RAT and cell values in R's raster-objects

Question

I really struggle to understand some basic concepts in the structure of how R's raster-objects save values. Namely the RasterLayer, the RasterBrick and the RasterStack.

I understand that the values (when manually creating a raster) can just be of type numeric, integer, logical or factor. When I create a raster with numeric values like this:

raster_numeric = raster(nrows = 6, ncols = 6, res = 0.5, 
               xmn = -1.5, xmx = 1.5, ymn = -1.5, ymx = 1.5,
               vals = seq(0.1,3.6,0.1))

I get this structure:

class      : RasterLayer 
dimensions : 6, 6, 36  (nrow, ncol, ncell)
resolution : 0.5, 0.5  (x, y)
extent     : -1.5, 1.5, -1.5, 1.5  (xmin, xmax, ymin, ymax)
crs        : +proj=longlat +datum=WGS84 +ellps=WGS84 +towgs84=0,0,0 
source     : memory
names      : layer 
values     : 0.1, 3.6  (min, max)

And apparently no attributes-slot.

Whereas, when I use factorial values (copied from the amazing Lovelace et al. book) I get, what to my understanding is a Raster-Attribute-Table (RAT). So why is this?

class      : RasterLayer 
dimensions : 6, 6, 36  (nrow, ncol, ncell)
resolution : 0.5, 0.5  (x, y)
extent     : -1.5, 1.5, -1.5, 1.5  (xmin, xmax, ymin, ymax)
crs        : +proj=longlat +datum=WGS84 +ellps=WGS84 +towgs84=0,0,0 
source     : memory
names      : layer 
values     : 1, 3  (min, max)
attributes :
 ID VALUE
  1  clay
  2  silt
  3  sand

What exactly is this RAT and can raster-objects really save just one value because in the end each cell can be references to a row in the RAT and thereby hold many values of different type?

Answer 1

A RAT is just a table that crosswalks the numeric values contained in an integer raster to descriptive attributes. It is not realistic to have a RAT for a true floating-point raster (eg., elevation) and really does not make much sense to. If one wants to represent a process as nominal than you just reclassify it into the desired ranges.

A good way to understand Raster Attribute Tables (RAT) is to work through creating one. The ID represent the unique values in a given raster and are fixed but, you can have multiple columns representing different attributes. In this example we create a raster with four values, ratify the raster and then define a few attributes that may emulate attribution between say, landcover a soil type that could result from a combination between two rasters representing these two nominal class types.

library(raster)
soil <- raster(extent(571823, 616763, 4423540, 4453690), resolution=100)
  soil[] <- sample(c(1:4), ncell(soil), replace = TRUE)
    soil

Here we ratify our soil raster, pull the resulting data.frame, add some attributes to it and then add it back to the data.

( soil <- ratify(soil) )    # create a raster with a RAT
  rat <- levels(soil)[[1]]  # pull the data.frame from the raster RAT

# modify data.frame object
  rat$covertype <- c("riparian", "wetland", "wetland", "riparian")
    rat$soil <- c("clay", "silt", "sand", "sand")

# Write data.frame back to raster RAT and print results
      levels(soil) <- rat
soil

Honestly, other than for plotting, I find RAT rasters of very limited use. You can accomplish the same types of analysis by just leveraging a stand alone data.frame object (that is not buried in the raster) to query the values of the raster.

Answer 2

"each cell can be references to a row" - no, in a single Raster Layer each cell is a reference to exactly one row in the attribute table. So in your example each cell is either clay, silt, or sand (or possible NA missing data if not in the corresponding numeric range).