Modelling the roles of climate, landscape and biotic

Modelling the roles of climate, landscape and
biotic factors in the distribution and maximum
trap catch of Culicoides vectors across Europe.
S. Withenshaw1, K. Searle1, A. Butler2, A. Allepuz3, J. Barber4, S. Carpenter4, E. Denison4, K. Labuschagne5, P. Mellor4, M. Miranda6, S. Napp3, N. Pages3, C. Sanders4,
F. Stubbins7, B. Purse1
1. Centre for Ecology and Hydrology 2. Biomathematics and Statistics Scotland 3. Centre de Recerca en Sanitat Animal 4. The Pirbright Institute 5. PVVD, ARC-Onderstepoort Veterinary Institute 6. Laboratory of Zoology, University of the
Balaeric Island 7. School of Agricultural, Forest and Environmental Sciences, Clemson University
Table 1: Covariates investigated in models of Culicoides female maximum annual trap
abundance. All covariates were extracted for the location of each trap site for model fitting,
and the centroids of 1km grid squares across Europe for predictions.
Background
Culicoides biting midges (Diptera: Ceratopogonidae) transmit
bluetongue virus (BTV), which has caused significant
disruption to ruminant production systems in northern Europe
over the past decade. The presence and abundance of
female Culicoides vectors are important determinants of the
risk of BTV infection within a region, therefore understanding
the drivers of these spatial patterns will aid the prediction and
Figure 1: Culicoides spp.
mitigation of future BTV incursions.
Here we present preliminary results of a novel Bayesian modelling framework for
determining the climate, landscape and biotic factors driving the maximum annual
trap catch of three Culicoides species groups sampled across the UK and Spain.
The resulting models were used to make predictions of this variable across west
and central Europe for past years.
Methods
1. Data
• Midges sampled weekly at sites in UK and Spain
(2005-2010).
• 404 site-year combinations (Fig 2).
• Light-suction traps.
• Species identified morphologically and counted (C.
imicola, C. obsoletus complex, C. pulicaris complex).
• Maximum size of weekly female catch per species
per year at each site was determined.
Covariate
Culicoides
species
C. imicola
Precipitation Summed precipitation in mm (ppt) per year
Land cover
Covariates included
in the 20 best
models
0.58
Temperature (20)
Sheep density (16)
Summed NDVI (15)
Elevation (12)
Cattle density (8)
Roe deer density (5)
Red deer density (4)
Number of dry days (dd) per year
Coefficient of variation in summed precipitation (cvppt)
per year
Categorical variable indicating the type of land cover
present
MODIS 8-day 1km
Land Surface
Temp
ENSEMBLES
daily 0.25°
precipitation
MODIS annual
500m land cover
(Crop, Crop/Natural Vegetation Mosaic, Deciduous Broadleaf/Mixed
Forest, Evergreen Needleleaf Forest, Grassland, Open Shrubland,
Savanna, Woody Savanna, Urban)
NDVI
Summed values for the Normalised Difference
Vegetation Index per year
Elevation
Metres above sea level
MODIS 8-day
250m Vegetation
Indices
WORLDCLIM
1km elevation
Deer density Modelled density per km2 of roe deer, red deer and total [1] [2]
deer (roe + red)
Livestock
density
Modelled density per km2 of cattle, sheep, goats and
total (cattle+sheep+goats)
[3]
Results and Discussion
• The covariates included in models of maximum trap catch differed between the Culicoides spp.
groups considered here (Table 2).
• The model predicts the distribution of C. imicola to be restricted to southern Europe (Fig 3a), in
line with previous prediction maps (Fig 3d) and sampled presence (Fig 3e)[4,5]. However, there is
under-prediction of presence in Spain and over-prediction in Italy[6] (Fig 3f) and Bulgaria[7].
• The model predicts higher abundance of C. obsoletus complex and C. pulicaris complex in
northern Europe compared to the south. However, the southern range limits for both of these
groups are known to stretch down to north Africa[5] (Fig 3e). The prediction of zero abundance in
the south Mediterranean region is therefore likely to be an artefact of the sampling range (Fig 2).
• The predicted distribution of the C. pulicaris complex in Portugal looks more similar to the known
distribution of a closely-related species, C. punctatus, which penetrates less far south[8]. This may
indicate complications related to morphological identification of some midge specimens.
• Models will be further refined, including a comparison of performance using an alternative
source of temperature data (ENSEMBLES). The maximum catch of further Culicoides species will
also be modelled and predicted (C. newsteadii and C. impunctatus).
• Models of the timing and size of seasonal peaks of Culicoides abundance are also in development
using the same data set and will be used to predict the seasonal dynamics of midges across the
same geographical range.
C. obsoletus complex
C. imicola
a
b
Key to midge
abundance
No prediction
d
C. pulicaris complex
0-10
c
0
0
00 00 ,0
0 000 , 0 0, 000
0 000 ,00 0, 000 ,00 0,
0
00 ,00 0, 00 ,00 0, 00 ,00
-1 -1 1 -1 -1 -1 -1 -1 -1
0 10 0 - 0
0 0
0
0 0
10 00 ,00 ,00 ,00 ,00 ,00
,
1
0
0 0
10 0
10 1,00 ,00 ,00
0 0
1
10
10 – 102
<VALUE>
Mean outof-sample
R2
Source
Temperature Summed degree days >12°C per year
2. Model fitting
• Poisson GLMM fit in INLA to identify the best-fitting
models using covariates listed in Table 1.
• Full subset model selection
• All models also included year as a fixed effect and
site-year as a random effect
• Out-of-sample model fit (R2) was used to identify the
best model, using a subset of data not used to fit the
model. This process was iterated 20 times, resulting
Figure 2: Location of midge trap sites.
in 20 best models per species (Table 2).
3. Predictions
• Predictions of the maximum annual abundance per species were made for each 1km pixel
across Europe (Fig 3).
• Predictions from each of the top 20 models (based on out of sample prediction performance)
were averaged for each species, weighted by their relative support in the data.
• Predictions were only made for pixels that were environmentally similar to the training data set,
defined as those with a Malhalanobis Distance (MD) falling within the 0-90% cumulative
distribution of MDs from the training data set.
Table 2: Covariates included in the 20 best models of maximum
annual Culicoides female trap abundance. Mean R2 of the 20 best
models for each species is given. Number of times a predictor
variable was included in the best model for a species is in brackets.
Description
102 – 103
C. obsoletus
complex
0.58
Elevation (20)
Precipitation (dd) (20)
Summed NDVI (20)
Roe deer density (12)
Precipitation (cvppt) (10)
Red deer density (10)
Sheep density (8)
Cattle density (4)
e
104 – 105
f
105 – 106
106 – 107
107 – 108
108 – 109
0
C. pulicaris
complex
0.65
103 – 104
Temperature (20)
Precipitation (cvppt) (11)
Precipitation (dd) (11)
Cattle density (8)
Sheep density (6)
Elevation (4)
Precipitation (ppt) (2)
Roe deer density (2)
Red deer density (1)
Figure 3: (a-c) Model averaged predictions of the mean maximum catch of three Culicoides species females in 2007 according to the
20 models with the best fit to the data. Predictions were made for 1km2 pixels across Europe. No predictions were made if a pixel was
too environmentally dissimilar from the training data set according to Malhalanobis Distances (d) Previously modelled predictions of C.
imicola presence taken from [4] (e) Previously identified range-limits of Culicoides spp. groups taken from [5], (f) Previously identified
distribution of C. imicola and C. obsoletus complex in Italy taken from [6].
References: [1] Wint et al. (2014) Open Health Data 2: e1 [2] Alexander et al. (2014) Open Health Data 2: e2 [3] Robinson et al. (2014) PlosOne 9: e96084 [4] Purse et al. (2007). Journal
of Applied Ecology 44: 1231-1242 [5] Purse et al. (2005) Nature Reviews Microbiology 3: 171-182 [6] Conte et al. (2007). Veterinary Parasitology 150: 333-344 [7] Purse et al. (2006).
Medical and Veterinary Entomology 20: 335-344 [8] Capela et al. (2003). Medical and Veterinary Entomology 17: 165-177.
WHY: develop models to predict the distribution and
abundance of Culicoides spp. across Europe?
BECAUSE: they are vectors of arboviruses that infect
livestock and knowing their distribution and population
dynamics will aid control efforts.