| eval_method {highriskzone} | R Documentation |
Evaluation of the procedures determining the high-risk zone.
Description
Evaluates the performance of the three methods:
Method of fixed radius
Quantile-based method
Intensity-based method
For further details on the methods, see det_hrz or the paper of Mahling et al. (2013)(References).
There are three ways to simulate data for the evaluation.
Usage
eval_method(
ppdata,
type,
criterion,
cutoff,
numit = 100,
nxprob = 0.1,
distancemap = NULL,
intens = NULL,
covmatrix = NULL,
simulate,
radiusClust = NULL,
clustering = 5,
pbar = TRUE
)
Arguments
ppdata |
Observed spatial point process of class ppp. |
type |
Method to use, can be one of |
criterion |
criterion to limit the high-risk zone, can be one of
|
cutoff |
Value of criterion (area, radius, quantile, alpha or threshold). Depending on criterion and type: If criterion = "direct" and type = "intens", cutoff is the maximum intensity of unexploded bombs outside the risk zone. If type = "dist" instead, cutoff is the radius of the circle around each exploded bomb. "If criterion = "indirect", cutoff is the quantile for the quantile-based method and the failure probability alpha for the intensity-base method. If criterion = "area", cutoff is the area the high-risk zone should have. |
numit |
Number of iterations |
nxprob |
Probability of having unobserved events. Default value is 0.1. |
distancemap |
(optional) distance map: distance of every pixel to the nearest observation
of the point pattern; only needed for |
intens |
(optional) estimated intensity of the observed process (object of class "im"),
only needed for type="intens". If not given,
it will be estimated using |
covmatrix |
(optional) Covariance matrix of the kernel of a normal distribution, only needed for
|
simulate |
The type of simulation, can be one of |
radiusClust |
(Optional) radius of the circles around the parent points in which the cluster
points are located. Only used for |
clustering |
a value >= 1 which describes the amount of clustering; the
adjusted estimated intensity of the observed pattern is divided by
this value; it is also the parameter of the Poisson distribution
for the number of points per cluster. Only used for |
pbar |
logical. Should progress bar be printed? |
Details
The three simulation types are:
- Data-based simulation
-
Here a given data set is used. The data set is thinned as explained below. Note that this method is very different from the others, since it is using the real data.
- Simulation of an inhomogeneous Poisson process
-
Here, an inhomogeneous Poisson process is simulated and then that data is thinned.
- Simulation of a Neyman-Scott process
-
Here a Neyman-Scott process is simulated (see
sim_nsppp,rNeymanScott) and this data is then also thinned.
Thinning:
Let X be the spatial point process, which is the location of all events and let Y
be a subset of X describing the observed process. The process of unobserved events
then is Z = X \ Y , meaning that Z and Y are disjoint and together
forming X.
Since Z is not known, in this function an observed or simulated spatial point pattern
ppdata is taken as the full pattern (which we denote by X') comprising the
observed events Y' as well as the unobserved Z'.
Each event in X' is assigned to one of the two processes Y' or
Z' by drawing independent Bernoulli random numbers.
The resulting process of observed events Y' is used to determine the high-risk zone.
Knowing now the unobserved process, it can be seen how many events are outside and inside the
high-risk zone.
type and criterion may be vectors in this function.
Value
A data.frame with variables
Iteration |
Iterationstep of the result |
Type, Criterion, Cutoff, nxprob |
see arguments |
threshold |
determined threshold. If criterion="area", it is either the distance (if type="dist") or the threshold c (for type="intens"). If criterion="indirect", it is either the quantile of the nearest-neighbour distance which is used as radius (if type="dist") or the threshold c (for type="intens"). If criterion="direct", it equals the cutoff for both types. |
calccutoff |
determined cutoff-value. For type="dist" and criterion="area", this is the quantile of the nearest-neighbour distance. For type="intens" and criterion="area", it is the failure probability alpha. For all other criterions it is NA. |
covmatrix11, covmatrix12, covmatrix21, covmatrix22 |
values in the covariance matrix. covmatrix11 and covmatrix22 are the diagonal elements (variances). |
numbermiss |
number of unobserved points outside the high-risk zone |
numberunobserved |
number of observations in the unobserved point pattern Z' |
missingfrac |
fraction of unobserved events outside the high-risk zone (numbermiss/numberunobserved) |
arearegion |
area of the high-risk zone |
numberobs |
number of observations in the observed point pattern Y' |
See Also
det_hrz, rNeymanScott, thin, sim_nsppp, sim_intens
Examples
## Not run:
data(craterB)
# the input values are mainly the same as in det_hrz, so for more example ideas,
# see the documentation of det_hrz.
evalm <- eval_method(craterB, type = c("dist", "intens"), criterion = c("area", "area"),
cutoff = c(1500000, 1500000), nxprob = 0.1, numit = 10,
simulate = "clintens", radiusClust = 300,
clustering = 15, pbar = FALSE)
evalm_d <- subset(evalm, evalm$Type == "dist")
evalm_i <- subset(evalm, evalm$Type == "intens")
# pout: fraction of high-risk zones that leave at least one unobserved event uncovered
# pmiss: Mean fraction of unobserved events outside the high-risk zone
data.frame(pmiss_d = mean(evalm_d$missingfrac),
pmiss_i = mean(evalm_i$missingfrac),
pout_d = ( sum(evalm_d$numbermiss > 0) / nrow(evalm_d) ),
pout_i = ( sum(evalm_i$numbermiss > 0) / nrow(evalm_i) ))
## End(Not run)