(a) Model Inputs

We collected with 121 Malaise-trap samples for seven days into 100% ethanol at 89 sampling points in and around the H.J. Andrews Experimental Forest (HJA), OR, USA in July 2018 (figure 1). Sites were stratified by elevation, time since disturbance, and inside and outside the HJA (inside, a long-term research site with no logging since 1989; outside, continued active management). HJA represents a range of previously logged to primary forest, but with notably larger areas of mature and old-growth forest reserves than the regional forest mosaic, which consists of short-rotation plantation forests on private land and a recent history of active management on public land.

Sampling design and taxonomic diversity of the Malaise trapping campaign. (a) Sampling points in and around the H.J. Andrews Experimental Forest (red line), OR, USA. The study area consists of old-growth and logged (grey patches) deciduous and evergreen forest under different management regimes. Arthropods were sampled with Malaise traps at 89 sampling points in July 2018, with one trap at 57 points (white circles) and with two traps 40 m apart at 32 points (white squares). Elevation indicated with a green to white false-colour gradient. (b) Taxonomic distribution of all detected operational taxonomic units (OTUs) from the samples. Node size and colour are scaled to the number of OTUs. See the electronic supplementary material, figure S4 for a heat tree of the 190 included OTUs.

We extracted the DNA from each Malaise-trap sample by soaking the arthropods in a lysis buffer and sent it to Novogene (Beijing, China) for whole-genome shotgun sequencing.

On the output fastq files, we carried out ‘in silico’ PCR using Kelpie 2.0.11 [45] and the BF3 + BR2 primers from [46], outputting 5560 unique DNA-barcode sequences. After 97%-similarity clustering and filtering for erroneous sequences, we were left with 1225 operational taxonomic units (OTUs) as the reference barcode set.

We then mapped the reads of each sample to the reference barcodes, creating a 121 − sample × 1225 − OTU table. A species was accepted as being in a sample if reads mapped at high quality along more than 50% of its barcode length, following acceptance criteria from Ji et al. [47].

To predict species occurrences in the areas between the sampling points, we collected 58 continuous-space predictors (electronic supplementary material, table S1), relating to forest structure, vegetation reflectance and phenology, topography, and anthropogenic features, restricting ourselves to predictors that can be measured remotely. The forest-structure variables were from airborne LiDAR data collected from 2008 to 2016, which correlate with forest structure in US Pacific northwest coniferous forests, such as mean diameter, canopy cover and tree density [48]. The vegetation-related variables came from Landsat 8 individual bands, plus standard deviation, median, 5% and 95% percentiles of those bands over the year, and indices of vegetation status, e.g. normalized difference vegetation index. Both the proportion of canopy cover and annual Landsat metrics were calculated within radii of 100, 250 and 500 m, given that vegetation structure at different spatial scales is known to drive arthropod biodiversity [49]. The topography variables were calculated from LiDAR ground returns, including elevation, slope, eastness and northness split from aspect, topographic position index, topographic roughness index (TRI) [50], topographic wetness index [51] and distance to streams, based on a vector stream network (http://oregonexplorer.info, accessed 24 October 2019). The anthropogenic variables include distance to nearest road, proportion of area logged within the last 100 and within the last 40 years, within radii of 250, 500 and 1000 m, and a categorical variable of inside or outside the boundary of the HJA. They are not directly derived from remote-sensing data, but we included them because they could be derived from remote-sensing imagery. We then reduced our 58 environmental covariates to 29, removing the covariates that were most correlated with the others (as measured by variance inflation factor). The 29 retained covariates include six anthropogenic activities, two raw Landsat bands, seven indices based on annual Landsat data, six canopy/vegetation-related variables from LiDAR, and eight topography variables (electronic supplementary material, table S1 and figure S5), which we mapped across the study area at 30 m resolution.