train.spLearner-methods: Train a spatial prediction and/or interpolation model using...

In Envirometrix/landmap: Automated Spatial Prediction using Ensemble Machine Learning

Train a spatial prediction and/or interpolation model using Ensemble Machine Learning

Description

Automated spatial predictions and/or interpolation using Ensemble Machine Learning. Extends functionality of the mlr package. Suitable for predicting numeric, binomial and factor-type variables.

Usage

## S4 method for signature 'SpatialPointsDataFrame,ANY,SpatialPixelsDataFrame'
train.spLearner(
  observations,
  formulaString,
  covariates,
  SL.library,
  family = stats::gaussian(),
  method = "stack.cv",
  predict.type,
  super.learner = "regr.lm",
  subsets = 5,
  lambda = 0.5,
  cov.model = "exponential",
  subsample = 10000,
  parallel = "multicore",
  oblique.coords = TRUE,
  nearest = FALSE,
  buffer.dist = FALSE,
  theta.list = seq(0, 180, length.out = 14) * pi/180,
  spc = TRUE,
  id = NULL,
  weights = NULL,
  n.obs = 10,
  ...
)

Arguments

observations	SpatialPointsDataFrame.
formulaString	ANY.
covariates	SpatialPixelsDataFrame.
SL.library	List of learners,
family	Family e.g. gaussian(),
method	Ensemble stacking method (see makeStackedLearner) usually stack.cv,
predict.type	Prediction type 'prob' or 'response',
super.learner	Ensemble stacking model usually regr.lm,
subsets	Number of subsets for repeated CV,
lambda	Target variable transformation (0.5 or 1),
cov.model	Covariance model for variogram fitting,
subsample	For large datasets consider random subsetting training data,
parallel	logical, Initiate parallel processing,
oblique.coords	Specify whether to use oblique coordinates as covariates,
nearest	Specify whether to use nearest values and distances i.e. the method of Sekulic et al. (2020),
buffer.dist	Specify whether to use buffer distances to points as covariates,
theta.list	List of angles (in radians) used to derive oblique coordinates,
spc	specifies whether to apply principal components transformation.
id	Id column name to control clusters of data,
weights	Optional weights (per row) that learners will use to account for variable data quality,
n.obs	Number of nearest observations to be found in meteo::near.obs (by default 10),
...	other arguments that can be passed on to mlr::makeStackedLearner,

Value

object of class spLearner, which contains fitted model, variogram model and spatial grid used for Cross-validation.

Note

By default uses oblique coordinates (rotated coordinates) as described in Moller et al. (2020; doi: 10.5194/soil-6-269-2020) to account for geographical distribution of values. By setting the nearest = TRUE, distances to nearest observations and values of nearest neighbors will be used (see: Sekulic et al, 2020; doi: 10.3390/rs12101687). This method closely resembles geostatistical interpolators such as kriging. Buffer geographical distances can be added by setting buffer.dist=TRUE. Using either oblique coordinates and/or buffer distances is not recommended for point data set with distinct spatial clustering. Effects of adding geographical distances into modeling are explained in detail in Hengl et al. (2018; doi: 10.7717/peerj.5518) and Sekulic et al. (2020; doi: 10.3390/rs12101687). Default learners used for regression are: c("regr.ranger", "regr.ksvm", "regr.nnet", "regr.cvglmnet"). Default learners used for classification / binomial variables are: c("classif.ranger", "classif.svm", "classif.multinom"), with predict.type="prob". When using method = "stack.cv" each training and prediction round could produce somewhat different results due to randomization of CV. Prediction errors are derived by default using the forestError package method described in Lu & Hardin (2021). Optionally, the quantreg (Quantile Regression) option from the ranger package (Meinshausen, 2006) can also be used.

Author(s)

Tom Hengl

References

• Moller, A. B., Beucher, A. M., Pouladi, N., and Greve, M. H. (2020). Oblique geographic coordinates as covariates for digital soil mapping. SOIL, 6, 269–289. doi: 10.5194/soil-6-269-2020

• Hengl, T., Nussbaum, M., Wright, M. N., Heuvelink, G. B., and Graler, B. (2018) Random Forest as a generic framework for predictive modeling of spatial and spatio-temporal variables. PeerJ 6:e5518. doi: 10.7717/peerj.5518

• Lu, B., & Hardin, J. (2021). A unified framework for random forest prediction error estimation. Journal of Machine Learning Research, 22(8), 1–41. https://jmlr.org/papers/v22/18-558.html

• Meinshausen, N. (2006). Quantile regression forests. Journal of Machine Learning Research, 7(Jun), 983–999. https://jmlr.org/papers/v7/meinshausen06a.html

• Sekulic, A., Kilibarda, M., Heuvelink, G. B., Nikolic, M. & Bajat, B. (2020). Random Forest Spatial Interpolation. Remote. Sens. 12, 1687, doi: 10.3390/rs12101687

Examples

library(rgdal)
library(mlr)
library(rpart)
library(nnet)
demo(meuse, echo=FALSE)
## Regression:
sl = c("regr.rpart", "regr.nnet", "regr.glm")
system.time( m <- train.spLearner(meuse["lead"],
      covariates=meuse.grid[,c("dist","ffreq")],
      oblique.coords = FALSE, lambda=0,
      parallel=FALSE, SL.library=sl) )
summary(m@spModel$learner.model$super.model$learner.model)
## Not run: 
library(plotKML)
## regression-matrix:
str(m@vgmModel$observations@data)
meuse.y <- predict(m, error.type="weighted.sd")
plot(raster::raster(meuse.y$pred["response"]), col=plotKML::R_pal[["rainbow_75"]][4:20],
   main="Predictions spLearner", axes=FALSE, box=FALSE)

## Regression with default settings:
m <- train.spLearner(meuse["zinc"], covariates=meuse.grid[,c("dist","ffreq")],
        parallel=FALSE, lambda = 0)
## Ensemble model (meta-learner):
summary(m@spModel$learner.model$super.model$learner.model)
meuse.y <- predict(m)
## Plot of predictions and prediction error (RMSPE)
op <- par(mfrow=c(1,2), oma=c(0,0,0,1), mar=c(0,0,4,3))
plot(raster::raster(meuse.y$pred["response"]), col=plotKML::R_pal[["rainbow_75"]][4:20],
   main="Predictions spLearner", axes=FALSE, box=FALSE)
points(meuse, pch="+")
plot(raster::raster(meuse.y$pred["model.error"]), col=rev(bpy.colors()),
   main="Prediction errors", axes=FALSE, box=FALSE)
points(meuse, pch="+")
par(op)
while (!is.null(dev.list())) dev.off()
## Plot of prediction intervals:
pts = list("sp.points", meuse, pch = "+", col="black")
spplot(meuse.y$pred[,c("q.lwr","q.upr")], col.regions=plotKML::R_pal[["rainbow_75"]][4:20],
   sp.layout = list(pts),
   main="Prediction intervals (alpha = 0.318)")
while (!is.null(dev.list())) dev.off()

## Method from https://doi.org/10.3390/rs12101687
#library(meteo)
mN <- train.spLearner(meuse["zinc"], covariates=meuse.grid[,c("dist","ffreq")],
        parallel=FALSE, lambda=0, nearest=TRUE)
meuse.N <- predict(mN)
## Plot of predictions and prediction error (RMSPE)
op <- par(mfrow=c(1,2), oma=c(0,0,0,1), mar=c(0,0,4,3))
plot(raster::raster(meuse.N$pred["response"]), col=plotKML::R_pal[["rainbow_75"]][4:20],
   main="Predictions spLearner meteo::near.obs", axes=FALSE, box=FALSE)
points(meuse, pch="+")
plot(raster::raster(meuse.N$pred["model.error"]), col=rev(bpy.colors()),
   main="Prediction errors", axes=FALSE, box=FALSE)
points(meuse, pch="+")
par(op)
while (!is.null(dev.list())) dev.off()

## Classification:
SL.library <- c("classif.ranger", "classif.xgboost", "classif.nnTrain")
mC <- train.spLearner(meuse["soil"], covariates=meuse.grid[,c("dist","ffreq")],
   SL.library = SL.library, super.learner = "classif.glmnet", parallel=FALSE)
meuse.soil <- predict(mC)
spplot(meuse.soil$pred[grep("prob.", names(meuse.soil$pred))],
        col.regions=plotKML::SAGA_pal[["SG_COLORS_YELLOW_RED"]], zlim=c(0,1))
spplot(meuse.soil$pred[grep("error.", names(meuse.soil$pred))],
         col.regions=rev(bpy.colors()))

## SIC1997
data("sic1997")
X <- sic1997$swiss1km[c("CHELSA_rainfall","DEM")]
mR <- train.spLearner(sic1997$daily.rainfall, covariates=X, lambda=1,
         nearest = TRUE, parallel=FALSE)
summary(mR@spModel$learner.model$super.model$learner.model)
rainfall1km <- predict(mR, what="mspe")
op <- par(mfrow=c(1,2), oma=c(0,0,0,1), mar=c(0,0,4,3))
plot(raster::raster(rainfall1km$pred["response"]), col=plotKML::R_pal[["rainbow_75"]][4:20],
    main="Predictions spLearner", axes=FALSE, box=FALSE)
points(sic1997$daily.rainfall, pch="+")
plot(raster::raster(rainfall1km$pred["model.error"]), col=rev(bpy.colors()),
    main="Prediction errors", axes=FALSE, box=FALSE)
points(sic1997$daily.rainfall, pch="+")
par(op)
while (!is.null(dev.list())) dev.off()

## Ebergotzen data set
data(eberg_grid)
gridded(eberg_grid) <- ~x+y
proj4string(eberg_grid) <- CRS("+init=epsg:31467")
data(eberg)
eb.s <- sample.int(nrow(eberg), 1400)
eberg <- eberg[eb.s,]
coordinates(eberg) <- ~X+Y
proj4string(eberg) <- CRS("+init=epsg:31467")
## Binomial variable
summary(eberg$TAXGRSC)
eberg$Parabraunerde <- ifelse(eberg$TAXGRSC=="Parabraunerde", 1, 0)
X <- eberg_grid[c("PRMGEO6","DEMSRT6","TWISRT6","TIRAST6")]
mB <- train.spLearner(eberg["Parabraunerde"], covariates=X,
   family=binomial(), cov.model = "nugget", parallel=FALSE)
eberg.Parabraunerde <- predict(mB)
plot(raster::raster(eberg.Parabraunerde$pred["prob.1"]),
   col=plotKML::SAGA_pal[["SG_COLORS_YELLOW_RED"]], zlim=c(0,1))
points(eberg["Parabraunerde"], pch="+")

## Factor variable:
data(eberg)
coordinates(eberg) <- ~X+Y
proj4string(eberg) <- CRS("+init=epsg:31467")
X <- eberg_grid[c("PRMGEO6","DEMSRT6","TWISRT6","TIRAST6")]
mF <- train.spLearner(eberg["TAXGRSC"], covariates=X, parallel=FALSE)
TAXGRSC <- predict(mF)
plot(raster::stack(TAXGRSC$pred[grep("prob.", names(TAXGRSC$pred))]),
    col=plotKML::SAGA_pal[["SG_COLORS_YELLOW_RED"]], zlim=c(0,1))
plot(raster::stack(TAXGRSC$pred[grep("error.", names(TAXGRSC$pred))]),
    col=plotKML::SAGA_pal[["SG_COLORS_YELLOW_BLUE"]], zlim=c(0,0.45))
while (!is.null(dev.list())) dev.off()

## End(Not run)

Envirometrix/landmap documentation built on June 10, 2022, 10:12 p.m.