DEFORESTATION MAPPING

Deforestation Mapping Using Landsat Imagery

Introduction

Background—Global Deforestation
International resource agencies, governments, university researchers, and conservation
organizations have repeatedly attempted to estimate the world’s forest cover and the
rate at which it is changing. While these assessments have produced widely varying
estimates of forest cover, one pattern emerges. Forest cover has declined between 38
% and 50 % since the last ice age and these declines have dramatically accelerated in
the past decades.

Claims by the United Nations Food and Agriculture Organization (FAO), that
deforestation is slowing, have recently been demonstrated to be misleading.
Deforestation still seems to be accelerating in the African and Asian tropics, while we
see slow reforestation in some developed countries. Global forest cover is estimated to
be between 3.44 billion and 3.95 billion hectares. During the past decade, between 9
and 11 million hectares of land have been deforested and the annual deforestation rate
is about 0.2-0.3 %.

However, these numbers are deceptive because deforestation does not occur evenly
across forested lands and can also vary considerably among years. For example, while
there is little net forest loss in the United States, the annual deforestation rate for
Southeast Asia is estimated to be close to 1.9 %.

Objectives
In this chapter you will learn how to assess forest cover change using several satellite
images of one area, collected over a specified period of time. The teaching objectives of
this chapter are to introduce course participants to:
ƒ major concepts used to determine changes in forest cover at a global, regional
and local scale.
ƒ techniques used to determine landcover change from satellite imagery.
ƒ strengths and limitations of deforestation studies using satellite imagery.

Why Do Deforestation Estimates Vary So Widely?

Until relatively recently, forest cover and deforestation assessments were based mostly
on non-spatial data compiled from forest inventories produced by the respective
resource agencies of nations. Differences in forest definitions, in methods used to
inventory forests, and unreliable data sources were major handicaps for comprehensive
forest resource assessments.

Systematic analysis of satellite imagery can provide valuable information for forest cover
and deforestation assessments, but is subject to misinterpretation if, for example, the
forest area sampled is too small to provide an accurate representation of the situation on
the ground. Usually, financial resources for satellite mapping are restricted and only
subsamples can be collected. How this should be done has been hotly debated. Some
groups propose systematic sampling, others propose random sampling. A recent study

1
DEFORESTATION MAPPING

estimates that in order to derive a deforestation estimate from Landsat TM imagery that
is within +/- 20 % of the actual deforestation 90 % of the time, a sampling rate of
between 80 to 90 % is required.

Deforestation Mapping
To date, the largest deforestation mapping project is probably the NASA-sponsored
Landsat Pathfinder Humid Tropical Deforestation Project. The project was a
collaboration between the University of Maryland, Michigan State University and NASA
Goddard Space Flight Center. The goal was to map deforestation in the humid tropics
using Landsat Multispectral Scanner System (MSS) and Thematic Mapper TM data for
the mid-1970s, 1980s and 1990s. The project produced the first systematic deforestation
assessments covering vast forest areas, despite the fact that problems in data
acquisition and analysis were hampered by many technical difficulties including:
• acquisition of satellite data for most areas was not systematic. As a result, many
areas had few useful images and often images from different decades were
collected during different seasons.
• little or no ground control data was available to accurately register imagery.
• little or no ground truth data was available to train the classification of the data
sets and to assess their accuracy.
• processing power of image processing workstations was limited. Because
computing power was not sufficient for analyzing co-registered and stacked
images, post classification had to be used for change detection.

More recently, multitemporal classification and change detection have become possible
by coregistering and combining two or more images into a larger, multiband,
multitemporal image. This newly stacked image is then clustered into hundreds of
different spectral clusters using unsupervised classification methods. Some of these
clusters represent areas that remain unchanged, others will be characteristic of areas
that changed in landcover between the dates chosen for the image combination.

Phenological changes should hamper the results little since they will occur consistently
across the image. True landcover changes, however, occur in distinct clusters and
should display prominently when using the multidate clustering.

Steps in Image Processing for Change Detection

To produce a landcover change map the following steps need to be performed:
1. An image pair is co-registered using 15-40 control points depending on the
topographic variation in the scene. The resampling should be performed
using a cubic convolution algorithm.
2. Masks for no data, cloud, water, and shadow are developed from a coarse
unsupervised classification. Data under these masks is excluded from further
analyses.
3. Unsupervised classification is performed on the combined spectral bands
from both images. The results from the classification are used to separate
the remaining landcover and landcover change classes.
4. Clusters produced from isodata classification are compared to the original
raw images. Based on this visual inspection landcover classes, landcover
change classes and cross-over classes are identified. Cross-over classes
are clusters of pixels that occur in more than one of the classes specified in
Table 1.

2
 DEFORESTATION MAPPING

5. Cross-over clusters are reentered into an isodata clustering and the new
categories resulting from this new unsupervised classification are again
assigned to the classes outlined under Step 4 and in Table 1. This iterative
procedure is repeated until all pixels are assigned.

Table 1. Landcover Class Definitions

Code Landcover Category
1 Water
2 Cloud
3 Shadow
4 Snow/Ice
5 Forest
6 Deforestation
7 Reforestation
8 Non-forest

Mapping Landcover Change in Boreal and Tropical Forests

The production of a Landcover Change Map from multidate Landsat imagery can take
an experienced image interpreter between 1 and 4 weeks. On-the-ground knowledge of
forest phenology and ecology is essential for an accurate classification of forested,
deforested, regenerating, and plantation areas. In addition, this knowledge is essential
to recognize different types of forest cover losses on the basis of multispectral
characteristics, as well as spatial patterns and context. These characteristics vary with
geographic region, cultural background, methods of forest harvesting and clearing, and
type of forest ecosystem. For example, deforestation from logging is characterized by
very different patterns in the boreal forests of Canada than in the mixed deciduous,
moist deciduous and evergreen forests of Burma. In this exercise, you will have the
opportunity to investigate deforestation patterns in both ecosystems.

We suggest that you start with the Images from Saskatchewan and attempt to classify
forest cover changes in the boreal forests of Canada. Later, you may want to examine
forest cover changes in the Rakhine Yoma Mountain Range of southwestern Burma. As
explained earlier, Landsat TM Images can be referenced by the location of the image tile
on the globe and the date the Image was accessed. In this exercise all images
referenced as p037r022 (Path 037 Row 022) are taken at one location in Saskatchewan,
while all Images referenced as p133r048 (Path 133, Row 048) display an area in the
Rakhine Yoma Mountain Range.

3
DEFORESTATION MAPPING

Footprint of satellite Image used for change mapping in Saskatchewan, Canada

Footprint of satellite Image used for change mapping in the Rakhine Yoma Mountain Range,
Burma.

4
 DEFORESTATION MAPPING

How Many Trees Make A Forest?

The International Union of Forestry Research (IUFR) has compiled over 90 different
definitions of forestland that are currently used for forest inventories and monitoring (Lund,
1999). These definitions vary among countries, international organizations, and even among
agencies within countries. Variables most frequently used to define forests include (i)
minimum area of the forest stand, (ii) minimum crown cover in the forest stand and (iii)
minimum tree height. Table 2 provides information on different definitions used for
forestland.
Note the range in the variable crown cover. The thresholds are necessary but also
essentially arbitrary. Sometimes the boundaries are visible but in many cases ecosystems
gradually transform into other ecosystems over a transition area or ecotone. What is a forest
to a German forester might not be a forest to a Brazilian forester and vice versa.
A minimum required crown cover of 10 % is low enough that tundra, savannas and
shrublands may be included. A recent United Nations Food and Agriculture Organization
(FAO) forest resource report uses 10 % crown cover to define forests. As a result, forest
cover estimates include areas in Australia that were classified desert by the Australian
government’s official maps.
Table 2

To estimate deforestation one needs to have an accurate baseline, i.e., a reasonable

estimate of the original forest cover that is based on the same forest definition used for the
deforestation assessment. Differences in the forest definition between baseline and
deforestation estimates will result in large biases.
Differences between baseline and endpoint forest cover also represent a problem for
deforestation estimates derived from maps or satellite imagery. For example, in
postclassification landcover change assessment, forest maps are produced from satellite
imagery for each of the two dates. However, if the images were taken during different
seasons, or the image interpreter used a different forest cover definition, the differences in
forest cover (i.e. deforestation and aforestation) may be over- or underestimated.
The approach used in this chapter to delineate deforestation is more robust. Because actual
differences in spectral reflectance are measured and compared among all locations, only
areas that show larger differences in spectral characteristics register as change. Thus,
deforestation and aforestation mapped with this approach is real and not due to differences in
forest definition.

5
DEFORESTATION MAPPING

Exercise
In this Exercise you will learn how to:
ƒ combine several multidate Landsat TM Images into one data file.
ƒ run an unsupervised classification of the multidate, stacked Landsat Image.
ƒ assign output classes to landcover types and change categories.
ƒ recode the landcover and change the Image to represent only major landcover
and change classes.
ƒ export an ERDAS Imagine Image to an ArcView Grid.
ƒ compare landcover and change classes between two very different geographic
regions.

Stacking Multidate Satellite Images

1. Start ERDAS Imagine by double clicking on the Imagine icon on your Windows
desktop.

2. In the ERDAS Imagine Icon Panel click on the Interpreter icon.

3. Click Utilities/LayerStack… The Layer Selection and Stacking dialog will open.

4. You will be stacking the layers from two different Images. Usually, Landsat ETM+
Images consist of 9 layers including two variations of the thermal layer—each at 60-m
spatial resolution, and one panchromatic layer—at 15-m spatial resolution. In this
exercise we are only utilizing six layers of the ETM+ Images (bands 1,2,3,4,5, and 7;
blue, green, red, near-ir, mid-ir, and far-ir).

The Landsat TM Images consist of seven layers but in this exercise we will exclude band
6—the thermal band with 60-m spatial resolution.

Under Input File: (*.img) select the TM Image called p037r022_1987.img from the
C:\course directory. Under Layer, select 1 and click on the Add button, then continue
adding the rest of the layers one-by-one except layer 6. Layer 6 is the thermal layer and
will not be included in the classification.

6
 DEFORESTATION MAPPING

5. Select the second input Image. Under Input File: (*.img) select p037r022_1999.img.
Under Layer select All and click on the Add button.

6. Under Output File: (*.img) select the C:\course directory and name the output file
p037r022_stacked_course.img.

7. Under Data Type, output should be Unsigned 8 bit. This refers to the combinations of
colors available per pixel. An 8 bit image can use up to 28 = 256 colors, whether
actual colors or varying shades of grey. An unsigned image allows the pixel values to
vary from 0-256, a signed image would vary from -256 to 256, doubling the number
of colors available.

8. Under Output Options, select Intersection and check the box for Ignore Zero in Stats.
Click OK. When the operation begins, the Modeler dialog box will appear. Click on the
Cancel button to stop it. Stacking Images can take a couple of hours depending on the
speed of the machine. We have prepared a stacked Image (p037r022_stack87-99.img)
prior to this exercise to expedite the process.

Question 1: Which bands in the stacked Images are blue bands?

7
DEFORESTATION MAPPING

Creating a Subset of the Stacked Images

If you open both raw Images at the same time you will notice that they are not the same
size.

9. In the Imagine Icon Panel click on the Viewer icon.

10. Select File/Open/Raster Layer... Select p037r022_1999.img. Then open up

p037r022_1987.img on top of it. Before pressing OK, click the Raster Options tab and
uncheck clear display and check background transparent. You should notice that the
newer ETM+ is larger than the TM.

If we stack these Images and then classify them, overlapping and the non-overlapping
areas will get classified. Pixels shared by both images may get mixed with pixels in the
border region that belong to only one of the images. This can cause problems when you
are assigning classes later on. To prevent this, you need to make a subset of the
stacked Images that excludes the entire border region.

8
 DEFORESTATION MAPPING

11. Click on AOI/Tools. Select the Create Polygon AOI tool.

12. Click and drag a polygon around the smaller Image. Make sure to exclude the
boundary region when drawing the polygon. It is probably best if you create the polygon
just inside the boundary of the smaller Image.

9
DEFORESTATION MAPPING

13. Click on File/Save/AOI Layer As… Save it in the C:/Course/ch4-deformap directory

as Canada.aoi.

14. Select the Interpreter icon from the Imagine icon bar. Then select Utilities/Subset...
The Subset dialog box should open.

15. In the C:\Course directory select p037r022_stack87-99.img for the Input File: (*.img).
Under Output File: (*.img) name the subset p037r022_stacksub_course.img.

16. Check the Ignore Zero in Output Stats box.

17. Click on the AOI button of the subset box. The Choose AOI dialog box will open.
Under AOI Source: check on the File circle. Select Canada.aoi from the
C:/Course/ch4-deformap directory. This will restrict the subset to the AOI box that
you created earlier. Click OK.

10
 DEFORESTATION MAPPING

18. Click OK in the Subset dialog box. When the operation begins click Cancel to stop
it. Creating a subset of an Image can take up to an hour depending on the speed of the
machine. We have prepared a subset of the stacked Image (p037r022_stacksub.img)
prior to this exercise.

Unsupervised Classification of Combined Spectral Bands

19. In the Imagine Icon Panel click on the Classifier icon.

20. Choose Unsupervised Classification. The Unsupervised Classification (Isodata)

dialog box should open.

21. Under Input Raster File: (*.img) select p037r022_stacksub.img.

11
DEFORESTATION MAPPING

22. Select directory and filename for Output Cluster Layer (*.img) and Output Signature
Set (*.sig). Call the output files p037r022_class_course.img and
p037r022_class_course.sig.

23. Choose Initialize from Statistics, and set the Number of Classes to 100. Sometimes
it is necessary to choose more than 100 classes. If you get a lot of crossover mixing of
classes (e.g. cloud and developed areas are in the same class), then you might want to
choose more classes. The maximum number of classes you can select is 256. For this
exercise we chose 100 classes because of time constraints. You will probably find it
very difficult to assign a landcover class to 100 classes let alone 256 in just a couple of
hours.

Question 2: Why is the maximum number of classes 256? (Hint: it has to do with the
color display system of computers).

24. Under Processing Options, set Maximum Iterations to 30, Convergence Threshold
to 0.95, and Skip Factors for X and Y should be set to 1 and 1. Do not select the box for
classifying zeros. Click OK to perform the isodata cluster. When the operation begins
the Isodata dialog box will appear, click on the Cancel button to stop it. Classifying
Images, like subsetting, can take several hours depending on the speed of the machine.
We have prepared a classified Image (p037r022_class.img) prior to this exercise to
expedite the process.

Assigning Classes and Determining Change

The results of the unsupervised classification should include an output Image with an
attribute table consisting of 100 classes. The cluster analysis that ERDAS Imagine
performs groups pixels together that have a similar spectral reflectance based on
statistical patterns.

25. In the Imagine Icon Panel click on the Viewer icon.

26. Select File/Open/Raster Layer... Select the Image on which you performed the
classification, p037r022_class.img.

27. Open a second Viewer. In Viewer #2 open both raw Images. Open the older Image
first p037r022_1987.img, followed by the newer one p037r022_1999.img. You will use
these Images as a reference when interpreting the results of the classification.

12
 DEFORESTATION MAPPING

Global Deforestation and Missing Carbon Sinks (Adapted from: Skole, D. L., W. A.
Salas, and C. Silapathong. 1998. Interannual Variation in the Terrestrial Carbon Cycle: Significance
of Asian Tropical Forest Conversion to Imbalances in the Global Carbon Budget. Pp. 162-186 in
Galloway, J. N. and J. M. Melillo, Eds, Asian Change in the Context of Global Change. Cambridge:
Cambridge University Press)

Large-scale commercial deforestation may be an important source for greenhouse gases. However,
studies on carbon concentration in the atmosphere do not really add up. One to two billion metric
tons of carbon that are released each year disappear from the atmosphere. Humans annually
release about 7 billion tons of carbon worldwide. Of that amount, 3 billion tons remain in the
atmosphere, 2 billion tons are absorbed into the oceans and the rest is presumably absorbed into
land vegetation.

Tropical deforestation is an important source of greenhouse gases, but large uncertainties in the
exact magnitude exist due to imprecise data on the rates of deforestation in a few key regions of the
tropics, particularly Southeast Asia. In recent years there has been much progress in acquiring and
analyzing satellite data from Landsat and Spot to better quantify the average annual rate of
deforestation over the last 10-20 years. NASA’s Landsat Pathfinder project, the UN Food and
Agriculture Organization’s Tropical Forest Assessment, and the European Union’s TREES project
have all contributed to improving the measurements of the rate of deforestation using decadal or
semi-decadal estimates. The Landsat Pathfinder Project, for instance, estimated the rate of
deforestation in Southeast Asia between 1973-1985 and 1985-1992 using complete inventories of
Landsat data during each of these periods.

The global budget for sources and sinks of carbon deduced from changes in annual atmospheric
CO2 or O2 measurements suggest that the tropics are neither a strong source nor a strong sink for
carbon. Some researchers suggest that the northern tropics are a carbon source, but the southern
tropics are a small carbon sink. They propose two explanations for these north-south variations: (i)
an increase in tropical net ecosystem production to offset deforestation sources and (ii) evidence for
a combination of reduced rates of deforestation and increased regrowth of previously cleared land,
"which underscore the uptake of carbon by recovering forests on land abandoned by shifting
agriculture and also suggest reduced clearing rates based on satellite analyses.”

Such observations are somewhat inconsistent with interdecadal deforestation data from satellites
which suggest that tropical deforestation is contributing significantly to the net source carbon. One
important explanation for this is that there are significant interannual differences in the rates of
deforestation leading to periods of asynchrony in the relative contribution to the net flux from
clearing vs. regrowth. In this scenario, two important conditions must be met: (i) large interannual
departures in rates of deforestation from the decadal mean values reported by Pathfinder results
and (ii) abandonment of land to secondary forests which accumulate carbon must be an important
part of the deforestation process, occurring over large areas. During years in which deforestation
rates are much less than the decadal mean and there is a large abundance of areas with secondary
succession and forest regrowth due to high rates of abandonment in previous years, we would
expect to see a significant dampening of the carbon source computed from the decadal mean data
alone.

13
DEFORESTATION MAPPING

28. GeoLink the Viewers by right clicking the mouse in Viewer #1 and selecting Geo.
Link/Unlink… Go to Viewer #2 and left click. The Viewers are now linked.

29. In Viewer #2 select Utility/Swipe. Swipe the new Image across the old Image and
look at the difference. Look for obvious change (e.g. large areas of deforestation or
change in colors).

30. Open Raster/Attributes… for the classified Image.

14
 DEFORESTATION MAPPING

31. You will need a column in the attribute table to enter the name of each class. Click
on Edit/Column Properties... Click on the New button and then put the new column at
the top of the list by clicking on Top. Check the Editable box and change the Title: to
Landcover Class. Under Type: select String. Under Display Width: select 10. Click OK.

At this point you can begin examining the classes to determine what type landcover or
feature each one represents. It is important to pay close attention to the raw Images -
look at them, swipe back and forth, and try to identify areas that look the same and
areas that appear to have changed. Many classes will represent the same type of
landcover and will be combined into a single class in the end. Some classes will be
mixed (i.e., different landcover types can have similar spectral reflectance and end up in
the same class).

Ancillary data can be used to assist in the classification. Elevation, roads, rivers,
railroads, population centers, mines and logging concessions are just a few examples of
helpful information. Shapes, colors and sizes are also helpful indicators to use. A
clearing created by deforestation usually looks different from a natural clearing or
wetland.

32. By default, the Raster Attribute table is in grayscale. You can change the colors to
represent the classes you identify. In the Raster Attribute table, right-click on the color
patch of the row you are interested in. A color table should appear and you can select
the appropriate color for each class. After changing the color, you can refer back to the
image to see what pixels changed color.

Alternately, you can look at the classified Image and point and click on a pixel that you
believe to be a certain class. When you refer back to the attribute table the class you
chose will be highlighted. You can then change its color to represent the landcover
class you are interested in. When you are finished assigning classes go to File/Save.

15
DEFORESTATION MAPPING

Question 3: What are the different kinds of changes that you can detect?

Question 4: Are all of these changes the result of human activities? If not, what other
kinds of change can you see?

Question 5: How do you treat areas covered by clouds in one or both of the images?
Are these areas of landcover change?

Recoding of the Landcover Change Map

After assigning a landcover type to all the classes, you need to recode the Image.

Deforestation/Reforestation—How many trees are there really?

From the Webster Dictionary:

Deforestation : the action or process of clearing of forests; also :

the state of having been cleared of forests

Reforestation : the action of renewing forest cover by planting seeds

or young trees

Aforestation : the act or process of establishing a forest especially

on land not previously forested

Loss of forest cover can have many causes ranging from human activities such as logging,
conversion of forest land into agricultural or urban land and surface mining to more natural type
disturbances such as fire, wind and ice damage to forest canopies, flooding and land slides.
Arguably, many of the natural disturbances may increase in frequency and intensity because of
broad scale global changes resulting from human activities. For example, the greenhouse effect
is thought to increase fire frequency and intensity in the boreal region and lead to more severe
storm patterns, both likely to have major effects on global forest distribution.

Similarly, most cleared or deforested lands may convert back either by natural regeneration or
by planned replanting. Sometimes natural forestlands are converted into timber plantations or
even palm oil plantations. However, the value of some of these reforested lands for preserving
biodiversity or other ecosystem services may be lowered by the intensity of management and
resource extraction.

Recoding is a way to combine multiple classes together that are the same landcover
type. The result of this operation will be a classified Image with an attribute table
representing just the number of classes you specified.

33. Click on the Interpreter icon in the icon bar.

34. Select GIS Analysis.../Recode...

35. Under Input File: (*.img) select p037r022_class.img.

16
 DEFORESTATION MAPPING

36. Under Output File: (*.img) select the C:\course directory and name the file
p037r022_recode.img.

37. Click on Setup Recode... The Thematic Recode dialog box will appear.

38. In the New Value column, assign a unique number to each of the landcover classes
(i.e., No Data = 0, Cloud = 1, Water = 2, Non-Forest = 3, Forest = 4, Deforestation = 5,
Reforestation = 6). To select each of the rows that are in the same landcover class, start
by right-click in the Value column. Choose Criteria…from the menu, then create a
Boolean expression to select all “Landcover class” = “forest”, etc. The selected rows
should be highlighted in yellow.

39. Enter a unique number for that class in the New Value box. Then click on the
Change Selected Rows box. When you are finished renumbering all of the classes go to
the Value column and right click, then choose Select None. Click OK.

40. In the Thematic Recode dialog box, check the Ignore Zero in Stats box. Click OK.
The recode operation will begin.

17
DEFORESTATION MAPPING

Question 6: How would you quantify the landcover changes you have observed?

Exporting the Classified Image as an Arc/Info or ArcView Grid

After you have completed the classification, assigned all the classes to a landcover type
and recoded, you need to export the Image. Converting from an Image to a Grid is an
important step in making a satellite Image useful in a GIS.

41. Click on the Import icon on the Imagine icon bar.

42. Select Export, under Type select GRID, and under Media select File.

43. Under Input File: (*.img) select p037r022_recode.img. Under Output File (*.grid)
select p037r022_chng. Click OK.

44. The Export Grid Data dialog box will open. Click OK.

18
 DEFORESTATION MAPPING

Tropical Deforestation in the Rakhine Yoma Mountain Range of Burma

The mechanisms behind deforestation vary considerably with geographic location and
forest ecosystem type, as well as the economy, legal system, culture and even religious
beliefs of local people in the place where the deforestation is taking place. To
demonstrate this point, we have included a satellite Image of a forested mountain range
in Burma in this exercise. These areas are among those we have identified as potential
wildlands that may support wild Asian elephant populations in the southern part of
Myanmar—the Rakhine Yoma Mountain Range.

45. Repeat steps 25-40 and replace p037r022 with p133r048 in all file names.

Question 7: What landcover types can you distinguish in the satellite Image?

Question 8: What are the differences in landcover types between the Images showing
areas in Canada and Images showing the Rakhine Yoma?

Question 9: What kinds of landcover changes can you detect in the Rakhine Yoma and
the surrounding landscape?

Question 10: Which of the two regions (Canada/Burma) suffers from greater
deforestation and forest loss?

Question 11: What are the main differences in deforestation patterns between the two
countries?

19