project.docx - knight geospatial science group |...
TRANSCRIPT
University of Minnesota.
The Pagami Creek Wildfire
Adam Burger
John Habib
Remote Sensing of Natural Resources and EnvironmentFNRM 3262
December 7, 2014
INTRODUCTION
Wildfires are some of the most prolific examples of nature’s fury. Destroying life, habitat
and property, they are at the same time an essential component of forest ecosystems. Remote
sensing provides us with powerful tools we can use to learn more about wildfires: where they
are, and how they change the areas that they affect. The Pagami Creek Wildfire was started by a
lightning strike approximately 13 miles east of Ely, MN and was first detected on August 18,
2011 [1]. Due to low humidity, dry conditions, and shifting winds, the fire had grown to 93,000
acres by September 13 [1], spreading so fast that fire shelters were deployed for the first time in
Minnesota.[2] This fire was particularly notable in that it took place in the Boundary Waters
Canoe Area Wilderness (BWCAW). The BWCAW is a 1,000 acre wilderness area located
within the Superior National Forest, containing over 1,200 miles of canoe routes and
approximately 2,000 designated campsites [3]. It is a popular destination for Minnesotans and
visitors alike, and contains some of the most treasured natural landscapes in the state. Thus the
Pagami Creek Wildfire was not only a threat insomuch as it might escape the BWCAW and
threaten private property, but the very fuel it burned was an immensely valuable natural and
recreational resource.
The goals of this project are to determine the extent of land cover change caused by the
Pagami Creek Wildfire (Goal 1), and to assess the progress of recovery in the affected area
which has occurred since the fire subsided (Goal 2). No man-made efforts, no matter how
intensive, can cause land cover to change as quickly over a large area as a wildfire can, so it is
expected that drastic change will be observed over the duration of the fire. The ability to use
remote sensing to assess the extent of a wildfire and also recovery of vegetation is very important
for forest managers as they make important decisions regarding forest restoration. Limited
though it may be, remote sensing offers a far more cost-effective toolbox to complete our
objectives than boots-on-the ground field observation of several thousand acres, to be sure.
MATERIALS
In order to assess the parameters which define the goals set forth in this project it was
first necessary to obtain imagery of the area affected by the Pagami Creek Wildfire. Landsat data
is available for all of Minnesota for many decades and is available at no cost, se we chose to use
Landsat imagery to perform our analysis. The USGS Global Visualization Viewer
(glovis.usgs.gov) offers a user-friendly way to browse satellite imagery available for a particular
place and time. Using the general location of the fire and an image of its footprint provided by
the US Forest Service, we were able to locate the affected area in the Glovis browser, and we
procured three images, one before the fire, one immediately after, and one of the present day.
The “before” image was taken on May 15th, 2011, by Landsat 4, and had 0% cloud cover. The
“after” image was taken on October 6th, 2011, by Landsat 4, and had 0% cloud cover. The
“present day” image was taken on October 22nd, 2014, by Landsat 8, and had 0% cloud cover.
For image processing we chose to use ERDAS Imagine 2014, due to its powerful and diverse set
of tools, and also its widespread availability and industry-recognized reputation.
METHODS
The USGS Glovis browser delivers each image in a compressed format. In order to
process each image using ERDAS Image it was first necessary to unzip each image into
individual TIFF files, each representing one band, and then use the Layer Stack function in
Imagine to compile each image into a single, multiband IMG file, recognizable by Imagine. The
difference in number of bands between Landsat 4 and Landsat 8 (Landsat 8 has more) made it
necessary to eliminate some of the bands available when stacking the “present” image in order to
ensure spectral homogeneity in the image pool.
After the images were rendered, the task of accomplishing the first goal of determining
what percentage of forest cover was eliminated by the fire was at hand. This involved classifying
the images using a supervised classification. Advantages of supervised classification over
unsupervised include the ability of the user to specify classes of interest prior to classification,
which eliminates the need for classes which are not of consequence to the task at hand. This
method is useful if the classes in the images are known and apparent, which was the case here.
Disadvantages include the time and manpower needed to train the classifier, but given the small
amount of images and the small area they cover that was hardly a concern for our project. We
used ERDAS Imagine software to classify the images. In order to perform a supervised
classification it was first necessary to train the classifier to recognize the classes of interest. We
chose four classes which comprised the bulk of the imagery: forest, burned forest, water, and
other. It was essential to have water classified as there are several small lakes interspersed in the
imagery as well as a large portion of Lake Superior. We created an AOI layer and drew polygons
to identify 10 training sites for each class on the October 6th (after) image, and then ran a
supervised classification on the ‘before’ and ‘after’ images using a minimum-distance algorithm.
The class delineations can be seen in ‘Results’. Since the ‘fire’ class is the main focus of our
analysis, both the ‘forest’ and ‘other’ class are simply represented by one color, which is green
on the map, while ‘water’ is blue and the 'fire' is red.
The next step was to perform a change detection. Post-classification change detection
offers the advantage of knowing what changed, and how much, as opposed to only the
magnitude of change offered by pre-classification, which is particularly important to our
objectives as stated here. We applied the “thematic change” function in Imagine to the classified
images of the burned area before and immediately after the fire, as well as to the images
immediately after the fire and of the present day. To obtain hard statistics describing the
destruction and regrowth of forest cover, we used the “Summary Report of Matrix” function to
obtain percentages and total hectares of class changes between images. This provided us with a
mathematical depiction of the natural processes at play. To calculate the total area affected by the
fire, the total areas that changed to from a class to 'fire' were summed, with the exception of the
area that was classified as 'fire' in the ‘before’ image and also in the ‘after’ image. This area was
considered to be a classification error because the ‘before’ image should not have any 'fire' and
was therefore subtracted from the final total of area affected by the fire.
Next the second goal presented itself: determining how much of the area decimated by
the wildfire has “grown back” between the cessation of the flames and the present day. “Grown
back” is hardly quantifiable, but fortunately a widely used spectral metric presents itself in the
form of the Normalized Difference Vegetation Index, or NDVI. NDVI is the ratio of the
difference between near infrared reflected light and visible reflected light to the sum of near
infrared reflected light and visible reflected light [4]. Mathematically, NDVI = (NIR —
VIS)/(NIR + VIS). Due to the spectral characteristics of healthy vegetation, a higher NDVI value
corresponds to a higher density of plant growth [4]. Hence, a higher NDVI value in the area
affected by the wildfire is an indicator of increased plant growth, affording us a quantifiable way
of determining how much of the area has indeed “grown back”.
ERDAS Imagine offers an NDVI function, which accepts a multiband spectral image and
outputs a single-layer image in which the DN value of each pixel is the NDVI value calculated
for that same pixel on the original image. This function was applied to each image in the pool,
generating three NDVI image files. In order to determine how NDVI has changed between the
three time periods, we applied the Image Difference change detection function in Imagine to the
“before” and “after” images, as well as to the “after” and “present” images. While only the latter
offers progress toward our second goal, the former acts as a sort of crude accuracy assessment:
the “ground truth” of vegetation decrease caused by the fire is as well known to us as it is to the
wildland firefighters who felled countless snags, but does the NDVI-image difference combo
corroborate their testimony?
Image difference simply subtracts the pixel DN values of the second image from the first
image. The result is a quantification of change. For example, a positive DN value for a pixel on
the “‘present’ minus ‘after’” change detection image indicates that the pixel in question had a
greater NDVI value in 2014 then it had in 2011, immediately following the fire. A DN value of
0.30 for that same pixel would indicate that the NDVI value has increased by 30% over that time
period. In this way we were able to quantify change in NDVI, on a pixel-by-pixel basis, for the
entire area of our imagery. This accomplished, we still needed to confine our analysis to the area
of interest, that is, the area affected by the wildfire. This area is visibly distinct on all imagery
following the fire, including that of the present day. Thus it was a small task to digitize an AOI
layer which represented the boundary of the affected area. A perfect match between the area’s
boundary and the layer is impossible, not to mention the fact that the boundary is likely not as
stark as it appears in satellite imagery, so care was taken to ensure that the AOI layer remained
within the area’s visual boundary. More important than including every pixel affected by the fire
in our analysis was the necessity of excluding pixels unaffected by the fire, which could easily
wreak havoc with our results. Finally, we used the Subset function in Imagine to generate new
image difference files composed of only the pixel values of each image contained within the AOI
layer. We then computed the statistics of the DN values contained within these two subset
images. The end result: the average increase in NDVI in the area affected by the Pagami Creek
Wildfire between 2011 and 2014. A single number, from so much data.
RESULTS
Below are the two unclassified Landsat 4 images used for the change detection before and after
the fire with the area of the fire clearly visible in the ‘After’ image.
Unclassified- Before
Unclassified- After
The classified images below show the
‘forest’ and ‘other’ classes as green, the
‘water’ as blue, and the ‘fire’ as red.
Supervised Classification- Before
Supervised Classification- After
Summary Report
The summary by zone report shows
the pixel count, percentage, and
total area in hectares of the class changes before and after the fire. The ‘Zone name’ represents
the classification in the ‘before’ image and data for each class that changed from that
classification is shown in the table. For example, in the first table below, 8009 pixels, or 720.81
hectares, were classified as ‘fire’ in the ‘before’ image and then classified as ‘fire’ in the ‘after’
image. The total area affected by the fire was found to be 33,453 hectares.
NDVI Processing
After Layer Stack
After NDVI
After Image Difference
“Highlight Change” file: Pixels with more than 10% increase are green, pixels with more than
10% decrease are red
Statistics for subset of image difference file generated from “after” and “present” NDVI files.
The “meat” of NDVI analysis is the Mean Value.
Statistics for subset of image difference file generated from “before” and “after” NDVI files.
Nonessential to NDVI analysis, rather a “crude accuracy assessment.”
DISCUSSION
As seen in the unclassified ‘before’ and ‘after’ images, the area of the burned forest is
relatively apparent. Once the supervised classification was applied, the area of the burned forest
is very well defined. Initially, the supervised classification was run using ‘maximum likelihood’
as the algorithm. This produced images with significantly more classification error than
‘minimum-distance’ and thus it was not used. However, some error is still apparent in both the
‘before’ and ‘after’ images. In theory, the ‘before’ image should not have any area classified as
'fire'. However, this is not the case as some small areas, especially near the margins of the photo
are seen as red. It is assumed that the error that occurred along the margins of the images are
likely due to sensor error. Classification error is also seen in the ‘after’ image where areas are
classified as 'fire' that are certainly outside the extent of the fire. These errors, however, were
mitigated by subtracting the areas misclassified in the ‘before’ image from the total area
calculation. While some misclassification still exists in the classified ‘after’ image, it is relatively
small with respect to the total area of the fire. The total area of the fire was calculated to be
33,453 hectares, which is fairly comparable with the 37,636 hectares reported by the US Forest
Service [1]. Some possible ways to reduce the amount of misclassification error in the future
would be to include more classes or to subset the image to focus more closely on the area of
interest.
The mean value for NDVI increase was 4%. What is the significance of this? To analyze
this statistic we can turn to what was previously referred to as a “crude accuracy assessment”: the
decrease in NDVI values caused by the fire. This number was 17%. On the surface, we can
conclude our analysis here: the reduction in NDVI values caused by the fire validates our
method, so we have confirmation that the area is indeed growing back, although slowly. But let
us go deeper. 4% increase, over three years, is an increase of approximately 1.3% per year. In
theory, 17% increase would constitute that area having achieved its previous level of vegetation.
Mathematically, this translates to 10 years from 2014, assuming a constant growth rate. Less
superficially, then, we could say: in 10 years, the acres affected by the Pagami Creek wildfire
will have regained their recreational and ecological value. Caveats, however, abound. First, the
data itself is questionable. The Inquire cursor in Imagine allows the user to browse pixel values
in an image. Browsing the values in the 4% file, we found an “observational average” of around
20% . The statistics reported a maximum of 40%. Why such a low number? Lakes, of which
there are no shortage, typically had values around zero, occasionally below. A more accurate, but
far more time-consuming analysis, would have involved eliminating lakes from the imagery.
Furthermore, with the countless factors affecting forest regeneration, how can we assume a
constant growth rate? We only measured two points on the line: in between, all sorts of things
may have happened. Even if the data is sound, however, can we trust it entirely? NDVI indicates
“greenness”. In 10 years we may have an area as “green” as it was before the fire, but it will not
be old growth forest. “Greenness” doesn’t translate directly to a resurgence in ecological and
recreational value. How much of the area’s value is dependent on its pure age? This is a question
for biologists and foresters.
REFERENCES
1. US Forest Service. Pagami Creek Wildfire. Available at
http://www.fs.usda.gov/Internet/FSE_DOCUMENTS/stelprdb5346343.pdf
2. James Wilson, MN-DNR Forester
3. US Forest Service. The Boundary Waters Canoe Area Wilderness. Available at
http://www.fs.usda.gov/detail/superior/specialplaces/?cid=stelprdb5202169
4. NASA Earth Observatory. Measuring Vegetation (NDVI & EVI). Available at
http://earthobservatory.nasa.gov/Features/MeasuringVegetation/printall.php
DIVISION OF WORK
● Location of imagery, initial processing trials, and research: Adam and John
● Processing, report, and presentation formation related to Goal 1: Adam
● Processing, report, and presentation formation related to Goal 2: John