Supervised Classification

Supervised classification both for classification and regression mode based on vector training data (points or polygons).

Usage

superClass(
  img,
  trainData,
  valData = NULL,
  responseCol = NULL,
  nSamples = 1000,
  nSamplesV = 1000,
  polygonBasedCV = FALSE,
  trainPartition = NULL,
  model = "rf",
  tuneLength = 3,
  kfold = 5,
  sampling = NULL,
  minDist = 2,
  mode = "classification",
  predict = TRUE,
  predType = "raw",
  filename = NULL,
  verbose,
  overwrite = TRUE,
  ...
)

Arguments

img: SpatRaster. Typically remote sensing imagery, which is to be classified.
trainData: sf or sp spatial vector data containing the training locations (POINTs,or POLYGONs).
valData: Ssf or sp spatial vector data containing the validation locations (POINTs,or POLYGONs) (optional).
responseCol: Character or integer giving the column in trainData, which contains the response variable. Can be omitted, when trainData has only one column.
nSamples: Integer. Number of samples per land cover class. If NULL all pixels covered by training polygons are used (memory intensive!). Ignored if trainData consists of POINTs.
nSamplesV: Integer. Number of validation samples per land cover class. If NULL all pixels covered by validation polygons are used (memory intensive!). Ignored if valData consists of POINTs.
polygonBasedCV: Logical. If TRUE model tuning during cross-validation is conducted on a per-polygon basis. Use this to deal with overfitting issues. Does not affect training data supplied as SpatialPointsDataFrames.
trainPartition: Numeric. Partition (polygon based) of trainData that goes into the training data set between zero and one. Ignored if valData is provided.
model: Character. Which model to use. See train for options. Defaults to randomForest ('rf'). In addition to the standard caret models, a maximum likelihood classification is available via model = 'mlc'.
tuneLength: Integer. Number of levels for each tuning parameter (see train for details).
kfold: Integer. Number of cross-validation resamples during model tuning.
sampling: Character. Describes the type of additional sampling that is conducted after resampling (usually to resolve class imbalances), from caret. Currently supported are up, down, smote, and rose. Note, that smote requires the packages themis and rose the package ROSE. Latter is noly for binary classification problems.
minDist: Numeric. Minumum distance between training and validation data, e.g. minDist=1 clips validation polygons to ensure a minimal distance of one pixel (pixel size according to img) to the next training polygon. Requires all data to carry valid projection information.
mode: Character. Model type: 'regression' or 'classification'.
predict: Logical. Produce a map (TRUE, default) or only fit and validate the model (FALSE).
predType: Character. Type of the final output raster. Either "raw" for class predictions or "prob" for class probabilities. Class probabilities are not available for all classification models (predict.train).
filename: Path to output file (optional). If NULL, standard raster handling will apply, i.e. storage either in memory or in the raster temp directory.
verbose: Logical. prints progress and statistics during execution
overwrite: logical. Overwrite spatial prediction raster if it already exists.
...: further arguments to be passed to train

Value

A superClass object (effectively a list) containing:

$model: the fitted model
$modelFit: model fit statistics
$training: indexes of samples used for training
$validation: list of
1. $performance: performance estimates based on independent validation (confusion matrix etc.)
2. $validationSamples: actual pixel coordinates plus reference and predicted values used for validation
3. $validationGeometry: validation polygpns (clipped with mindist to training geometries)
$map: the predicted raster
$classMapping: a data.frame containing an integer <-> label mapping

Details

Note that superClass automatically loads the lattice and randomForest package. SuperClass performs the following steps:

Ensure non-overlap between training and validation data. This is neccesary to avoid biased performance estimates. A minimum distance (minDist) in pixels can be provided to enforce a given distance between training and validation data.
Sample training coordinates. If trainData (and valData if present) are polygons superClass will calculate the area per polygon and sample nSamples locations per class within these polygons. The number of samples per individual polygon scales with the polygon area, i.e. the bigger the polygon, the more samples.
Split training/validation If valData was provided (reccomended) the samples from these polygons will be held-out and not used for model fitting but only for validation. If trainPartition is provided the trainingPolygons will be divided into training polygons and validation polygons.
Extract raster data The predictor values on the sample pixels are extracted from img
Fit the model. Using caret::train on the sampled training data the model will be fit, including parameter tuning (tuneLength) in kfold cross-validation. polygonBasedCV=TRUE will define cross-validation folds based on polygons (reccomended) otherwise it will be performed on a per-pixel basis.
Predict the classes of all pixels in img based on the final model.
Validate the model with the independent validation data.

Examples

library(RStoolbox)
library(caret)
library(randomForest)
#> randomForest 4.7-1.2
#> Type rfNews() to see new features/changes/bug fixes.
#> 
#> Attaching package: ‘randomForest’
#> The following object is masked from ‘package:gridExtra’:
#> 
#>     combine
#> The following object is masked from ‘package:ggplot2’:
#> 
#>     margin
library(e1071)
#> 
#> Attaching package: ‘e1071’
#> The following object is masked from ‘package:ggplot2’:
#> 
#>     element
#> The following object is masked from ‘package:terra’:
#> 
#>     interpolate
library(terra)
train <- readRDS(system.file("external/trainingPoints_rlogo.rds", package="RStoolbox"))

## Plot training data
olpar <- par(no.readonly = TRUE) # back-up par
par(mfrow=c(1,2))
colors <- c("yellow", "green", "deeppink")
plotRGB(rlogo)
plot(train, add = TRUE, col =  colors[train$class], pch = 19)

## Fit classifier (splitting training into 70\% training data, 30\% validation data)
SC       <- superClass(rlogo, trainData = train, responseCol = "class",
model = "rf", tuneLength = 1, trainPartition = 0.7)
#> 09:46:58 | Begin sampling training data
#> 09:46:58 | Starting to fit model
#> 09:46:58 | Starting spatial predict
#> 09:46:58 | Begin validation
#> ******************** Model summary ********************
#> Random Forest 
#> 
#> 21 samples
#>  3 predictor
#>  3 classes: 'A', 'B', 'C' 
#> 
#> No pre-processing
#> Resampling: Cross-Validated (5 fold) 
#> Summary of sample sizes: 16, 18, 15, 17, 18 
#> Resampling results:
#> 
#>   Accuracy  Kappa
#>   1         1    
#> 
#> Tuning parameter 'mtry' was held constant at a value of 1
#> [[1]]
#>   TrainAccuracy TrainKappa method
#> 1             1          1     rf
#> 
#> [[2]]
#> Cross-Validated (5 fold) Confusion Matrix 
#> 
#> (entries are average cell counts across resamples)
#>  
#>           Reference
#> Prediction   A   B   C
#>          A 1.4 0.0 0.0
#>          B 0.0 1.4 0.0
#>          C 0.0 0.0 1.4
#>                        
#>  Accuracy (average) : 1
#> 
#> 
#> ******************** Validation summary ********************
#> Confusion Matrix and Statistics
#> 
#>           Reference
#> Prediction A B C
#>          A 3 0 0
#>          B 0 3 0
#>          C 0 0 3
#> 
#> Overall Statistics
#>                                      
#>                Accuracy : 1          
#>                  95% CI : (0.6637, 1)
#>     No Information Rate : 0.3333     
#>     P-Value [Acc > NIR] : 5.081e-05  
#>                                      
#>                   Kappa : 1          
#>                                      
#>  Mcnemar's Test P-Value : NA         
#> 
#> Statistics by Class:
#> 
#>                      Class: A Class: B Class: C
#> Sensitivity            1.0000   1.0000   1.0000
#> Specificity            1.0000   1.0000   1.0000
#> Pos Pred Value         1.0000   1.0000   1.0000
#> Neg Pred Value         1.0000   1.0000   1.0000
#> Prevalence             0.3333   0.3333   0.3333
#> Detection Rate         0.3333   0.3333   0.3333
#> Detection Prevalence   0.3333   0.3333   0.3333
#> Balanced Accuracy      1.0000   1.0000   1.0000
SC
#> superClass results
#> ************ Validation **************
#> $validation
#> Confusion Matrix and Statistics
#> 
#>           Reference
#> Prediction A B C
#>          A 3 0 0
#>          B 0 3 0
#>          C 0 0 3
#> 
#> Overall Statistics
#>                                      
#>                Accuracy : 1          
#>                  95% CI : (0.6637, 1)
#>     No Information Rate : 0.3333     
#>     P-Value [Acc > NIR] : 5.081e-05  
#>                                      
#>                   Kappa : 1          
#>                                      
#>  Mcnemar's Test P-Value : NA         
#> 
#> Statistics by Class:
#> 
#>                      Class: A Class: B Class: C
#> Sensitivity            1.0000   1.0000   1.0000
#> Specificity            1.0000   1.0000   1.0000
#> Pos Pred Value         1.0000   1.0000   1.0000
#> Neg Pred Value         1.0000   1.0000   1.0000
#> Prevalence             0.3333   0.3333   0.3333
#> Detection Rate         0.3333   0.3333   0.3333
#> Detection Prevalence   0.3333   0.3333   0.3333
#> Balanced Accuracy      1.0000   1.0000   1.0000
#> 
#> *************** Map ******************
#> $map
#> class       : SpatRaster 
#> size        : 77, 101, 1  (nrow, ncol, nlyr)
#> resolution  : 1, 1  (x, y)
#> extent      : 0, 101, 0, 77  (xmin, xmax, ymin, ymax)
#> coord. ref. : +proj=merc +lon_0=0 +k=1 +x_0=0 +y_0=0 +ellps=WGS84 +towgs84=0,0,0,0,0,0,0 +units=m +no_defs 
#> source(s)   : memory
#> name        : class_supervised 
#> min value   :                1 
#> max value   :                3 

## Plots
plot(SC$map, col = colors, legend = FALSE, axes = FALSE, box = FALSE)
legend(1,1, legend = levels(train$class), fill = colors , title = "Classes", 
horiz = TRUE,  bty = "n")

par(olpar) # reset par

Usage

Arguments

Value

Details

See also

Examples