1
Forest Landscape Change in East Texas: 1974–2009
I-Kuai Hung1,*, Daniel Unger1, Yanli Zhang1, Jeff Williams1, Jason Grogan1,
Dean Coble1, and Jimmie Yeiser2
Abstract - We used Landsat satellite imagery to monitor forest-landscape change within
a 4-county area (Angelina, Nacogdoches, San Augustin, and Shelby) in East Texas. We
consulted images from the Multispectral Scanner (60-m resolution), Thematic Mapper
(30-m resolution), and Thematic Mapper Plus (30-m resolution) from 1974, 1980, 1984,
1987, 1992, 1997, 2002, 2004, 2006, 2007, 2008, and 2009. We classified each image into
1 of 2 land-cover types: forest and non-forest. For data-quality assurance, we assessed 2
of the classified maps for accuracy. We assessed accuracy of the 2002 land-cover map using
field validation (overall map accuracy of 98.46%), and the 2009 land-cover map using
National Agriculture Imagery Program (NAIP) 2009 aerial photos as a reference (overall
map accuracy of 90.77%). To determine forest contagion and fragmentation and their effects
on the local landscape, we calculated landscape metrics including PPU (patch per unit)
and SqP (square pixel) based on landscape patches identified within each classified map.
Results of the 12 classified maps showed a trend of forest-area increase from the 1970s to
the early 2000s. Although this East Texas region supports large and small forest stands,
we observed habitat fragmentation on non-forest lands; both the total number of patches
and total perimeter increased, resulting in smaller patch-size and greater shape complexity
on non-forest lands. These changes are influencing timber production and socioeconomic
activities in the area, as well as the plant community, wildlife habitat, and water resources
of the entire ecosystem.
Introduction
The landscape of East Texas is known for its vast areas of forests that account for
19.6% of the state’s forestlands and 83% of the state’s timberlands that are managed
for timber production (TFS 2012). The majority of the forestlands in East Texas are
Pinus (pine) plantations that repeatedly go through the rotation cycle of site preparation,
planting, thinning, and harvesting. Forests provide different habitats for
plant and wildlife species as the stands grow to maturity and their structure changes.
It is important to understand regional landscape composition at any given time so
that best practices can be applied to manage the land as a complete ecosystem.
Different approaches have been applied to monitor landscape change over time.
Streng and Harcombe (1982) examined local land-use records and aerial photos
taken in 1930, 1941, 1952, and 1956 for the same area to determine why forests
did not develop on 2 East Texas savannas. Glitzenstein et al. (1986) studied disturbance,
succession, and maintenance of species diversity in a forest in the Big
Thicket of southeast Texas through tree-ring chronologies. They were able to
1Arthur Temple College of Forestry and Agriculture, Stephen F. Austin State University,
Nacogdoches, TX 75962. 2University of Arkansas at Monticello, Monticello, AR 71655.
*Corresponding author - hungi@sfasu.edu.
Manuscript Editor: Jerry Cook
Proceedings of the 6th Big Thicket Science Conference: Watersheds and Waterflow
2016 Southeastern Naturalist 15(Special Issue 9):1–15
Southeastern Naturalist
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016
2
Vol. 15, Special Issue 9
reconstruct the stand history dating back to 1800s. For the continuous assessment
of forest recourses, the Forest Inventory and Analysis (FIA) National Program
(McRoberts et al. 2005) established sample plots throughout the nation which are
surveyed periodically. The FIA program has expanded to measure tree species, size,
and health, as well as soils, understory vegetation, woody debris, and lichens in
order to better understand the dynamics of forest ecosystems.
On-the-ground surveys of large areas are costly and time consuming; however,
satellite remote-sensing provides an alternative to monitor landscape change over
time. The first Landsat satellite was launched in 1972, and the imagery produced by
this system has been instrumental in observing the Earth’s resources. Sensors on a
satellite consistently capture images of the Earth’s surface, making it is possible to
detect landscape changes over time. Satellites capture images at different bands of
the electromagnetic spectrum, which are reflected or emitted from the Earth’s surface.
Various surface objects respond to the electromagnetic wavelengths differently
to reveal a unique pattern across different wavelengths known as a spectral signature.
Based on similarities in spectral signatures of surface features, a land-cover map can
be derived from a satellite image by classifying each pixel of the image into a landcover
class (e.g., water, agriculture, forest, urban) by grouping pixels based on their
similar spectral signatures. The historical Landsat-data archive dates back more than
4 decades, and the consistency in sensor specifications allows for analyses across a
time span that no other satellite system provides. Landsat data provide global coverage
and can be acquired free of charge or at a nominal cost.
Landsat data have been used in several projects at the national scale. The National
Land-Cover Database (NLCD) was establiushed in 1992 as a joint effort
of multiple federal agencies to provide baseline land-cover information based on
Landsat satellite imagery for the entire US. The most recent database, NLCD 2011,
is the first one capable of making spatially explicit comparisons to the previous
30-m-resolution database using the same classification scheme (Jin et al. 2013).
Another use of Landsat data is the National Biomass and Carbon Dataset (NBCD).
Using Landsat satellite imagery as a data source, the NBCD provides a baseline
estimate of basal-weighted canopy height; above-ground live, dry, biomass; and
standing carbon-stock for the conterminous US (Kellndorfer et al. 2012). To map
the landscapes of Texas, the Texas Parks and Wildlife Department (TPWD) developed
its own Texas Ecological Mapping Systems Project (TPWD 2014). This
ecological-systems classification incorporated Landsat imagery and data on the
state’s soils, topography, and hydrology. This product is at a finer spatial resolution
(10 m) than the NLCD, and employs 15 land-cover classes and 398 vegetation
classes; the NLCD program has 16 land-cover classes. The Texas A&M Forest
Service developed a forest-distribution map that can be accessed by both forest
industry personnel and the general public at http://texasforestinfo.tamu.edu/map/
fd via the Texas Forest Information Portal. This interactive forest-distribution map
displays the spatial distribution of forest-cover density or biomass at a user-selected
location. The data set was created by processing and integrating satellite imagery
with field-sampled FIA data.
Southeastern Naturalist
3
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016 Vol. 15, Special Issue 9
In 2012, the Texas forest sector contributed $17.8 billion in industry output
and directly employed over 59,400 people (TFS 2012). Of the 5,746,536 ha (14.2
million ac) of timberland in Texas, 83% (estimated volume = 495.5 million m3
[17.5 billion ft3]), is located in East Texas and encompasses 43 counties. Primary
solid-wood products from East Texas, mainly from pine forests, account for 73%
of Texas’s wood-product production. The Forest Resources Institute (FRI) at Stephen
F. Austin State University in Nacogdoches, TX, developed a standardized,
cost-effective, and repeatable remote-sensing methodology to quantify forested
resources, provide baseline information on forests, and promote economic development
in Texas (Unger et al. 2008). FRI created a 2002 baseline land-cover map
of a 4-county area (Angelina, Nacogdoches, San Augustine, and Shelby counties)
in East Texas that was derived from Landsat imagery. This data set had a 72.78%
overall land-cover map accuracy (Unger et al. 2008). FRI further estimated the age
of each resulting forest land-cover class using Landsat satellite imagery in a time
series (1974, 1980, 1984, 1987, 1992, 1997, and 2002) that resulted in a 58.69%
overall land-cover map accuracy for the age by land-cover map.
Forests remain an important natural resource in East Texas; this study temporally
extended the FRI study from 2002 to 2009, and used remote-sensing methodologies
and Landsat data to further quantify forest resources in the same area of East Texas.
However, in this study we particularly focused on how the forest landscape changed
over time through the use of landscape-metrics analysis.
Study-area Description
We selected Angelina, Nacogdoches, San Augustine, and Shelby counties in east
Texas for this study. The study area (total area = 848,237 ha (3275 mi2), is located
within the region known as the Piney Woods of East Texas and it also represents the
westernmost pine forests of the southern US. Timber production is one of the major
agriculture activities in the region, due to its humid subtropical climate with more
than 150 cm (60 in) of annual precipitation. Part of the 2 national forests, Angelina
National Forest and Sabine National Forest, fall within the 4-county study area
(Table 1, Fig. 1). In 2012, Angelina and Nacogdoches counties ranked 6th and 8th,
respectively, among all East Texas counties in terms of their forest-output value,
whereas, in San Augustin and Sabine counties, the forest-based industry was the
largest manufacturing sector in each county (TFS 2012).
Table 1. Land area and population data for the 4-county study area.
Land within
Population national forest
County Land area (km2) Population 2010 density (per km2) administration (%)
Angelina 2242.4 83,810 37.37 25.36
Nacogdoches 2540.1 63,010 24.81 9.51
San Augustine 1537.3 8709 5.66 43.63
Shelby 2162.5 26,610 12.31 31.89
Southeastern Naturalist
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016
4
Vol. 15, Special Issue 9
Methods
We used a series of Landsat images to create land-cover maps. We retrieved
the images from the US Geological Survey (USGS) EarthExplorer online system
(earthexplorer.usgs.gov), and images from different sensors, inlcuding Landsat
Multispectral Scanner (MSS), Thematic Mapper (TM), and Thematic Mapper Plus
(ETM+) data sets. Table 2 shows a list of all images used ranging from 1974 to 2009.
Some of the early images did not cover the entire study area, or a cloud-free scene
was not available. In those cases, we used multiple images for a particular target year.
We geo-referenced all images to the universal transverse mercator (UTM) coordinate
system (Zone 15) with the world geodetic system (WGS) 1984 datum. To match the
spatial resolution of the TM and ETM+ data sets, we resampled the MSS data from 60
m to 30 m. We separately processed each data-band per image using histogram subtraction
to reduce atmospheric haze prior to classifcation (Jensen 2004).
Baseline land-cover maps
We created land-cover maps for the years 2002 and 2009 to establish baseline
information for map-accuracy assessment and forest age-estimation comparisons.
The classified 2002 land-cover map also served as the spatial reference for other
images in order to assure positional integrity on a pixel-to-pixel basis. We classified
vegetation on the the 2002 image into 5 classes—non-forest, pine forest,
hardwood forest, mixed forest, and regeneration—using the ISODATA (iterative
self-organizing data-analysis technique) unsupervised-classification method described
in Unger et al. (2008), where water bodies were coded as non-forest during
the process. The 2009 baseline land-cover map-classification process followed the
Figure 1. Land-cover map (2002) for the Texas 4-county study area with national forest
(NF) boundaries (red).
Southeastern Naturalist
5
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016 Vol. 15, Special Issue 9
same classification procedure to ensure repeatability. We performed an accuracy assessment
on both the 2002 and 2009 baseline maps. For the assessment of the 2002
land-cover map, we ground-truthed a total of 518 reference points. Our budget and
workforce were reduced for the 2009 land-cover map assessment; thus we validated
a total of 520 points using National Agriculture Imagery Program (NAIP) aerial
photos from 2009 (USDA FSA 2015) for reference. We applied a stratified randomsampling
process to select reference locations for validation for the 2002 and 2009
baseline maps. We established an error matrix for each classified map and assessed
accuracy by calculating overall map accuracy, producer’s accuracy for each landcover
class, and kappa statistics for comparison.
Landscape-patch identification
To detect forest-landscape change over time, we reclassified the 5-class landcover
maps for 2002 and 2009 to only 2 classes—non-forest and forest—by
collapsing the original 3 forest classes (pine forest, hardwood forest, and mixed
forest) to 1 forest class, and merging the original regeneration class to the nonforest
class because a return to forest at these sites was not guaranteed. Once we
reclassified both land-cover maps, we recalculated accuracy-assessment statistics.
Table 2. List of Landsat images used for the study.
Year Date Landsat Sensor Resolution (m) Path Row ID Number
1974 10 December 1973 1 MSS 60 27 38 1027038007334490
13 July 1974 1 MSS 60 26 38 1026038007419490
19 August 1974 1 MSS 60 27 38 1027038007423190
14 February 1975 1 MSS 60 26 38 1026038007504590
1980 7 July 1980 2 MSS 60 26 38 2026038008019090
23 August 1980 3 MSS 60 27 38 3027038008023690
6 October 1980 2 MSS 60 26 38 2026038008028090
12 November 1980 2 MSS 60 27 38 2027038008031790
29 November 1980 2 MSS 60 26 38 2026038008033490
1984 17 August 1984 5 TM 30 25 38 5025038008423010
7 December 1984 5 TM 30 25 38 5025038008434210
1987 30 January 1987 5 TM 30 25 38 5025038008703010
25 July 1987 5 TM 30 25 38 5025038008720610
1992 25 November 1991 5 TM 30 25 38 5025038009132910
6 July 1992 5 TM 30 25 38 5025038009218810
1997 25 January 1997 5 TM 30 25 38 5025038009702510
18 June 1997 5 TM 30 25 38 5025038009716910
21 August 1997 5 TM 30 25 38 5025038009723310
2002 15 January 2002 7 ETM+ 30 25 38 7025038000201550
18 July 2002 5 TM 30 25 38 5025038000219910
3 August 2002 5 TM 30 25 38 5025038000221510
28 September 2002 7 ETM+ 30 25 38 7025038000227150
2004 14 December 2004 5 TM 30 25 38 50250382004349EDC00
2006 18 January 2006 5 TM 30 25 38 50250382006018EDC00
2007 6 February 2007 5 TM 30 25 38 50250382007037EDC00
2008 9 February 2008 5 TM 30 25 38 50250382008040EDC00
2009 26 November 2009 5 TM 30 25 38A 50250382009330CHM01
Southeastern Naturalist
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016
6
Vol. 15, Special Issue 9
Assuming the same level of accuracy would be achieved, we reclassified each
satellite-image dataset in the time series from 1974 to 2009, except 2002 and 2009,
to a 2-class (non-forest and forest) land-cover map using the same classification
scheme utilized when creating the 2002 and 2009 land-cover maps.
The pixel value of each 2-class land-cover map then represented 2 possible landcover
classes, non-forest or forest. In order to eliminate island pixels embedded in
each class, we applied a 9 x 9 window to each land-cover image to assign each focal
pixel the majority value based on its 9 pixels x 9 pixels neighborhood (Star and Estes
1990). The elimination of island pixels embedded within a more representative
and homogenous area resulted in a smoother image that better represented the land
cover at the landscape scale.
We identified each group of contiguous pixels (including diagonal neighbors)
with the same pixel value (non-forest or forest) within each 2-class land-cover map;
each group represented a piece of land on the ground of the same cover type. This
group of same-value pixels represented a patch (Forman and Godron 1986). For each
2-class land-cover map in the time series, we identified multiple patches and assigned
a unique pixel value to all pixels of a patch. We calculated the area and perimeter of
each patch through a zonal operation, where each patch was considered a unique
zone, and the calculation was performed by counting all pixels within a zone.
Landscape-metrics calculation
The terms contagion and fragmentation are often used in landscape ecology
to compare landscape compositions at different locations or the same location at
different times. Contagion is the tendency of land-cover to cluster or clump into
a few large patches (Wickham et al. 1996); fragmentation is the tendency of landcover
to break up into many small patches (Forman 1995). In this study, we used
2 landscape-metrics—patch per unit (PPU) and square pixel (SqP)—introduced by
Frohn (1997) to identify landscape fragmentation and patch-shape complexity. PPU
is calculated based on the equation:
PPU = 1 - (m / [n × λ])
where m is the total number of patches, n is the total number of pixels in the study
area, and λ is a scaling constant equal to the area of a pixel. For this study, PPU
was the number of patches per km2. PPU is contrasted to traditional contagion or
aggregation of patches for quantifying landscape clumping and it increases as the
landscape becomes more fragmented.
The other metric, SqP, is calculated as: SqP = 1 - ([4 × √–A] / P)
where A is the total area of all pixels and P is the total perimeter of all pixels in
the study area. SqP is an alternative to traditional fractal dimension for quantifying
patch shape complexity. It is unitless and constrains values to between 0 (for
squares) and 1 (maximum perimeter, edge, deviation from that of a perfect square).
For this study, we used the following alternative form of SqP: SqP = 1 / (1 - SqP) = P / (4 × √A–)
Southeastern Naturalist
7
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016 Vol. 15, Special Issue 9
It gives a scaled perimeter-to-area ratio that ranges from 1 (for square) to infinity
where SqP value increases as a patch shape becomes more complex.
Results
To assess the accuracy of the 2002 five-class land-cover map, we deployed a
ground-measurement team to visit each of the 518 sample plots. This in-situ assessment
resulted in an overall map accuracy of 72.78% and a kappa statistic of
62.51%, which suggested that the accuracy of the map was 62.51% better than
expected by random chance (Table 3). For the 2009 land-cover map, we conducted
the accuracy assessment by comparing each randomly selected sample point to
NAIP 2009 1-m-resolution aerial photos. Compared with the 2002 map, the overall
accuracy of the 2009 land-cover map increased to 78.70% and the kappa statistic
increased to 72.26%. For the 2002 map, the mixed-forest class had the lowest producer’s
accuracy (13.51%). This particular issue was clarified or improved in the
2009 map, where producer accuracy was 65.40%. At the same time, the producer’s
accuracy for the hardwood class also increased from 70.49% for the 2002 map to
88.40% for the 2009 map. However, the producer’s accuracy for the regeneration
class decreased from 100.00% (2002) to 48.80% (2009).
As expected, when we reclassified the land-cover type map from 5 to 2 classes,
both overall map accuracy and its associated kappa statistic increased. Table 4
shows a very high overall accuracy of 98.46% for the 2002 map and 90.77% for
the 2009 map; kappa statistic = 95.66% (2002) and 78.62% (2009). Only the nonforest
class on the 2009 map scored less than 90.00% when we assessed producer’s
accuracy for individual classes.
The total land-area by cover-type showed an overall increase in forest land
from the 1970s to the early 2000s for both land-cover maps classified as 2 classes
(Fig. 2). In 2006, ~75% of the 4-county land area remained forested. When patches
were identified on each of the 2-class maps, we found that the total numbers of
forest patches in the 4-county area decreased over time, with the greatest decrease
in 1984 (1645 patches) and the lowest decrease in 2009 (784 patches). In contrast,
Table 3. Accuracy-assessment statistics for the 5-class land-cover maps of 2002 and 2009.
Producer’s accuracy Overall
Year Non-forest Pine Hardwood Mixed Regeneration accuracy Kappa
2002 87.36% 85.37% 70.49% 13.51% 100.00% 72.78% 62.51%
2009 85.50% 81.60% 88.40% 65.40% 48.80% 78.70% 72.26%
Table 4. Accuracy-assessment statistics for the 2-class land-cover maps of 2002 and 2009.
Producer’s accuracy
Year Non-forest Forest Overall accuracy Kappa
2002 95.87% 99.24% 98.46% 95.66%
2009 84.85% 93.52% 90.77% 78.62%
Southeastern Naturalist
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016
8
Vol. 15, Special Issue 9
non-forest patches increased from ~2000–3000 range during 1970s and 1980s, to
more than 3000 total patches in 1990s (Fig. 3).
Figure 2. Total land-area change of non-forest vs. forest from 1974 to 2009.
Figure 3. Total number of patches by land-cover type from 1974 to 2009.
Southeastern Naturalist
9
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016 Vol. 15, Special Issue 9
Mean patch-size, a combination of total land-area and number of patches,
increased for forest and decreased for non-forest (Fig. 4). However, when we
combined non-forest and forest, the mean patch-size in the 4-county area remained
steady from 1974 to 2009. At any given time, the mean patch-size of forest was
greater than that of non-forest, and the size ratio between forest and non-forest
became larger over time. When we calculated landscape contagion with the PPU
metric, we observed a reversed trend compared with the mean patch-size (Fig. 5).
Figure 4. Mean patch-size by land-cover type from 1974 to 2009.
Figure 5. Patch-per-unit by land-cover type from 1974 to 2009.
Southeastern Naturalist
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016
10
Vol. 15, Special Issue 9
From 1974 to 2009, the PPU values of non-forest continued to increase through
2009 (1.659 patches per km2), while those of the forest decreased through 2007
(0.118 patches per km2). Landscape-contagion values remained steady across time
when non-forest and forest were counted together.
Figure 6, which shows total patch-perimeter by cover type, demonstrated that
the total patch-perimeter for forest decreased over time, whereas that of nonforest
areas remained about the same. When we calculated SqP values, forest and
non-forest showed opposite trends (Fig. 7); the shape of non-forest patches was
Figure 6. Total patch perimeter by land-cover type from 1974 to 2009.
Figure 7. Patch-shape complexity by land-cover type from 1974 to 2009.
Southeastern Naturalist
11
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016 Vol. 15, Special Issue 9
becoming more complex and forest patches formed less-irregular shapes. Forest
patches reduced overall landscape-shape complexity when combined with
non-forest patches in the calculation. This overall landscape change is shown in
Figure 8, which compares a focus area between 1974 and 2009; the total area of
forest increased and the number of forest patches decreased, while the total area
of non-forest decreased and the number of non-forest patches increased, resulting
in smaller patch sizes.
Discussion
Our assessment of total forest land-area change within the 4-county area was
in agreement with what was reported in the Texas Statewide Assessment of Forest
Resources (TFS 2009) as well as with Rosson (2000) and Wear and Greis (2002).
The forest-land recovery from 1970s to 1990s in the 4-county area was part of a
regrowth process of the southern forests following wholesale land abandonment
after extensive logging from 1900s to 1930s. The reversion from agricultural land
to forest also contributed to the total forest-area increase as stated in the 1992 Forest
Resources of East Texas report (Rosson 2002). As a result, forestry remains an
important economic sector in this area; over 75% of the land is currently covered
with forests, which also play an important role in providing goods and services for
the East Texas ecosystem.
During the forest-regrowth process, forest lands aggregated to form larger and
more-homogenous land coverages. Therefore, forest lands not only increased in
total area, but also increased in mean patch-size. On the other hand, the mean
patch-size of non-forest lands became smaller at the same time that its total
area decreased. This situation arose as forests re-established on croplands, thus
Figure 8. Landscape-patch comparison on a focus area between 1974 and 2009.
Southeastern Naturalist
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016
12
Vol. 15, Special Issue 9
causing the mean patch-size of croplands (non-forest) to decrease. Also, the development
of many small urban patches within forested areas also reduced the
mean patch-size of the non-forest lands. This landscape-composition change
altered the habitats of this region, which would likely impact the distribution of
flora, both native and invasive. The increased number of oil-drilling pads within
this 4-county area due to expanded oil exploration in the region in recent years
also contributed to the increase of the total number and decrease of mean patch
size of non-forest patches (Unger et al. 2015). Consequently, we found that nonforest
patches were more fragmented, with PPU values remaining high since 2006
(Fig. 5). This situation also led to the highest shape-complexity for the non-forest
patches observed in 2009 (Fig. 7).
The high percentage of forest cover and increasing forest-patch size created
buffers between non-forest areas, thus improving conditions for species that travel
between forest patches through non-forest corridors. The forests also maintained
the overall landscape contagion and complexity that contributed to the total (nonforest
and forest combined) PPU and SqP values that have remained low since
1974. Although the overall landscape contagion and complexity seemed to be
contained due to the large area of forest-lands within the 4-county area, the fact
that non-forest land cover is becoming more fragmented and complex in shape
should not be overlooked. Species that inhabit non-forest cover-types, such as
shrublands and grasslands, will be confined to smaller and more-isolated patches.
When new housing developments were nested in a forest area, they created a
wildland-urban interface (WUI), where increased human influence and land-use
conversion are changing natural-resource goods, services, and management (Macie
and Hermansen 2002). In recent years, severe wildfires throughout the country have
demonstrated the challenge imposed on the WUI. When a deadly wildfire strikes,
there are losses of human property as well as within the ecosystems upon which
diverse organisms rely. Wildfire suppression can cause fuel accumulation, diversity
alteration, severe insect outbreaks, and other disturbances, even though the total
area of forest remains high.
Using satellite-data–based remote-sensing techniques to quantify forest resources
has proven to be cost-effective and repeatable (Unger et al. 2008). In a similar study
on 6 East Texas counties using Landsat ETM+ data, Sivanpillai et al. (2005) achieved
an overall accuracy (85%) similar to this study. They recommended that mediumresolution
(10–30-m) satellite data can be used to obtain county-level forest-cover
estimates. Landscape-metrics calculations are based on a land-cover map that is classified
from satellite imagery; thus, any error on a land-cover map will propagate to
the landscape metric. An acceptable overall-accuracy standard for land-cover maps
should be established based on a commonly used threshold such as 85% (Anderson et
al. 1976) or 75% (Thomlinson et al. 1999). Also, caution should be taken when comparing
the accuracy of 2 land-cover maps. It is particularly important that the 2 cover
maps are classified using the same methodology and similar initial data sets. For the
2002 land-cover map, the overall accuracy increased from 72.78% for the 5-class
map to 98.46% for the merged 2-class map. For the 2009 map, the accuracy only
Southeastern Naturalist
13
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016 Vol. 15, Special Issue 9
increased 12.07%, from 78.70% for the 5-class map to 90.77% for the 2-class map,
likely due to the very low producer’s accuracy of the regeneration class (48.80%)
(Tables 3, 4). This issue might be related to the 2 different reference-data sets used
for accuracy assessment in this study: ground measurement for the 2002 map and
aerial photos for the 2009 map. When viewing an aerial photo, it is sometimes unclear
if a site with low vegetation will grow to a forest (regeneration) or other crop
(non-forest); it is usually easier to make this determination during ground surveys.
Using high-resolution aerial photos as a reference for land-cover map-accuracy assessment
is a common practice, and it costs much less than ground measurement.
Ground measurements can be conducted only when budgets and time allow. These
surveys must also be completed in a limited time-frame in order to match the time of
image capture. When studying landscape change over time, outdated photographs are
often the only choice for reference data.
When comparing landscape metrics between 2 studies, researchers should verify
that the land-cover map used for landscape-metric calculation from one study has
the same classification scheme as the land-cover map in the other study. In his
study on the Angelina National Forest from 1974 to 1997, Li (2000) reported that
the landscape changed from large clumped patches to smaller, more-fragmented
patches. This result is consistent with our findings for non-forest patches, but not
forest patches. Unlike the single forest-class used in this study, Li classified forested
land-cover into 3 more-detailed categories: pine forest, pine regeneration, and
hardwood forest. Another factor to consider is the perimeter-measurement methodology
of each landscape patch. In this study, we measured patch perimeter along
the outline boundary of a group of pixels with the same value. If the perimeter of
each inner hole (i.e., the perimeter of a non-forest patch nested in a forest patch if
the patch measured is forest) were also counted, the result of the calculated shapecomplexity
metric would be different.
This study demonstrated the use of remote sensing to study forest-landscape
change in an area of East Texas. Based on the landscape metrics calculated, we
found that from 1974 to 2009, the total forest-area and patch size increased, and
the total area of non-forest cover and patch size decreased with patches becoming
more fragmented and complex in shape. This information is relevant to studying
plant communities, wildlife populations, landscape ecology, biodiversity, carbon
sequestration, etc. With the successful launch of Landsat 8 in February 2013, the
Landsat program will continue to provide global coverage-imagery for the study of
the Earth resources. With a well-established classification scheme, forest-landscape
change can be monitored continuously into the foreseeable future as new satellite
imagery is acquired.
Acknowledgments
This research was supported by McIntire-Stennis funds administered by the Arthur
Temple College of Forestry and Agriculture, Stephen F. Austin State University, Nacogdoches,
TX.
Southeastern Naturalist
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016
14
Vol. 15, Special Issue 9
Literature Cited
Anderson, J.R., E.E. Hardy, J.T. Roach, and R.E. Witmer. 1976. A land-use and land-cover
classification system for use with remote-sensor data. US Geological Survey Professional
Paper 964. 28 pp.
Forman, R.T.T. 1995. Landscape Mosaics: The Ecology of Landscapes and Regions. Cambridge
University Press, Cambridge, UK. 633 pp.
Forman, R.T.T., and M. Godron. 1986. Landscape Ecology. John Wiley and Sons, New
York, NY. 619 pp.
Frohn, R.C. 1997. Remote Sensing for Landscape Ecology: New Metrics for Monitoring,
Modeling, and Assessment of Ecosystems. Levis Publishers, Boca Raton, FL. 101 pp.
Glitzenstein, J.S., P.A. Harcombe, and D.R. Streng. 1986. Disturbance, succession, and
maintenance of species diversity in an east Texas Forest. Ecological Monographs
56(3):243–258.
Jensen, J.R. 2004. Introductory Digital-Image Processing: A Remote Sensing Perspective,
3rd Edition. Prentice-Hall, Upper Saddle River, NJ. 544 pp.
Jin, S., L. Yang, P. Danielson, C. Homer, J. Fry, and G. Xian. 2013. A comprehensive
change-detection method for updating the National Land Cover Database to circa 2011.
Remote Sensing of Environment 132:159–175.
Kellndorfer, J., W. Walker, E. LaPoint, J. Bishop, T. Cormier, G. Fiske, M. Hoppus, K.
Kirsch, and J. Westfall. 2012. NACP aboveground biomass and carbon baseline data
(NBCD 2000), USA, 2000 data set. Available online at http://dx.doi.org/10.3334/ORNLDAAC/
1081. Accessed 11 May 2015.
Li, Q. 2000. Monitoring changes in landscape patterns of the Angelina National Forest, East
Texas. Ph.D. Dissertation, Stephen F. Austin State University, Nacogdoches, TX. 81 pp.
Macie, E.A., and L.A. Hermansen. 2002. Human influences on forest ecosystems: The
southern wildland–urban interface assessment. Pp. 26–42, In S.W. Vince, M.L Duryea,
and E.A. Nacie (Eds.).CRC Press, Boca Raton, FL. 312 pp.
McRoberts, R.E., W.A. Bechtold, P.L. Patterson, C.T. Scott, and G.A Reams. 2005. The
enhanced forest inventory and analysis program of the USDA Forest Service: Historical
perspective and announcement of statistical documentation. Journal of Forestry
103(6):304–308.
Rosson, J.F. 2000. Forest resources of East Texas, 1992. Research Bulletin SRS-53, USDA
Forest Service, Southern Research Station, Asheville, NC. 70 pp.
Sivanpillai, R., C.T. Smith, R. Srinivasan, M.G. Messina, and X.B. Wu. 2005. Estimating
regional forest-cover in East Texas using Enhanced Thematic Mapper (ETM+) data.
Forest Ecology and Management 218:342–352.
Star, J., and J.E. Estes. 1990. Geographic Information Systems: An Introduction. Prentice
Hall, Englewood Cliffs, NJ. 303 pp.
Streng D.R., and P.A. Harcombe. 1982. Why don't east Texas savannas grow up to forest?
American Midland Naturalist 108(2):278–294.
Texas A&M Forest Service (TFS). 2009. Texas statewide assessment of forest resources,
2009. Texas A&M Forest Service, College Station, TX. Available online at http://txforestservice.
tamu.edu/uploadedFiles/Sustainable/assessment/Texas%20State%20Assessment.
pdf. Accessed 11 May 2015.
TFS. 2012. Economic impact of the Texas forest sector, 2012. Texas A&M Forest Service,
College Station, TX. Available online at http://tfsfrd.tamu.edu/economicimpacts/
Texas%20Flyer/EconomicImpact2012.pdf. Accessed 11 May 2015.
Southeastern Naturalist
15
I-K. Hung, D. Unger, Y. Zhang, J.Williams, J. Grogan, D. Coble, and J. Yeiser
2016 Vol. 15, Special Issue 9
Texas Parks and Wildlife Department (TPWD). 2014. Ecological-mapping systems of
Texas: Summary report. Texas Parks and Wildlife Department, Austin, Texas. Available
online at https://drive.google.com/folderview?id=0B32g5sG2VKbgUFp6OHc5cDZZU
VU&usp=sharing. Accessed 11 May 2015.
Thomlinson, J.R., P.V. Bolstad, and W.B. Cohen. 1999. Coordinating methodologies for
scaling landcover classifications from site specific to global: Steps toward validating
global map products. Remote Sensing of Environment 70(1):16–28.
Unger, D., J. Kroll, I. Hung, J. Williams, D. Coble, and J. Grogan. 2008. A standardized,
cost-effective, and repeatable remote sensing methodology to quantify forested resources
in Texas. Southern Journal of Applied Forestry 32(1):12–20.
Unger, D.R., I. Hung, K.W. Farrish, and D. Dans. 2015. Quantifying land-cover change due
to petroleum exploration and production in the Haynesville Shale region using remote
sensing. International Journal of Applied Geospatial Research 6(2):1–17.
USDA Farm Service Agency (USDA FSA). 2015. National Agriculture Imagery Program
(NAIP) information sheet. Available online at http://www.fsa.usda.gov/Internet/FSA_
File/naip_info_sheet_2015.pdf. Accessed 24 September 2015.
Wear, D.N., and J.G. Greis. 2002. The southern forest-resources assessment: Summary
report. General Technical Report SRS-54, USDA Forest Service, Southern Research
Station, Asheville, NC. 103 pp.
Wickham, J.D., K.H. Ritters, R.V. O’Neill, K.B. Jones, and T.D. Wade. 1996. Landscape
“contagion” in raster and vector environments. International Journal of Remote Sensing
10:891–899.Climbing Fern
Triadica sebifera (L.) Small Chinese Tallow-tree