Weather Radars Reveal Environmental Conditions for High Altitude Insect Movement Through the Aerosphere

Abstract

High-flying insects that exploit tropospheric winds can disperse over far greater distances in a single generation than species restricted to below-canopy flight. However, the ecological consequences of such long-range dispersal remain poorly understood. For example, high-altitude dispersal may facilitate more rapid range shifts in these species and reduce their sensitivity to habitat fragmentation, in contrast to low-flying insects that rely more on terrestrial patch networks. Previous studies have primarily used surface-level variables with limited spatial coverage to explore dispersal timing and movement. In this study, we introduce a novel application of niche modelling to insect aeroecology by examining the relationship between a comprehensive set of atmospheric conditions and high-flying insect activity in the troposphere, as detected by weather surveillance radars (WSRs). We reveal correlations between large-scale dispersal events and atmospheric conditions, identifying key variables that influence dispersal behaviour. By incorporating high-altitude atmospheric conditions into niche models, we achieve significantly higher predictive accuracy compared with models based solely on surface-level conditions. Key predictive factors include the proportion of arable land, altitude, temperature, and relative humidity.

1. Introduction

Insects play a critical and multifaceted role in ecological systems. They serve as vital intermediaries within trophic networks, bridging the gap between producers and large-bodied consumers [1,2,3]. The role that insects play in resource processing is central to nutrient cycling, facilitating decomposition and subsequent nutrient assimilation [3,4]. Moreover, insects act as keystone species that regulate plant community composition through herbivory and pollination services [1,2,3,4]. Due to their wide-ranging involvement in ecological processes, changes in insect population structure or distribution (e.g., range shifts, see Doak and Morris, 2010; Tomiolo and Ward, 2018 [5,6]) can produce impacts that have cascading effects across the entire ecosystem [2,7,8]. Ecosystem change brought about by insects is also not confined to the local population but can impact other nearby populations and habitats through the flows of individuals (and associated nutrients and biomass) among habitats [2,7,8].

The flow of insects and associated nutrients and functions is driven by the capacity for movement, of which flight is a common and highly effective mode. The ability of flying insects to enter the air column and traverse meaningful distances is a significant factor in insect dispersal capabilities and determines the exchange of individuals among habitats. However, there is a wide range of powered flight ability between families of insects with different morphotypes [9,10,11,12]. For example, odonates and lepidopterans have been found to be capable of long-distance migrations in a single generation [13,14,15], and lepidopterans have been shown to carry out directed flight while accounting for crosswinds [16,17]. Smaller, shorter-winged insects like aphids (3.2–4.3 mm, see Bell and Shephard (2024) [18]) and thrips are also found in large numbers in the air column during mass emergence events [18,19,20], although their mode of flight leaves them highly vulnerable to loss of flight control (aphids fly at up to 0.70 m·s⁻¹ in laboratory conditions, see Bell and Shephard (2024) [18]), even under fairly slow windspeeds [21,22,23,24,25].

Regardless of flight ability, when insects fly above the canopy layer, even weak fliers become capable of travelling great distances by exploiting the wind. In contrast, dispersal through the terrestrial environment is limited to powered flight alone. Dispersal at higher altitudes is underexplored and rarely included by range shift and habitat connectivity studies, which have been used as the basis for future conservation and pest management policy [26,27,28,29,30]. For example, it is unclear to what extent tropospheric winds contribute to dispersal, how wind-assisted flight contributes to the functional connectivity of habitats that might be considered unconnected under conventional models of dispersal, and the taxonomic breadth of species exploiting high-altitude dispersal routes. Thus, the consequences of high-altitude dispersal remain unexplored, but such a phenomenon raises the possibility of complex responses to habitat fragmentation and resulting climate change-induced range shifts [31,32,33,34]. As high-flying insects (active at heights above the ground > 100 m) could be capable of continuing dispersal or migration between increasingly isolated terrestrial patches, using the aerosphere as a corridor, the functional connectivity of habitat patches could be greater than for below-canopy fliers [33,35,36,37]. However, the ecological modelling literature has largely neglected the aerosphere as a distinct habitat due to traditionally constrained observational capabilities, leading to the assumption that atmospheric interactions constitute only a brief phase in organism life histories [38,39]. Therefore, there is a need to recognise the ecological importance of ‘flight niches’ in insect dispersal ecology, a concept already applied to a few pioneering avian studies [40,41].

One key limitation to studying flight niches is the difficulty of observing organisms at high altitudes. Previous aeroecological studies have addressed this challenge chiefly through vertical-looking radars (VLRs) and traditional field sampling. VLRs have been able to observe the stratification of insect activity into layers at given altitudes [16,42,43,44], as well as favouring of specific windspeeds (the flight boundary layer concept) and directions during flight [15,24,42,45]. However, VLR studies have been restricted to studying the vertical distribution of insects at single sites, over short timescales [42,46,47]. Field studies, on the other hand, have provided richer diversity and population data on flying insects over wide spatial scales, but with poor spatial resolution, and they are typically restricted to low-altitude observation [10,15].

Compared with VLR and field studies, weather surveillance radars (WSRs) like the NCAS mobile X-band [48] can provide greater coverage of insect aerial presence (the X-band typically detects insects up to a ~30 km horizontal radius around the radar, and from 100 to 3000 m altitude above sea level) and over annual to decadal timescales [9,13,49,50,51]. Unlike VLRs, which detect individual scatterers [52] in a thin, vertically pointing beam, WSRs have the advantage of a wide beam that scans along the horizontal as well as the vertical, allowing applications such as the estimation of the spread of insect outbreaks [9]. A further advantage to using WSR data is that a wide range of devices are already deployed globally as part of national meteorological services, several of which make their data accessible to academics [50,53]. Observations from WSRs have been used to demonstrate the weather-dependence of seasonality in insect and winged vertebrate mass emergences over large spatial scales [41,53,54,55]. Thus, WSRs represent an underexplored opportunity to correlate high-altitude insect presence with meteorology to better our understanding of long-distance insect dispersal and aerial habitat connectivity [36,56,57,58,59].

Studies based on observations from WSRs and VLRs [38,60,61,62] have shown that passive dispersal models (e.g., the Lagrangian movement) do not offer an adequate description of insect flight behaviour, which appears to contain elements of both active and passive dispersal (see [63,64] for a comparison of active/passive dispersal concepts). Further, insects appear to prefer certain atmospheric conditions for flight, as shown in VLR studies.

Aeroecological studies have previously used the relationship between animal presence and environmental conditions to characterise active dispersal behaviour. For example, Wang et al. (2023) [65] draw on several documented behavioural responses to environmental thresholds, which they use to inform their dispersal model of fall armyworm [66,67]. Aralimarad et al. (2011) [42] also used correlations with windspeed direction to explain common orientation and clustering of individual insects around specific altitudes. However, establishing consistent correlations requires significant research efforts per species, which limits the generalisability of this approach [28,65,68,69]. Furthermore, the atmospheric variables that have been used in previous ecological studies are often restricted, with the majority focusing on more easily measured variables with an obvious biological impact, such as temperature and relative humidity [21,70,71,72]. These variables are also generally measured at the surface level or under controlled conditions, where vertical variability and multivariate collinearity are difficult to observe.

Here, we combine WSR-based observations of insects in flight with contemporaneous three-dimensional, high-resolution atmospheric data to characterise presence–environment correlations using environmental niche modelling (ENM). We demonstrate our radar-based niche modelling approach during the deployment of an X-band radar system over a year. We test the hypothesis that the incorporation of high-altitude meteorological variables enhances our ability to predict insect presence compared with models containing only surface-level predictors. If supported, these findings would indicate that insects are monitoring and responding to their environment after take-off to ensure they remain in suitable conditions for continued flight (active dispersal). We further investigate which atmospheric variables are the primary correlates of aerial insect activity within the region and time of our study and contrast the relative importance of surface variables with high-altitude variables. Identification of key predictor variables would allow us to make mechanistic inferences about the link between environmental conditions and the behavioural strategies of flying insects. Finally, we illustrate the effect of the dynamic atmosphere on the spatial distribution of suitable tropospheric habitats over time.

2. Materials and Methods

2.1. Data Used in This Study

2.1.1. The NXPol-1 Weather Surveillance Radar

WSR observations were sourced from the NCAS (National Centre for Atmospheric Science) mobile dual-polarisation Doppler X-band WSR (NXPol-1, Meteor 50DX model, made by Selex-Gematronik, Neuss, Germany; [48]), which was deployed at the NERC Facility for Atmospheric Radar Research (NFARR) at Chilbolton (near Hampshire in the United Kingdom; 51.14N, −1.44E; [48]). The deployment took place from November 2016 to June 2018.

NXPol-1 was deployed on a scanning regime of 10 elevations over 0.5 to 20 degrees. A complete scanning regime took approximately 5 min, with a single 360-degree scan at one elevation taking close to 30 s. NXPol-1 collected observation data from 75 m to 150 km in range in 150 m bins, at an azimuthal resolution of 1 degree. The WSR data had also been pre-filtered based on a signal-to-noise ratio to remove non-significant echoes. This scanning regime was repeated over the course of every active day during the deployment. We made an initial selection of hourly data from the 1 January to the 31 December 2017.

2.1.2. The ECMWF Operational Forecast Model

For this study, we chose to represent the atmospheric environment with data sourced from the European Centre for Medium Range Weather Forecasting (ECMWF) MARS operational archive (© 2017 European Centre for Medium-Range Weather Forecasts (ECMWF), www.ecmwf.int (https://apps.ecmwf.int/datasets/licences/general/, accessed 26 November 2023), CC BY 4.0, ECMWF does not accept any liability whatsoever for any error or omission in the data, their availability, or for any loss or damage arising from their use). The ECMWF Operational Forecast provides high spatial and temporal resolution data, with a wide range of surface and tropospheric variables [73]. The data were downloaded at a resolution of 0.07 by 0.07 degrees (~4 × 8 km, longitude × latitude) across a bounding box of 49°–60°N and 12°W–5°E. Atmospheric variables were downloaded on pressure levels 1000, 900, 800, and 700 hPa (roughly corresponding to 0, 1000, 2000, and 3000 m altitude above sea level) at an hourly interval for a 0–11 h range for forecasts started at 00:00 and 12:00 h (the initial conditions, equivalent to an analysis step; see ECMWF (2023) [73]) for the entire year of 2017. We also downloaded surface variables on the same grid, temporal resolution, and time range. The ECMWF operational forecast data effectively represents instantaneous values taken at every hour of the simulation [73], so we were able to consider the closest radar file in time as near-simultaneously observed with the conditions of the atmosphere.

The ECMWF operational forecast offers variables that vary over pressure levels (altitude) and ‘Surface’ variables that are independent of pressure levels, either because they affect the entire air column or are modelled to a fixed altitude (e.g., 2 m, 10 m). To test our hypothesis that atmospheric variability within the air column is a better predictor than surface-level measurements alone, we downloaded variables from both the pressure levels and surface datasets. We divided them into ‘Aerial’ (pressure levels) and ‘Terrestrial’ (surface) categories for comparison in later niche modelling; see

Table 1. Categorised list of ECMWF atmospheric variables and CEH land cover types used in this study. A description of all variables (except altitude band and time) in the ECMWF forecast model can be found at https://codes.ecmwf.int/grib/param-db/, accessed 26 November 2023 [73]. The land cover types are described in detail at https://catalogue.ceh.ac.uk/documents/f6f86b1a-af6d-4ed8-85af-21ee97ec5333, accessed 26 November 2023, and in Morton et al. (2020) [74].

Variable Category	Variable Type	Variable Full Name	Units
Aerial (Pressure Levels)	Wind	Zonal Wind	m s−1
		Meridional Wind	m s−1
		Vertical Velocity	Pa s−1
	Stability and Flow	Divergence	s−1
		Relative Vorticity	s−1
		Potential Vorticity	K° m−2 kg−1 s−1
	Temperature	Temperature	C°
	Precipitation	Relative Humidity	%
	Geometry	Altitude Band	m (±500 m)
		Time	Hours
Terrestrial (Surface)	Wind	10 m U-component of Wind	m s−1
		10 m V-component of Wind	m s−1
		Instantaneous 10 m Wind Gust	m s−1
	Temperature	2 m Temperature	C°
		Skin Temperature	C°
	Precipitation	Convective Rain Rate	Kg m−2 s−1
		Large Scale Rain Rate	Kg m−2 s−1
	Geometry	Time	Hours
	Land Cover Type	Broadleaf Woodland	%
		Coniferous Woodland	%
		Arable	%
		Grassland	%
		Urban–Suburban	%

for a complete list of the variables we selected.

The seven ‘Terrestrial’ variables were selected to include temperature and wind conditions at ground level, to act as a control against vertical variability (Table 1). Surface-level conditions have been frequently associated with the timing of insect mass dispersal events, such as soil temperature and humidity in the seasonal emergences of winged Lasius niger [71,75,76], and wind direction in dragonflies [15].

2.1.3. Centre for Ecology and Hydrology Land Cover Maps

In addition to the above atmospheric data, we included the proportion of land cover types inside the ECMWF grid cells. The Land Cover Data used in this study is owned by UK Centre for Ecology & Hydrology (© Database Right/Copyright UKCEH). Knop et al. (2023) [77] recently conducted an analysis that demonstrated a correlation between aerial insect abundance and the type of habitat at the surface of the atmospheric column, where the insects were observed. To test whether this produced similar effects on the spatial extent of aerial insect presence, we included a list of these habitat types from the Centre of Ecology and Hydrology (CEH) land cover maps [74] in the terrestrial variables we supplied for niche modelling (Table 1).

The CEH land cover map was downloaded for the year of 2017 on a 1 km resolution. The CEH maps are provided on the British National Grid (EPSG: 27700; [74]), so we reprojected the data onto longitude–latitude coordinates at a resolution of 0.04 × 0.04 degrees (approximately similar to the ECMWF grid, 0.07 × 0.07) using QGIS [78]. We transformed the data from the categorical land cover types to percentages of land cover contained in the new grid cells. Then, we averaged these percentages of land cover per ECMWF grid cell, to find the proportions of land cover type at surface level. As the Chilbolton study site was unlikely to contain certain habitat types in abundance, we limited our selection of land cover types to broadleaf and coniferous woodland, arable land, grassland and urban–suburban (built-up areas); see Morton et al. (2020) [74] for the full list of land cover types.

2.2. Methodology

2.2.1. Overview

To address these research aims, we have developed a methodology that makes use of paired observations of insect aerial activity derived from WSR observations and archived data describing the atmospheric and other environmental conditions. In this and the following sections, we summarise the approach we have used to create this paired data (presence-absence-environment tables) and subsequent ENM analysis of their covariance. The data analysis and storage of this project used JASMIN, the UK collaborative data analysis facility.

The first step was to acquire WSR and atmospheric data. WSR observations had to be sourced from a dual-polarisation WSR system that recorded equivalent reflectivity factor in the horizontal and vertical (Z_H/Z_V), differential reflectivity (Z_DR), and correlation coefficient (ρ_HV) (Figure 1, Section 2.1.1). Here, we used a signal-to-noise ratio filtered radar returns from the NCAS NXPol-1 X-band radar system [48]. Atmospheric data had to cover as much of the WSR’s spatial and temporal range and resolution as possible. We used weather forecast models from the ECMWF Integrated Forecasting System (IFS) as a source (Figure 1, Section 2.1.2). Weather forecasts were favoured over reanalysis as IFS reanalysis is limited to 6 h intervals. We rejected the hourly reanalysis, ERA5, due to poor spatial resolution (0.25° × 0.25°) compared with IFS forecasts (0.07° × 0.07°). Land cover data were included to test for the effect of habitat types that might have acted as sources of or destinations for insects moving through the air column. We sourced land cover data from the UK Centre for Ecology and Hydrology UK land cover maps (see Section 2.1.3, CEH).

The second step was to classify and filter the radar data so only insect presence–absence observations remained. The filtering procedure had to exclude meteorological scatterer and correspond to known patterns of insect signatures in WSR observations. The filtering steps produced a spatiotemporal record of true/false presence–absence for a defined biological scatterer class (in the case of this study, Insecta) on the spherical coordinate system of the radar, defined by its range gates and scanning regime (which defined the azimuthal resolution and elevation angles).

Next, the spherical presence–absence data were paired with atmospheric data by gridding them to the geographic coordinate system used by the atmospheric data (Figure 1, Section 2.2.3). The presence–absence data points were treated as single voxels (a ‘voxel’ being equivalent to a radar range bin in 3D space), representing the centroid of their spherical coordinate voxel. The voxels were then counted per atmospheric grid cell. Then, we assigned each cell a single value dependent upon whether it contained absences, presences, or neither (see Section 2.2.3). The gridding process was performed for each timestep of atmospheric data to get paired presence–absence and environment observations as close in time as possible. Gridding resulted in a presence–absence–environment table including all the observations of true presence and true absence with the corresponding states of the environmental variables of interest.

The presence–absence–environment tables need to be reduced due to the high resolutions of the radar and atmospheric data. High resolution and range (compared to biological field surveys) result in a large number of observations even after filtering, so the tables are often too big to analyse efficiently with R implementations of niche modelling (see Thuiller et al. (2023) [79] for a package list) and contain a large number of redundant observations (identical data). Therefore, our next step was to reduce the number of rows in the table through random subsampling or another equivalent method (Figure 1, Section 2.2.3). We implemented a constrained random subsampling process, which preserved the proportion of presences to absences and the distribution of values for the environment variables. In this way, we subsampled the data to much smaller tables, while limiting introduced bias from random sampling.

The final step was to carry out niche modelling on the subsampled presence–absence–environment tables (Figure 1, Section 2.2.4). We implemented niche modelling using the ‘biomod2’ package, using the ‘tree’, ‘glm’, ‘rpart’, and ‘randomForest’ R packages [79]. ENMs produced several types of assessment statistics to investigate the quality of the model and the roles of predictor variables. In this study, we addressed our research aims by analysing model assessment statistics under different variable selections (surface level, high-altitude, and both combined, see Section 2.2.4), followed by a comparison of variable importance metric, and then the response curves of the most important variables (Figure 1, Section 2.2.4). These models could then be projected onto gridded environment data to picture how the atmospheric data could influence spatial distributions of flying organisms (Figure 1, Section 2.2.4).

Figure 1. A high-level, generalised overview of our WSR-ENM procedure. The procedure can be considered to comprise two principal stages, radar filtering into insect presence–absence data and the pairing of 3D gridded atmospheric data with insect presence–absence [80]. This procedure produces Species with Data tables which can be used with a range of niche modelling approaches.

Figure 1. A high-level, generalised overview of our WSR-ENM procedure. The procedure can be considered to comprise two principal stages, radar filtering into insect presence–absence data and the pairing of 3D gridded atmospheric data with insect presence–absence [80]. This procedure produces Species with Data tables which can be used with a range of niche modelling approaches.

2.2.2. WSR Filtering Procedure

We first selected processed WSR data (see Section 2.1.1) from the recordings to create an hourly resolution dataset, spanning 2017 from the 1 January to the 31 December. All radar voxels in each of these files were then transformed from a list of radar measurements to a classification of presence or absence under the following procedure.

First, we used the 95% confidence thresholds reported by Kilambi et al. (2018) [80] for classifying radar signals as meteorological scatter and non-meteorological scatter (Figure 2a,b). These thresholds were based on a maximum threshold for reflectivity (Z_H < 35 dBZ) and the depolarisation ratio (DR < −12), which was calculated from the linear scale differential reflectivity (Z_dr) and correlation coefficient (ρ_HV) using Equation (1) (reproduced from Kilambi et al., 2018 [80]).

$$\mathrm{DR}={\displaystyle \frac{{\mathrm{Z}}_{\mathrm{dr}}+1-2{\mathrm{Z}}_{\mathrm{dr}}^{\frac{1}{2}}{\mathsf{\rho }}_{\mathrm{HV}}}{{\mathrm{Z}}_{\mathrm{dr}}+1+2{\mathrm{Z}}_{\mathrm{dr}}^{\frac{1}{2}}{\mathsf{\rho }}_{\mathrm{HV}}}}$$

(1)

After this classification was applied to the radar voxels, we removed all isolated voxels of non-meteorological scatter, using the despeckling filter provided in the Py-ART Python package. This was applied directly to the classified data, using a threshold of 0.5 (where a non-meteorological voxel is represented as 1, see

Table 2. Final designations of signal types in our WSR-ENM procedure, the corresponding data code, and abbreviated rules for determining type. Az—azimuth, El—elevation, DR—depolarisation ratio, Z_H—horizontal reflectivity, SNR—signal-to-noise ratio, NA—not available, ‘|’—or, ‘&’—and.

Signal Type	Presence/Absence Code	Determinants
Non-Meteorological (TP)	1	DR > −12 & ZH < 35 & SNR > 0.5
No Signal (TA)	0	ZH = NA | SNR < 0.5
Weather (FA)	−1	DR < −12 | ZH > 35 | SNR > 0.5
Beam Blockage (FA)	−2	54° < Az < 65° | 82° < Az < 91° | 128° < Az < 155° | 175° < Az <185° | El < 2°
Indeterminate Scatter (FA)	−3	ZH! = NA & ZV = NA

) and a group size of 30 [81]. Beam blockages were manually removed within sets of predefined azimuths (54–65°, 82–91°, 128–155°, 175–185°), and ground clutter was removed by excluding voxels from elevations below 2°. Finally, we manually filtered the data of extraneous events that could be misidentified as insect scatter. This included melting layers, sun spikes, interference from other radio sources, chaff from aircraft over the site, and ice-bearing clouds. Unfortunately, as there is no dependable method to remove reflectivity caused by birds from X-band radar scans, we have to admit the possibility of contamination. The outcome of our filtering was a collection of clear-air days containing spatially congruent, non-meteorological scatter, which bloomed from the centre of the radar in accordance with diurnal cycles. An example plan positional indicator (PPI) plot is shown in Figure 2, and we further provide a video overview of all radar moments filtered from over our study period in the Supplementary Materials.

Figure 2. Outcome of radar filtering applied to NXPol-1 observations (Section 2.1.1) on 10 May 2017 at ~12:00, demonstrated with plan position indicator (PPI) plots at 2.0° (a,c,e,g) and 4.5° elevation (b,d,f,h) in the radar antenna. See Table 2 for the list of classification rules per signal type. See Figure 1 of Lukach et al., (2022) [82] for a visual depiction of a PPI in real space. (a,b) Z_H. (c,d) Z_DR. (e,f) ${\rho }_{HV}$. (g,h). Classifications based on DR (Table 2) (Kilambi et al., 2018) [80]. ‘Indeterminate’ scatter beyond the range of insect presence is due to a lack of Z_V and consequently Z_dr, resulting from attenuation in the vertical polarisation, which prevents classification by DR.

Figure 2. Outcome of radar filtering applied to NXPol-1 observations (Section 2.1.1) on 10 May 2017 at ~12:00, demonstrated with plan position indicator (PPI) plots at 2.0° (a,c,e,g) and 4.5° elevation (b,d,f,h) in the radar antenna. See Table 2 for the list of classification rules per signal type. See Figure 1 of Lukach et al., (2022) [82] for a visual depiction of a PPI in real space. (a,b) Z_H. (c,d) Z_DR. (e,f) ${\rho }_{HV}$. (g,h). Classifications based on DR (Table 2) (Kilambi et al., 2018) [80]. ‘Indeterminate’ scatter beyond the range of insect presence is due to a lack of Z_V and consequently Z_dr, resulting from attenuation in the vertical polarisation, which prevents classification by DR.

2.2.3. Pairing of WSR and Atmospheric Data

WSR and atmospheric data were paired using a simple grid-based approach. We used the longitude latitude grid from the atmospheric data to partition the radar voxels (centroids) into raster grid cells. We then counted the number of true presences and true absences (Insects, No Signal; Table 2) per grid cell. Based on the counts of true presence–absence, we then gave each grid cell a single classification: where any true presence data were found, the cell was classified as ‘present’, regardless of the number of true absences recorded (

Table 3. Designations of presence or absence for ECMWF PA tables, corresponding data code, and abbreviated rules for each type.

Raster Box Designation	Presence/Absence Code	Determinants
True Presence (TP)	1	TP count > 0
True Absence (TA)	0	TP count = 0 & TA count > 0
No Data	NA	TP count = 0 & TA count = 0

). Where no presence data were found, but there were true absences, the grid cell was classified as ‘absent’ (Table 3). Where no presence or absence data were found (i.e., the cell was outside the radar’s effective range or only contained false presences/absences), the cell was classified as ‘NA’ (Table 3).

We gridded the WSR data to the atmospheric grid. However, there may still be data redundancy (repetitive presence–environment observations) to consider when employing datasets with extensive spatial and temporal range and resolution (e.g., for our selection of data, this resulted in ~24 h × 31 days × 12 months × 1600 cells = ~14.2 million table rows). To reduce redundancy, we used a random subsampling approach to reduce the presence–absence table length by a set of three pre-determined fractions, which we refer to as subsampling factors. The method randomly subsampled presences and absences independently to preserve their proportions in the new table for a user-defined number of iterations (n = 100 in this study). After random subsampling, we selected the table that was the closest fit to the distributions of the atmospheric variables in the original table, based on a least-squares difference between quartile values (see Supplementary Materials for a complete description of this method). To check whether our random subsampling method significantly affected later niche modelling, here we used subsampling factors of 0.01, 0.005, and 0.001.

2.2.4. Niche Modelling with Biomod2

We organised the experimental design for our niche modelling using the biomod2 R package [79]. We selected four widely used algorithms from biomod2 to use in niche modelling to test whether the model assessment statistics (see following paragraph) were influenced by the choice of algorithm, thus inappropriately biasing our ecological interpretations. We selected modelling algorithms to represent two simple implementations of the decision tree family of niche models, as well as the generalised linear model family, and further included a complex algorithm from the decision tree family to allow for a range of species response curve relationships (e.g., thresholds, splines, gradients). These algorithms included Classification Tree Analysis (CTA), Random Forest (RF), and the Generalised Linear Model (GLM). All three algorithms were applied with 100 replications to each subsampled presence–absence–environment table, per selection of environment variables (Aerial, Terrestrial, Aerial and Terrestrial; see Table 1). We therefore produced niche models under three different treatment groups—model algorithm, subsampling factor, and variable types. The model algorithm and subsampling factor treatment groups both served to detect bias in the models resulting from our analytical process, whereas assessment of the effect of the variable types included was used to answer our primary and secondary science questions. We did not alter any other settings from their defaults in biomod2.

Biomod2 includes a range of model assessment statistics. We chose to use the area under the curve for Receiver Operator Curve (ROC) and True Skill Statistic (TSS) to assess model accuracy, which we used to answer our primary science question by conducting a statistical analysis of these metrics across our three treatment groups, via an analysis of variance (ANOVA). We included TSS alongside ROC as, although ROC is the more traditional metric, it has faced criticism, with TSS being proposed as a more reliable alternative [83,84,85]. The ANOVA, coupled with a Tukey HSD post hoc test, revealed the treatments associated with the highest accuracy per group and the relative contribution of each treatment group to model predictive skill. We also assessed variable importance—the relative contribution of each variable to model predictions—using biomod2’s variable importance metric.

The variable importance metric was based on the response in predicted probability of presence, as the variable of interest was allowed to vary. The variable’s values were then ‘shuffled’ with their spatial coordinates to randomise the variable’s spatial distribution. The new values were then fed back into the niche model to obtain a new response value. The response from this random spread was then tested using Pearson’s correlation against the original response in probability of presence, using the true data. The resulting Pearson’s correlation coefficient was subtracted from 1 to give the variable importance value. Therefore, the more the predictions changed due to spatial randomisation, the higher the variable’s importance in the model. The whole process was carried out for a set number of repetitions by the user, with another variable selected randomly for comparison each time (See Thuiller et al. (2023) [79] for more details). Biomod2’s variable importance statistic was used to answer our second science question by using a similar ANOVA procedure to the one described for ROC and TSS.

We also produced response curves for variables scoring highly in importance to identify critical limits of presence probability. Finally, we created a simple spatial projection using the GLM onto four days from our dataset, chosen to provide a range in seasonal meteorology. The days chosen were 17 February 2017, 17 April 2017, 17 July 2017, and 17 September 2017. We used these projections in conjunction with our response curve plots for a speculative analysis of the mechanisms influencing insect flight.

3. Results

After we applied our radar filtering procedure to the radar returns from NXPol-1 (Figure 2), we found that insect aerial activity detected across the radar range was highly variable by season in 2017 (Figure 3). While some background activity remained over autumn (particularly October), most activity took place from April to August (on average, at least 1500 forecast grid cells registered presence at midday, Figure 3), with July proving to be the most active month. Insect activity was also observed to be highly diurnal, the majority taking place between 0800 and 1700 (UTC). The diurnal cycle of activity was largely preserved between months. However, activity appeared to be prolonged in June over the morning and evening (Figure 3).

In our analysis of the outcomes of niche modelling, which compared response of insect aerial activity to high-altitude (‘Aerial’) versus surface (‘Terrestrial’) level conditions, we identified that the strongest source of variability in model predictive skill (in terms of the ROC and TSS statistics) was the level of subsampling performed prior to modelling (ANOVA, F_2,2693 = 223.67, p < 0.001; Tukey HSD, 0.01 vs. 0.001, diff = 0.117, p < 0.001; 0.01 vs. 0.005, diff = −0.068, p < 0.001; 0.001 vs. 0.005, diff = 0.185, p < 0.001). The model algorithm also produced a relatively strong impact on predictions (ANOVA, F_2,2693 = 154.05, p < 0.001; Tukey HSD, RF vs. CTA, diff = 0.138, p < 0.001; RF vs. GLM, diff = 0.131, p < 0.001; CTA vs. GLM, diff = 0.008, p = 0.629) compared with our choice of variables, which overall produced a weak impact on the average model accuracies (ANOVA, F_2,2693 = 55.91, p < 0.001; Tukey HSD, Aerial vs. Terrestrial, diff = 0.05, p < 0.001; Aerial vs. Aerial + Terrestrial, diff = 0.09, p < 0.001; Terrestrial vs. Aerial + Terrestrial, diff = −0.043, p < 0.001). Next, we tried to predict insect aerial presence with niche models based on atmospheric variables sampled from high-altitude (aerial) versus surface conditions (terrestrial). Comparing findings from the aerial and terrestrial models, we found that niche models using only terrestrial variables (Table 1) outperformed those built with aerial variables during model evaluation (Table 1) (one-tailed Welch’s t-test, T = −5.43, p < 0.001). However, a model combining both aerial and terrestrial conditions outperformed models containing only aerial or terrestrial variables in both instances (one-tailed Welch’s t-test, Aerial + Terrestrial vs. Aerial, T = 9.57, p < 0.001; Aerial + Terrestrial vs. Terrestrial, T = 9.14, p < 0.001). The average predictive skill for niche models, including aerial variables, was 0.785 for ROC and 0.557 for TSS statistics, while the average predictive skill for models built only using terrestrial variables was 0.819 ROC and 0.607 TSS (

Table A4. Mean niche model accuracy scores from validation. Scores included are the area under the curve (AUC) value from Receiver Operator Curves (ROCs) and the True Skill Statistic (TSS). This study compares the impact of the model algorithm (rows), variables used, and subsampling factor (columns). Scores are coloured from red (0.5/0) to yellow (0.65/0.3) to green (0.9/1) according to the skill of the model (worst to poor to best). See also Figure 6. CTA—Classification Tree Analysis, GLM—Generalised Linear Model, RF—Random Forest.

	Environment Variable Types		Aerial			Aerial + Terrestrial			Terrestrial
Model Algorithm		Subsampling Fraction	0.001	0.005	0.01	0.001	0.005	0.01	0.001	0.005	0.01
	Evaluation Metric
CTA	ROC		0.628	0.798	0.767	0.720	0.871	0.804	0.744	0.819	0.811
	TSS		0.277	0.602	0.531	0.467	0.760	0.612	0.518	0.612	0.627
GLM	ROC		0.701	0.900	0.845	0.642	0.871	0.863	0.699	0.836	0.820
	TSS		0.526	0.719	0.598	0.293	0.696	0.667	0.430	0.599	0.550
RF	ROC		0.739	0.861	0.825	0.884	0.950	0.927	0.839	0.897	0.907
	TSS		0.544	0.664	0.553	0.783	0.833	0.745	0.700	0.722	0.709

). Although both surface-level conditions and aerial conditions contain good predictors of aerial insect activity, when taken independently (Figure 4, ROC >> 0.5, TSS >> 0), the models with the best predictive skill in our analysis combined variables from both the surface and the aerial environments (Table A4, Figure 4). The average predictive skill for these combined models was 0.837 for ROC and 0.651 for TSS statistics. We also observed an average gain of 0.018 ROC and 0.043 TSS, whenever aerial variables were included in addition to terrestrial, in the ENM.

Of our chosen model algorithms, GLMs performed only marginally better than CTAs (Tukey HSD, GLM versus CTA, diff = 0.008, p = 0.629), whereas RFs were significantly more accurate models than GLMs and CTAs on average (Tukey HSD, RF versus GLM, diff = 0.131, p < 0.001; RF versus CTA, diff = 0.139, p < 0.001; Table A4; Figure 4). There were no substantial differences in predictive skill between decision tree models and linear regression in this analysis (Table A4; Figure 4). Thus, we do not observe significant effects from algorithm complexity or type on model predictive skill for aeroecological data, meaning we cannot offer a conclusive recommendation on which algorithm to favour in future work. Algorithms will need to continue to be decided on a case-by-case basis [88,89].

We note that the above statistical tests of niche model evaluation metrics (ROC and TSS) were limited to TSS, as TSS and ROC evaluations proved to be highly correlated (Spearman’s ρ > 0.90 and p < 0.001 in all cases). TSS values were further confirmed to be nonnormal through a Kolmogorov–Smirnov test against the normal distribution (D = 0.52, p < 0.001).

We identified four primary atmospheric correlates of aerial insect activity using averages of biomod2’s variable importance metric across the proportion of arable land cover (~0.47,

Table A5. List of mean variable importances by algorithm and variable. See Figure 5 for a visual representation of these values. The variable importances are presented here in rank order of the mean of all runs across algorithms (column 1). See Section 2.2.4 for a description of the biomod2 variable importance metric and supporting citations. CTA—Classification Tree Analysis, GLM—Generalised Linear Model, RF—Random Forest.

Explanatory Variable	Mean Variable Importance	Mean CTA Variable Importance	Mean GLM Variable Importance	Mean RF Variable Importance
Arable Land Proportion	0.468	0.697	0.445	0.262
Altitude Band (+500 m)	0.257	0.317	0.327	0.129
2 m Temperature	0.195	0.131	0.379	0.073
Temperature	0.155	0.204	0.150	0.112
Relative Humidity	0.152	0.188	0.182	0.086
Skin Temperature	0.117	0.079	0.186	0.087
Zonal Wind	0.060	0.046	0.103	0.031
Coniferous Wood Proportion	0.046	0.042	0.034	0.062
10 m Zonal Wind	0.046	0.008	0.116	0.014
Urban Suburban Land Proportion	0.045	0.027	0.089	0.019
Broadleaf Wood Proportion	0.042	0.020	0.086	0.020
Instantaneous 10 m Wind Gust	0.040	0.009	0.096	0.016
Large-Scale Rain Rate	0.038	0.022	0.079	0.015
Time (h)	0.037	0.027	0.060	0.024
Meridional Wind	0.033	0.018	0.049	0.032
10 m Meridional Wind	0.032	0.008	0.065	0.021
Vertical Velocity	0.031	0.014	0.058	0.020
Grassland Proportion	0.029	0.007	0.071	0.010
Atmospheric Divergence	0.025	0.008	0.055	0.013
Potential Vorticity	0.021	0.009	0.018	0.035
Relative Vorticity	0.020	0.010	0.030	0.021
Convective Rain Rate	0.013	0.000	0.025	0.013

), the ±500 m altitude band (~0.26, Table A5), surface and air temperatures (2 m, ~0.20; at height, ~0.16; Table A5), and the relative humidity (~0.15, Table A5). Notable variables in the middle ranks of variable importance (~0.03–0.04, Table A5) included the zonal windspeed and coniferous wood proportion (Table A5).

We found significant variability in biomod2’s variable importance metric, depending on the model algorithm used (ANOVA, p < 0.001 for all contributing factors,

Table A3. Summary table of F and p values from analysis of variance tests carried out on biomod2’s estimates of variable importances, per model type (combined/aerial/terrestrial). In all estimates of variable importance for all models, the environment variable used provided the greatest source of variation.

ANOVA Statistic	F-Value			p-Value
Model Type	Combined	Aerial	Terrestrial	Combined	Aerial	Terrestrial
Variable Name	10,604.9	2295.61	5993.0	<2e−16	<2e−16	<2e−16
Subsampling Factor	311.9	78.86	268.2	<2e−16	<2e−16	<2e−16
Algorithm	15,168.6	382.29	3604.5	<2e−16	<2e−16	<2e−16

). While the rank order of the variable importances was clear (Table A5), this indicated a degree of uncertainty in the magnitude of relative contributions. Hence, caution was applied when making inferences about the specific magnitude of the effect of different variables. The primary source of variability in calculated importance was found to be the selection of environmental variables in aerial and terrestrial models (Table A3). However, in models combining aerial and terrestrial variables, the choice of algorithm proved to be a more substantial contributor to variable importances than to model evaluation metrics and became the primary source of variability (Figure 5, Table A3). The subsampling factor was also a significant but comparatively weak contributor to variable importance (Figure 5, Table A3; p < 0.001).

The response curves of the ENMs allowed us to explore critical environmental thresholds for flight, as estimated probability of presence varied (Figure 6). However, we were unable to identify unambiguous critical thresholds for flight, as estimated probability of presence varied substantially by the ENM algorithm. This was caused by the underlying modelling approaches in each algorithm and were beyond the scope of this study to explore. However, we were able to determine some approximate thresholds by noting the broad patterns in response curve shape across algorithms (Figure 6, coloured lines per panel). The model response curves predicted an increased probability of presence within lower altitude bands (0–500 m, 500–1500 m), although activity at higher bands was expected to remain above 0.25 probability of presence, on average (Figure 6). By comparing the elevated probability of presence associated with increased surface and aerial temperatures (>20 °C, Figure 6) in ECMWF UK forecasts, we posited that common patterns of habitat suitability could be linked to seasonal and diurnal temperature trends across the UK, with limited regulation from national-scale bands of rain and moisture (Figure 7. We further noted that elevated probability of presence in voxels associated with increased arable land proportion (>60% Figure 6) appeared to be responsible for localising (‘clumping’) patches of high probability of presence in the projection map (Figure 7). However, as the Chilbolton study site consists of mostly arable land, this model may not hold true for all UK insect populations. The models on which the response curves were based can also be used to visualise how the spatial distribution of environmental variables may alter the suitability of a habitat, as shown through our niche model projections of the whole of the UK (Figure 7). However, we emphasise that this analysis is speculative, as the niche models can only identify areas similar to those where we have observed flight on the WSR and are not a substitute for physiological and behavioural studies.

In summary, our analysis showed that aerial variables are strong predictors of activity when used independently of terrestrial predictors, but terrestrial predictors are the stronger of the two, and the models with the best predictive skill combined both. Our analysis of the three treatment groups (Aerial, Terrestrial, Aerial + Terrestrial) demonstrated that niche models with atmospheric variables are well suited to modelling insect aerial activity, and models combining both terrestrial and aerial conditions performed significantly better than any model using terrestrial or aerial only. We have highlighted the primary abiotic covariates of insect flight for the Chilbolton community in 2017, using our comparison of variable importances: where insect activity was positively associated with increasing arable land proportion, temperature at-height and at-surface, and relative humidity. We also noted strong seasonal patterns in the diurnal cycle of insect flight activity, and a speculative connection between environmental stimuli of insect flight and national-level weather patterns.

4. Discussion

In this study, we describe a new method for extending environmental niche modelling to high-altitude biological observations, which has been strongly argued for by the aeroecological community [39,62]. Such an approach theoretically allows researchers to describe insect active dispersal, the consequences of dispersal for meta-populations, and conservation under climate change [90,91,92,93]. We demonstrate that niche models incorporating atmospheric variables outperform those using information from the surface level only. We also note that insect occurrence in the aerosphere is most closely associated with arable land and temperature at the surface below, as well as temperature and relative humidity at-height.

TSS and ROC values from terrestrial niche models were significantly increased when the model included aerial variables (on average, +0.018 ROC and +0.043 TSS). Therefore, the model accuracy and variable importance metrics show that insect presence at high altitude may be better explained by including atmospheric conditions above the surface level, but ENMs achieve their highest predictive power when combining predictor variables from both the surface and the aerial environment (Figure 4). The outcome was that our primary hypothesis is inconclusive. We expect this finding was due to conflating factors such as the inability to distinguish between atmospheric influence on take-off and flight adjustment behaviours, which theoretically were more affected by conditions at surface and in the air column, respectively. To show definitively that high-altitude (aerial) variables are more accurate predictors of insect aerial activity compared with surface (terrestrial) variables, analysis of a wider range of radar sites will be required. Model variability resulting from model algorithm will also require further consultation with the ecological modelling community, to understand the uncertainties in each approach [88,89].

Our niche model analysis has indicated that atmospheric humidity perhaps plays a more dominant role in enabling insect flight than has been suggested by the current literature [21,70,72]. By contrast, windspeed and ground temperature are well documented as limiting factors in insect flight activity [45,70,72,94,95]. However, there is inconsistency in and disagreement on the importance of relative humidity, as well as the mechanism by which it affects insect flight: several papers claim that flight behaviour is affected by changes in relative humidity [96,97], that humidity is a secondary regulator and modulates behavioural responses to temperature [21,72], and that humidity plays no significant role in flight behaviour at all [70,98]. The entomological literature further suffers from a lack of comparable results due to the wide range of study species, locations, and experimental designs (e.g., controlled conditions, ground-level field experiment, tracking of individual flight) [72,94,96,98]. As a species’ dispersal strategy determines flight behaviour, geographical differences in the structure of insect communities could also influence the importance of predictor variables. Further investigation of the physiological and behavioural mechanisms of flight will be needed to validate future niche models, as well as prospective insect forecast tools.

Our niche model analysis further indicated a highly positive relationship between arable land cover and the probability of presence. By contrast, the previous literature has described a negative association between arthropod diversity and intensively farmed land, with a similar relationship reported for abundance [99,100,101]. However, studies have shown that the consequences for arable land on insects can depend on the land management practises, e.g., the use of chemicals and fertiliser, or grazing by sheep and cattle [99,102]. Our findings can be explained by the dominant habitats in the vicinity of the radar, which comprise arable land with patches of broadleaf wood, with coniferous woodland, grassland, and urban–suburban land found in relatively small proportions (<20% cover). The prevalence of insect activity over arable land is possibly caused by the intermingling of natural communities and pest species associated with crops. However, the CEH arable land classification, according to Jackson (2000) [103], comprises a range of cropland types, including intensively farmed land alongside annual leys and fallow land. Therefore, we are not able to distinguish the effect of different land management practises in this study. Further, as the Chilbolton site offers limited spatial heterogeneity, and the CEH land cover is far more granular than our radar observations, our niche models are unable to meaningfully analyse the effect of habitat gradients on high-flying insect abundance (see Henrys and Jarvis (2019) [104] for further discussion of the importance of spatial resolution when integrating ground survey and remote sensing data). However, by adopting the methodology we demonstrate here with a national-scale weather surveillance radar network, it would be possible to sample a broader range of habitat types and gradients.

Whether working at the local or national scale, an ongoing challenge of studying insects with radars is the filtering procedure. A shortcoming of our observations of insect activity is that several of the filtering stages are based on fixed thresholds, such as the non-meteorological and meteorological filter from Kilambi et al. (2018) [80]. Fixed thresholds ignore potential overlap between radar signatures attributed to different scatterers, especially for birds and insects [105,106,107], leading to false presences and potentially to false correlations in later niche modelling. Certain meteorological phenomena can also be responsible for false presences (e.g., melting layers; see Baldini and Gorgucci, 2006; Ryzhkov and Krause, 2022 [108,109]), which we noted on certain days of NXPol-1 observations, where large rainclouds were present over the radar site. Single events such as sun spikes, interference from other radio sources, and chaff from military exercises can contribute to erroneous presences identified in filtering, which continues to necessitate manual data cleaning before further analysis. Radar instrumentation is also inherently range-biased by the sensitivity of the system, as the power density of the beam decreases over distance, and the return signal suffers attenuation in the atmosphere (see Rocchini et al., 2023 [110] for a discussion of the impact of spatial bias in ecological modelling). This and the need to carry out manual data cleaning (where confirmation bias can be introduced) with no supporting at-height validation data introduces uncertainty into any spatial analysis performed downstream. This points to an urgent need for further, closer collaboration between the radar science and ecology communities to ensure radar observations are properly validated and understood in the ecological context that future aeroecology work can meet the best practises of both fields. WSR observations of presence could benefit in future from improvements to classification accuracy and error quantification, either through machine learning algorithms of existing radar data, to better distinguish target classes [82,105,111], or ideally through cooperative field campaigns led jointly between atmospheric scientists and ecologists. As noted previously, there is also a lack of species specificity in WSR biological observations, including those produced by our filtering of NXPol-1. Recent work with VLRs has shown promise of remedying this problem, as they are capable of identifying individual insect morphotypes from their ventral reflectivity signatures, as well as accurately determining their headings as they pass through the radar beam [16,56,112,113]. Thus, a potential solution to the limitations of both WSR and VLR systems could be to carry out joint radar experiments with WSR and VLR, backed up with field ground truthing.

Our results illustrated an association between insect presence–absence over time and conditions tied to local weather and ground habitat (arable land cover, temperature, and humidity). Previous work has further indicated that national-level weather patterns play an indirect role in insect flight and reproductive behaviour by regulating local temperature, humidity, and groundwater [114,115,116,117]. These variables are particularly shaped by the distribution of frontal zones between cold and warm air masses, which implies that insect reproduction and mass emergences could be responsive to these weather patterns. If true, this relationship would render local insect activity more predictable and would require little alteration to existing weather forecasting infrastructure. Further tools that adapt existing niche-model approaches from the terrestrial and marine environments will allow future aeroecological studies to quickly build and describe probabilistic relationships between many collinear variables and the probability of presence of aerial organisms, without the need for regular, wide-ranging, and costly field campaigns. However, it remains prudent to continue to carry out localised field investigations of insect flight and behaviour in the aerosphere to continue to ground-truth observations from radars. Controlled experiments will also be essential to understand the mechanistic causes of in-flight habitat selection in insects, which cannot be definitively described with remote sensing alone (see [72,98]).