Hidden Markov modelling of occupancy data
O. Gimenez, L. Blanc, A. Besnard, R. Pradel, P. Doherty, E. Marboutin, R. Choquet
Montpellier, July 4th, 2014
My conversion to occupancy models
• I’m more of a capture-recapture (CR) guy
• Can we use what we know from CR for occupancy?
• Individuals in CR = Sites in occupancy
• Two sides of same coin: hidden Markov models (HMM)
• Suggested by M. Kéry, I. Fiske, W. Challenger, …
• Flexible framework, well developed in other areas
Dynamic occupancy models as HMMs
1
O U
O = occupied; U = unoccupied
1 = species detected; 0 = species undetected
0 1
extinction
e
1-pp p
Dynamic occupancy models as HMMs
1
O U
O = occupied; U = unoccupied
1 = species detected; 0 = species undetected
0 1 0 0 0
1-pp
extinction
0
e
p 0 0
Dynamic occupancy models as HMMs
0
U O
O = occupied; U = unoccupied
1 = species detected; 0 = species undetected
0 0 1 1 1
colonization
g
0 0 0 p p p
Single-season occupancy models as HMMs
• No colonization ( =0g ) and no extinction (e=0) – closure assumption
Single-season occupancy models as HMMs
Initial statesState
processObservation
process
• No colonization ( =0g ) and no extinction (e=0) – closure assumption
Advantages of the HMM formulation
1. (Almost?) all occupancy models in a unified framework
• Single-season, dynamic models
• Mixtures and random effects (see case study)
• Multistate models, with uncertainty
• Multispecies models
• False-positives (Chambert & Miller, submitted)
• …
2. Formal link between occupancy and CR communities
E-SURGE and occupancy models
• E-SURGE, software developed to analyze CR data with HMMs
Rémi Choquet Roger Pradel
• Model specification via user-friendly syntax
• Numerical tools (random effects, identifiability)
E-SURGE and occupancy models
Case study with Eurasian lynx in
France
• Signs of presence between 2002 and 2006
• 197 sites, 5 1-y periods
• Single-season occupancy & detection heterogeneity
Van Gogh
Detection heterogeneity
Random effect Finite mixture
Royle (2006), Gimenez & Choquet (2010) Pledger et al. (2003), Pradel (2005)
Perspectives – based on CR experiences
1. Distribution mapping: unobserved states via Viterbi
2. Accounting for lack of independence:
• Trap-dependence
• Spatial autocorrelation (memory model)