First, determine the fraction of the total area that each class takes up.
This is best done by just evaluating the total number of pixels with the given value, easily done iteratively. Perhaps by just reading all pixels into a long vector and subsetting that:
inputRaster <- raster(result_from_classification)
classifed_values <- values(inputRaster)
data_from_1s <- classified_values[which(classified_values==1)
number_of_1s <- NROW(data_from_1s)
Second, you need to determine the number of samples you want per class and extract that number of samples from the class:
number_of_samples <- 10000
number_of_pixels <- 3750000
number_of_samples_of_1s <- number_of_1s / number_of_pixels * number_of_samples
representative_sample_of_1s <- sample( data_from_1s , number_of_samples_of_1s )
All the code above has not been tested for errors, but should be fairly close to functioning.