I have binary raster data (deforestation hotspot or not) that I would like to plot spatial correlograms for in R, to look at spatial autocorrelation over different spatial lags. I came across this website explaining several packages that would plot spatial correlograms, but they all seem to be for continuous data (to calculate Moran's I for spatial autocorrelation).

I have calculated spatial autocorrelation for my binary data using the joincount.test function in the spdep package, but cannot find any package that allows plotting of spatial correlograms using jointcount. Does anyone know of any?


What is the hypothesis of spatial process that you are testing? I am dubious over the logic of your analysis. What you describe is inherently a meta-model problem where error and spurious spatial process is propagated, in unknown forms, through each step. This makes it impossible to quantify any associated significance. If this raster is the result of a local autocorrelation statistic (eg., LISA), applied to a change detection, then what you are proposing is completely invalid.

Why does this need to be a binary problem? If you calculate a multiscale evaluation of proportional change you can perform a valid global or local Geary's-C or Moran's-I (LISA).

I have never seen a joins count correlogram and am not sure if there is a testable hypothesis supported by this proposed test statistic. It would also be computationally very expensive over large spatial lags. To test significance you would have to derive an expected null following a Poisson CSR process for each spatial lag. For a raster, I am not even sure that this is possible in the R environment.

  • I'm trying to determine what's an appropriate neighbourhood distance to calculate the distance-weighted autocovariate to input into an auto-log model. I was intending to use the spatial correlogram to see what the distance over which spatial autocorrelation occurs is. I thought that joincount gives a spatial autocorrelation value similar to Moran's I but for binary data, so it can be similarly used for correlograms. – Jocelyne Sze May 26 '16 at 14:35
  • It's binary because in my prior steps, I determined what cells are considered hotspots or not based on the deforestation rate (above/below a certain threshold value). – Jocelyne Sze May 26 '16 at 14:52
  • 1
    I can tell you now that this is an autocorrelated process. How can it not be? Deforestation happens across discrete areas, therefore a joins count should tell you that the process is clustered from random. I think that it may be prudent to start exploring landscape metrics to quantify spatial process. This is why I asked what hypothesis you are testing. Without this, I cannot recommend candidate metrics. You should always have your hypothesis/question at the forefront because it invariably defines the appropriate method, not the other way around. – Jeffrey Evans May 26 '16 at 16:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.