add predict method for unsuperClass objects

NAMESPACE

@@ -4,6 +4,7 @@ S3method(fortify,RasterBrick)
 S3method(fortify,RasterLayer)
 S3method(fortify,RasterStack)
 S3method(predict,superClass)
+S3method(predict,unsuperClass)
 S3method(print,mapValidation)
 S3method(print,superClass)
 S3method(print,unsuperClass)

NEWS

@@ -4,10 +4,12 @@ New:
 
 * function `oneHotEncode`: splits a single rasterLayer into multiple layers (one per class)
   with one-hot encoding, i.e. pixel value = 1 for matches, pixel value = 0 for all other classes (background).
-* ggR() can now display more than one layer. Each layer can be plotted to a subplot in a multi-panel grafic.
+* `ggR()` can now display more than one layer. Each layer can be plotted to a subplot in a multi-panel grafic.
+* `encodeQA()`, `decodeQA()` and `classifyQA()` can now deal with the new QA format introduced with Landsat Collection data. Legacy QA designations can still be interpreted by setting the `legacy` argument.
+* new `predict()` method for unsupervised classification models (`unsuperClass`).
 
 Changes: 
-* all `radCor` DOS methods now work for Landsat 8 OLI 
+* all `radCor` DOS methods now work for Landsat 8 OLI
 
 Fixes:
 * fix `unsuperClass` for algorithms other than Hartigan-Wong (reported by Alex Ilich)

R/unsuperClass.R

@@ -90,6 +90,40 @@ unsuperClass <- function(img, nSamples = 10000, nClasses = 5, nStarts = 25, nIte
 
 
 
+
+#' Predict a raster map based on a unsuperClass model fit.
+#' 
+#' applies a kmeans cluster model to all pixels of a raster.
+#' Useful if you want to apply a kmeans model of scene A to scene B.
+#' 
+#' @method predict unsuperClass
+#' @param object unsuperClass object
+#' @param img Raster object. Layernames must correspond to layernames used to train the superClass model, i.e. layernames in the original raster image.
+#' @param ... Further arguments passed to writeRaster.
+#' @export 
+#' @examples 
+#' ## Load training data
+#' data(rlogo)
+#' 
+#' ## Perform unsupervised classification
+#' uc  <- unsuperClass(rlogo, nClasses = 10)
+#' 
+#' ## Apply the model to another raster
+#' map <- predict(uc, rlogo)
+predict.unsuperClass <- function(object, img, ...){
+  stopifnot(inherits(object, c("RStoolbox", "unsuperClass")))
+  model <- object$model
+  wrArgs <- list(...)
+   out   <- .paraRasterFun(img, rasterFun=raster::calc, args = list(fun=function(x, kmeans=force(model)){
+    if(!is.matrix(x)) x <- as.matrix(x)
+    predKmeansCpp(x, centers=kmeans$centers)}, forcefun=TRUE), wrArgs = wrArgs)
+
+return(out)
+}
+
+
+
+
 #' @method print unsuperClass
 #' @export 
 print.unsuperClass <- function(x, ...){

tests/testthat/test-unsuperClass.R

@@ -47,4 +47,15 @@ test_that("kmeans fail detection", {
             expect_warning(unsuperClass(lsat, nSamples = ncell(lsat), nStarts = 1, nClasses = 20), "doesn't converge properly")
         })
 
-
+## Predict S3 method
+test_that("predict.unSuperClass", {
+    skip_on_cran()
+    uc <- unsuperClass(lsat, nSamples = ncell(lsat), nClasses = 2)
+    expect_s4_class(pred <- predict(uc, lsat), "RasterLayer")
+    expect_equal(unique(uc$map - pred), 0)
+    tmpFile <- tempfile(fileext = ".grd")
+    pred <- predict(uc, lsat, filename = tmpFile )
+    expect_false(inMemory(pred))
+    expect_equal(filename(pred), tmpFile)
+    file.remove(tmpFile, gsub("grd", "gri", tmpFile))
+  })