How to randomly stratify sampling within landcover classes?

Question

I'm working on an accuracy assessment of a classified landsat image and I'd like input on the best possible way to generate random points across both the classified image and a ground truth image I created by processing vector data from aerial photo interpretation.

I've processed both images and converted them to vectors in R and would like to generate random points in specific landcovers across both images so I can compare the predicted vs. actual landcover value at each random site.

My methodology so far has been to compare both vectors (in their entirety) and create a confusion matrix on them but the images this would include the training pixels in the supervised classification so I'd rather not have them as part of the accuracy assessment. A smaller subset of assessment pixels is probably best but I'd like feedback on those who have done this in R in the past as to best practices.

What approach should I take to generate random pixels in each image so I can compare the landcover values?

Answer 1

It s best to avoid converting to vector to do the sampling if you do not have to. You can use the sampleStratified() function in the raster package to generate randomly stratified points in your land cover classes. For example:

library(raster)

# Generate some sample landcover raster data
r <- raster(ncol=10, nrow=10)
names(r) <- 'stratum'
values(r) <- round((runif(ncell(r))+0.5)*3)

# Randomly sample within each class (do not sample in NA)
s <- sampleStratified(r, size=3, na.rm=TRUE)

# Convert cell number to XY coords
xy <- xyFromCell(r,s) 

# Plot data
plot(r, col = topo.colors(3))
points(xy)