Remote Sensing - Lab 5 Supervised classification by slappypappy123


									                                Remote Sensing - Lab 5
                               Supervised classification

From this lab you will learn how to perform supervised classification. This is performed when we
know something about the landscape that we are trying to classify. Typically we classify it into
land-cover categories (not to be confused with land-use).

There are 2 major types of classification:
   • Supervised classification −We know what our classes are to start with
   • Unsupervised classification − we let the computer choose spectrally similar clusters for us

What to submit for this class
        1. A supervised classified image, with annotations, including a legend (this will
           be the first time you create a legend), for the upper Cuyahoga dataset.

        2. A table with three columns: Column 1 should contain “class”, Column 2 should
           contain the # of pixels in that class, and Column 3 will represent the number of
           pixels as an area in km2. Most of this will be given to you by the computer once
           you perform the classification.

For this lab we will use the same dataset used for LAB 3. This is one of the most recent
datasets that we have for the region. Note that the instructions below use a different dataset,
however you need to learn to use different datasets than those given in the instructions.

Adding layers to perform the classification

Presently, there are 7 raster layers in your dataset which represent the 7 Landsat bands. In order
to perform the classification we need to add two new layers to the dataset, one to “train” the
dataset”, and one to output the classification on.

Open the file you wish to classify in FOCUS. Switch the left frame to “files” view, right click on the
file name and go to “new”- “raster layer” then “add” 2 new raster layers.

Add 2 8 bit layers.

You should now have a total of 9 layers in your pix file, although the two layers that you added
will be empty. Switch back to “Maps”, in the left frame. Right click on either the red, green or blue
channel. You will see that there are now 2 additional empty layers.

    2 new empty channels

Setting up the classification
From the Analysis drop-down menu choose “image classification” and then “supervised”.

A new window will open asking you open a file. Choose your downloaded file.

Again, a new window will open. Choose “new session”.

A new image classification window will open. The red, green and blue columns can be set to
anything you want as this is for display purposes only. You can change these at any point in
time. Below, I changed the display to a 432 image.

Input channels are those that you will use to perform the classification. For this exercise we will
use all bands. Click on all of them as “input channels”. Choose the first empty layer, and click on
it to choose it as your “training channel”. Choose the last empty layer as the “output channel”,
meaning the results will ultimately be displayed here.

   Note that once you hit OK, your display will change to the RGB combination that you chose, a
  “new area” will appear in the left hand window. Below that is “Classification MetaLayer” which
  has all the classification information.

This MUST be highlighted
when entering training data

This is the RGB satellite
image that you opened in
the classification process

When the classification is
complete, it will be placed
on this layer

            This is the original satellite image that you
            had open when you added your layers. It
            is the same dataset as above but may be
            displayed differently than your RGB setting
            in the classification window. Click it off
            and you’ll see your RGB satellite image

Adding classes
Meanwhile, a second window should have also opened. Under the “class” drop down menu,
choose “new”, which will allow to add “classes” to the dataset.

Add 6 new classes. Our classes are: 1. Water; 2. Grass; 3. Trees; 4. Urban; 5. Other; and 6. No

To change, for example “Class-01” to “Water” simply click on the box and it will let you type over
what is already there. Likewise, to change the color, simply click on it and a choice will be
provided for you. Change the colors to something that makes sense (e.g. Blue for water; green
for grass etc.)

Training the dataset
You are ready to start training the dataset. To do this, choose the “vector edit tool” from the icons
in FOCUS. Here you can choose lines, polygons, rectangles etc. to highlight areas of the image
and assign them to your chosen class.

                                                                 You can use any of these to
                                                                  help identify areas on the
                                                                  satellite image for training

An example: The Kent area has lots of trees, so I am going to use that to train the program to
recognize trees. Note: the further you zoom in, the easier it is to see things.

         Choose the appropriate
      selection type for highlighting
       the training area. Make sure
         the button is depressed

                  Training areas must be chosen
                     at all times while training

                                                                                    Draw shapes
                                                                                   directly on the
                                                                                 features you know
                                                                                    belong to the
   The class you are training                                                       class you are
   must be chosen here (e.g.                                                        trying to map
       trees in this case)

Continue to choose areas on the dataset that you know fall within the class categories (usually
you will use GPS coordinates and ground truthing information).

A very useful tool is the “raster seeding” tool found in the polygon window. This allows the
choosing of a particular pixel. The computer then will assign the same class to all surrounding
pixels that have similar characteristics.

                                                 Example of raster seeding.

Enter training information for EVERY class.

Previewing the classification
From the “tools” menu choose “classification preview”, then one of the algorithms to preview the
data. Maximum likelihood is the most useful algorithm to use.

This will show a preview of the dataset.

Cross-check your classification with the underlying satellite image carefully by zooming in, and
clicking off and on the classified layer. It should become pretty obvious fairly quickly what
represents roads, what represents fields etc. If it does not look correct, add more training
information. Note, sometimes you may have chosen too large of an area during training, and
sometimes you need to go back and erase an area. You can do this by choosing the erase tool,
and highlighting the areas you want to erase.


Running the classification

If you are satisfied with what you see in your first or any subsequent preview, then right click on
the “classification metalayer” and run the classification. Choose, maximum likelihood, and ensure
“show report” is clicked on.

When you run the classification you will also generate a table of values. Save the table as a .txt
file and use it to answer Q 2 on p. 1 of this lab.

  Note: when between sessions, make sure you
  save this project, as this will maintain all your
               training information.
To test the accuracy of the classification you can generate a random sample of points from the
post-classification analysis window. You would then identify whether these points belong to the
correct class or not.

              Some additional useful information
                  You do not need to do this specifically for this lab.

Performing a simple accuracy assessment

Choose accuracy assessment

The following window will open – choose “select classified image”.

Make sure you choose your classified layer

A new window will open that contains your classes. Now choose “Generate random sample”

Choose 50 as your number of samples.

Again, now 50 samples will be generated. As you click on each number generated the point will
be shown on your image.

Using the actual satellite image (not the classification) identify the landcover at that point. Simply
click on the landcover type in the left frame and transfer it over to the right frame. Make sure all
50 points have a landcover type assigned to them.

One complete, choose Accuracy assessment and “generate report”

Choose “save report”.

To reopen your classification if you close your project down
    1. Save the entire project as a GPR file, then open that; or
    2. From the Analysis / Image Classification/ Supervised menu, choose the file you are
       working on, and choose your training channel rather than <new session>.

This will bring you back to your training layer.


To top