Modelling the spatial distribution of Culicoides biting midges at the local scale

Authors


Correspondence author. E-mail: g.kluiters@liverpool.ac.uk

Summary

  1. Culicoides midges (Diptera: Ceratopogonidae) are ubiquitous on farms in the United Kingdom (UK), but little research has explored their abundance, an important determinant of disease risk. Models to explain and predict variation in their abundance are needed for effective targeting of control methods against bluetongue virus (BTV) and other Culicoides-borne diseases. Although models have been attempted at the national scale (e.g. Scotland), no investigations have taken place at a finer spatial scale.
  2. Midge abundances were estimated using light traps on 35 farms in Bala, north Wales. Culicoides catches were combined with remotely sensed ecological correlates, and on-farm host and environmental data, within a GLM model. Drivers of local-scale variation were determined at the 1-km resolution.
  3. Local-scale variation in abundance exhibited an almost 500-fold difference (74–33 720) between farms in maximum Obsoletus Group catches. The Obsoletus Group model explained 81% of this variance and was dominated by normalized difference vegetation index (NDVI). This is consistent with previous studies suggesting strong impacts of forest cover and vegetation activity on distribution, as well as shaded breeding site requirements.
  4. The variance explained was consistently high for the Pulicaris Group, C. pulicaris and C. punctatus (80%, 73% and 74%), the other probable BTV vector species in the United Kingdom. The abundance of all vector species increased with the number of sheep on farms, but this relationship was missing from any of the non-vector models. This is particularly interesting given that none of the species concerned are known to utilize sheep-associated larval development sites. Performance of the non-vector models was also high (65–87% variance explained), but species differed in their associations with satellite variables.
  5. Synthesis and application. At a large spatial scale, there is significant variation in Culicoides Obsoletus Group abundance, which undermines attempts to record their nationwide distribution in larger-scale models. Satellite data can be used to explain a high proportion of this variation and, if shown to be generalizable, they may produce effective predictive models of disease vector abundance. We recommend undertaking a prior survey for farms with high Culicoides catches within the sampling area and checking stability in catch size between seasons and years.

Introduction

Bluetongue (BT), a vector-borne, viral disease of ruminants has undergone an unprecedented emergence following the entry of bluetongue virus (BTV)-8 into northern Europe in 2006. The Palaearctic Obsoletus and Pulicaris Groups of Culicoides biting midges are implicated in virus transmission (Carpenter, Wilson & Mellor 2009). Although BT has long been notifiable to the World Organization for Animal Health (OIE), the recent introduction and establishment of BTV-8 into northern Europe has intensified interest in understanding the drivers of Orbivirus outbreaks.

Recent advances have been made in understanding larval development sites, taxonomy and molecular recognition of Culicoides. Little is currently known, however, about certain ecological characteristics of the vector and non-vector species (Conte et al. 2007a), including their distribution and abundance, making disease risk assessment and management difficult. As BT and other Culicoides-borne diseases are transmitted almost entirely by the bites of their vector species (Mellor, Boorman & Baylis 2000), their distribution and infection intensity are dependent on the distribution and abundance of Culicoides.

Statistical models facilitate informing opinions where ecological knowledge is lacking, by identifying relationships between climatic and environmental factors, and the known distribution of vectors (Baylis & Rawlings 1998; Baylis, Meiswinkel & Venter 1997). Suitability maps can then be produced for vector occurrence across regions where vector distribution is unknown, on the basis of their climate and environment (Rogers & Randolph 2003). Analysing patterns in Culicoides abundance with different ecological characteristics in the same model framework (Conte et al. 2007a; Calvete et al. 2008) can aid understanding of biological mechanisms underlying the sensitivities of midge species to particular environmental factors, despite limited ecological knowledge of these species.

The spatial scale at which these techniques have been used varies greatly, but larger-scale studies have primarily been the focus, while local-scale, high-resolution models are less common. Country-wide risk mapping, as was undertaken in Switzerland prior to the BTV-8 incursion, highlights the use of regression analyses together with GIS techniques in determining vector suitability maps for the major European vector group C. obsoletus (Racloz et al. 2007), whilst discriminant analysis has been used in terms of continent-wide mapping in Europe and Africa (Tatem et al. 2003). Discriminant analysis and regression have also been used in terms of country-wide vector distribution and risk mapping in Sicily (Purse et al. 2004b), various regions in the Mediterranean (Baylis et al. 2001) and Morocco (Baylis et al. 1998), Iberia (Baylis & Rawlings 1998; Wittmann, Mellor & Baylis 2001) and Italy (Conte et al. 2003), respectively.

The spatial scale and intensity of a study impacts the temporal frequency of trapping, because of the limited resources, mostly time, available for processing catches. At the extreme ends of the scales, national-level surveys have undertaken single nights of catches at a large number of sites (Meiswinkel et al. 2008; Hartemink et al. 2009), while other studies report daily catches over several months at a single site (Birley & Boorman 1982; Gerry & Mullens 2000). Many studies adopt protocols in between these extremes: in Morocco, 22 trap sites were sampled twice weekly for 2 years (Baylis et al. 1997). The trade-off between spatial and temporal resolution employed by such studies has rarely been overcome, except in recent years where national-level BT concerns led, in some instances, to country-wide surveillance using government resources (Calvete et al. 2006; Conte et al. 2007a).

The lack of local-scale modelling is likely to be due to climate-driven models performing poorly when validated at the local scale with independent data (Capela et al. 2003), as non-climatic factors such as landscape configuration, farm husbandry, host availability and microclimate also influence abundance. Local-scale distribution of Culicoides and BTV are best explained by models that incorporate landscape (Guis et al. 2007), topographical (Conte et al. 2007b) or host factors (Calvete et al. 2008) alongside climate.

Until now, few attempts were made to model the relationship between climate, host and environmental factors and the distribution of current and potential bluetongue vectors within the United Kingdom. Although all Culicoides species share the same basic habitat requirements, that is, the presence of host for bloodmeals and breeding sites (Mellor, Boorman & Baylis 2000), they differ in their life-history characteristics and therefore the extent to which their distribution and abundance is affected by environmental factors. Purse et al. (2012) investigated landscape, host and climate on Culicoides in Scotland, yet failed to produce a strong model for the major BT vectors, the Obsoletus Group, highlighting the need for strong models of this species group.

Seasonal climatic variables derived by temporal Fourier processing are good predictors of vectors and vector-borne disease patterns, including tsetse flies and trypanosomiasis (Rogers 2000), malaria (Rogers et al. 2002), tick-borne diseases (Randolph et al., 2000) and bluetongue and its vectors (Tatem et al. 2003; Purse et al. 2007). Previous models indicate that climatic determinants of distribution differ between Culicoides species, with Purse et al. (2004a,b) finding the distributions of C. obsoletus and C. newsteadi were primarily related to remotely sensed temperature variables [land surface temperature (LST), air temperature (TAIR)], while normalized difference vegetation index (NDVI) was the most important for C. pulicaris, and C. imicola was modelled using a combination of LST, NDVI and middle infra-red reflectance (MIR) in Sicily.

This study aimed to measure the light trap catch of Culicoides at a local, farm-level, scale and identify determinants of the abundance. The Bala region of north Wales was selected as the study site. Here, the distribution and abundance of known BT vectors and other Culicoides species are modelled in relation to satellite-derived ecological correlates, and host and environmental variables at the 1-km scale, with determinants of distribution compared between species. Specific objectives included building a model for Obsoletus Group (C. chiopterus, C. dewulfi, C. obsoletus and C. scoticus), determining whether predictors could be generalized across midge species with particular ecological characteristics (groups) and investigating the use of satellite imagery at the local spatial scale.

Materials and methods

Trapping design

This study was undertaken in the Welsh province of Bala, situated in Snowdonia National Park. This area primarily consists of extensive sheep and beef cattle farming, with a very hilly landscape comprising a mixture of forests and field. A 6 × 6 km grid was overlaid on a 1 : 25 000 ordinance survey map of the area just north of Bala lake, north Wales. In each grid square (36 in total), one farm or smallholding was selected to participate in the study, although due to the nature of the terrain, two squares contained no properties. One further control farm, outside the gridded region, also participated (35 farms in total). All farms were recruited via personal contact.

A thrice-replicated randomized trapping grid allocated the selected trapping sites to a 12-night trapping schedule between 7 July and 18 July 2008. Each farm was therefore sampled for a total of three trapping nights.

Questionnaire design

A short questionnaire was designed to capture data on (i) host characteristics – number and distance of dairy, beef, sheep, horses and other animals; use and frequency of insecticide administration on animals, buildings and dung heaps; (ii) Farm surroundings/environment – presence of breeding sites (dung or manure heaps, leaf litter and food heaps) within 250 m radius of traps; water sources (standing and running water, waterlogged ground, artificial sources such as troughs) within 250 m of traps. Altitude of each trapping site was measured using a Garmin eTrex® H GPS receiver (Olathe, Kansas, USA).

Traps and midge collection

Trapping was undertaken using 15 Onderstepoort-type down draught black light traps connected to either a mains power source or car battery. Such traps measure a mixture of Culicoides abundance in an area as well as their activity and attraction to light. Traps were positioned as close to livestock as possible, and the number of livestock within 50 m was recorded each night. Midges were collected in 500-ml beakers containing c. 200 ml of water and a small amount of washing-up liquid to break the surface tension of the water. Traps were set between 1600 and 1800 h and collected between 0800 and 1000 h the following morning when collections were transferred to 70% ethanol for storage.

Insect sorting

Culicoides sorting and counting was undertaken at the Onderstepoort Veterinary Institute Agricultural Research Council, South Africa. Large collections were subsampled (Van Ark & Meiswinkel 1992), and females were age-graded into nulliparous, parous, gravid or blood-fed based on abdominal pigmentation (Dyce 1969). Males were also counted, but all other insects were stored uncounted. For the Obsoletus Group, the females of four constituent species (C. chiopterus, C. dewulfi, C. obsoletus and C. scoticus) were counted together, while males were counted separately. Only females were considered in the analyses as males do not take blood meals and consequently do not act as vectors of disease between vertebrates. For the Pulicaris Group, C. pulicaris, C. punctatus and C. impunctatus catches were modelled together, as well as separately. In Europe, members of the Obsoletus Group, as well as the C. pulicaris and C. punctatus members of the Pulicaris group, are considered the most important vectors (Mellor & Pitzolis 1979; Carpenter, Wilson & Mellor 2009); we have considered the other Culicoides species trapped to be non-vectors.

Satellite-derived climate data

Seventy remotely sensed variables were derived from MODerate-resolution Imaging Spectroradiometer (MODIS) imagery from the NASA Terra satellite (Scharlemann et al. 2008). Five variables with environmental significance were extracted at 1-km grid resolution between 2001 and 2005: NDVI, MIR, day and night land surface temperature (dLST and nLST) and enhanced vegetation index (EVI). NDVI is a measure of chlorophyll abundance, correlated with soil moisture, rainfall and vegetation biomass, coverage and productivity (Campbell 1996). MIR is correlated with water content, surface temperature and vegetation canopy structure (Boyd & Curran 1998). EVI is similar to NDVI, measuring vegetation activity correlated with levels of soil moisture (Chen et al. 2006; Waring et al. 2006), but has improved sensitivity in wet zones with high biomass. For each of these factors, 14 temporal Fourier-processed (Rogers 2000) predictors were produced (Table 1).

Table 1. Temporal Fourier-processed predictors of five environmentally significant variables derived from MODIS imagery
MODIS variableExplanation
a0Overall mean amplitude
a1Amplitude of the annual cycle
a2Amplitude of the biannual cycle
a3Amplitude of the triannual cycle
p1Phase (peak timing) of the annual cycle
p2Phase (peak timing) of the biannual cycle
p3Phase (peak timing) of the triannual cycle
d1Proportion of variance explained by the annual cycle
d2Proportion of variance explained by the biannual cycle
d3Proportion of variance explained by the triannual cycle
daProportion of variance explained by the annual, biannual and triannual cycles combined
mnMinimum of the seasonal cycle
mxMaximum of the seasonal cycle
vrVariance

Data analysis

For Pearson Product-Moment correlations between trap catches of different Culicoides spp., the critical value for significance was adjusted to a lower threshold using the Bonferroni correction to take account of multiplicity of P values.

Nightly species, or group, catches were log10(n + 1) transformed, and the maximum of the three catches per farm was used in model building (hereafter, log-max catch). The maximum catch was preferred to the mean because Culicoides catches are readily reduced by weather conditions, and, arguably, the maximum provides a better measure of abundance over a short time period (Baylis et al. 1997). Abundance models were not parameterized for seven of the 19 species due to low catches. For the 12 other species, none of the distributions differed significantly from normality (Anderson-Darling Test for Normality,  0·4 for all species). Log-max catches for these species were related to satellite-derived ecological correlates, host, and environmental variables using General Linear Models (McCullagh & Nelder 1989) in r version 2.8.1 (R Development Core Team, 2011).

Predictor sets were (i) host factors – on farm sheep, dairy, beef, horse and other animal density; insecticide use – on animals, buildings and dung heaps; (ii) environmental factors – presence of breeding sites composed of leaves, dung heaps, food piles and number of breeding sites within 250 m; presence of running, standing, poor draining, artificial or other water sources and number of water sources within 250 m and (iii) 70 remotely sensed variables – 14 NDVI, 14 EVI, 14 MIR, 14 dLST and 14 nLST variables derived from MODIS satellite data (as described by Scharlemann et al. 2008).

The number of explanatory variables available for multivariable modelling was reduced, in order to minimize the risk of overfitting the model. Explanatory variables were examined for univariable Pearson Product-Moment correlations with log-max catch. Only variables with a probability of correlation <0·2 were retained. Collinear variables least correlated with log-max catch were removed.

For model development, a best subsets approach was first used to select a subset of variables (within each of the three predictor sets) that best explained the variation in abundance of each species, where the R2 and adj-R2 were the highest possible, while each variable was significant at the  0·05 level individually. Finally, the subsets of variables were combined across predictor sets into a global model and the selection procedure, based on R2 (adj-R2) values and individual variable significance, repeated to produce a final model for each species. When multiple models fulfilled these criteria, the final model was determined on the basis of having a lower Akaike's Information Criterion value (AIC, Akaike 1973).

The small number of sampled sites precluded partitioning the data into a calibration and evaluation data set. Therefore, to evaluate likely generalization errors of the final models (i.e. overfitting), leave-one-out regression analysis allowed cross-validation to occur, whereby each data point in turn is left out of the analysis and the final model refitted. The stability of variance explained, coefficients and fitted abundance values were evaluated across leave-one-out models.

Variograms were computed from the models' residuals to identify residual spatial autocorrelation, or second-order (local) effects. Second-order effects describe small-scale variation due to the interactions between neighbours (Pfeiffer et al. 2008). The Moran's I correlation coefficient (Moran 1950) was employed, with fixed neighbourhood sizes of 1·5, 2 and 3 km, to evaluate spatial patterns and examine residual spatial autocorrelation between farms. Moran's I is one of the most established indicators of spatial autocorrelation, and like the correlation coefficient, its values range from 1 (strong positive spatial autocorrelation), through 0 (a random pattern), to -1 (strong negative spatial autocorrelation).

Results

Trapping

The 175 trap catches produced a total of 357 233 Culicoides of 19 species in the Bala region. The single largest catch was 65 763 midges in one trap over one night, while the mean of the maximum catches was 2706 midges per trap per night. One catch contained zero Culicoides due to trap malfunction. A total of 61·9% of the Culicoides trapped across sites belonged to the Obsoletus Group and 31·6% to the Pulicaris Group. Of the latter, C. punctatus and C. impunctatus were the most abundant species, making up 15·4 and 14·4% of the total Culicoides sampled, whilst C. pulicaris comprized 1·8%. Of the other species, C. achrayi contributed 5·5%, while the others made up <1% each (Table 2). Due to the low catches of C. brunnicans, C. circumscriptus, C. kibunensis, C. minimus, C. nubeculosis, C. pictipennis and C. stigma models of abundance were not parameterized.

Table 2. Culicoides species trapped around Bala
Species trappedFemale (% of total catch)Male (% of total catch)
Obsoletus group total211 927 (64·92)9246 (30·02)
By Species
 C. chiopterus 14
 C. dewulfi 1694
 C. obsoletus 6576
 C. scoticus 962
Pulicaris group total97 262 (29·80)15 615 (50·71)
By species
 C. impunctatus 37 22914 225
 C. pulicaris 5992574
 C. punctatus 54041816
Other Culicoides17 250 (5·28)5933 (19·27)
By Species
 C. achrayi 14 1115458
 C. albicans 1135346
 C. brunnicans 230
 C. circumscriptus 50
 C. delta 4021
 C. fascipennis 124956
 C. festivipennis 16826
 C. kibunensis 4541
 C. minimus 10
 C. nubeculosus 1015
 C. pictipennis 90
 C. stigma 10
Total326 439 (100)30 794 (100)

Spatial variation in maximum abundance of the Obsoletus and Pulicaris Groups, along with the three Pulicaris Group species individually, can be seen in Fig. 1. The Obsoletus Group exhibited an almost 500-fold difference in maximum catches (74–33,720) between farms, but C. punctatus displayed the highest variation with an almost 4000-fold difference (6–23 656) across sites, while the Pulicaris Group exhibited the lowest 330-fold difference (85–28 423). Similar spatial patterns are seen between the male Obsoletus Group members in Fig. 2.

Figure 1.

Spatial variation in maximum log(n + 1) abundances of the (a) Obsoletus Group, (b) Pulicaris Group and the Pulicaris Group constituent species: (c) C. pulicaris, (d) C. punctatus, (e) C. impunctatus, across trapping sites. Abundances have been centred in each grid square for anonymity.

Figure 2.

Spatial variation in maximum log(n + 1) abundances of the male Obsoletus Group species (a) C. dewulfi, (b) C. obsoletus, (c) C. scoticus; and the Pulicaris Group constituent species: (d) C. pulicaris, (e) C. punctatus, (f) C. impunctatus, across trapping sites. Abundances have been centred in each grid square for anonymity. No map was drawn for the C. chiopterus member of the Obsoletus Group as there were only 11 males trapped across four sites in the maximum catches.

High correlation emerged between trap catches of females of certain Culicoides spp. on farms (Table 3). Obsoletus Group abundance was highly correlated to that of the vectors C. pulicaris and C. punctatus, but also the non-vector C. achrayi, which was associated with five of the eight species examined. The vector species, C. pulicaris and C. punctatus, were significantly associated. The non-vector species C. impunctatus, although a member of the same subgenus as C. pulicaris and C. punctatus, was most strongly associated with the other non-vector species C. achrayi. Males of the individual species were highly correlated to the females (except C. pulicaris where = 0·08), apart from C. fascipennis and C. festivipennis which were not correlated. Males of the Obsoletus Group constituent species (C. chiopterus, C. dewulfi, C. obsoletus and C. scoticus) were also highly correlated to each other (= 0·001) except for C. chiopterus with only 11 males in the total maximum catch.

Table 3. Pearson Product-Moment correlation coefficients of the abundance of females of different species of Culicoides spp. on farms around Bala
 Obsoletus Group C. pulicaris C. punctatus C. impunctatus C. delta C. fascipennis C. festivipennis C. albicans
  1. Significance (determined using the Bonferroni correction) is given where * 0·001.

C. pulicaris 0·828*       
C. punctatus 0·745*0·838*      
C. impunctatus 0·4010·4080·414     
C. delta 0·3720·3120·326−0·288    
C. fascipennis 0·3360·3930·606*0·2930·306   
C. festivipennis 0·3950·2780·2680·1390·5030·261  
C. albicans 0·3370·3480·4820·5310·1240·3850·369 
C. achrayi 0·637*0·605*0·593*0·703*0·1040·4070·3420·582*

Mean parous rate and range for the eight species modelled were for the Obsoletus Group 0·84 (0·4–0·94), C. achrayi 0·83 (0·36–0·94), C. albicans 0·64 (0–1), C. fascipennis 0·79 (0–1), C. festivipennis 0·31 (0–1), C. impunctatus 0·83 (0·49–0·99), C. pulicaris 0·34 (0–0·72) and C. punctatus 0·78 (0·31–0·99). The parous rates of four of these species were significantly correlated to their abundance (C. achrayi: r = 0·46, = 0·008; C. albicans: r = 0·35, = 0·06; C. impunctatus r = 0·39, = 0·03; C. pulicaris r = 0·35, = 0·05).

Questionnaire data

Of the 35 farms, 22 (62·9%) kept sheep, 14 (40%) beef cattle and two (5·7%) horses, while only one kept dairy cattle (Table 4). Other animals included pigs, dogs and chickens. Three farms had sheep present, but did not own them and were unaware of the exact numbers. As the ‘number of sheep’ was determined as an important variable early on, it was decided that data from those three farms would be omitted from model building, leaving 32 observations.

Table 4. Host animals on farms around Bala
Host variableNumber of farms (%)Meana number of animalsSDRange
  1. SD, standard deviation.

  2. a

    Mean of those farms with animals.

  3. b

    Not applicable.

Sheep22 (62·86)541·41688·928–2600
Beef Cattle14 (40)90·5791·292–350
Horses2 (5·71)31·412–4
Dairy Cattle1 (2·86)70NAbNAb
Pigs1 (2·86)40NAbNAb

Nineteen farms (54·3%) used insecticides regularly on animals, but none on buildings or for dung management, although 21 (60%) farms had dung piles within 250 m of the trapping site (mean 18·29 m, range 1–100 m). In terms of other breeding sites, 24 (68·57%) farms had leaf litter between 1 and 75 m (mean 22 m), and eight (22·86%) had food heaps within 20–150 m (mean 54 m) of the traps (Table 5). The mean altitude on the farms was 237·9 m, but varied between 169·1 and 335·1 m.

Table 5. Breeding sites and water sources on farms around Bala
Predictor variableNumber of farms (%)Mean distance (m)aSDDistance range (m)
  1. SD, standard deviation.

  2. a

    From light trap.

  3. b

    Within 250 m of the trapping site.

Breeding sitesb
Dung21 (60)18·2925·851–100
Leaf Litter24 (68·57)21·8321·801–75
Food Heaps8 (22·86)54·2944·0120–150
Water sourcesb
Running24 (68·57)104·2291·122–250
Standing9 (25·71)197·22104·795–250
Wet Ground18 (51·43)70·1586·062–250
Artificial22 (62·86)1216·771–50

MODIS satellite data

In the 2001–2005 MODIS data, day-time temperatures peaked in Bala in early to mid-June, and night-time temperatures in late June. Site temperatures ranged from −0·8 to 20·8 °C (mean 10·9 °C) in the day-time and −6 to 9·2 °C (mean 0·9 °C) at night-time. The peak of NDVI occurred between late June and early August, and the seasonal range in this index varied from 1·617 to 1·871. The peak of EVI occurred between early and late June, and the seasonal range at sites varied from 1·140 to 1·739. For MIR, a peak occurred across most sites between early March and May, although the peak on one farm occurred in late February. The seasonal range varied from 0·036 to 0·129.

Overall performance of model variables

The variables included in the final models are shown in Table 6. Overall, NDVI was the only explanatory variable selected in all models, while MIR was seen in six. The only host variable within the final models was ‘number of sheep’ on a farm, while for environmental variables the number of water sources, food heaps and dung heaps were included. The vector group models (Obsoletus and Pulicaris Groups) both contained a host variable (sheep), as did the two individual vector species models (C. pulicaris and C. punctatus), while this remained absent in the non-vector models. See Table S1 in Supporting Information for the coefficients of the model parameters.

Table 6. Abundance models for each species or group, including percentage of variance explained (R2 and adjusted R2) and model AIC (Akaike's Information Criterion). After each variable, the () indicates the operator of the correlation coefficient
Species/GroupNDVIEVIMIRLSTOtherAIC (Null Model AIC)Mean R2(%) (adjusted R2(%))
dLSTnLST
  1. Remotely sensed variables: NDVI, normalized difference vegetation index; EVI, enhanced vegetation index; MIR, middle infra-red reflectance; dLST, day land surface temperature; nLST, night land surface temperature.

  2. MODIS variables: a = amplitude; p = phase (peak timing); d = proportion of variance explained by the 1 = annual cycle, 2 = biannual cycle and 3 = triannual cycle. da = proportion of variance explained by all three cycles combines; mn = minimum of the seasonal cycle; mx = maximum of the seasonal cycle; vr = variance.

Obsoletus Group

p3 (+)

d1 (+)

da (−)

mx (+)

a3 (+)

d3 (−)

d2 (−)p3 (+)

Sheep (+)

Water (−)

40·035 (73·587)81·23 (72·29)
Pulicaris Group

p3 (+)

d1 (+)

da (−)

a3 (+)

d1 (+)

d2 (+)

da (−)

p2 (+)Sheep (+)28·462 (62·533)80·41 (72·39)
C. impunctatus p2 (−)

a1 (−)

a3 (+)

d3 (−)

mx (−)

a0 (+)

p1 (+)

p3 (−)

d2 (+)

d3 (−)

35·565 (68·759)80·93 (71·85)
C. pulicaris

p3 (+)

d3 (−)

mx (+)

p1 (−)p3 (−)p3 (+)Sheep (+)48·876 (76·84)73·10 (65·25)
C. punctatus

p3 (+)

d1 (+)

mx (+)

p1 (−)

d1 (−)

Sheep (+)44·518 (75·851)74·17 (67·97)
C. achrayi

a1 (+)

a3 (−)

d1 (−)

d3 (+)

da (+)

vr (−)

a3 (+)

p2 (−)

d1 (+)

d2 (+)

da (−)

mx (−)

p1 (+)

p2 (+)

p3 (−)

Breeding Environments (−)31·712 (58·435)84·10 (67·15)
C. albicans

a1 (−)

p3 (+)

d1 (+)

d3 (−)

vr (+)

mn (+)

a0 (−)

mx (+)

p1 (+)

da (−)

mx (+)

p3 (+)Water (+)41·030 (50·362)65·78 (41·07)
C. delta

vr (+)

mx (+)

a1 (−)

p2 (−)

mn (−)

mx (+)

a0 (+)

a1 (+)

mx (−)

a3 (+)

p3 (−)

mn (−)

28·469 (63·01)84 (73·90)
C. fascipennis a1 (+)

a1 (−)

d1 (+)

da (−)

vr (+)

mx (+)

a1 (−)

a1 (+)

d1 (−)

d2 (+)

vr (−)

15·671 (61·5)87·97 (81·35)
C. festivipennis

a1 (−)

p1 (+)

d1 (+)

da (−)

vr (+)

mn (+)

a0 (+)

da (+)

p1 (−)

a1 (+)

a2 (+)

a3 (−)

vr (−)

p3 (+)23·014 (44·433)78·57 (60·93)

Model evaluation

Moran's I tests for spatial autocorrelation were insignificant at all neighbourhood sizes for all species and groups, except C. festivipennis which exhibited negative spatial autocorrelation at the 1·5-km and 2-km neighbourhood sizes (in a 1·5-km neighbourhood I = −0·39, = 0·003). The Obsoletus Group model verged on, but did not reach significance at the 1·5-km neighbourhood size (I = −0·2596, E[I] = −0·0323, = 0·06). The insignificant Moran's I values suggest spatial autocorrelation has little, or no, influence on patterns of midge trap catches at these scales.

Leave-one-out regression analysis was used for cross-validation due to the small number of sampled sites. Figure 3 shows the fitted abundance values from the leave-one-out regressions for the six most abundant species. Due to the stability of the fitted abundance values and variance explained, low generalization errors would be expected if our predictions were extended to another region of north Wales.

Figure 3.

Predicted vs. observed log(n + 1) abundance for the (a) Obsoletus and (b) Pulicaris Groups and for (c) C. pulicaris, (d) C. punctatus, (e) C. impunctatus and (f) C. achrayi. Error bars indicate the magnitude of standard errors, solid black line indicates line of best fit and dashed black line indicates line of equality.

Discussion

This study represents the first attempt to explain and predict trap catch patterns of the Obsoletus Group, Pulicaris Group species and other potential UK vectors at a high resolution (1 km) in relation to a range of ecologically relevant satellite, environmental and host factors. Culicoides inhabit a wide range of moist microhabitats in agricultural and natural ecosystems (Mellor, Boorman & Baylis 2000). As such, mapping potential areas of disease risk and targeting disease control measures requires models that can explain and predict local-scale variation in abundance, rather than simply occurrence, of species.

The Obsoletus and Pulicaris groups were the most abundant Culicoides caught in the Bala area, in agreement with previous studies in north Wales (McCall & Trees 1993; Baylis et al. 2010). Culicoides achrayi and C. impunctatus were the most abundant non-vector species. Culicoides achrayi is also highly prevalent on Belgian farms (Haubruge 2008). In contrast, relatively few C. achrayi were trapped at Chester Zoo (Vilar et al. 2011) and instead a high-proportion C. kibunensis. The reasoning behind this is unclear, but apparent opposing preferences for a farm or zoo environment are likely due to differences in these species' life-history characteristics.

Large differences between Culicoides catches on neighbouring farms highlight the fact that catches on one farm should not be deemed representative of a region. This raises questions about the validity of nationwide entomological surveillance schemes which, inevitably, trap at coarse resolutions. Our results suggest large-scale surveillance should consist of multiple trapping sites in each sampling area, or undertaking a prior survey of multiple farms in each region and proceeding with those that yield the highest catches.

A trade-off exists between temporal and spatial aspects of most surveys. In our study, the intense spatial detail (one trap per km2) and high-frequency nightly trapping required a degree of effort and generated a volume of midges that precluded undertaking trapping longer than three nights per site and 12 nights in total. This may limit our ability to generalize findings in both space and in time. We cannot be sure that the same farm-to-farm heterogeneity occurs elsewhere in the United Kingdom or at Bala at other times of year or in other years. However, a study near Bala in May–June 2007 reported significant heterogeneity in Culicoides catches between just four farms over 12 nights trapping on each farm, suggesting that our findings are robust, at least for Bala. If verified, the small-scale variability observed in this study should highlight an area of concern for those interpreting large-scale studies with scarce sampling points.

High local spatial variation may also explain the difficulty of building strong large-scale models for the Obsoletus Group (Purse et al. 2012). We have successfully modelled high spatial variation at the small spatial scale for several species, including the Obsoletus Group. One reason for the success of high-resolution models is that nearby farms may differ in the levels of important explanatory variables while distant sites spread across large areas may differ in the nature of those explanatory variables.

Abundances of several vector species (Obsoletus Group, C. pulicaris and C. punctatus) were correlated, suggesting common predictors of abundance due to similar life-history characteristics, such as the presence of live hosts. Similarly, male abundances of three of the Obsoletus Group species were correlated, again suggesting similar ecological factors favour abundance.

Calvete et al. (2008) found incorporating host variables into bioclimatic models vastly increased the variance explained in BTV-4 occurrence in Spain. In Italy, both biotic (forest and vegetation activity) and abiotic (topography, temperature and aridity index) axes were found to govern the occurrence of the C. obsoletus group (Conte et al. 2007a). Half of our models contained satellite variables only, and there was no significant difference between the explanatory power of these or the mixed predictor models. This concurs with Calvete et al. (2008) who, from the superior performance of climate only models, inferred that bioclimatic variables were the main ecological factors driving BTV occurrence across Spain. Satellite-derived ecological correlates dominated in number in the final models for all species in Bala, highlighting that the importance of these ecological drivers extends to the local scale.

For the Obsoletus Group, four species (C. chiopterus, C. dewulfi, C. obsoletus and C. scoticus) with a mix of host and breeding-habitat preferences were modelled together. Even so, the model explained a high amount of variance (81%) and was consistent with other studies detecting large impacts of landscape factors, such as forest cover and vegetation activity, on the distribution patterns of the Obsoletus Group or Complex (Purse et al. 2004a; Conte et al. 2007a; Calvete et al. 2008). The goodness of fit of the group model could either indicate that one species is highly dominant (and thus the model most influenced by its requirements) or that all species have similar habitat requirements. Further work should be undertaken to explore whether these species' habitats differ.

The distributions of the Obsoletus Group were dominated by remotely sensed NDVI variables. Most NDVI variable coefficients were positive, indicating a preference for microclimates with high levels of moisture, favouring vegetation growth. This is supported by previous observations that C. obsoletus breeds preferentially in forest litter (Amosova 1956; Dzhafarov 1964). Conte et al. (2007a) also found significant correlation between areas of deciduous and mixed broadleaved/coniferous forests and the Obsoletus Complex. The inclusion of two MIR variables, also correlated with vegetation levels (Boyd & Curran 1998), also supports this theory. The correlation of ‘number of water sources’ with the Obsoletus Group indicates that moist habitats are favourable, likely due to their semi-aquatic larval stage, and, indeed, C. obsoletus has been reared from wet areas of Scotland (Kettle & Lawson 1952).

Obsoletus Group trap catches also increased with the number of sheep on farms. This is in agreement with Garcia-Saenz, McCarter & Baylis (2010) who found a linear increase between C. obsoletus trap catches and sheep number.

Molecular identification was deemed unnecessary as there is currently insufficient information to determine vector competence of the Obsoletus Group species individually. If it becomes clear that there are important differences between vector competences, molecular identification of the Obsoletus Group species would provide further evidence of their ecological differences. Whilst sufficient numbers of males were caught and could be used to model the Obsoletus Group species, there is little evidence their relative abundance is proportional to the females of each species.

Like the Obsoletus Group, C. pulicaris and C. punctatus abundances were positively correlated with NDVI variables and sheep number. Catches of C. pulicaris, a wet-soil and bog species, have been associated with high, stable, levels of moisture (high, less variable NDVI) elsewhere in Europe (Purse et al. 2004b, 2005). The C. punctatus model is similar to that of C. pulicaris, only lacking in LST, highlighting the close relationship of the two species. The inclusion of LST variables in the C. pulicaris model may highlight the species' need for more stable temperatures, as highlighted by Parker (1950) who found that C. punctatus eggs are less adversely affected by above-normal temperatures than Cpulicaris eggs.

The C. impunctatus model, unlike the other Pulicaris Group species models, was dominated by LST variables. The annual mean of the dLST indicates a preference for warmer temperatures and a later peak in day-time temperatures (positive coefficient for p1). EVI featured heavily in this model, with an increased trap catch in locations with low variation in vegetation activity throughout the year and, in converse to the Obsoletus Group model, less densely covered areas with more access to sunlight. This is likely to represent this species' preference for organically enriched, soil-breeding sites with high water content (Blackwell, Young & Mordue 1994; Blackwell et al. 1999). Kettle & Lawson (1952) describe that immature stages of C. impunctatus are commonly found in bogland sites in wetter areas of moorland where Sphagnum and Polytrichum moss growth is thin enough to permit penetration by Juncus articulates, a species of rush that thrives in hot overhead sunlight. The exclusion of ‘number of sheep’ in the C. impunctatus model may be attributable to its autogenous nature (Boorman & Goddard 1970), making the species less reliant on host blood meals than anautogenous vector species. In turn, the only models incorporating host factors are those that have, currently, been implicated as BT vector species.

Four of the species had parous rates that were significantly correlated to their abundance. A high parous rate is an indication of survivorship, and the correlation for these four species suggest that survivorship, as opposed to proximity to breeding sites, is the reason for the high population sizes, reinforcing the reliability of this data.

Analysing a large number of explanatory variables relative to the number of data points creates a danger of overfitting. Leave-one-out regression analysis is a useful technique for examining overfitting as overfitted models often show poor predictive ability. The leave-one-out regressions of our models show good predictive ability, and therefore, while we accept there was a risk of overfitting, our models show no evidence of it.

Our study highlights the high variation in Culicoides abundance that can occur between neighbouring farms, but work still needs to be undertaken to determine the level of variation present at the smaller, within-farm scale and the factors driving that variation. Given the remarkable heterogeneity detected, we recommend that large-scale surveillance includes multiple sites per region or that a prior survey of each region is undertaken to determine those farms with the highest Culicoides catches.

Acknowledgements

We would like to thank the farmers for allowing us to undertake trapping on their land, Dr G. R. Wint and Professor. D. J. Rogers, University of Oxford, for supplying the MODIS satellite data and Professor P. J. Diggle, Lancaster University, for his statistical input. G. K. was supported by a BBSRC DTG-funded PhD studentship awarded to M.B. and Dr Jon Read. D.S. was supported by a DEFRA-VTRI M.Sc. award. The project was conceived by M.B. and designed by M.B. and H.G. Fieldwork was led by D.S., with assistance from H.G., K.M.M. and M.J.V. Insect identifications were led by K.L and statistical analysis, modelling and interpretation were by G.K. G.K wrote the manuscript, with support from H.G. and M.B.

Ancillary