Forest Disturbance Detection Using Remote Sensing

This is a sample report from a Remote Sensing class I took in the winter of 2011.  Please note that an introduction has been omitted.  For more background information on the project, see the lab instructions at Dr. David Wallin’s website.

Using Unsupervised Classification to Detect Forest Disturbances in Northwestern Washington

Lauren Vanderlugt

February 15th, 2011

Abstract

The purpose of this study was to conduct a change detection analysis of forested areas in a subsection of Whatcom County, Washington.  My research focused on change between two time periods: 1988 through 1992 (88-92) and 1992 through 1995 (92-95).  I used a series of Landsat images that had been previously transformed using the tassel cap method (Campbell, 2007).  The first image is a compilation of four layers that represent differences in brightness and greenness between the two time periods of interest.  I also used three images, from 1988, 1992 and 1995, to distinguish change patterns between these years.  I used ISODATA unsupervised classification to divide the first image into 50 spectral classes, and assigned those classes to one of three information classes based on the three images from specific years.  The three informational classes were “no change,” “timber harvest between 1988 and 1992,” and “timber harvest between 1992 and 1995.”  Then, I quantified the amount of harvested area for each period.  I also determined the rate of harvest within different ownership categories.  Overall, the amount of harvested land decreased from 88-92 to 92-95.  Harvest rates increased only in wilderness areas.  This process highlights the usefulness of tassel cap transformation in understanding disturbance rates and patterns on agricultural lands.

Methods

Data

The study area is a section of northwest Washington State, encompassing the area between Bellingham Bay and Mt. Baker.  The entire area is 237,825 hectares and each pixel is 25 x 25 meters.

All of the images used for this study, except the ownership image, were previously transformed using the tassel cap transformation method (Campbell, 2007).  The pixel values of these images represent levels of brightness and greenness. All the images are from Landsat and have been georectified.

The first image I used, the bakerbay-change image, captured overall change in brightness and greenness between 1988 and 1995.  The image was a compilation of four difference channels.  Each difference channel represented either a change in greenness or a change in brightness between 1992 and 1995 or 1988 and 1992.  I used three more images, each representing brightness and greenness values at a specific instant in time.  The bakerbay-88 image, the bakerbay-92 image and the bakerbay-95 image were taken in the summers of 1988, 1992 and 1995, respectively.

I used one other image that showed land ownership within the study area.  The four different land ownership types are The U.S. Forest Service (USFS), The Department of Natural Resources (DNR), wilderness areas, and private lands.

Analysis

I began the analysis by running the bakerbay-change image through an ISODATA unsupervised classification.  The resulting image had 50 spectral classes, each representing changes in brightness and greenness between 1988 and 1995.  I used the three images, from 1988, 1992 and 1995, to assign the spectral classes of the ISODATA image into three information classes.  The three information classes are “no change,” “timber harvest between 1988 and 1992” and “timber harvest between 1992 and 1995.”  I used the three images to locate areas of increased brightness.  For example, if the 1992 image showed high brightness in an area that had low brightness and high greenness values in the 1988 image, than I classified those pixels as “timber harvest between 1988 and 1992” in the ISODATA image.  High brightness values are typical of areas that have been clear-cut, leaving behind exposed soil.  High greenness values indicate abundant vegetation (Campbell, 2007).

After classifying all spectral classes into one of three information classes, I combined like classes to produce an image with only three spectral classes.  Next, I created a mask to reassign all pixels below 100 meters and above 1700 meters in elevation to “no change.”  This was necessary because many low-elevation agricultural lands and high-elevation areas covered in snowpack were improperly classified as harvest.

Then I used a sieve tool to remove patches of harvest that were smaller than a typical harvest unit.  Although typical harvest units are generally at least 10 hectares in size, I eliminated all units that were smaller than two hectares (Wallin, 2009).  I also used a smoothing window of 5 x 5 pixels to reassign isolated patches of harvest to the value of the majority of pixels around them.  These two techniques created more uniform harvest blocks typically associated with timber harvest patterns (Fig 1).

From this final image, I calculated the total amount of harvest for each time period, the overall rate of harvest per year and the harvest rate in forested areas (Table 1).  I quantified forested areas by using the land cover image I produced in a previous lab (Vanderlugt, 2011).  I also calculated rates of harvest within the forested areas of each ownership category (Table 2).  Figure 2 shows the different land ownership areas overlaid by the harvest units.

For a detailed step-by-step guide of these methods, see Wallin, 2009.

Results

Comparison of timber harvest by time period

Overall, the amount of harvested land decreased between the 88-92 and 92-95 periods (Table 1).  The amount of land harvested during 92-95 was only 36.4 percent of the land harvested during the previous four years.  The rate of harvest in forested areas, measured as percent of harvested forested land per year, decreased from 1.53 percent during 88-92 to 0.74 percent during 92-95.  Forested land represents 66 percent of the total image.

Comparison of timber harvest, by ownership

Three of the four ownership categories reflected a decrease in harvest from 88-92 to 92-95 (Table 2).  Harvest rates in forested USFS lands decreased from 0.89 percent to 0.33 percent over the successive time periods.  The harvest rate on private lands also decreased dramatically, by 71 percent, between 88-92 and 92-95.  On DNR land, the rate of harvest decreased by 79 percent between the two study periods.

Harvest rates increased only in wilderness areas (Table 2).  The harvest rate of forested wilderness areas doubled between 88-92 and 92-95.

Patterns in the final image

A visual analysis of the harvest units within different land ownership categories indicates that units are typically larger on private lands and DNR lands, than on wilderness areas or USFS lands (Fig 2).  Timber plots harvested during 88-92 are generally larger than plots harvested during 92-95.

Discussion

Advantages of using tassel cap transformed images

The purpose of this exercise was to identify disturbances in forested areas over time.  The result was two unique pieces of information that describe the locations and rates of disturbances.  The final image shows us where these disturbances are located, and when they took place (Fig 1).  Table 1 and 2 quantify the image into numbers that provide a clear analysis of the data.  Using tassel cap transformed images was a vital part of the process.  The advantage of this method is that the pixel values only reflect greenness and brightness.  This minimizes noise from non-agricultural features.  Greenness conveys information about the vigor of living vegetation and brightness values are highest for exposed, bare soil (Campbell, 2007).  By using a tassel cap transformation, the non-thermal Thematic Mapper bands are manipulated specifically for use in agricultural applications.

Advantages and disadvantages of quantifying disturbances with remotely sensed images

This method of detecting patterns in forest disturbances is advantageous for the different agencies and persons that manage forest lands, as well as forest ecologists.  First, this method is more efficient than tediously pouring over historical records of clear-cuts and examining air photos to garner data on disturbances.  Because Landsat images are employed in this technique, a similar detection analysis can be done for any period of interest since 1972.  Another advantage is the visual aspect of the data.  Figure 2 shows characteristics of the harvest units that may otherwise be overlooked in the tables of data.  For example, the relative size, shape and location of each unit can be observed in the images (Fig 1 and 2).

These characteristics may help scientists decipher the cause of each disturbance unit.  A more thorough investigation would include parameters to distinguish different types of disturbances, such as timber harvests, wildfires or climate-induced disturbances.

Validity of results

The weakness of this exercise is the lack of an accuracy assessment on the analysis.  To further test the validity of these results, an accuracy assessment should be conducted using historical records, air photos and applicable ground-truth data.  Any ground-truth data used would need to have been collected around the time that each photo was taken.

Tables and Figures

Table 1.  Amount of harvested land and rates of harvest for 88-92 and 92-95 periods.  Overall rate of harvest refers to the harvested land as a proportion of the entire image.  Rate of harvest in forested area refers to the harvested land as a proportion of the forested area.

area (ha)

overall rate of harvest (%/yr)

rate of harvest in forested area (%/yr)

88-92 Harvest

9669.9

1.02

1.53

92-95 Harvest

3521.3

0.49

0.74

Entire Image

237825.0

Forested Area

157647.9

Table 2.  Rates of harvest in forested areas within each land ownership class.

rate of harvest in forested area, by ownership (%/yr)

Land Ownership

forested area (ha)

88-92 Harvest

92-95 Harvest

Wilderness Areas

17790.1

0.63

1.28

National Forest

42631.4

0.89

0.33

Private Lands

59906.8

2.23

0.65

Dept. of Natural Resources

37279.3

1.59

0.33

Figure 1.  Final image produced by unsupervised classification, showing timber harvests from 1988 to 1992 (green), timber harvests from 1992 to 1995 (purple), and areas without disturbances between 1988 and 1995 (black).  The lower left image shows the entire study area.  The upper image is a zoomed portion which better depicts the general size and shape of harvest units.  The lower right image is zoomed on a portion of a harvest unit.

Figure 2.  Harvests units from 1988 through 1992 (green) and 1992 through 1995 (purple), on land symbolized by ownership.  Land ownership types include private (black), Department of Natural Resources (brown), national forest (lavender) and wilderness areas (white).  This image shows the entire study area.

Literature Cited

Campbell, James B. Introduction to Remote Sensing. 4th ed. New York: Guildford, 2007.

Vanderlugt, Lauren M. 2011.  Unsupervised Classification of Land Cover Types Based on Ground-Truth Points and Conditional Raster Calculation Methods in a Subsection of Whatcom County, Washington. Unpublished raw data.

Wallin, David O. “Lab V: Change Detection Using Unsupervised Classification with ENVI.” 17 Sept. 2009. Web. 10 Feb. 2011. <http://faculty.wwu.edu/wallin/envr442/ENVI/442_change_lab_ENVI.htm&gt;.

Advertisements

Discussion

No comments yet.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

Categories

  • No categories

Archives

%d bloggers like this: