Making the most out of different sources and
qualities of data in species distribution modeling:
an example on the distribution of brown bears
(Ursus arctos, L.) in Greece.
Anne-Sophie Bonnet-Lebrun, Alexandros A. Karamanlidis,
Miguel de Gabriel Hernando, Olivier Gimenez
INTRODUCTION – the brown bear (Ursus arctos)
Distribution:
- Holartic
- Greece: southernmost distribution in Europe
Conservation status:
- Globally: Least Concern (IUCN Red List status)
- Locally in Europe: small and isolated populations
Threats:
- Habitat loss and fragmentation
- Human-bear conflicts (livestock depredation, agricultural damage…)
 map its distribution in Greece -> map potential problematic interactions
INTRODUCTION
Brown bears are difficult to monitor:
- Cryptic and solitary
- Low density in very large areas
-> Build models (Species Distribution Models)
-> Use multiple data sources
METHODS – Species Distribution Models
Partial information on
the species’ presence
+
Probabilities of presence in
the overall area of interest
Environmental variables
METHODS – Environmental variables (1km² grain size)
- Physical (digital elevation model)
Mean slope
Cumulated length of
rivers in the pixel
Altitude
METHODS – Environmental variables (1km² grain size)
- Land cover (CORINE land cover maps)
Percentage of forest
Percentage of agricultural land
METHODS – Environmental variables (1km² grain size)
- Human-related
Distance to the roads
Population density
METHODS – Datasets
- Power poles: collection of
signs of bear presence (hair
deposits, marks, mud smears,
scats, tracks, …)
-> presence-absence
-> systematic
METHODS – Power poles
METHODS – Datasets
- Opportunistic data: citizen calls
when signs of bear presence
-> presence-only
-> opportunistic
=> observer bias
(more observations
in places that are
more accessible by
more people)
METHODS – Power poles data
Signs at power poles:
- Presence/absence = binomial response
 GLM possibly with spatial autocorrelation
(autologistic model; Augustin et al. 1996)
Presence/Absence ~ Mean slope (-)
+ Population density (-)
+ Rivers length (+)
+ Mean altitude (+)
+ % of agricultural land (+)
METHODS – Opportunistic data
-presence-only data
 inhomogeneous Poisson point process (Warton & Shepherd 2010)
- Poisson point process: random process to generate points scattered in space
- Intensity: average number of points per unit area
Homogeneous
Intensity = constant
Inhomogeneous
Intensity = f(spatial variable)
METHODS – Opportunistic data
-opportunistic data
 Model observer bias
1. Use both ecological (forest cover, altitude, …)
and observer bias (distance to the roads, …)
variables to build the model
2. condition on a common level of bias when
projecting the model (Warton & Renner 2013)
METHODS – Opportunistic data
Very much alike ANCOVA, propensity score method in biostatistics
RESULTS
Model based on power poles data
Probabilities of presence
Model based on opportunistic data
Average expected
number of observations
The results of the two models seem coherent
DISCUSSION
Outputs of the models:
Power poles data (GLM) -> probabilities of presence
Opportunistic data (PPM) -> intensities (average number of points per unit area)
How to combine the results; reconcile the outcomes?
How to evaluate the relative contribution of each dataset?
Model selection for each dataset / integrated dataset
How to assess the goodness-of-fit of integrated model to the data?
Are Poisson Point Processes the way to go?
DISCUSSION
DISCUSSION
ISEC 2014
DISCUSSION
ISEC 2014
Thank you for your attention!
Thank you for your attention!