Dear Naeem,

there is no need for sampling or transforming your raster data into vector format, or using expensive ArcGIS software.

Sampling is not good because you want to test correlation of entire rasters, rather than sparse pixels.

Instead, I recommend you to use open-source QGIS 2.18 (https://www.qgis.org/en/site/) with plugin called MOLUSCE (https://plugins.qgis.org/plugins/molusce/).

You should load entire rasters into QGIS, and load them into MOLUSCE as spatial variables. The next step is Evaluating correlation, where you can calculate Pearson's or Cramer's correlation coefficient or Joint information uncertainty. This document may help you with detailed instructions (https://github.com/nextgis/molusce/blob/master/doc/en/QuickHelp.pdf).

If some of the layers with spatial variables have high correlation (over 0.7), you should exclude one of them from your final model.

The other software is GRASS GIS (https://grass.osgeo.org/).

It has function r.covar that calculates correlation matrix between input rasters (https://grass.osgeo.org/grass70/manuals/r.covar.html). The computation time is much faster than in QGIS, but be aware that it is not user-friendly software.

Dear Naeem,

I usually do something like this:

1. Generate a suitable number of random sampling points covering the area (they can be further subdivided in train, test and validation subsets afterwards if needed).

2. Sample the raster layer stack on the random points with a command available in your GIS platform (such as "sample" in ArcGIS) to create an attribute table. Do not forget to add point ID, X, Y coordinates to the table to be able to map back results for geostatistical analysis)

3. Export your table to a convenient statistical package for collinearity testing. On that account, I would like to remember you that in case you want to study environmental variables role in some physical process, collinearity may sometimes mask some important effects. So, try to separate pure collinearity from independent effects. One method I use is to study relative parameter importance by using bootstrapping techniques (out-of-bag statistics in machine learning).

Hope this helps,

best

Filippo

I am assuming that each of the 20 layers have vector data (not raster). If their spatial extent is the same one can join all to get a large attribute table containing all the 20 variables. You may also use a Join method (or "Identity" operation). Then one can use Exploratory Regression (or GWR) that can check for colinearity and other limitations

Dear Naeem,

there is no need for sampling or transforming your raster data into vector format, or using expensive ArcGIS software.

Sampling is not good because you want to test correlation of entire rasters, rather than sparse pixels.

Instead, I recommend you to use open-source QGIS 2.18 (https://www.qgis.org/en/site/) with plugin called MOLUSCE (https://plugins.qgis.org/plugins/molusce/).

You should load entire rasters into QGIS, and load them into MOLUSCE as spatial variables. The next step is Evaluating correlation, where you can calculate Pearson's or Cramer's correlation coefficient or Joint information uncertainty. This document may help you with detailed instructions (https://github.com/nextgis/molusce/blob/master/doc/en/QuickHelp.pdf).

If some of the layers with spatial variables have high correlation (over 0.7), you should exclude one of them from your final model.

The other software is GRASS GIS (https://grass.osgeo.org/).

It has function r.covar that calculates correlation matrix between input rasters (https://grass.osgeo.org/grass70/manuals/r.covar.html). The computation time is much faster than in QGIS, but be aware that it is not user-friendly software.

Correlation can be calculated between rasters directly. Simplest approach is using grass toolbox in QGIS. Just search for r.covar in QGIS processing toolbox and follow the screen and pop up.