GEOG 372: Introduction to Remote Sensing
Fall 2007
October 31st and November 1st
Laboratory 9
Image Classification
No
late labs can be accepted. Please
type the answers just below the
questions and hand the printouts at
the beginning of the next lab:
November 7th (0102) and Nov 8th
(0101).
Also
do not forget to mention your
MAJOR
on
your labs
The purpose of this
laboratory exercise is to learn the basic approaches to image
classification and compare the outputs for several image
classifications. In this lab we are going to work on a small subset of
the Landsat ETM+ scene located to the northeast of the city of
Baltimore. The image subset
is located in U:\\g372\372-fall07\lab9data. Load ENVI 4.2 and open
“subset” (the filename without an extension)
Load the NIR, red, and
green bands into RGB and load the display. Examine the image.
1.
What are the 3 main types of landcover presented in the image?
Read this entire paragraph before you start!
The first step in the image classification is to collect “samples”
representative of each land cover type. On the display toolbar go to
Tools – Region of Interest – ROI Tool. You will see a blank
region line in the “#1 ROI Tool” pop-up window. In the “ROI name” column
type in “urban” and hit “Enter” button on your keyboard. (If you don’t
hit the enter button the name will go back to the “Region #1”). Select
the radio button for “Zoom” above. Move the red square in the “scroll”
and “image” windows to select an urban area. In the “Zoom” window click
within the area you consider urban. Using the left click button on the
mouse create a perimeter around the urban area. Make sure to select only
pixels representative of urban areas, avoid pixels of other kinds! When
you are ready to close out your polygon right-click the mouse and you
will see a little diamond in the middle of your polygon. Right-click
again, and your polygon will become filled with red color. Make sure
again that the polygon is completely within the urban area. If you are
not satisfied with the polygon, delete it by clicking on the center
button on your mouse. Make several selections for urban areas in
different parts of the image.
When you are done with
selecting “urban” samples, repeat the process to collect samples for
forest and water. Click on the “New Region” button in the “#1 ROI Tool”
window. Always remember to name your ROI appropriately! Enter “forest”
in the ROI #2 name and “water” in the ROI #3 name.
When you are done
selecting your ROIs, go to Classification – Supervised –
Minimum Distance. In the “Select Input file” window select “subset”.
Click OK. In the “Minimum Distance Parameters” pop-up window click on
“Select All items” under the “Select Classes from Regions” window. Click
on the “None” radio button under “Set Max stdev from Mean” and “Set Max
Distance Error”. Use the arrow button to change the “Output Rule Image”
to “No”. Navigate to your directory and give your file a meaningful
name, e.g. “min_dist” so that you don’t have any trouble later
remembering what classification type you used. Click OK. You will see
the output classification file appear in your “Available bands” window.
Open the classification output in a new display. Link the displays and
examine how well the minimum distance algorithm performed.
Repeat these steps to
create parallelepiped and maximum likelihood classifications. In the
parameters widow for these classification types, don’t change anything,
just use the default values, and under “Output Rule Image” make it No.
Then, open the output results in a new window each time. Visually
examine and compare the result of the classifications.
2. In your
opinion, which of the classifications provides the best result? Why do
you think so?
3. In your
opinion, which of the classifications provides the worst result? Why do
you think so? What could be the reason behind it?
Take a screenshot of
the classification results next to the original image.
Now we will use
separately collected ROIs which will represent “ground truth” and make a
statistical evaluation of the mapping accuracy of each. Open the ROI
Tool on Display 1. Delete all your original ROIs. Go to File –
Restore ROIs. Navigate to your instruction drive lab9_data folder
and select true_pixel.roi file. Click OK in the pop-up window.
On the ENVI tool bar go
to Classification – Post Classification – Confusion Matrix –
Using Ground Truth ROIs. Select your minimum distance classification
image. In the “Match Classes Parameters” window click on “true_urban”
under the “Ground Truth ROI” and “urban” under the “Select
Classification image”. Click on the “Add Combination” button. Continue
the process to match true forest with forest and true water with water.
Click OK. Click OK again in the “Confusion Matrix Parameters” window.
Examine the confusion matrix for the Minimum Distance algorithm.
4.
Complete the table:
|
Parameters |
Minimum Distance |
Parallelepiped |
Maximum likelihood |
|
overall
classification accuracy (%) |
|
|
|
|
value of Kappa
coefficient |
|
|
|
|
urban areas
classified as urban (%) |
|
|
|
|
forest areas
classified as forest (%) |
|
|
|
|
water
classified as water (%) |
|
|
|
|
total % of
urban area in the image |
|
|
|
|
total % of
forests in the image |
|
|
|
|
total % of
water in the image |
|
|
|
|
Commission
error: urban (%) |
|
|
|
|
Commission
error: forest (%) |
|
|
|
|
Commission
error: water (%) |
|
|
|
|
Omission error:
urban (%) |
|
|
|
|
Omission error:
forest (%) |
|
|
|
|
Omission error:
water (%) |
|
|
|
|
Producer
accuracy: urban (%) |
|
|
|
|
Producer
accuracy: forest (%) |
|
|
|
|
Producer
accuracy: water (%) |
|
|
|
|
User accuracy:
urban (%) |
|
|
|
|
User accuracy:
forest (%) |
|
|
|
|
User accuracy:
water (%) |
|
|
|
5.
Using the “Help” button on the ENVI toolbar find out what these
parameters mean, how they are calculated and explain in your own words
(please do not copy and paste from the help files!):
a)
Classification accuracy
b) Kappa
coefficient
c) Commission
error
d) Omission
error
e) Producer
accuracy
f) User
accuracy
6.
In conclusion, which of the classifications in your opinion
provided the best results? Explain why. Try to combine both visual and
statistical evaluation of classification performance.
Copy your
screenshots to a CD or USB jump drive, email them to yourself, or print
them using your print account.
Log off your
computer and turn off the monitor when you are finished.
Typed answers
with screenshots are due November 7th (0102) and Nov 8th (0101), 2007