Supplementary Material 1: Instructions for using program E

advertisement
1
Supplementary Material 1: Instructions for using program E-SURGE for implementing
2
models of habitat dynamics in the presence of misclassification.
3
4
It is possible to implement analysis of hidden markov chain models in program E-SURGE
5
(Choquet and Nogue 2010), freely downloadable at http://www.cefe.cnrs.fr/biom/En/
6
softwares.htm. Program E-SURGE works by decomposing dynamics into parameters associated
7
with initial states, transitions between states, and classification probabilities that relate
8
observations (events) to true states.
Occasion
t
t +1
Transition
True states
Events
Events
Observations
Transition
True states
Observations
9
10
The basic structural form needed to run analyses is specified with the following “pattern"
11
matrices in the “GEPAT" module in E-SURGE. Here, “*” entries denote the complement of the
12
sum of positive row entries, and “-“entries denote zeroes.
13
Note that E-SURGE was developed as software to analyze multi-state mark-recapture data in
14
wildlife populations in order to estimate survival and transition parameters in the face of state
15
uncertainty (Pradel 2005). Because of this, it includes an absorbing state called “Dead” in the
16
structure of transition and events matrices and a state Not Observed in the event matrix. The
1
17
three other states correspond to Forest (F), Managed (M) and Urban (U) for the pixel level
18
analysis or to Dominant forest (F), Average forest (A) and Low forest (L) for the cell level
19
analysis.
20
The matrix notation used in E-SURGE is different than the matrix notation used in matrix
21
models of demography: in E-SURGE, departure states (or FROM) are defined in rows and
22
arrival states (or TO) in columns. In matrix models used in demography (the notation we used in
23
this study), departure states (or FROM) are defined in columns and arrival states (or TO) in rows.
24
25
We first describe the pattern matrices in GEPAT
26
1-For the initial state vector, we have
27
𝜋𝑡 = [𝜋
28
2-For the transition parameter matrix, we have
29
∗
𝜓
𝜓𝑡 = [
𝜓
−
30
The fourth row and the fourth column correspond to the state Dead, an absorbing state from
31
which transition elsewhere is not possible. In models of habitat dynamics, this might correspond
32
to habitat that becomes completely unavailable (for example a parking lot). However, since these
33
instances do not occur in our data, we constrained transitions to and from Dead to 0.
34
3-For the event process (the habitat classification probabilities in our study), denoted as B:
35
− ∗ 𝛽
− 𝛽 ∗
𝐵𝑡 = [
− 𝛽 𝛽
∗ − −
36
The fourth row corresponds again to the state Dead, the first column to the state Not Observable.
37
Thus corresponding classification parameters are constrained to 0 in our models.
𝜋
∗]
𝜓 𝜓 −
∗ 𝜓 −
]
𝜓 ∗ −
− − ∗
𝛽
𝛽
]
∗
−
2
38
39
After these structures are specified in “GEPAT," the user must symbolically formulate
40
design matrices to specify linear models for each parameter. This can be done using the
41
“GEMACO" function in E-SURGE.
42
43
44
We analyzed 3 types of models in this study:
1- Models where classification is not accounted for. This is equivalent to constraining all 𝛽
45
parameters previously defined in GEPAT to 0 in GEMACO in order to set all
46
classification probabilities to either 1 or 0 (no misclassification) ( in “GEPAT”, this
47
matrix is slightly different from a diagonal matrix because of the presence of the Dead
48
and Not Observable states).(Fig.1)
49
50
51
52
53
54
Fig. 1: Schematic representation of type 1 models in E-surge : at each time step (or occasion), β
55
parameters of the matrix B (so-called event probabilities in E-surge or classification probabilities
3
56
in this study) are fixed to 0. Note that constraining the β (misclassification) parameters to 0
57
results in the diagonal (correct classification) elements being 1.
58
2- Models where information on true states at the level of a cell is obtained from computing
59
expected true states based on pixel accuracy information. Dummy occasions are created
60
at each time step in order to incorporate this additional information. Between dummy
61
occasions and real occasions, all 𝜓 parameters of the transition matrix previously defined
62
in GEPAT are fixed to 0 in GEMACO (resulting in an identity transition matrix) as well
63
as all 𝛽 parameters of the event matrix fixed (Fig. 2)
64
65
66
67
68
69
70
71
72
73
74
75
76
Fig. 2: Schematic representation of type 2 models in E-surge : for each time step (or occasion), a
77
dummy occasion is created. Between dummy occasion and real occasion, both transition and
78
event probabilities defined in GEPAT are fixed to 0.
4
79
3- Lastly, we also analyzed a model where ground-truth data are available to determine true
80
states, but only for a sample of habitat histories and only for the last occasion. Note that 𝛽
81
parameters are estimable for the last time period. Application of these classification
82
probabilities to data from previous time periods requires the assumption that these
83
classification probabilities are constant over time.
84
85
86
87
Fig. 3: Schematic representation of type 3 models in E-SURGE : in these models a dummy
88
occasion is created only for the last time step, where ground-truth data are available for a sample
89
of habitat histories (their habitat histories have a last column not equal to 0). Between the last
5
90
occasion and the dummy occasion, both transition and event probabilities defined in GEPAT are
91
fixed to 0. Note that when coding the habitat histories in E-SURGE, the group with no ground
92
truth has a negative number in the last column (indicating no data for the last occasion)
93
94
Reference
95
Choquet, R., and E. Nogue. 2010. E-SURGE 1.7 User’s Manual. CEFE, Montpellier, France.
96
6
Download