eks (version 1.1.1)

tidyst_kda: Tidy and geospatial kernel discrimination analysis (classification)

Description

Tidy and geospatial versions of kernel discrimination analysis (classification) for 1- and 2-dimensional data.

Usage

tidy_kda(data, ...)
st_kda(x, ...)

Value

--For tidy_kda, the output is an object of class tidy_ks, which is a tibble with columns:

x

evaluation points in x-axis (name is taken from 1st input variable in data)

y

evaluation points in y-axis (2-d) (name is taken from 2nd input variable in data)

estimate

weighted kernel density estimate values

prior_prob

prior probabilities for each group

ks

first row (within each group) contains the untidy kernel estimate from ks::kda

tks

short object class label derived from the ks object class

label

estimated KDA group label at (x,y)

group

grouping variable (same as input).

--For st_kda, the output is an object of class st_ks, which is a list with fields:

tidy_ks

tibble of simplified output (ks, tks, label, group) from tidy_kda

grid

sf object of grid of weighted kernel density estimate values, as polygons, with attributes estimate, label, group copied from the tidy_ks object

sf

sf object of 5%, 10% ..., 95% contour regions of weighted kernel density estimate, as multipolygons, with attributes contlabel derived from the contour level; and estimate, group copied from the tidy_ks object.

Arguments

data

grouped tibble of data values

x

sf object with grouping attribute and with point geometry

...

other parameters in ks::kda function

Details

A kernel discriminant analysis (aka classification or supervised learning) assigns each grid point to the group with the highest density value, weighted by the prior probabilities.

The output from *_kda have the same structure as the kernel density estimate from *_kde, except that estimate is the weighted kernel density values at the grid points (weighted by prior_prob), and label becomes the KDA grouping variable that indicates to which of the groups the grid points belong. The output is a grouped tibble, grouped by the input grouping variable.

For details of the computation of the kernel discriminant analysis and the bandwidth selector procedure, see ?ks::kda. The bandwidth matrix of smoothing parameters is computed as in ks::kde per group.

Examples

## tidy discriminant analysis (classification)
library(ggplot2)
data(hsct, package="ks")
hsct2 <- dplyr::as_tibble(hsct)
hsct2 <- dplyr::filter(hsct2, PE.Ly65Mac1 >0 & APC.CD45.2>0 &
    (subject==6 | subject==12))
hsct2 <- dplyr::select(hsct2, PE.Ly65Mac1, APC.CD45.2, subject)
hsct2 <- dplyr::mutate(hsct2, subject=factor(paste0("Subject #", subject), 
    levels=c("Subject #6", "Subject #12")))
hsct2 <- dplyr::group_by(hsct2, subject)
hsct1 <-dplyr::select(hsct2, PE.Ly65Mac1, subject)

## tidy 1-d classification
t1 <- tidy_kda(hsct1) 
gt1 <- ggplot(t1, aes(x=PE.Ly65Mac1)) 
gt1 + geom_line(aes(colour=subject)) + 
    geom_rug(aes(colour=label), sides="b", linewidth=1.5)

## tidy 2-d classification
t2 <- tidy_kda(hsct2)
gt2 <- ggplot(t2, aes(x=PE.Ly65Mac1, y=APC.CD45.2))
gt2 + geom_contour_ks(aes(colour=subject)) + 
    geom_tile(aes(fill=label), alpha=0.2) + coord_fixed()

## geospatial classification
data(wa)
data(grevilleasf)
grevillea_gr <- dplyr::filter(grevilleasf, species=="hakeoides" |
    species=="paradoxa")
grevillea_gr <- dplyr::mutate(grevillea_gr, species=factor(species))  
grevillea_gr <- dplyr::group_by(grevillea_gr, species)
s1 <- st_kda(grevillea_gr)
s2 <- st_ksupp(st_kde(grevillea_gr))
s1$grid <- sf::st_filter(s1$grid, sf::st_convex_hull(sf::st_combine(s2$sf)))

## base R plot
xlim <- c(1.2e5, 1.1e6); ylim <- c(6.1e6, 7.2e6)
plot(wa, xlim=xlim, ylim=ylim)
plot(s1, which_geometry="grid", add=TRUE, alpha=0.1, border=NA, legend=FALSE)
plot(s1, add=TRUE, lwd=2)

## geom_sf plot
gs1 <- ggplot(s1) + geom_sf(data=wa, fill=NA) + ggthemes::theme_map() + 
    geom_sf(data=s1$grid, aes(fill=label), alpha=0.1, colour=NA) 
gs1 + geom_sf(data=st_get_contour(s1), aes(colour=species), fill=NA) +
    coord_sf(xlim=xlim, ylim=ylim)