ES 775 Lab 6

UNSUPERVISED CLASSIFICATION

Ft. Leavenworth, Kansas

Landsat TM datasets are provided for the vicinity of Ft. Leavenworth, northeastern Kansas--see map. Portions of the military reservation include forests in two topographic situations--Missouri valley bottomland and upland bedrock ridges.

See Ikonos images.

Growth of trees in these forests is little affected by human activities. Seasonal and interannual variations in forest vigor depend mainly on climatic events. The upland forest consists mostly of oak species, hackberry, walnut, and other deciduous hardwood trees. This forest is thought to be particularly susceptible to yearly soil-moisture conditions that fluctuate due to changes in precipitation.

Kite aerial photograph of study forest at Ft. Leavenworth.
More KAP examples of Ft. Leavenworth.
Ground view of opening in study forest at Ft. Leavenworth.
Forest is composed mainly of various oak species.
Missouri River valley near Wathena, KS.
Agriculture is the primary landuse in this scene.
Missouri River and bridge at Atchison, KS.
All photos © by J.S. Aber

The objective of this lab exercise is to isolate an upland forest study site at Ft. Leavenworth and to compare yearly variations in forest vigor, as documented by the normalized difference vegetation index (NDVI). Isolation of the study forest involves unsupervised classification techniques that will separate the upland deciduous forest from all other kinds of vegetation cover--agriculture, prairie and bottomland forest. A major problem in vegetation classification is the similarity in spectral signatures for deciduous trees, grass, and active crops.

You will work with Landsat TM datasets for July scenes in three years. These three years represent pre-drought (1987), drought (1988), and post-drought (1990) conditions. Download the following files into your student directory. These scenes have been haze corrected and digitally resampled into the UTM zone 15 coordinate scheme. Note the relatively small size of the datasets (533 columns by 300 rows).

Scene Date TM Bands File Name
July 1987 1, 2, 3, 4, 5, 7 LV87
July 1988 3 and 4 LV88
July 1990 3 and 4 LV90

Exercise

Begin with the 1987 scene, for which TM bands 1-5 & 7 are provided. Make a false-color composite based on bands 3, 4 & 5 for the July 1987 scene. This particular composite has proven most useful for general display of vegetation cover in a naturalistic manner. This year was particularly good for vegetation, as documented by tree rings, in terms of favorable climatic conditions. Notice the appearance of different kinds of vegetation cover depicted in green and yellow-green colors.

Landsat TM false-color composite (bands 3, 4, 5) of Ft. Leavenworth, Kansas, July 1987. Vegetation appears in green and yellow-green colors. Note the dark-green and olive shades of the the study forest and other forested areas. Click on the small image to see a full-sized version.

1. Describe the appearance of forests and agricultural fields.

2. On what basis are you able to discriminate between forests and agricultural crops?

The dark-green coloration of forest areas is visually similar to many agricultural fields, which are evident because of their rectangular shape. Furthermore, upland and bottomland forests display the same spectral signatures in this image. On this basis, the false-color composite is not effective in separating the upland forest from other types of vegetation cover in the scene.

Next you will derive the normalized difference vegetation index (NDVI) based on bands 3 (red) and 4 (near-infrared). Use the OVERLAY module (rightmost icon). Enter LV87-4 as the first image and LV87-3 as the second; select overlay option: First - Second / First + Second. Give an appropriate title, and list "raw ndvi" as the value units. Give the output file a "temp1" name. The display will come up automatically; change to the NDVI256 palette. Vegetation appears in green; non-vegetated surfaces are yellow and brown.

3. What is the data type, and what is the range of values for your NDVI image?

The raw NDVI values are fractional real numbers that range between -1.0 (no vegetation) to +1.0 (max vegetation). For further processing, change the raw NDVI values into the byte-binary data type as follows.

You may now delete the "temp" files. Display LV87-NDVI with the NDVI palette, and examine the patterns of vegetation depicted on the image.

4. From the LV87-NDVI image, can you distinguish between forest and agricultural fields on the basis of NDVI values (green color)? Explain your answer.

Another approach to vegetation analysis is called "tasselled cap." Read about the TASSCAP module in the Idrisi Kilimanjaro Help section (under Program modules, Analysis, Image processing, Transformation). This approach utilizes all Landsat TM bands (except 6) for a more in-depth analysis of vegetation attributes. The Idrisi TASSCAP module will derive three values (3 files).

Now run the TASSCAP module. Check TM, and enter files for July 87 bands 1-5 and 7. Enter "LV" as the output prefix. Display the resulting files with autoscaling and the default (Idrisi256) palette.

5. Describe the appearance of the study forest on the three images. Do any of these images effectively separate upland forest from other types of vegetation cover?

It should be apparent at this point that ordinary image enhancement will not be sufficient to separate the upland study-forest area from other kinds of vegetation cover present in the scene. You will now work with some more advanced techniques. Read about the ISOCLUST module (under Image processing, Hard classifiers). This module is an "iterative self-organizing unsupervised classifier" based on Landsat TM visible, near-infrared and mid-infrared bands (1-5 and 7).

Now run the ISOCLUST module. Indicate 6 as the number of bands to process, and enter Landsat TM bands 1-5 and 7 (July 1987). Click "next" and a histogram will appear. Notice that most of the data are contained in clusters 1-16 (left side of chart). Clusters beyond number 16 are minor components of the scene. This is quite typical of isocluster operation on TM datasets. On this basis, indicate 3 iterations and 16 clusters, and name the output image LV87-ISOC. Click OK, and wait a while for processing (even on a fast PC). The results will display automatically with the Qualitative palette. You may delete the ISOTMP file that was created during processing.

6. Which clusters represent the forested areas? Do these clusters also include any agricultural fields?

Note: It appears that clusters and cluster numbering of ISOCLUST are subject to some minor random variation. In most cases, the study forest is composed of clusters 1 and 5. Examine your results carefully and compare to the TM 345 composite to identify the study forest clusters. Notice the study forest has some narrow openings for roads and other clearings and buildings, which will show up as different clusters.

This is about the best that unsupervised classification is able to do with a scene of this complexity in terms of similar kinds of vegetation cover. To further isolate the upland study-forest area, you will utilize a procedure similar to that of Lab 4. This involves reclassification to produce a boolean image, followed by grouping and another reclassification. In the end, the study forest will be left (value = 1) and all other areas of the image will be blank (value = 0). The general procedure is outlined below. You will have to work out details for each step.

Name your final result LV87-FOREST, and display it with the Qualitative palette. The study forest should appear in bright red and all other areas are black. What you have created is called a mask. It shows only a single feature and everything else is blacked out. You have managed to isolate the study forest within the Landsat TM scene.

Mask of the upland study forest (red) extracted from the isocluster analysis of the July 1987 Landsat TM dataset. The area shown on this image corresponds exactly to previous examples. Click on the small image to see a full-sized version.

7. How does the study-forest mask look in comparison to the study-forest area identified on the sample images?

8. What is the surface area (in km²) of the study forest as depicted in the mask?

Your next objective is to determine the status of vegetation within the study-forest area based on NDVI values. Use OVERLAY to multiply LV87-FOREST times LV87-NDVI. Name the output file SF87-NDVI. The image should display automatically with the default palette; change to the NDVI256 palette. The study forest should appear in green colors, while the rest of the image is black.

Now utilize the HISTO module to look at the statistical spread of NDVI values for the SF87-NDVI image. Select "graphic" output type, and reset the "class width" at 1. Reset the "minimum value for display" = 200 (in order to eliminate the zero background values and to rescale the histo display).

9. What are the mean and standard deviation values for SF87-NDVI (excluding zero background)?

Save the resulting histogram to turn in. Use the "save to clipboard function," then paste into Paint as a bmp file. Save it as a 256-color image with the name HISTO87. Further reduction in file size is possible in "gif" or "jpg" format. Make sure you save the final file in your student directory, not elsewhere!

Climatic conditions and forest growth

The summer of 1987 was the culmination of several years of favorable climate and good tree growth, as documented by tree-ring samples taken from oaks in the study forest. The NDVI statistics derived from SF87-NDVI, thus, represent a healthy forest with a well-developed canopy of leaves. A severe drought began in 1988 and lasted through 1989. By 1990, precipitation had returned to near normal for the region. The tree-ring record marks a strong response to the drought; tree-rings of 1988 are the smallest of the last 30 years. By 1990, tree-ring growth had returned to a near-normal level. In other words, tree-ring growth is in phase with annual climatic conditions. (Aber, Wallace and Nowak 2002).

It is often assumed that NDVI values reflect current climatic and environmental conditions that may affect vegetation vigor and growth. While this is usually true for annual agricultural crops and perennial prairie grass, it is not necessarily the case in forests. Mature trees have large food reserves in their bodies (above and below ground) and are able to access deep soil moisture, thus they may not respond immediately to annual climatic events such as droughts. The growth of trees in a particular year may depend on antecedent conditions; that is, conditions of previous years. Each tree species may react differently to antecedent and current conditions.

You will now determine how NDVI values of the study forest varied in response to climatic events of the period 1987-90. Prepare NDVI values for the July 1988 and 1990 images Convert the file data type into the byte-binary scale, as you did for the 1987 image. Now multiply the NDVI files by the LV87-FOREST mask. Use HISTO to determine statistics for NDVI values in the study-forest area for 1988 and 1990. Save the resulting histograms to turn in (HISTO88 & HISTO90).

10. What are the mean and standard deviation values for the study forest in July 1988 and 1990 (excluding zero background)? Using the July 87 NDVI mean as 100%, what are the percentage values for July '88 and '90 mean values?

11. Offer an explanation for the year-to-year variations in NDVI values in relation to climatic events and tree-ring growth.

Turn in

Reference


Return to course schedule.
ES 775 © J.S. Aber (2007).