rfe4rainfall: Performs recursive feature selction

In environmentalinformatics-marburg/Rainfall: Functions for training, predictiion and validation of an MSG based optical rainfall retrieval

Description

Performs recursive feature selction

Usage

rfe4rainfall(predictors, response, sampsize = 1, threshold = 0.06,
  out = "Rain", varSize = c(1:5), tuneGrid = list(.size = 2:5, .decay =
  c(0.05, 0.07)), method = "nnet", seed = 20)

Arguments

predictors	Either a data.frame with each column is one predictor and each row represents one pixel. Or (if only one scene is used for training) a RasterStack with one Raster is one Predictor Variable.
response	A vector of either Rainfall area or rainfall rates for the corresponding pixels in predictors. If only one scene is used for model training, "response" may also be a RasterLayer of the response variable.
threshold	if response is Rainfall rate: pixels larger than the threshold are used for rainfall rate training
out	Either Rain or RInfo indicating weather rainfall rates or rainfall areas should be used.
varSize	integer vector indicating the numbers of variables to consider in rfe.
tuneGrid	list of tuning parameters to be supplied to model training. See https://topepo.github.io/caret/modelList.html for tuning values
method	ML algorithm to be applied. default is nnet
seed	Any integer number. Used to produce reproducable results
sampsize.	Number of data points used for training

Details

Predictors are centered and scaled according to mean and sd values. If the day of the year is used as predictor, this variable is scaled considering max=365 and min=1

Value

A rfe model

Author(s)

Hanna Meyer

References

rfe Function in the caret package

See Also

rfe

Examples

# stack the msg scenes:
msg_example <-getChannels(inpath=system.file("extdata/msg",package="Rainfall"))

# raster the sunzenith 
sunzenith<-getSunzenith(inpath=system.file("extdata/msg",package="Rainfall"))

#get Date
date <- getDate(inpath=system.file("extdata/msg",package="Rainfall"))

#calculate variables (takes some time...)
pred <- calculatePredictors(msg_example,
sunzenith=sunzenith,
spectral=c("VIS0.6","NIR1.6","T0.6_1.6"),
texture=expand.grid(c("NIR1.6","T6.2_10.8"),
c("variance", "contrast"),c(3,5,9)),
filterstat=expand.grid(c("NIR1.6","T6.2_10.8"),
c("sd","min"),c(3,5,9)),
 further=NULL,
 date=date)
 
response <- raster(system.file("extdata/radar",
"201007121650_radolan_SGrid.rst",package="Rainfall"))

#Train small rfe model with 0.1% of the pixels (takes around 1 minute...)
rfeModel <- rfe4rainfall(predictors=pred,
response,
out="Rain",
sampsize=0.01,
varSize=c(2,5,10,15,20))

# Show results:
print(rfeModel)
plotRfeCV(rfeModel)
predictors(rfeModel)
plot(varImp(rfeModel$fit,scale=TRUE))

environmentalinformatics-marburg/Rainfall documentation built on May 16, 2019, 7:49 a.m.