[GIS] Use QGIS to identify and count pixel colour in a photo

qgis

I am trying to use QGIS to analyse the colour component of a photo. I have imported the photo as a raster layer, created a polygon of the area in question and have tried to use zonal statistics but this doesn't appear to work. When I use the zonal statistics function, I fill in all the relevant boxes and click ok. Then the message window closes. That's it. When I have tried using the function on maps I already have, after clicking ok the progress bar came up and after it finished I could find the resulting stats. When I tried it on my photo, absolutely nothing happens. The window just closes, there are no stats anywhere.

Ideally, the stats I would like to produce are the number of pixels in the polygon, and the number of pixels of each colour. Or, to assign the colours into similar groups say 0-25, 25-50, 50-75 etc up to 255. Of course, this is made more challenging by the fact that a photo used as a raster layer is made up of 3 colour bands and, as yet, I haven't managed to separate the bands.

Best Answer

You want to count the number of discrete pixel colours inside an area? As far as I know, that can't be done with zonal statistics.

e.g.

Polygon 1 - area is 1000 pixels
500 pixels are (122,135,21)
400 pixels are (22,132,178)
100 pixels are (2,156,99)

Doing that on an raw (unclassified) rgb image will give a HUGE number of possible colours. A polygon with an area of 1000 pixels could easily have 900 unique colours, for example. Most pixel colours will appear only once, and a few will appear twice.

A python/gdal script could be written to count unique colours, but it would be slow and memory-intensive for that very reason.

You can get around this by using a paletted image to "merge" similar colours into a single value. This means that greens would all appear as one value, greys another value, and so on.

(If you're familiar with Photoshop/GIMP, this is the same as reducing an image to a fixed palette size e.g. 16 colours, where each pixel is 'rounded' to the nearest colour from a representative palette of 16 colours)

Try "convert rgb to paletted" (raster > conversion > rgb to pct). More info here

This will classify pixels into groups of similar colour.

Now the problem is easier to address.

I took an aerial image and reduced it to 4 colours.

Then I ran gdalinfo to get a histogram of the values. In Linux, use the terminal, in Windows, use the OSGeo4W shell.

gdalinfo -hist /path/to/my.tiff

This gave me

313730340 154657 11795972 160083079 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

So there are 313,730,340 pixels with value 0. And only 154,657 with value 1.

You'll need to work out which value equates to which colour, though.

There's also a K-Means raster classifier in Orfeo Toolbox, this does a simplification of colours but uses a different algorithm.

Related Question