nnErrMap: Generate a map of the distance - or error - from a pixel to...

In henkelstone/NPEL.Classification: Classifaction and data-handling routines for NPEL Caribou Project

Description

Generate a map of the distance — or error — from a pixel to it's nearest neighbour.

Usage

nnErrMap(model, idLayer, rData, outFilename, kernel = NULL, weight = NULL)

Arguments

model	the model used to generate the ID layer
idLayer	the layer identifying which neighbour is the nearest of each pixel
rData	a Raster* object with the 'other' raster data, that is all the data used to generate the model, and hence the distance from this pixel to the nearest neighbour. Note: this must be the same extents and alignment as idLayer
outFilename	the filename to which to write the data as it is generated.
kernel	(optional) the function used to compute the distance, see details.
weight	(optional) a function used to convert the distance to an error, see details.

Details

One way to estimate the accuracy/confidence of a nearest neighbour map, is to compute the distance from a pixel to it's nearest neighbour. It is important to understand that this 'distance' is not the spatial distance, i.e. the distance along the earths surface, but rather the distance in the space in which this was the nearest neighbour, that is, the distance in the statistical space used. For example, this might be the so-called ‘Tasselled Cap’ variables.

Since it is possible to measure distance through parameter space in non-Euclidean ways, the kernal parameter is provided. If this argument is specified it should be a function in the form f(a,b) in which a and b are the vectors of data for this pixel and it's nearest neighbour. If kernel is not provided it will default to the Euclidean distance, that is:

distance = sqrt(sum( (a-b)^2 ))

Distance, however, is not accuracy; in some cases though the two may be linked. There could be a number of ways to determine this relationship: linear modelling, piecewise linear modelling, logit models, etc. Once a model is developed, it is relatively simply to predict accuracy using (other) measures of distance. Example code for doing this is included below.

Value

a raster* object representing distances/kernel values, or the error from weight function if specified.

See Also

generateModels, and writeTile for more information on building models for imputation purposes.
impute for imputing other values from the map of nearest neighbours.

Examples

fx <- formula('siteID ~ brtns + grnns + wetns + dem + slp + asp + hsd')
nnData <- cbind(siteID=factor(1:nrow(siteData)), siteData)
models <- generateModels(nnData, suppModels[!suppModels %in% contModels], fx)

fNN <- paste0(dirname(tempfile()),'/Tmp_nn.tif')
egTile <- readTile(file.path(system.file("extdata", "egTile", package = "NPEL.Classification"),''),
                   layers=c('base','grnns','wetns','brtns','dem','slp','asp','hsd'))
egData <- writeTile (models[[1]], egTile, fNN, layers='class')
names (egData) <- 'siteID'

fEMap <- paste0(dirname(tempfile()),'/Tmp_nnEMap.tif')
nnEMap <- nnErrMap(models[[1]], egData, egTile, fEMap)
plot (nnEMap)

## Generate a distance/accuracy relationship
inputClass <- cbind(ecoType=siteData$ecoType, getData(models[[1]]))
predClass <- inputClass[getFitted(models[[1]]),]

dfIn <- cbind(inputClass[,-(1:2)], predClass[,-(1:2)])
nl <- ncol(dfIn)/2
kernel <- function(a,b){ sqrt(sum((a-b)^2)) }
df <- data.frame(dist=apply(dfIn,1, function(x){kernel(x[1:nl],x[(nl+1):(2*nl)])}),
                 correct=as.numeric(inputClass$ecoType == predClass$ecoType))
rm (dfIn,nl,kernel)

plot (df$dist,df$correct)
corrLm <- lm ('correct~dist', df)
summary (corrLm)                # No significant relationship in this case
abline (corrLm)

# Convert map to estimated accuracies afterwards; not meaningful with non-significant model!
Func <- function(x){ corrLm$coefficients[1] + x*corrLm$coefficients[2] }
plot(calc(nnEMap, Func))

unlink (fNN)
unlink (fEMap)

henkelstone/NPEL.Classification documentation built on May 17, 2019, 3:42 p.m.