Skip to main content

Creating PCA images in QGIS

PCA image of Wyvern hyperspectral imagery

Image of Saudi Arabia agriculture: Natural colour on left, 3-band PCA on right

In this tutorial, we're going to use QGIS to apply Principal Component Analysis (PCA) to a Wyvern hyperspectral image.

PCA is a form of unsupervised learning. When applied to hyperspectral imagery, it creates a new set of bands (or components) which are:

  • unrelated to each other (orthogonal)
  • and ordered by the amount of image variation they explain

Because the bands are unrelated to each other, each band contains unique information about the original image. And because the bands are ordered by the image variation they explain, the first few bands will contain the most information about the image; you can focus your attention on those.

The end result is an image that can be used to reveal hidden correlations and similarities within the original image, and highlight areas for further exploration.

Before you start...​

Before starting this tutorial, please ensure you have done the following:

Note: Apple silicon devices may have trouble installing QGIS.

For best results, use a Windows or Linux (x86) machine, or follow the instructions on the QGIS website to install the OSX version.

Generating PCA images in QGIS​

Note: OTB and NODATA values

OTB does not currently have a simple way to handle NODATA values. Because of this, we're going to crop our raster to exclude NODATA values, and perform PCA on that. More info can be found in this OTB issue.

Cropping the image​

Ensure the image is selected in the Layers pane. Click on the Raster menu, then Extraction, then Clip raster by Extent. In the dialog box, click on the Clipping extent menu and select Draw on map canvas.

Clip raster by Extent dialog

Clip raster by Extent dialog

Draw on the map, being careful to stay within the borders of the displayed layer, then click Run. You should now see a new layer named Clipped (extent) with all 23 bands:

Clipped (extent) layer

Clipped (extent) layer

Generate the PCA​

Click on the Processing menu, then Toolbox. In the "Search..." bar, type DimensionalityReduction. You should see it listed under OTB -> Image Filtering.

QGIS Dimensionality Reduction Menu

QGIS Dimensionality Reduction Menu

Click on that, and you should see a new dialog window pop up:

QGIS Dimensionality Reduction Dialog

QGIS Dimensionality Reduction Dialog

  • For Input Image, ensure the Clipped (extent) layer is selected.

  • For Output image, click on the 3 dots and select the directory where you're keeping the original image. Otherwise, it will use a temporary file, and you may run out temp space depending on your computer. Name the file clipped_pca.tif for now.

  • For Inverse Output image, Output files containing eigenvalues and Transformation matrix, select Skip output.

Click Run. Within a minute or two, the processing should be done and you should see a new layer in the Layers window.

Now we can start to examine the data!

Different ways of viewing the PCA images​

RGB image​

By default, QGIS will select the first three bands (ie, the first three components, or eigenrasters) as the Red, Green and Blue bands. We've set the Min/ Max Value Settings to Mean +/- standard deviation to make the layer distinct, but you can experiment with these to see what works best for you.

PCA Multiband Colour

PCA Multiband Colour

You can also experiment with selecting different bands (ie, different eigenvectors) to see what this brings out in the image. Note that as you choose higher and higher bands, the effects will decrease -- you're effectively looking at lesser and lesser significant sources of information within the image.

Singleband pseudocolour​

As with the RGB preview, picking the correct single band pseudocolour for your application will require some exploration. Here, we're showing Band 1 (ie, the most significant eigenraster) in the Turbo colour ramp:

PCA Single Band Pseudocolour

PCA Single Band Pseudocolour

We can see a clear differentiation between not just the irrigated plots (blue), non-irrigated plots (yellow), and the empty plots (green), but the buildings showing up as smaller rectangular areas (red) within the plots at the top.

What does PCA show us?​

Now that we've found helpful ways to display the PCA bands, let's dig into what PCA might tell us. We'll focus on a few areas.

Striking similarities between different sections of the image​

Here we have a zoomed-in area showing fields, roads and paths:

Natural colour

PCA - Bands 1, 2, 3

Roads and fields: Natural colour vs PCA Bands 1, 2, 3

We'll highlight a few areas of similarity here from the PCA image:

  • The yellow areas in the PCA image consist of field sections and borders, roads, and irregular paths.

  • The purple border of the field in the top left quadrant matches field sections in the bottom right.

  • The greens in the three fields in the top left quadrant are quite close, despite being readily distinguishable in natural colour.

If we compare these parts of the PCA image to the natural colour image, these correspondences are either not apparent or not nearly as striking. They suggest a relationship between disparate areas that are not immediately visible in the original natural colour image.

Differences within similarly-coloured areas are much more apparent​

Here, we take a closer look at four adjacent fields. Pay attention to the one in the bottom left:

Natural colour

PCA - Bands 1, 2, 3

Roads and fields: Natural colour vs PCA Bands 1, 2, 3

PCA has brought out differences within that field, particularly the structure in the top half, that are not nearly as obvious in the original natural colour image.

Different PCA bands contain different information​

Finally, we take a look at another set of fields -- but this time, we are changing the bands we use. Below on the left, you can see the image composed of PCA Bands 1, 2 and 3; on the right, it's composed of PCA Bands 2, 3, and 4:

PCA - Bands 1, 2, 3

PCA - Bands 2, 3, 4

PCA: Bands 1, 2, 3 vs Bands 2, 3, 4

Here are a few things the change in bands bring out:

  • The structure of the borders between field sections, and the outer borders of the fields

  • Additional detail in the centre of each circular field

  • We lose some features in the rectangular field on the right but highlight others

This highlights the value of exploring different bands to see what each one reveals.

Note

PCA bands are ordered by how much information they contain about the original image. A common technique for simplifying PCA results is choosing to discard some of the lower bands, as they're "less important" than the higher ones. Choosing where to draw that line is often subjective, and outside the scope of this tutorial -- but we'll point you to scree plots and the elbow method to get started.

Next steps​

We've focused here on one image (Saudia Arabia agriculture), and on the first three components produced by PCA -- but that leaves much more to explore. Here are some questions to investigate:

  • What does the RGB image look like if you select a different set of bands (components) from the PCA layer? What differences are highlighted?

  • What happens as you select higher and higher bands from the PCA layer? What does this tell you about the image? How would you decide which bands to ignore?

  • Try looking at pseudocolour images of individual bands. How is this useful compared to an RGB image?

  • What happens if you pick another image from our open data program? What does PCA highlight if you switch from agriculture to urban, forest, mining, or maritime scenes?

Troubleshooting​

The first band is all one colour!​

This can happen if you run PCA on the entire original image, rather than selecting an extent and running PCA on that.

The reason is that our image contains NODATA values surrounding the image; when PCA sees that, it decides that the first and most important way the data varies is by whether it's NODATA or not.

I get an error saying there was a problem processing the image!​

Symptoms: a message that says The following layers were not correctly generated.

Check your tmp directory to see if it's filled up -- this may be left over from previous attempts to generate the PCA image. You should be able to remove the files within the directory without problems, and try again.

You can also try setting a different tmpdir for this process. Click on Advanced, then Parameters, and set Temporary Folder as needed.