Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Determining Information about Timber Harvests in the Baker to Bay Area from 1988-2011 Using
Change Detection through ISODATA Unsupervised Classification, Masking, Sieving, and Clumping
Tori Niewohner
Abstract
Using Landsat images of the Baker to Bay study area from 1988, 1992, 1995, 2000, 2005, and
2011 and an image created by doing Tasseled Cap Transformations, I performed change detection
analysis on the created image by running an ISODATA unsupervised classification and assigning each
spectral class to a class of No Change or a timber harvest class between the different years. I used an
elevation and forest mask, a sieve function, and a clump function on the classified image to further
refine the image. 10.19% of the image was in a timber harvest class and 1988-1992 had the highest
percent of forest harvested per year. Using ownership data, I determined that from 1988 to 2011,
private lands had the highest percent of forest harvested per year at 1.13% and National Forests had the
lowest at 0.39%. Change detection is a useful tool for detecting disturbances and although I do not know
what my accuracy of my image is, the USDA and Forest Service Report on Washingtons Forest
Resources supported my findings that private land provides a lot of timber and that the timber harvest
during this time period would have been highest between 1988 and 1992.
Methods
All of the methods that follow and the images used were provided by Dr. Wallin on his website
(Wallin 2016). All of the images used are of the Baker to Bay study area with a pixel size of 25 meters
by 25 meters. The six initial images were created using data from August 31, 1988 (Landsat 5), August
10, 1992 (Landsat 5), July 18, 1995 (Landsat 5), September 25, 2000 (Landsat 7), July 29, 2005 (Landsat
5), and July 30, 2011 (Landsat 5). The 1995 image was georectified to the correct UTM coordinates and
the other images were georectified to the 1995 image. The seventh image was created by doing Tassel
Cap Transformations on each of the six images to get greenness and brightness indices which were
then used to create ten difference channels where each years greenness and brightness is
2
subtracted from the next years greenness and brightness. This image of the change from 1988 to
Using the image representing the change from 1988 to 2011, I ran an ISODATA unsupervised
classification in ENVI for 50 spectral classes. I then had to assign each of the 50 classes to one of 6
information classes: No Change, Timber Harvest between 1988 and 1992, 1992 and 1995, 1995 and
2000, 2000 and 2005, and 2005 and 2011. In order to determine which class went to which information
class, I opened the six images from the different years in color-IR. I compared 1988 to 1992, 1992 to
1995, and so on, in order to determine where there were clear-cuts that were not in the previous image.
I then used the Cursor Location Value tool to identify which classes were being displayed in the clear-
cut. Then I could assign the spectral classes to the information classes. For example, if a clear-cut
showed up between 2000 and 2005, whatever spectral class was within that clear-cut could be classified
as Timber Harvest between 2000 and 2005. Any spectral classes that did not show up in the clear-cuts
were classified as No Change. The classes were then combined so as to get 6 classes, each representing
Next I created an elevation mask to alter change polygons from below 100 meters and above
1700 meters into the class No Change since those are changes in snowpack or agriculture. Following
this mask, I created a forest mask which used my final unsupervised classification image from last lab
and distinguished between forest and non-forest (Niewohner 2016). I multiplied these two masks by
each other in order to only get change polygons that were forest. I then used a sieve function on the
image to delete all patches that were smaller than 2 hectares so that most of the patches would
represent a timber harvest. Finally, I used a clump function with a 3 pixel by 3 pixel smoothing window
to reclassify isolated patches into the class that surrounds them. To determine ownership information, I
performed several band math functions on ownership data provided by Dr. Wallin (Wallin 2016).
3
Results
The resulting image of timber harvests from 1988 to 2011 in the Baker to Bay study area was
overwhelmingly made up of land that had no timber harvests occur (Figure 1). 89.81% of the image had
no timber harvests (Table 1). The remaining 10.19%, or 24,239 hectares, was timber harvests between
1988 and 2011 (Table 1). The time period with the largest area of timber harvest was from 1988 to 1992
(Table 1). Out of the five time periods, the one with the largest percent of the study area harvested per
year and percent of the forest harvested per year is 1988 to 1992 (Table 1). 2005 to 2011 had the
smallest percent of the study area and forest harvested per year (Table 1). During the entire time period
(1988-2011), 0.80% of the forest in the study area was harvested per year (Table 1).
Legend
No Change
Timber Harvest 1988-1992
Timber Harvest 1992-1995
Timber Harvest 1995-2000
Timber Harvest 2000-2005
Timber Harvest 2005-2011
Figure 1. The result of change detection for timber harvests on the Baker to Bay study site from 1988-
2011. To get to this point the image went through ISODATA unsupervised classification, an elevation and
forest mask, a sieve function, and a clump function. Each polygon represents a timber harvest unit.
4
Table 1. The data in this table is based off of Figure 1. For each period of timber harvest between 1988
and 2011, as well as the overall timber harvest in this time period, is listed the area in hectares, the
percent of the image, the percent of the study area harvested per year, and the percent of forest area
harvested per year. The total amount of forest area in the study area was calculated using my final
unsupervised classification image from the previous lab (Niewohner 2016).
During the 1988 to 2011 time period, timber harvests occurred at different rates per year in
areas of different ownership. Within the Baker to Bay study area there are 10,118 hectares of forest in
the wilderness, 37,089 hectares of forest in National Forest lands, 51,469 hectares of forest on Private
Lands, and 36,438 hectares in Department of Natural Resources Land (Table 2). From 1988 to 2011, the
land with the highest percent of forest harvested was on private lands at 1.13% and the lowest percent
of forest harvested was in National Forests at 0.39% (Table 3). Private lands had the highest percent of
forest harvested per year during each time interval except 2005-2011, where wilderness areas had the
highest percent (Table 3). 1988 to 1992 had the highest percent of forest harvested in the wilderness,
National Forests, and private lands, while 2000 to 2005 had the highest rate of forest harvest for
Department of Natural Resources land (though 1988 to 1992 was the second highest) (Table 3). 1992 to
1995 had the lowest harvest rate for the wilderness, 2000 to 2005 had the lowest harvest rate for
5
National Forests, and 2005 to 2011 had the lowest harvest rate for private lands and Department of
Table 2. The area of non-forest, forest, and the area overall for lands in the Baker to Bay study area in
the ownership categories of Wilderness lands, National Forest lands, Private lands, and Department of
Natural Resources Land (DNR).
Wilderness National Forest Private DNR
Area of Non-
17,282 7,650 75,089 2,647
Forest (hectares)
Area of Forest
10,118 37,089 51,469 36,438
(hectares)
Total Area
27,400 44,739 126,558 39,084
(hectares)
Table 3. The data in this table is based off of Figure 1 as well as additional statistics on land ownership.
For each period of timber harvest between 1988 and 2011 as well as the overall timber harvest in this
time period is listed the percent of the forest harvested in four different ownership categories:
Wilderness, National Forests, Private Lands, and Department of National Resources (DNR).
Discussion
One of the issues with the process of change detection through unsupervised classification is
that assigning certain spectral classes to information classes can be subjective. Sometimes it is very clear
which spectral class should go to which information class. This occurred several times when I found a
very evident clear-cut with no vegetation left inside if that was almost completely composed of one
spectral class. However, sometimes there are clear-cuts which still appear to have some vegetation left
inside them. When I examined which spectral classes were inside that type of a clear-cut, there were
often several of them. I initially assigned a few spectral classes to the information class Timber Harvest
between 1988 and 1992 which after re-examination I determined to be vegetation within the clear-cut
and re-assigned them to No Change. However, this was up to my own discretion and another person
could potentially leave them in the 1988 and 1992 change class.
The image that resulted after the assignment of classes and the combination of those classes
was on the right track because it was primarily classified as no change. This is to be expected because
from 1988 to 2011, most of the image should not have been harvested. However, much more of the
image was in timber harvest classes than seemed reasonable and some of the change was in areas that
did not make sense for timber harvests to have occurred in, such as on Mt. Baker or near Bellingham.
One reason for this is because the change detection would detect changes in snow, clouds, and
agricultural land, none of which is a timber harvest. This issue was more or less resolved with the
application of the forest and elevation masks. Most changes in snow were eliminated because they are
above 1700 meters and most changes in agriculture were removed because they are below 100 meters.
Changes in snow or agriculture left-over in between 100 meters and 1700 meters or any changes due to
clouds were removed because the forest mask removed any changes that were not in deciduous forest,
coniferous forest, or a clear-cut. After these masks were applied, the image looked significantly better
because more of it was in the no change class and the change polygons that were still in the image,
7
were in areas that seemed appropriate. The only issue with this process is that the forest mask was
drawn from my unsupervised classification from last lab which had an overall accuracy of 57.3%
(Niewohner 2016). This means that there is a possibility that some of the change polygons could be non-
forest area misclassified as forest or some of the areas in no change could be forest misclassified as
non-forest.
The sieve function also helped my image a lot because it removed any polygons smaller than 2
hectares. According to Dr. Wallins website, timber harvests are generally at a minimum of 10 hectares,
so by sieving out small polygons, I removed areas of change that could be from some disturbance other
than a timber harvest (Wallin 2016). If I did this again, I would probably set the limit at 5 hectares
instead of 2 hectares because 2 hectares seems overly conservative when the typical minimum harvest
area is 10 hectares. The final step, the clump function, made the image much easier to interpret by
changing small, isolated patches to the class that surrounds them and by smoothing out the polygons.
This allows for a user to much more easily find individual change polygons.
The results of this change detection are intriguing. In order to help analyze my results, I looked
over the USDA and Forest Service report, Washingtons Forest Resources, 2002-2006, by Sally
Campbell, Karen Waddell, and Andrew Gray. The result that 1988 to 1992 had the highest harvest rate
of any of the time periods was not too shocking as Washingtons volume of timber harvest has been
declining since 1989 (Campbell, Waddell, and Gray 2010). However, total lumber production increased
between 2000 and 2006, which I would have expected to see in my results, but did not (Campbell,
Waddell, and Bray 2010). This could be an error in my change detection process or it is possible that the
lumber production did not increase in the Baker to Bay study area, but in other areas of Washington.
I initially expected the highest rate of forest harvest per year to be on Department of Natural
Resources Land. While DNR land did have higher harvest rates than the wilderness and National Forests
8
during most time periods and over the whole time period, harvest rates on private lands were always
higher. However, after thinking about it a bit more and reading the USDA and Forest Service Report, it
makes sense. DNR lands are all under the same group and thus timber harvest is controlled, whereas
private lands are under many different owners and do not have an overarching body determining what
to harvest. To flush this out further, federal forest management (such as DNR) have transitioned more
towards resources other than timber, so now private lands have to provide most of the timber
(Campbell, Waddell, and Gray 2010). Also, private lands tend to have a higher proportion of productive
forest land which helps to explain the higher rate of timber harvest (Campbell, Waddell, and Gray 2010).
Forested wilderness and national forests typically having the smallest harvest rate per year makes sense
as these are areas that are protected and unable to be harvested or are simply not owned by anyone
One remaining question of this analysis is how accurate the change detection is. I would only be
able to assess this accuracy with reliable information on the location, size, and timing of historic timber
harvests which I do not have. I can say that the accuracy of my unsupervised classification from the
previous lab was 57.3% which definitely will have influenced the accuracy of this change detection
image (Niewohner 2016). If I was to do this process over again for a different image or location, I would
attempt to find some document on timber harvests or sales to get an accuracy for my final image.
This technique could be very applicable to many situations. Timber harvests were not always
reliably documented, so change detection can be very useful to narrow down when timber harvests
happened, what size they were, and other information that might not be documented elsewhere.
Change detection also can be applied to other disturbances besides timber harvest by tweaking the
procedure. It could be used to find out when certain disturbances such as fire, windthrow, bug-killed
trees and many other types of disturbances occurred and during which time periods they occurred in.
However, the primary issue that I can see with using change detection on other disturbances is
9
determining what a given disturbance is. With timber harvests, it is fairly clear what they are because
they are so large, but a change polygon of a smaller size could be windthrow event or could be a tree
killed by bugs. Overall, change detection is an excellent tool for learning about disturbances of the past
Literature Cited
Campbell, S. K. Waddell, and A. Gray. 2010. Washingtons Forest Resources, 2002-2006: Five-Year Forest
Inventory and Analysis Report. Gen. Tech. Rep. PNW-GTR-800. U.S. Department of Agriculture, Forest
Service, Pacific Northwest Research Station. http://www.fs.fed.us/pnw/pubs/pnw_gtr800.pdf
Cohen, W.B. et al. 1998. An efficient and accurate method for mapping forest clearcuts in the Pacific
Northwest using Landsat imagery. Photogrammetric Engineering and Remote Sensing 64: 293-300.
cohen_etal_1998_apr_293-300.pdf
Niewohner, T. 2016. Using ISODATA Unsupervised Classification, Ground Truth Data, and ArcGIS Models
to Classify the Baker to Bay Image of Bellingham, Washington.
Wallin, D. 2016. Lab V: Change Detection Using Unsupervised Classification with ENVI
http://faculty.wwu.edu/wallin/envr442/ENVI/442_change_lab_ENVI.htm