fields (version 11.6)

smooth.2d: Kernel smoother for irregular 2-d data

Description

An approximate Nadaraya Watson kernel smoother is obtained by first discretizing the locations to a grid and then using convolutions to find and to apply the kernel weights. The main advantage of this function is a smoother that avoids explicit looping.

Usage

smooth.2d(Y, ind = NULL, weight.obj = NULL, setup = FALSE, grid = NULL,
    x = NULL, nrow = 64, ncol = 64, surface = TRUE, cov.function =
gauss.cov, Mwidth = NULL, Nwidth = NULL, ...)

Arguments

Y

A vector of data to be smoothed

ind

Row and column indices that correspond to the locations of the data on regular grid. This is most useful when smoothing the same locations many times. (See also the x argument.)

weight.obj

An object that has the FFT of the convolution kernel and other information ( i.e. the result from calling this with setup=TRUE).

setup

If true creates a list that includes the FFT of the convolution kernel. In this case the function will return this list. Default is false.

grid

A list with components x and y being equally spaced values that define the grid. Default are integers 1:nrow, 1:ncol. If x is given the ranges will be used to define the grid.

x

Actual locations of the Y values. Not needed if ind is specified.

nrow

Number of points in the horizontal (x) axis of the grid. Not needed if grid is specified the default is 64

ncol

Number of points in the vertical (y) axis of the grid. Not needed if grid list is specified the default is 64

surface

If true (the default) a surface object is returned suitable for use by image, persp or contour functions. If false then just the nrowXncol matrix of smoothed values is returned.

cov.function

S function describing the kernel function. To be consistent with the other spatial function this is in the form of a covariance function. The only assumption is that this be stationary. Default is the (isotropic) Gaussian.

Nwidth

The size of the padding regions of zeroes when computing the (exact) convolution of the kernel with the data. The most conservative values are 2*nrow and 2*ncol, the default. If the kernel has support of say 2L+1 grid points then the padding region need only be of size L+1.

Mwidth

See Nwidth.

…

Parameters that are passed to the smoothing kernel. ( e.g. the scale parameter theta for the exponential or gaussian)

Value

Either a matrix of smoothed values or a surface object. The surface object also has a component 'ind' that gives the subscripts of the image matrix where the data is present.

Details

The irregular locations are first discretized to a regular grid ( using as.image) then a 2d- FFT is used to compute a Nadaraya-Watson type kernel estimator. Here we take advantage of two features. The kernel estimator is a convolution and by padding the regular by zeroes where data is not obsevred one can sum the kernel over irregular sets of locations. A second convolutions to find the normalization of the kernel weights.

The kernel function is specified by an function that should evaluate with the kernel for two matrices of locations. Assume that the kernel has the form: K( u-v) for two locations u and v. The function given as the argument to cov.function should have the call myfun( x1,x2) where x1 and x2 are matrices of 2-d locations if nrow(x1)=m and nrow( x2)=n then this function should return a mXn matrix where the (i,j) element is K( x1[i,]- x2[j,]). Optional arguments that are included in the ... arguments are passed to this function when it is used. The default kernel is the Gaussian and the argument theta is the bandwidth. It is easy to write other other kernels, just use Exp.cov.simple as a template.

Examples

# NOT RUN {
# Normal kernel smooth of the precip data with bandwidth of .5 ( degree) 
#  
look<- smooth.2d( RMprecip$y,  x=RMprecip$x, theta=.25)

# finer resolution used in computing the smooth 
look3<-smooth.2d( RMprecip$y, x=RMprecip$x, theta=.25, nrow=256, 
ncol=256,Nwidth=32,
Mwidth=32) 
# if the width arguments were omitted the padding would create a  
# 512X 512 matrix with the data filled in the upper 256X256 part. 
# with a bandwidth of .25 degrees the normal kernel is essentially zero  
# beyond 32 grid points from its center ( about 6 standard deviations) 
#
# take a look:

#set.panel(2,1)
#image( look3, zlim=c(-8,12))
#points( RMprecip$x, pch=".")  
#image( look, zlim =c(-8,12))
#points( RMprecip$x, pch=".")  


# bandwidth changed to .25, exponential kernel   
look2<- smooth.2d( RMprecip$y, x=RMprecip$x, cov.function=Exp.cov,theta=.25)
# 


 
# }

Run the code above in your browser using DataCamp Workspace