Lab exercise 10, due right
before the class on November 9, 2005
EES5053: Remote Sensing. Earth and
Environmental Science, UTSA
http://www.utsa.edu/LRSG/
Student Name: ___________________
Unsupervised classification and change detection
Purpose
In this lab, you will examine two images (April and July images) at the exactly the same area but in different crop seasons. You can see the changes visually. For example, changes from Fallow land to crop lands, or from growing alfalfa lands to mature alfalfa lands or fallow lands. You will first actually examined the spectral differences of a same pixel in different time (April and July). then use tools in ENVI to automatically quantify those changes, which we called change detection. In this lab, you will explore two very simple change detection methods based on ENVI: Computer Difference Map and Change Detection Statistics. This lab is designed to give you a fresh look of image classification and change detection which is one of the focuses in the image processing class.
Step 1. Preparation
Create a directory Lab10 under /EES5053 and Create data and MyWork subdirectories under Lab10. You will need to copy new data needed for this lab from the class website. The elpaso-3338apr26.img is a subset of path33/row38 image acquired on April 26, 2001. The elpas-3338jul15.img is a subset of path33/row38 image acquired on July 15, 2001. Images below are false color images of band compositions 742. Please examine these images with care by following the background information introduced below:
This study area (subset) is an agriculture area in the lower valley of El Paso County, Texas. In the winter time (see image of April 26), the main crop in the area is Alfalfa colored as bright green (growing) or dark green (just been cut). I hope you know why vegetation is green in this 742 composite. Most of the agricultural fields are fallow lands (bare soil, no green). Ground truth indicates that in the summer time (image of July 15), farmers planted a number of different crops such as grass, chili, corn, cotton, and alfalfa (see the image of ground truth results). Pecan is also a very important economic crop in the area, though it is not included in the ground truth area. In the July image, you can see most of the fields are active crop lands (green), a few of them are fallow lands. Rio Grande River is the international boundary between Mexico and USA in this region. Noticing most of the agricultural fields in the Mexico side (southwest portion of the image) were alfalfa fields (bright green) in April, while some of them became fallow lands in the July image, some of them are still alfalfa, but with different greenness. The most southwest and northeast portion of study area are desert lands.
.
Step 2. Spectra analysis of different crops and non-crops
Open and load two subset images into two display windows using 742 compositions. Link the two displays, click the Cursor Location/Value, move the mouse in the images, you will see the data values of these two images are less than 1 (floating point). Actually these two images are reflectivity images. I did the conversions for you: DN -> radiance -> reflectivity using the steps in Lab 6.
Question 1
See an example of spectral reflectivity plot below, please do two plots (one for April and one for July) to include the main agricultural crops such as grass, chili, corn, cotton, and alfalfa and other non-agricultural objects such as I-10, road, Rio Grande River, desert land, fallow land. Use the link tool to link two images and check individual pixel change from April to July: such as fallow land to cotton, or growing alfalfa to mature alfalfa. Based on your plots of spectral reflectivity, please do necessary analyses such as the spectral differences of different crops, spectral differences of crops and non-crops, and crops in different stages (growing and mature).
Step 3. Preparation for change detection
in the ENVI main menu, click Basic Tools -> Change Detection, you will find the Computer Difference Map and Change Detection Statistics.
Question 2
What does it means by Computer Difference Map? (please go to ENVI Online help, click Contents tab, then Using ENVI -> Basic Tools -> Change Detection Analysis)
What does it means by Change Detection Statistics? (the same place as above).
What are the main differences of these two methods?
Step 4 .Computer Difference Map
This method is very easy and simple, only processes one band of information in each time. For example band 4 of April image to band 4 of July image. For multispectral or hyperspectral imagery, this is not a good change detection tool. If you have only one band such as the temperature image, this is a good tool. So you can see the temperature difference of two day or two hour.
In this lab, I only need you do a practice with no question to ask you. Click the Basic Tools -> Change Detection Analysis -> Computer Difference Map. A new window called “Select ‘Initial State’ Image” will popup for you to pick the initial image; in this case, you can pick any band (say band 4) of the April image. Click OK, you will be asked to pick the ‘Final State’ image, which is the same band (here is band 4) of the July image. Click OK. A new window called Compute Difference Map Input Parameters will popup. Change Number of Classes to 5, click the Define Class Thresholds, you will see the classified image will be classified as 5 classes, change the class threshods values as the figure below. if the value of difference of any pixel between April and July is (-0.03, 0.03), it will be classified as Class 3, meaning no change; if it is between (0.03, 0.15], it will be Class 2, meaning reflectivity increase (could be from fallow land to crop land); If it is between (-0.03, -0.15] will be Class 4, …… Click OK to close the window, Click output result to Memory. A classified image will be created, open it and link it the original images to check them. The pixel values of this image will be class 1, 2, 3, 4, or 5.
Step 5. Change Detection Statistics
5.1 Unsupervised classifications
From the definition you found above, you should already knew that for the change detection statistics, you need two classification images. To get the classification images, ENVI provides a number of tools. In this lab, you will only play with the simplest tool: IsoData. IsoData calculates class means evenly distributed in the data (image) space and then iteratively clusters the remaining pixels using minimum distance techniques between a cluster and a pixel. Each iteration recalculates means and reclassifies pixels with respect to the new means. For a detail definition, please go to the ENVI online help. (You can also try the SAM that you used for lab 7. SAM is supervised classification method, you will first select region of interest (ROI), which is the known classes (such as fallow land, cotton, water). If you also do the SAM in this lab, you will get 20 points of extra credit)
In the ENVI main menu, Click Classification -> Unsupervised -> IsoData. This will bring you a window for you to select the image to be classified. You will first pick the April image, an ISODATA Parameters window will be popup. Keep all default numbers except change Maximum Iterations to 5. Output your result either to a File or Memory. Click OK to run the classification. (You can find what all these parameters mean from the ENVI online help if you would like, but it is not required for this lab). Load your classified image to a new Display window, link this classified image to the original April image.
Question 3
Find how many classes (colors) in classified image, and what object(s) each class (color) represents by comparing the linked two images? (hint, when you move you mouse in the classified image, the “Cursor Location/Value” window will tell you class number of each color. For example, the red color is class 1 and the yellow color is class 4, from the linked original image you should be able to tell the class 1 is mostly alfalfa)
Using the same method to classify the July image (Keep all default numbers except change Maximum Iterations to 5), and linking the classified image to the original July image.
Question 4
Find how many classes (colors) in classified image, and what object(s) each class (color) represents by comparing the linked two images?
3.2 change detection statistics
Click Basic Tools -> Change Detection -> change detection statistics. A new window called “Select ‘Initial State’ Image” popup for you to pick the initial image, in this case, you should pick the classified April image. Click OK, you will be asked to pick the ‘Final State’ image, which is the classified July image. Click OK. Then the Define Equivalent Classes window will popup (see table 1 below). Paired Classes such as Class 1 of the Initial State Class will be automatically paired with Class 1 of the Final State Class.
Click OK, you will have a new table called “Change Detection Statistics” popup. This table will tell you all information about changes (Pixel count change, Percentage change, Area change) for each class. Below is an example for helping you to read and interpret the table:
Taking class 1 in Initial State (April image) as an example to see the changes in Final State (July Image):
Class 1 (alfalfa) has no unclassified pixels (0 %)
32.318% of Class 1 did not change, still Alfalfa or grass
25.289% of Class 1 changed to Class 2 (fallow land 1?)
30.263% of Class 1 changed to Class 3 (fallow land 2?)
9.620% of Class 1 changed to Class 4 (desert or residential?)
2.510% of Class 1 changed to Class 5 (desert?)
67.682% of Class 1 has been changed from April to July
Image difference is calculated by:
(total pixels of Class 1 in Final State image – total pixels of Class 1 in Initial State image)
---------------------------------------------------------------------------------------------------------------------
total pixels of Class 1 in Initial State image
which is (from Pixel Count): (34244-22433)/ 22433 = 0.5265 = 52.65%. This means that total pixels of alfalfa or grass increase about 52.65% from April to July.
Question 5
(1) Copy and paste your Change Detection Statistics table to here, and do all analyses as the example above.
(2) And tell which class has the largest change in percentage, which class has the less change in percentage,
(3) Which class has the largest increase in pixels, and which class has the largest decrease in pixels.
(4) Tell your feeling about the overall quality of the unsupervised IsoData classification results.