The Chi-square test is applied when you have two or more categorical variables from a single population. It is used to determine whether there is a significant association between the variables. So I think the test is suitable in your study. But imo not run the test step by step. Analyse all of the data in one big file so that it reveals all interactions :)

I agree with the previous respondent's answer that this an appropriate test. The best explanation for a Chi-squared analysis that I have heard came from my Stats Prof. Dr. Alan Agresti. His textbook on Statistical Methods for Social Sciences was a great entry level book, and another one, Categorial Data Analysis, covers the same material but at a higher level. The way I usually approach these kids of questions is to carefully read through his explanation of the statistic in the book, revise the published example (SAS) code, and substitute my data. While examples are written for SAS, they are transferable to R (or any other software) once you understand the basic parameters of the test.

The problem with using a chi-square test of association in this case is that there isn't a natural way to count the land unit areas. That is, if your mapping is on 30m x 30m grid, you would get different counts than if it were on a 50m x 50m grid.

If you were to count up all the pixels and do a chi-square test, you would likely get a statistically significant result, even if the differences as a percentage were quite small. This is because there will be many pixels. This makes the p-value from the analysis of little practical value.

A few things to think about:

1) If you are looking at more than a few time periods, you might think about the time periods being ordered. That is, the more interesting thing may be the change in each land use category as a percent over time.

2) If you use something like a chi-square test of association, the p-value will probably have little value because the high counts of pixels. But an effect size measurement won't be affected by sample size. You might look at Cramér's V as way to gauge the extend of the differences.

Say your raster files (TIF or whatever the format is) are in 90*90 cell size

1. Create a surface reflectance raster file

2. Load the raster file in a convenient programming API (matlab, python, IDL whatever you're working with)

2. Mask the raster for each land use (clip the raster file using your each land cover class say, vegetation, as a mask)

3. For each land use,

-- Extract all representative pixel values (the surface reflectance values for example) for each time

4. Say you have two files for two timesteps, A11(t=1, landuse 1), and A21(t=2, landuse 1), then in Matlab you can extract the grid values (each pixel as,

X = A11(:); Y = A21(:)

5. Now you have two 1-D arrays stored in variables X and Y with values of each corresponding pixels. You can run the ChiSquare test with it. You can also perform a regression to find the p-value to see if their correlation is statistically significant.

I hope this helped.

Roland,

I see no reason why Chi-squared test cannot be used. The data you wish to collect is nominal data. The problem is the sampling procedure that would suit best for extracting your data. Do you apply systematic or random sampling. That depends largely on the nature of the spatial patterns present on the maps. Probably a method of random sampling would be preferable.

Certainly relate land use/types specific to similar time sets. Once data has been gathered and tabulated for running the Chi test, set up your null hypothesis, I.e. that there is NO relationship between . . . While composite tests can be done, I think testing pair by pair is better. Composite testing can obscure the main relationship.

Having said that, I am not sure but that other graphic and/or statistical methods may prove more useful, e.g. trend graphs or histograms. They require more interpretation than Chi values but can give much more useful information, especially concerning changes over time.

George