fit.regModel: Fits a regression model to spatial data

In GSIF: Global Soil Information Facilities

Description

Fits a regression or a trend model (e.g. a GLM) and, if not available, a variogram for the response residuals using the default settings.

Usage

## S4 method for signature 
## 'formula,data.frame,SpatialPixelsDataFrame,character'
fit.regModel(formulaString, rmatrix, predictionDomain, 
      method = list("GLM", "rpart", "randomForest", "quantregForest", "lme",
       "xgboost", "ranger"), 
      dimensions = NULL, fit.family = gaussian(), stepwise = TRUE, rvgm, 
      GLS = FALSE, steps, subsample, subsample.reg, ...)

Arguments

formulaString	object of class "formula" — regression model
rmatrix	object of class "data.frame"; regression matrix produced as a result of spatial overlay
predictionDomain	object of class "SpatialPixelsDataFrame"; spatial domain of interest
method	character; family of methods considered e.g. "GLM", "rpart" (regression trees), "randomForest" (random forest)
dimensions	character; "2D", "3D", "2D+T", or "3D+T"
fit.family	family to be passed to the glm (see examples below)
stepwise	specifies whether to run step-wise regression on top of GLM to get an optimal subset of predictors
rvgm	residual variogram (to avoid fitting the variogram set as NULL)
GLS	fit trend model using Generalized Least Squares implemented in the nlme package
steps	integer; the maximum number of steps to be considered for step-wise regression; see stats::step for more details
subsample	integer; maximum number of observations to be taken for variogram model fitting (to speed up variogram fitting)
subsample.reg	integer; maximum number of observations to be taken for regression model fitting (especially important for randomForest modelling)
...	other optional arguments that can be passed to gstat::fit.variogram

Details

Produces an object of class "gstatModel" that contains: (1) fitted regression model (e.g. a GLM, cubist model, or randomForest model), (2) fitted variogram, and (c) object of class "SpatialPoints" with observation locations. To combine overlay and model fitting operations, consider using fit.gstatModel.

Author(s)

Tomislav Hengl, Mario Antonio Guevara Santamaria and Bas Kempen

See Also

fit.gstatModel, stats::glm, gstat::fit.variogram, randomForest::randomForest

Examples

library(sp)
library(rpart)
library(nlme)
library(gstat)
library(randomForest)
library(quantregForest)
library(xgboost)
library(caret)
library(scales)
library(AICcmodavg)

## load the Meuse data set:
demo(meuse, echo=FALSE)

## prepare the regression matrix:
ov <- over(meuse, meuse.grid)
ov <- cbind(data.frame(meuse["om"]), ov)
## skip variogram fitting:
m <- fit.regModel(om~dist+ffreq, rmatrix=ov, meuse.grid, 
    fit.family=gaussian(log), method="GLM", rvgm=NULL)
m@regModel
m@vgmModel
plot(m)
## fit a GLM with variogram:
m1 <- fit.regModel(om~dist+ffreq, rmatrix=ov, meuse.grid, 
    fit.family=gaussian(log), method="GLM")
m1@vgmModel
plot(m1)
rk1 <- predict(m1, meuse.grid)
plot(rk1)

## fit a regression tree with variogram:
m2 <- fit.regModel(log1p(om)~dist+ffreq, rmatrix=ov, meuse.grid,
     method="rpart")
plot(m2)
rk2 <- predict(m2, meuse.grid)
plot(rk2)

## fit a lme model with variogram:
m3 <- fit.regModel(log1p(om)~dist, rmatrix=ov, meuse.grid,
     method="lme", random=~1|ffreq)
plot(m3)
#rk3 <- predict(m3, meuse.grid)
#plot(rk3)

## fit a randomForest model with variogram
## NOTE: no transformation required
m4 <- fit.regModel(om~dist+ffreq, rmatrix=ov, meuse.grid,
     method="randomForest")
plot(m4)
rk4 <- predict(m4, meuse.grid)
plot(rk4)

## RF is very sensitive to the 'mtry' argument:
m4b <- fit.regModel(om~dist+ffreq, rmatrix=ov, meuse.grid,
     method="randomForest", mtry=2)
plot(m4b)

## Not run: 
## RF with uncertainty (quantregForest package)
m5 <- fit.regModel(om~dist+ffreq, rmatrix=ov, meuse.grid,
     method="quantregForest")
plot(m5)
rk5 <- predict(m5, meuse.grid)
plot(rk5)
dev.off()
## plot prediction error map:
spplot(rk5@predicted["var1.var"])

## ranger
m6 <- fit.regModel(om~dist+ffreq, rmatrix=ov, meuse.grid,
     method="ranger", rvgm=NULL)
plot(m6)
rk6 <- predict(m6, meuse.grid)
plot(rk6)

## XGBoost
m7 <- fit.regModel(om~dist+ffreq, rmatrix=ov, meuse.grid,
     method="xgboost", rvgm=NULL)
plot(m7)
rk7 <- predict(m7, meuse.grid)
plot(rk7)

## End(Not run)

GSIF documentation built on May 2, 2019, 5:44 p.m.