FitRoi: Prediction at ungauged sites using region of influence and...

In martindurocher/floodStat: FloodNet Regional Frequency Analysis

Description

Return the prediction of a local regression model using region of influence (ROI) where the the residuals are further predicted by kriging.

Usage

FitRoi(x, xnew, k, trend, similarity, kriging = NULL, ker = TRUE, ...)

Arguments

x	Data for trainging the model.
xnew	Data at new locations (Validation set).
k	Number of sites in the neighborhoods.
trend	Formula defining the trend.
similarity	Formula defining covariates used for forming region of influence based on Euclidien distance.
kriging	Formula defining the spatial covariates . used for predicting residuals with spatial correlation.
ker	Should a kernel be used in the local regression model.
...	More arguments to pass to kriging function autoKrige
model	Variogram model. See vgm

Value

pred	Prediction at new sites.
pred.se	Standard deviation at new sites.
trend	Part of the prediction attributed to trend
trend.se	Standard deviation associated with the trend
fitted	Fitted values (training sites)
fitted.se	Standard deviaton of the fitted values
vgm	Sample variogram.
model	Fitted variogram model.

Examples

attach(canFlood)

## Transform data if necessary
xdf <- with(canFlood,
  data.frame(y      = log(L1),
             area   = scale(log(BASIN_AREA)),
             wb     = scale(log(WB_AREA_PC)),
             stream = scale(log(STREAM_DEN)),
             map    = scale(log(PPTAVG_BAS)),
             lon    = ELON,
             lat    = ELAT))

 # Note that the coordinates were transform to a Euclidean space using
 # multidimentional scaling

 ## select a validation and training set
 vid <- runif(nrow(xdf)) > .8
 tid <- !vid

 # formula of the relationship between flood quantile and descriptors
 ftrend <- y ~ area + map + wb + stream
 fsimilarity <- ~ area + map + wb + stream

 ## Fit a local regression model
 fit <- FitRoi(x = xdf[tid,], xnew = xdf[vid,], k = 50,
               ftrend, similarity = fsimilarity)
 print(fit)
 sd(xdf[vid,'y'] - fit$pred)

 ## Refit the model and perform the kriging of the residuals
 fitk <- FitRoi(x = xdf[tid,], xnew = xdf[vid, ], k =50,
               ftrend, loc = fsimilarity ,
               kriging = ~ lon + lat)
 print(fitk)
 sd(xdf[vid,'y'] - fitk$trend)
 sd(xdf[vid,'y'] - fitk$pred)

 ifold0 <- sample(floor((1:n)/(n+1)*fold)+1)

 ## Perform cross-validation. Long to compute
 ## This speed up the computation as there is no search for a best  model
 f <- FitRoi.cv(x = xdf, k = seq(20,150, 10), fold = ifold0,
               ftrend,  similarity = ~ lon + lat,
               verbose = TRUE, crit = 'mad')

martindurocher/floodStat documentation built on May 31, 2019, 12:42 a.m.