGAYATRGCHTTYCCHCG⁴² and BR2 CDGGRTGNCCRAARAAYCA⁴³. The primers were supplemented with 5′-end Illumina sequence adapters (forward: ACACTCTTTCCCTACACGACGCTCTTCCGATCT-3′, reverse: 5′-GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCT). To increase the complexity of the libraries, each primer was further complemented with variable length inserts (TGA, GA, A, or no base for the forward primer, and GAT, AT, T, or no base for the reverse primer)between the adapter sequence and the target-binding region, generating phased primers in equal proportions^44,45. PCR 1 reactions were carried out in a final reaction volume of 40 μL containing 20 μL Qiagen, Germany, Multiplex PCR Master mix, 1 μM of each primer, and 4 μL of template DNA (for mild lysis and preservative ethanol samples). As the DNA concentration for homogenates was relatively higher, the PCR reactions were carried out in a final volume of 10 μL with 1 μL of template DNA. The PCR conditions were 95 °C for 15 min, 25 cycles of 94 °C for 30 s, 50 °C for 90 s and 72 °C for 90 s, followed by a final elongation step of 72 °C for 10 min. PCR 1 products were cleaned with magnetic beads (Carboxyl-modified Sera-Mag Magnetic Speed-Beads, Hydrophobic, CYTIVA), using 2/1 (v/v) magnetic beads to sample ratio. In the second step (PCR 2, indexing PCR), indexes were added and Illumina adapters completed. The indexing primer design followed the Adapterama scheme^41,46 with 10 bp indexes as follows: index i5 (forward): AATGATACGGCGACCACCGAGATCTACACxxxxxxxxxxACACTCTTTCCCTAC index i7 (reverse): CAAGCAGAAGACGGCATACGAGATxxxxxxxxxxGTGACTGGAGTTCAG. Libraries were double-uniquely indexed – in other words, each forward and each reverse index was used for only one library in a given sequencing lane. PCR 2 conditions were 95 °C for 15 min, 7 cycles of 94 °C for 30 s, 50 °C for 90 s and 72 °C for 90 s, followed by a final elongation step of 72 °C for 10 min. To form the sequencing pool, all samples were pooled approximately equimolar based on the intensity of the band in an agarose gel, and the resulting sample pool was then purified with the Promega ProNex® Size-Selective Purification System, using 1/1.5 (v/v) pool to magnetic beads ratio. The quality of the library pool was then checked with an Agilent DNA High Sensitivity Kit on a 2100 Bioanalyzer instrument (Agilent Technologies). Library pools were sequenced on a NovaSeq 6000 instrument using the ‘NovaSeqXp’ workflow in ‘SP 500-cycle’ flow cells, with 384 double-uniquely indexed samples per lane, or a total of 768 libraries per flow cell (8 × 96 well plates). This sequencing was performed at the Swedish National Genomics Infrastructure (NGI) at SciLifeLab (Solna, Sweden). The detailed step-by-step protocol can be found in²⁹.

Processing sequencing data

Sequences were preprocessed (read trimming and filtering) using a Snakemake workflow available at Github: https://github.com/biodiversitydata-se/amplicon-multi-cutadapt. In this workflow, paired end reads were trimmed in a series of steps using the program cutadapt⁴⁷ (v3.1):

1. 1.

   Discard all reads with the Illumina TruSeq adapters in either the 5’ or 3’ end of sequences.

2. 2.

   Search for and trim primer sequences from the start of reads in R1 and R2 files using forward and reverse primers, respectively. Remove any untrimmed reads. This step is done with additional settings ‘–no-indels’ and ‘-e 0’ in order to only accept perfect matches.

3. 3.

   Discard any remaining reads that still contain primer sequences.

4. 4.

   Trim reads to a fixed length. This length is calculated by subtracting the length of the longest primer from the read length defined by the ‘expected_read_length’ parameter under the cutadapt: section in the config file (default value is 251).

Reads were then filtered to retain only sequences that were between 403 and 418 nt in length (in incremental steps of 3 nt) and that did not contain any in-frame stop codons. Preprocessed reads were denoised using the nf-core/ampliseq Nextflow workflow⁴⁸ (v2.4.0) which uses the DADA2 algorithm⁴⁹ to infer amplicon sequence variants (ASVs) from the preprocessed reads. Due to the indexing scheme used with unique dual indexes per sample it was not necessary to perform per-sample abundance filtering to correct for mistagging^50,51.

The ASVs were then processed using the HAPP pipeline (https://github.com/insect-biome-atlas/happ) described separately³¹ to taxonomically annotate the ASVs, remove chimeras, cluster the ASVs into OTUs, and remove NUMTs and other noise from the data. Specifically, ASVs were taxonomically annotated using SINTAX⁵², as implemented in vsearch⁵³, against a custom-made reference CO1 database available at Figshare⁵⁴. This reference database was assembled from sequences in the BOLD database³⁴ as follows. Firstly, nucleotide sequences and metadata linking record ids to BOLD BINs were downloaded from the GBIF Hosted Datasets (ibol_2022_01_17.zip). This was merged with taxonomic information for BOLD BINs obtained from the GBIF backbone (backbone.zip from 2022-11-23). The data was then filtered to only keep records annotated as ‘CO1-5P’ and assigned to a BOLD BIN ID. The taxonomic information was parsed in order to assign species names and resolve higher-level ranks for each BOLD BIN ID. Sequences were processed to remove gap characters and leading and trailing ‘N’s. After this, any sequences with remaining non-standard characters were removed. Sequences were then clustered at 100% identity using vsearch. This clustering was done separately for sequences assigned to each BOLD BIN ID. These steps were implemented in a Python package called coidb, available at Github https://github.com/insect-biome-atlas/coidb. Taxonomic assignments obtained from the kmer-based SINTAX classifier were refined using phylogenetic methods. Specifically, ASV sequences classified as Insecta or Collembola (class) but unclassified at order level by SINTAX were reassigned by phylogenetic placement with EPA-NG⁵⁵ into an insect phylogeny⁵⁶, followed by taxonomic assignment with gappa⁵⁷. Assignments obtained this way were used to update the SINTAX taxonomy, but only at the order level, leaving child ranks with the ‘unclassified’ prefix.

Following taxonomic assignments ASVs were further processed to remove chimeras with uchime⁵⁸ and the remaining non-chimeric sequences were clustered using Swarm⁵⁹ (v3.1.0) with setting ‘-d 15’. To choose the best setting for Swarm we evaluated the performance of different settings by calculating precision and recall values in Sundh et al.³¹. In the datasets, we provide taxonomic assignments for all ASVs using several different approaches. We also provide consensus annotations for OTU clusters using an abundance-based consensus approach. For each cluster, starting from the lowest rank (BOLD BIN here), each unique taxonomic name was weighted by the sum of reads across samples and ASVs with said name. These sums were then normalized to percentages. If a single taxonomic name made up at least 80%, then that name was assigned to the cluster, including the assignments of parent ranks. If no single name reached the 80% consensus threshold, the process was iterated for the parent rank. Ranks for which no consensus could be reached were prefixed with ‘unresolved.’ followed by the name of the most resolved consensus taxonomy. Using this procedure, it is, in theory, possible that OTUs are assigned a consensus taxonomic label from an unclassified ASV, or to an ASV with an ambiguous name. However, this happens very rarely in practice. Out of the 33,989 OTUs in the Swedish dataset, only 282 (0.8%) get an unclassified species level assignment even though an ASV with a properly assigned species label is present in the same OTU. For Madagascar this happens for 0.1% (77) of the 77,599 OTUs. For each cluster, the ASV sequence with the highest median of normalized read counts across samples was selected as representatives of the cluster. Ties were broken by taking the ASV with the highest mean.

The clustered data was further cleaned from NUMTs and other types of noise using the NEEAT algorithm, which takes taxonomic annotation, correlations in occurrence across samples (‘echo signal’) and evolutionary signatures into account, as well as cluster abundance³¹. We used default settings for all parameters in the evolutionary and distributional filtering steps, and removed clusters unassigned at the order level and with less than three reads summed across each dataset. Additionally, we removed clusters present in more than 5% of blanks.

Biomass and count data

To allow an assessment of how the wet biomass of a Malaise trap sample translates to the number of specimens, we counted all the specimens for 24 Malaise trap samples from Sweden. We complemented these data with wet biomass estimates and specimen counts for 224 Malaise trap samples from a separate Swedish Malaise trapping campaign in 2018–2019, the Swedish Insect Inventory Project (SIIP).

The data (Fig. 3) show that the number of specimens is only approximately proportional to the biomass of a sample (linear model without intercept, adjusted R² 0.81; Fig. 3a). Specifically, there is a slight tendency for the larger samples (in terms of biomass) to contain more specimens than if the relation was strictly proportional, as shown in a log-log model (R² 0.79, regression coefficient 1.09 ± 0.04; Fig. 3b). Fitting a linear model with intercept does not support the alternative explanation of a constant amount of alcohol residue causing this (adjusted R² 0.69, intercept positive and not negative as expected under the hypothesis). Using the proportional model, the Swedish Malaise trap material is estimated to contain 7.0 M specimens, and the Madagascar material 1.7 M specimens. Accounting for the deviation from proportionality by applying the log-log regression equation to each sample separately, the Swedish material is instead estimated to contain 5.6 M specimens and the Madagascar material 1.2 M specimens.

Linear regression between biomass of Malaise trap samples and the number of individuals (insect specimens). The data points are individual Malaise trap samples; the 224 black data points are from the SIIP project and the 24 red data points from the IBA project. The blue line is a fitted straight line with the 95% confidence interval marked in gray. The equations of the fitted lines are shown in blue. (a) The fitted line is forced through the origin, as the biomass is zero at zero individuals. The number of individuals per gram is estimated at 273 ± 17. The R-squared is 0.81. (b) The same dataset with logged axes. The slope is 1.09 + −0.07, that is, significantly larger than 1.0. This indicates that large samples (in terms of biomass) tend to have slightly more specimens in them than a strict proportional relationship between biomass and specimens would suggest. The R-squared is 0.79.