See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228783509

An assessment of Landsat TM band 6 thermal data for

analysing land cover in tropical dry forests

Article  in  International Journal of Remote Sensing · February 2004

DOI: 10.1080/0143116031000139917

CITATIONS READS

89 1,909

1 author:

Jane Southworth
University of Florida
132 PUBLICATIONS   5,467 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Changes in land-cover and vegetation and their implications in Ecuador View project

Dynamics and sustainability of the global savanna and grassland biomes View project

All content following this page was uploaded by Jane Southworth on 20 May 2014.

The user has requested enhancement of the downloaded file.

INT. J. REMOTE SENSING, 20 FEBRUARY, 2004,
VOL. 25, NO. 4, 689–706

An assessment of Landsat TM band 6 thermal data for analysing

land cover in tropical dry forest regions

J. SOUTHWORTH
Department of Geography & Land Use and Environmental Change Institute
(LUECI), University of Florida, PO Box 117315, 3141 Turlington Hall,
Gainesville, FL 32611-7315, USA;
e-mail: jsouthwo@geog.ufl.edu

(Received 2 August 2001; in final form 27 March 2003 )

Abstract. This research investigates the utility of integrating Landsat band 6

thermal data for land cover classification and, specifically, for the differentiation
between successional stages of forest growth. Such successful differentiation has
become critical for the assessment of land cover and land use change. The study area
is in the state of Yucatan, Mexico, which is a region of predominantly tropical dry
forest. Thermal band data measures the emission of energy from the Earth’s surface
and, as this is a function of the surface cover, it can be used as a determinant of land
cover type based on the temperatures measured. This research uses a technique of
land cover classification utilizing Landsat Thematic Mapper (TM) thermal band
data. Land cover classes were found to relate strongly (an eta correlation ratio of
0.804) to band six calculated black body temperatures. Satellite coverages of black
body temperatures also provided a unique spatial analysis of the image area using
continuous data, not discrete classes. Discriminant analysis revealed that band 6
contains considerable information for the discrimination of land cover classes in the
dry tropical forest ecosystem, with a coefficient of 0.354. Band 6 also outperformed
many of the individual spectral bands of information, despite their greatly improved
(616) resolution.

1. Introduction
Gaining a better understanding of the ways that land cover and land use practices
evolve is a primary concern for the global change research community. Changes in land
cover affect ecosystems function, biodiversity and climate. The ability to project future
states of land cover even over local scales, due to anthropogenic or natural causes, is a
requirement for making predictions about other potential changes which may occur
globally (Turner et al. 1993, Hansen et al. 2000).
In tropical regions, many land cover changes of ecological and climatic signi-
ficance are currently taking place, such as colonization of marginal lands, deforesta-
tion, drylands degradation, landscape fragmentation and rapid urbanization (Lambin
1999). Although tropical dry forests cover more area than humid forests, and are
greatly affected by human activity, there have been few studies of their functions,
structure and processes (Whigham et al. 1990). Such analyses require repetitive
surveys and in this context remote sensing can be a powerful tool (Estève et al.
1998, Hansen et al. 2000). In addition, these issues need to be monitored and
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online # 2004 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/0143116031000139917
690 J. Southworth

studied across a range of temporal scales and across local, regional and global
spatial scales (Hall et al. 1991).
One of the major advantages of using Landsat Thematic Mapper (TM) band 6
data is that this information, when converted into temperatures, can be used to link
directly to other processes (e.g. micrometeorological). Land surface processes are
of paramount importance for the redistribution of energy and moisture between
the land and the atmosphere. Surface temperature is an important parameter in
the characterization of energy exchange between the ground surface and the
atmosphere. Landsat TM thermal band data can be used for this purpose (Hurtado
et al. 1996). These exchanges of radiative, moisture and heat fluxes affect biosphere
development and the physical living conditions at the Earth’s surface (Hall et al.
1991, Bastiaanssen et al. 1998). Due to the spatial variations in land cover, soils and
moisture, representative meteorological conditions cannot always be obtained from
a limited number of climate stations. In such instances, remote sensing data provide
an excellent means of obtaining consistent and frequent observations of spectral
reflectance and emittance of radiation of the land surface on a micro to macro
scale. As such, remote sensing provides an instantaneous snapshot of the radiative
properties of the particular land surface (Roy 1997, Bastiaanssen et al. 1998).
Hence, with the advent of thermal remote sensing, it has become feasible to
undertake detailed analysis of landscape-scale surface temperature variability.
Many of the world’s remaining tropical forests exist in relatively inaccessible
regions where it is not feasible to study forest canopy thermal budgets thoroughly
using conventional micro-meteorological techniques (Luvall et al. 1990). In such
locations, remote sensing, and thermal remote sensing in particular, can serve as an
excellent information source.
Spatial information on existing land uses is important for the analysis of current
environmental conditions and to estimate and determine future land use activities.
Traditionally, land cover has been used as the principal surrogate for land use when
remotely sensed data have been used, although the two are very different. Accurate
analysis of land cover characteristics using remotely sensed data is a prerequisite for
determining land use, which is a necessary input for land use/land cover change
(LULCC) models. The use of Landsat 5 TM data for land cover classification of
vegetation types has become increasingly successful (Mausel et al. 1993, Li et al.
1994, Hansen et al. 2000) and with the launching of Landsat 7, continued
monitoring and higher resolution thermal and panchromatic band information are
now also available. The ability to monitor and measure changes in land cover
within these areas is critical to addressing questions concerning changing climate
and other issues pertaining to deforestation in many tropical regions (Li et al.
1994). In addition, we can measure land-surface temperature, which is one of the
most important land-surface parameters (Sobrino et al. 1991, Wittich 1997) due to
its effect on energy and moisture cycles. This research is concerned with monitoring
land cover, as derived from remotely sensed images.
The Landsat TM thermal band, which is designated as TM band 6, is sensitive
in the region 10.4–12.5 mm. It has lower radiometric sensitivity, and coarser spatial
resolution (120 m6120 m), relative to the reflective TM bands (approximately
30 m630 m). Most studies on land cover change using satellite data have not used
band 6 thermal data (Trisurat et al. 2000, Sader et al. 2001). Thermal data spatial
resolution of 120 m6120 m is advantageous when wishing to undertake a more
broad-scale classification scheme, as proposed here. The Landsat TM band 6
responds to varying temperatures and emissivities on the ground. This produces
 Landsat TM band 6 data for analysing land cover 691

images that show the relative differences in emitted thermal energy that correlate in
part with the effects of solar heating on surfaces of varying composition and
orientation. Landsat TM thermal imagery (despite its lower spatial resolution) both
supplements reflected radiation images and stands alone as an information source
(Short 1996, Roy et al. 1997). Compared to Landsat 5, Landsat 7, which is now in
orbit, has a band 6 resolution of 60 m660 m, which is a tremendous improvement
in resolution. This provides even greater reason for researchers to use this infor-
mation and incorporate surface temperature analyses into land cover classifications.
However, in order to undertake time-series analysis going back before 1999 (and
hence pre-Landsat 7), utilizing band 6 thermal information, one must also use
Landsat 5 TM band 6 data, as is illustrated here.
The thermal band, like vegetation indices, allows for continuous representation
of land cover, rather than the creation of discrete land cover classes such as in
classification techniques (Lambin 1999). This conventional allocation of each pixel
to a discrete land cover class is an issue of concern in land cover analyses,
particularly as it may frequently be more appropriate to model continuous
variations in land surface properties (Foody 1996, Andrefouet and Roux 1998,
DeFries et al. 2000). Land cover changes can occur in two forms: conversion of
land cover from one category to a completely different category (via agriculture,
urbanization, etc.), or modification of the condition of the land cover type within
the same category (thinning of trees, selective cutting, etc.) (Meyer and Turner
1992). The monitoring of land conversion from one category to another is well
documented and more easily undertaken than the monitoring of change within a
category. This is problematic as land cover modifications may be as significant
environmentally as land cover conversions (Lambin 1997, Foody 2001). As such, in
addition to more traditional land cover classifications it is also desirable to address
within-class variability of a given land cover type and also to evaluate such within-
class changes over time. This will allow for the analysis of forest degradation or
regeneration, even if the land cover class remains the same. As such, the thermal
band allows the monitoring of ‘within-class’ differences in a vegetated area because,
even when a new land cover class is not created, differences across a surface will be
evident from the temperature values (Roy 1997, Roy et al. 1997, Lambin 1999).
This is important, for example, when trying to determine differences in forest type,
age or structure.
The relevance of satellite derived surface temperature values for studies of forest
microclimate depends largely on the relationship between temperature at the
canopy and that within the forest. Satellite derived temperatures correspond to
surface temperature and in a forested area this means the temperature at the top of
the canopy (Nichol 1995). Within this dry topical forest region in the Yucatan,
Mexico, a lower temperature means a denser and more mature forest. This
relationship does not hold true for all forest types. For example, some tropical
moist forest areas have secondary successional forests which are denser, and hence
cooler, than the mature forest. So within a dry tropical forest environment the
canopy effectively regulates canopy temperature, which is a useful indicator of
forest microclimate (Nichol 1995). Hence, the differentiation of forest age or
successional class should be possible using band 6 data for this region.
The study area for this research is located in and around the city of Ticul, within
the state of Yucatan, Mexico. This region is dominated by tropical dry forest
vegetation, porous soils, exponential population growth and high rates of
deforestation (Sohn et al. 1999). The primary objective of this research is to
692 J. Southworth

increase our understanding of and our ability to use thermal band data in land
cover studies. More specifically, this research: (1) assesses the effectiveness of
thermal band information for vegetation studies of successional change; (2)
incorporates the use of band 6 information into studies which currently use only
data from the reflective bands; (3) evaluates the additional information provided by
the thermal band for vegetation studies; and (4) compares the effectiveness of band
6 versus other information using both statistical and image based techniques for use
in remote sensing of land cover.

2. Methods
2.1. Study area
The study area is centred around the city of Ticul located in Yucatan state, in
Mexico. The vegetation is dry tropical forest. Along the coasts there are mangrove
swamps, grassy shrublands, and tules (marshlands). In the northern portion of the
peninsula, where it is drier and the soil is porous, the tropical vegetation ranges
from scrublands to forest. Further south the vegetation becomes taller, with fewer
of the trees being deciduous. The peninsula is based on calcareous rock and there
are no surface rivers, but rather there are many cenotes (large waterholes) and
underground rivers connecting these together. Rainfall in the region is heaviest
during the summer months but does occur year round, and ranges from 500 mm
along the coast to 1500 mm in the southern portions of the state. Mean tem-
peratures remain high throughout the year, ranging from 20–30‡C.
The image used in this analysis was taken on 27 March 1995 and is from the
Landsat 5 TM (row 20, path 46). Meteorological data from local climate stations
are used to record large-scale temperature contrasts and precipitation events, which
may result in localized flooding or pooled water. Localized precipitation events are
much more difficult to determine from these (or any) data and are also spatially
distinct, occurring over a limited area for short durations. No areas of precipitation
were discerned across the study area for the week preceding the satellite image date.
In addition, the image was taken at the end of the dry season, so surface effects
relating to soil moisture differences are limited, as is dense undergrowth. Hence, the
signal obtained by the satellite relates mostly to the canopy characteristics. Climate
data reveal mean temperatures through the month of March of 21–27‡C with dew
points around 10–15.5‡C. The last measurable large-scale precipitation event
occurred on 17 March and varied in amount from 5–15 mm across the study area.
Hence, the image for 27 March 1995 is ideal for this analysis as no large-scale
precipitation event had occurred within the preceding 10 days. Precipitation can be
a problem in thermal image analysis as differences in temperatures may relate to
soil moisture differences, rather than vegetation differences.

2.2. Image conversion and calibration

Image conversion and calibration was undertaken for both the thermal band
and the reflective bands so as to eliminate as sources of variability and noise,
differences due to satellite instrumentation, Earth-Sun distance, solar elevation
angle, solar curve and atmospheric effects (Chavez 1996). This is especially impor-
tant for time-series analysis, but also when comparing across different satellite
sensors.
 Landsat TM band 6 data for analysing land cover 693

2.3. Geo-referencing
Geo-referencing registered the image into Universal Transverse Mercator
(UTM) coordinates, using 1 : 50 000 scale topographical maps of the study area,
thus allowing cross-referencing of images across space and time. Root mean square
(rms) error of v0.5 pixels was recorded, indicating a high degree of accuracy.

2.4. Classification
Supervised classification using the Gaussian Maximum Likelihood technique
was used to create seven classes: (1) urban/roads, (2) sink holes, (3) water/shade, (4)
bare/soil, (5) irrigated agriculture, (6) early–mid successional forest, and (7) mid–
late successional forest. Training sample data for each land cover class obtained
during the summer of 1998 were used to provide a detailed accuracy assessment of
the results. For each training sample the land cover class, mean vegetation height,
composition, structure, canopy closure and location are recorded. Of the 120
training sample points, 60 were used to create the initial land cover signatures, and
60 were used in the accuracy assessment. From these data the accuracy of the
created classified image was evaluated. The results of this accuracy assessment for
both of the classifications, the first using bands 1–5 and 7, and the second
classification using all seven bands were identical, with an overall classification
accuracy of 96.1% and an overall kappa index of 0.95. While separately collected
datasets are ideal (Jensen 2000), the training sites used in this analysis were
randomly located across the image prior to arrival on-site. As such, this random
location of training samples obtained does help refute the need for two separate
training and reference site datasets (Jensen 2000). In addition, field data and classi-
fication results were also obtained from fellow researchers working in this region
(Sohn et al. 1999), which allowed for an additional verification of the land cover
classifications created.

2.5. Surface temperatures of land cover classes

During fieldwork an additional measure of surface temperature was taken.
While these training samples were gathered in a different year and season than the
satellite image, comparison of the range of values obtained for each land cover
class, to enable a fieldwork to satellite comparison, was desired. In addition, the
field measured temperatures were all standardized to around 10:00 a.m. local time
in order to compare more closely with the satellite measures. Again, although the
absolute values measured are not directly comparable, the pattern of temperature
values across land cover classes can be compared for the field and satellite images
(table 1).
For the satellite image the band 6 raw data were converted to at-sensor
apparent or black body temperatures (Ta in K). These temperature values do not
account for the different land cover types. In order to determine an actual surface
temperature it is necessary to know the surface material and from this determine its
emissivity value. Using these black body temperatures and the known surface
emissivity (e) values, the actual surface temperatures (Ts in K) can be calculated:
esTS4 ~sTa4
where s is the Stefan Boltzmann constant (5.6761028 W m22 K24) and emissivity is
an average value for the land cover class (Nichol 1995). Using this formula, surface
temperatures for the study area can be obtained. Hence, the actual surface
694 J. Southworth
Table 1. Comparison of surface temperatures for each land cover class for image and field
based measures based on training sample data locations only.

At-satellite Land cover Calculated actual

Measured field black body emissivity surface
Class temperature* (‡C) temperature (‡C) (Oke 1987) temperature (‡C)
Water N/A 22.78–25.00 0.97 22.61–24.81
Agriculture N/A 25.56–25.90 0.94 25.17–25.50
Mid-late succession 31.67–33.89 26.11–27.22 0.99 26.04–27.15
Early-mid succession 33.89–35.56 27.78–28.89 0.98 27.64–28.74
Bare earth 35.56–43.33 29.44–32.78 0.95 29.06–32.36
Urban/roads 41.11–48.89 33.33–36.11 0.92 32.64–35.37

*Field measures of temperatures were obtained to determine whether the pattern of values
was similar to the calculated surface temperatures using the satellite data. However, these
field measurements were taken in a different year and in a different season; thus, the values
do show the same pattern but are not the same as the satellite measured values.
N/A indicates field samples with insufficient information to obtain surface temperatures,
e.g. access to site not available and so direct surface temperature measures were not obtained.

temperatures were calculated for the land cover classes based on the training sample
data (table 1). This allows for the comparison of the field and satellite based
measures. As can be seen from table 1, the inclusion of the emissivity to calculate
the actual surface temperature does not change the pattern of temperatures for each
land cover class, but it does act to accentuate the differences. Aside from this
analysis, which was used to verify the field versus satellite measures, the remainder
of this research uses only the black body temperature values. If emissivity values
based on land cover were used to calculate actual surface temperatures then the two
datasets to be compared (land cover and temperature) would already be correlated
due to the inclusion of land cover type in the calculation of actual surface
temperatures. In addition, as this research aims to highlight the usefulness of band
6 data as a stand-alone dataset (in addition to its usefulness when used in
combination with the reflective bands) only black body temperature images were
used. This analysis therefore allows us to evaluate the land cover classes, as
determined from the reflective bands and fieldwork, in terms of their biophysical
and meteorological properties, using band 6 as an independent data source.

2.6. Vegetation indices

Studies conducted on temperate forests have successfully related biophysical
properties with red and near-infrared radiance, particularly with the Normalized
Difference Vegetation Index (NDVI). NDVI has also been strongly correlated with
land cover types in urban, suburban and natural ecosystems (Lo et al. 1997). Boyd
et al. (1996) suggested that the relationship between forest biophysical properties
and middle and thermal infrared radiance may be stronger than that between forest
biophysical properties and visible and near-infrared radiance. Hence, vegetation
indices created for this analysis were the NDVI, a tasselled cap transformation
analysis and a number of thermal based vegetation indices.
The NDVI is created through a ratio of the TM bands 3 and 4 using the
equation: NDVI~(TM42TM3)/(TM4zTM3). NDVI was one of the first
successful vegetation indices created based on band ratioing. However, in many
tropical environments, researchers have found less than satisfactory results between
NDVI and ground based measures of vegetation (Boyd et al. 1996). Since the
 Landsat TM band 6 data for analysing land cover 695

development of this initial vegetation index, many more indices have been created
and researchers have attempted to relate these indices to different land cover or
vegetation characteristics. One such index is the tasselled cap transformation
developed by Kauth and Thomas in 1976, originally created with Landsat MSS
data (Jensen 2000). The analysis identifies six new axes (using Landsat TM data
bands 1–5 and 7). The first two axes represent soil brightness as a ‘brightness
index’, and green vegetation as a ‘greenness index’. The third axis is believed to
relate to moisture and the fourth to haze, but these have only been tested in a
limited capacity with TM data (Jensen 2000). Hence, in this analysis most of the
attention is on the first two bands created: brightness and greenness.
In addition to these more commonly used vegetation indices, a number of less
common or newly created indices also were used. All of these indices used the black
body temperature image data (referred to as TM6). The vegetation indices are: (1)
TM6/(TM3zTM5); (2) (TM22TM6)/(TM2zTM6); (3) (TM2*TM6)/TM7; (4)
TM6/NDVI; (5) GREENNESS*TM6; and (6) (B4*B7)/B6. The first three of these
indices were taken from Boyd et al. (1996) and the latter three were created for this
analysis. These latter three equations were based on knowledge of the reflective
patterns of vegetation and hence are the most useful bands to include within the
analysis. For tropical forest analysis, NDVI, tasselled cap analysis, band 4, and
bands 5 and 7 frequently outperform other reflective bands or indices (Boyd et al.
1996). In this research band 6 was simply added into these indices. All of these
different vegetation indices were evaluated in terms of their relationships with and
discrimination of the land cover classifications.

2.7. Statistical analysis of the land cover classification

Prior to any statistical analysis the band 6 data had to be reformatted so as
to have the same spatial resolution as the reflective bands. Band 6 pixels are
120 m6120 m, whereas the reflective bands are 30 m630 m. Hence, each band
6 pixel was subdivided into 16 individual 30 m630 m pixels, with the value recorded
in each of the 16 smaller pixels being identical such that the data can be compared
statistically but the actual band 6 values have not been altered.
The dependence of surface temperature on the magnitude of biomass can be
evaluated statistically by correlating the thermal band image data with any vege-
tation index, such as NDVI or the tasselled cap transformation variables, such as
greenness. The image bands were converted into an ASCII dataset, which was then
imported into a statistical package for analysis. Due to the land cover classes being
categorical data and the band information and vegetation indices being continuous
data, Pearson’s correlation coefficient (r) could not be used. In addition, there was a
definite question of whether the relationships across bands, classes and indices
could be assumed to be linear. Given these issues the eta correlation ratio was used
in this analysis. Eta is a coefficient of nonlinear association, which for linear
relationships equals the Pearson’s r correlation coefficient. For nonlinear relation-
ships it is greater. Eta2 is the percentage of variance in the dependent variable
explained linearly or nonlinearly by the independent variable. When categorical
data are used (e.g. land cover classes) these data must be the independent variable.
One limitation of eta is that it cannot prove causal direction and, hence, eta has no
sign and varies from 0 to 1.0 (Garson 2001).
A correlation matrix was calculated and this illustrates the interrelationships
between the reflected bands (bands 1–5 and 7). Thermal band data are gathered
696 J. Southworth

from a different instrument, making the relationships with band 6 statistically

meaningful since they are not merely a function of the measurement process.
Having an independent data source is of great importance within many statistical
analyses. In addition, the information from the seven bands of Landsat TM data
was correlated with the main vegetation indices discussed, as well as with the
classification analysis results, to evaluate the strengths of the various relationships.
A discriminant function analysis was used to determine the relative importance
of the different variables in terms of discrimination of the land cover classes. The
Wilks’s lambda coefficients were used to determine the relative importance of the
different information layers as discriminating variables of land cover. The magni-
tude of the lambda coefficient (0.0–1.0) is inversely proportional to the ability of an
information layer to discriminate between the classes (Boyd et al. 1996).

3. Results
3.1. Classification
The addition of band 6 data did not improve the classification accuracy
statistically, but did improve the overall appearance of the classified image by
decreasing graininess (figure 1(a) and (b)). Figure 1 clearly illustrates the differences
in resultant land cover as a result of the two classification techniques. While the
overall patterns are very similar across the image, the classification using band 6
(figure 1(b)) does have a loss of some detail, specifically some linear features such as
small roads etc. In addition, this classification results in more bare soil pixels and
fewer early to mid-succession pixels. This is due to the very similar temperature
values between bare soil (30–30.9‡C in figure 1(c)) and the early succession class
(28.0–29.9‡C). Hence, inclusion of band 6 data within the classification results in a
different boundary between these two classes. By looking at the temperature
coverage (figure 1(c)) one can easily determine this pattern. Aside from the dif-
ferences in the bare soil and early succession class, however, the remaining classi-
fications are very similar. In addition, figure 1(c) shows how well the temperature
coverage compares visually with land cover.
Adding band 6, while not affecting the overall accuracy of the classification
procedure, smoothes the surface in a manner similar to running a 363 filter pass
over the image. Different filters are frequently applied by researchers to improve
image accuracy or appearance (Sader et al. 2001). Individual pixels of a given class
may well be misclassified or alternatively the class of interest may be a set size and
applying a 363 or a 565 majority filter will remove pixels with size classes below
those desired. For example Sader et al. (2001) used a 565 majority filter after
classification to eliminate isolated pixels and to consolidate the boundaries of the
cleared patches. Since many post-classification methodologies do include such a
filter pass as an extra step after classification, the inclusion of band 6 in the classi-
fication procedure itself allows us to incorporate this step into a single method. In
addition, rather than applying a statistical filter with no relevance to the data being
smoothed, one can use an independent data source, at a coarser resolution, which
represents an additional source of information. The inclusion of band 6 is useful
and relevant as these data relate directly to land cover.

3.2. Surface temperatures

In the study area distinct spatial variations in thermal characteristics are evident
from the black body temperatures and are clearly observed to correspond to land
 Landsat TM band 6 data for analysing land cover 697

Figure 1. Image classification of the study area using (a) bands 1–5 and 7, (b) all 7 bands,
and (c) black body surface temperatures from band 6 only.

cover differences (figure 1). Mature forest has the lowest temperatures and cleared,
bare soil has the highest temperatures of the natural surfaces, with urban/roads
having the highest temperatures overall. In addition, when including the effect
of surface emissivity to calculate surface temperatures, these differences become
even greater across land cover types (table 1). Temperatures collected during the
698 J. Southworth

fieldwork analysis also showed similar temperature trends for the land cover classes.
This is important in the analysis and verification of the temperature coverage
(figure 1(c)). As can be seen in table 1, the relationship between land cover and
temperatures follows a pattern of the highest temperatures relating to urban and
roads, followed by bare earth and then early-mid succession. As temperatures
continue to decrease we find mid-late succession, intensive agriculture (in this region
this is frequently irrigated, hence the lower temperatures) and water. These
temperature patterns were constant across these land cover classes, which were
measured for each training sample. Hence, land cover types clearly responded to
specific temperature ranges, and the differentiation of forests based on successional
stage is feasible. The spatial variability of temperatures across the image (figure 1(c))
clearly illustrates the utility of this temperature based information layer. In addition,
the close relationship between temperatures and land cover class (table 1, figure 1)
highlights the importance of the information in this image band.
When comparing the classification undertaken using bands 1–5 and 7 (figure 1(a))
versus the temperature coverage (figure 1(c)), there is a noticeable similarity in the
coverages. This is of particular interest given the amount of information available
for the classification procedure: bands 1–5 and 7 not only contain 16 times more
information due to their 30 m630 m resolution (compared to band 6 resolution of
120 m6120 m), but also all six bands are used in combination for the analysis
(compared to only one band for the band 6 temperature coverage).

3.3. Statistical analysis of different band and vegetation indices contributions to

the classification analysis
3.3.1. Correlation analysis
A correlation matrix (table 2) computing eta or correlation ratios represents the
percentage of variance in the dependent variable explained linearly or nonlinearly
by the independent variable. Temperature values are most strongly correlated with
NDVI (0.893), band 7 (0.885), band 5 (0.875), band 3 (0.864) and land cover class
(0.804), all of which are very strongly related to vegetation. Given the differences in
band resolution of 16 reflective pixels to every band 6 pixel, these relationships are
impressive. There are a large number of data points included in the statistical
analyses (the study area was converted from a raster image to an ASCII file with

Table 2. Correlation matrix of the reflective bands, temperatures, tasselled cap and NDVI
vegetation indices, and land cover class.

Band Band Band Band Band Band

Bright Green Temp 1 2 3 4 5 7 NDVI Class
Bright 1.00
Green 0.531 1.00
Temps 0.646 0.840 1.00
Band 1 0.855 0.739 0.737 1.00
Band 2 0.908 0.541 0.519 0.870 1.00
Band 3 0.903 0.904 0.864 0.900 0.875 1.00
Band 4 0.294 0.929 0.732 0.612 0.285 0.810 1.00
Band 5 0.908 0.840 0.875 0.839 0.788 0.927 0.729 1.00
Band 7 0.895 0.919 0.885 0.856 0.801 0.945 0.830 0.977 1.00
NDVI 0.943 0.998 0.893 0.902 0.910 0.990 0.993 0.931 0.954 1.00
Class 0.704 0.904 0.804 0.693 0.604 0.803 0.862 0.834 0.852 0.905 1.00
 Landsat TM band 6 data for analysing land cover 699

more than 100 000 data points) and all relationships discussed here are statistically
significant at the 0.99 level of significance.
The correlation matrix also reveals the relationships between the TM bands, the
tasselled cap transformation images of brightness, greenness, moisture and haze,
NDVI, and the land cover classes. A correlation of 0.998 was calculated between
the tasselled cap transformation greenness index and the NDVI, indicating that for
this tropical dry forest ecosystem these two separate indices are near identical. Land
cover class is most strongly correlated with NDVI/greenness (0.905/0.904), band 4
(0.862), band 7 (0.852), band 5 (0.834), temperatures (0.804) and band 3 (0.803). It
is least correlated with band 2 (0.604), band 1(0.693) and brightness (0.704). Again,
these results indicate that the bands associated most with vegetation cover (bands 4,
7 and 5) or with biomass (NDVI and greenness) perform the best. What is
impressive is how well temperatures also relate to land cover class despite the 16 : 1
resolution difference.
A comparison between the NDVI and temperatures indicates a close spatial
correspondence between temperatures and land cover type. The dependence of
black body temperatures (and hence surface temperatures) on biomass was tested
by regressing the band 6 temperature image data with the NDVI image data. A
regression of NDVI and black body temperatures results in a negative relationship,
with increasing vegetation amounts producing a higher index value, and corres-
ponding to cooler temperatures (figure 2). An adjusted Pearson’s r2 correlation
coefficient value of 0.70 was calculated between temperatures and NDVI. This
value increases to 0.893 if we test for a nonlinear relationship. This relates well
to previous research (Nichol 1995). The pattern of high NDVI equals lower
temperatures. As NDVI decreases and temperatures increase, the pattern of land
cover changes from intensive agriculture, to mature forest, to mid-early successional
forest, and then to soils or bare earth. Urban/roads have the lowest NDVI values
and the highest temperatures (figure 2).
A comparison of some of the vegetation indices illustrates the location of tem-
peratures within the NDVI/brightness space (figure 3(a)) and within the brightness/
greenness space (figure 3(b)). Within the analysis of NDVI and temperatures, as
discussed above, there is now included a third variable. Hence, a high NDVI and
low brightness value both indicate lower temperatures. As brightness values
increase, temperatures increase. In addition, higher values of greenness indicate
lower temperatures. These perspectives (figure 3) allow us to determine more
accurately the existing relationships, and also allow us to see the value of the
temperature analyses, which so plainly differentiates the land cover type.

3.3.2. Wilks’s lambda coefficient

For all the indices and band information used in this analysis, the relative
strengths of the relationships between land cover classes and the different infor-
mation layers were evaluated. The Wilks’s lambda coefficients (table 3), as deter-
mined from the discriminant analysis, are used to determine the relative importance
of the different bands of information and vegetation indices as discriminating
variables for land cover analyses. The magnitude of the lambda coefficient obtained
is inversely proportional to the ability of the information layer to discriminate
between the land cover classes (Boyd et al. 1996). Here the rank order of the
individual information bands and indices (table 3) illustrates that data acquired by
band 6, while not the most significant of all the information layers, is one of the
700 J. Southworth

Figure 2. NDVI versus black body temperatures with colours representing land cover classes.

more important, with a lambda coefficient of 0.354. Looking at the performance of

the individual bands (table 3) bands 4 (0.257), 7 (0.274) and 5 (0.305) perform the
best, followed by temperatures (0.354) and band 3 (0.355). Bands 1 (0.519) and 2
(0.635) perform quite poorly. The most significant information layers are NDVI
(0.181), greenness (0.183) (these first two indices are highly correlated and hence are
really measuring the same thing) and greenness*TM6 (0.188, here including band 6
data actually decreases the performance of greenness alone). The remaining vege-
tation indices TM6/(TM3zTM5) (0.270) and (TM2*TM6)/TM7 (0.279) both
perform very well, with brightness (0.504), TM6/NDVI (0.556), (B4*B7)/B6 (0.559)
and (TM22TM6)/(TM2zTM6) (0.743) all having very weak relationships.
It is also of interest to note that the vegetation indices created here frequently
outperform the individual Landsat TM bands in classifying land cover. It appears
from this research that the incorporation of the individual TM band 6 into
vegetation indices acts to enhance their information content and thus increase the
strength of the relationships with land cover. For example, bands 3 and 5 are not
among the highest bands for discriminating land cover when viewed individually
but when used in the vegetation index with band 6 (TM6/TM3zTM5) becomes one
of the best discriminating indices (0.270).
 Landsat TM band 6 data for analysing land cover 701

Figure 3. Vegetation indices with (a) NDVI versus brightness, and (b) brightness versus
greenness, with colours representing temperatures.

4. Discussion
Satellite based thermal analysis offers an accurate, useful, novel and viable
means of monitoring forest environments and estimating spatial aspects of the
forest energy budget. The inclusion of band 6 data within vegetation indices can
increase the strength of the relationships between radiance values and land cover
classes of dry tropical forests (as for the index TM6/TM3zTM5). Therefore, the
potential for incorporating band 6 data into land cover analyses is great.
Specifically, the incorporation of band 6 into actual classification procedures and
vegetation indices is recommended both using Landsat 5 and 7 satellite products.
The data obtained using thermal band information from Landsat 5 is at a
resolution 16 times lower than that of the reflective bands, but this single band
provides an immense amount of information, as illustrated in this research. Other
reflective bands must be used in combination to maximize accuracy and usefulness.
702 J. Southworth
Table 3. Results from the discriminant analysis, to determine the ability of each information
layer to distinguish land cover class.

Wilks’ lambda
Individual bands
Band 4 0.257
Band 7 0.274
Band 5 0.305
Temperatures (Band 6) 0.354
Band 3 0.355
Band 1 0.519
Band 2 0.635
Vegetation indices
NDVI 0.181
Tasselled cap: greenness 0.183
Greenness*TM6 0.188
TM6/(TM3zTM5) 0.270
(TM2*TM6)/TM7 0.279
Tasselled cap: brightness 0.504
TM6/NDVI 0.556
(B4*B7)/B6 0.559
(TM22TM6)/(TM2zTM6) 0.743

Band 6 temperature data is also a useful stand-alone data source. As illustrated

here (figure 1(c), table 1) the different land cover classes have specific temperature
ranges (both for field based and satellite measures) and this results in a powerful
image analysis tool for land cover differentiation. Specifically, forest age or
successional stage is evident from such temperature surfaces, and such data are
invaluable in many global change studies, such as carbon sequestration research
where forest age is a major determinant of carbon storage.
Lambin and Ehrlich (1996) in their research on Africa argued that the use of
such remotely sensed variables as measurements of thermal infrared radiation or
derived surface temperatures may be as important as the many reflectance based
vegetation indices used in much of today’s research. Indeed, they found that the
combination of thermal band information with vegetation indices improved the
mapping and monitoring of land cover at broad scales. They found a combination
of vegetation indices, a thermal measure and seasonality (if relevant) produced the
best results in terms of land cover analyses.
Thermal remote sensing has been used most commonly to obtain meteorological
data for energy balance models (Mauser and Schädlich 1998, Snyder et al. 1998,
François et al. 1999, McVicar and Jupp 1999). This method allows for the
derivation of specific time-of-day components of the energy balance at the time of
the remotely sensed data acquisition. McVicar and Jupp (1999) found that climate
station data, which include maximum and minimum temperatures and daily
precipitation, can be accurately linked with thermal remote sensing (either Landsat
or Advanced Very High Resolution Radiometer (AVHRR) data). As such, these
data can be used to increase the spatial density of meteorological stations, thus
allowing better determination of the spatial energy balance. François et al. (1999)
compared the minimum air temperature to the thermal band satellite based surface
temperature and found a very strong correlation between the two. Such data can
also be linked with land cover data, thus providing a complete biophysical and
meteorological analysis of a landscape.
 Landsat TM band 6 data for analysing land cover 703

Although quite limited, there have been some studies undertaken linking thermal
remote sensing to land cover analysis. In a study undertaken in the Singapore
central catchment area nature reserves, Nichol (1995) found distinct spatial varia-
tions in thermal characteristics (using Landsat 5 band 6), which were observed to
correspond to land cover differences relating to forest, urban, water, and suburban
areas. Differences also were noticed in the temperatures calculated for areas of
forest, corresponding to primary and late secondary succession. These coolest areas
are buffered in most instances by a broad belt of slightly warmer forest or water.
Cool interior forest polygons were surrounded by slightly warmer forest polygons
rather than non-forest. Additionally, within the analysis, small interruptions in the
forest canopy were detected as points or ridges of higher temperatures. Such results
indicate that differences within vegetation types, and specifically across different
successional stages, can be picked up by thermal analysis. Nichol (1995) also tested
the dependence of surface temperature results on biomass by correlating the
thermal band emissivity corrected image data with NDVI using a random sample of
2000 pixels (excluding water). A Spearman’s rank correlation coefficient of 0.74
between temperature and NDVI (which she related to biomass) was found.
Likewise, Mauser and Schädlich (1998) found that NOAA AVHRR thermal
data could be used to differentiate land cover types. Their research demonstrated
the mesoscale heterogeneity of the Upper Bavaria region in Germany, revealing
landscape-like units relating to cities, forest, lakes, and agricultural areas. They
found a correlation of 0.80 (p~0.99) between evapotranspiration and temperatures.
These relationships may have been even stronger had the researchers also under-
taken nonlinear analysis, as was done in the present research.

5. Conclusions
The main objective of this research was to evaluate Landsat TM thermal data
for use both as a stand-alone dataset to produce temperature coverages and a
preliminary, broad-scale land cover analysis and as a supplement to other sensor
information such as in vegetation indices. It is an excellent data source both alone
and when used in conjunction with the other image bands (e.g. in colour composites
and vegetation analyses), and can be linked directly to land cover.
The following conclusions were drawn from the results of this research.
1. Eta correlation ratios were calculated between all information layers and
land cover classifications, verifying the usefulness of band 6 data both as a
single data layer and for use in combination with other bands.
2. Discriminant analysis revealed that band 6 data contain considerable infor-
mation for the discrimination of land cover classes in the dry tropical forest
ecosystem of Yucatan, Mexico.
3. Incorporation of band 6 into the classification procedure can replace the need
for a filter process, producing superior classification products with no loss of
accuracy using an understandable biophysical property of the land cover
under analysis.
4. Band 6 data are a continuous data source. Despite most researchers being
interested in land cover classes, the power of the continuous data (e.g. NDVI,
temperatures) allows for within-class differentiation, not just across-class
change analysis, e.g. forest degradation or thinning. This is especially useful
for the identification of successional stages of forest growth.
5. For the best results in remote sensing analyses of land cover change, multiple
704 J. Southworth

data sources should be used in order to achieve a comprehensive description

of surface processes. To do this both reflective and emitted data sources and
image products must be used together.
The thermal band information can also be used in conjunction with maps and
other regional information as a verification tool. These data are measured by a
different system than the reflective bands, allowing us to have a completely separate
and statistically independent data source for use in analyses in which we know the
sensors are ‘seeing’ and recording information for the same pixels. The thermal
band information may be used in association with the other TM bands when actual
field verification is not feasible, timely or cost-effective.
One of the main arguments against using band 6 data relates to its coarser
spatial resolution compared to the reflective bands. Despite the spatial resolution
differences, band 6 frequently outperformed the individual reflected bands, and also
performed well in combination with reflective bands despite the 16:1 resolution
difference. As such, the issue of resolution cannot really be used as a valid
argument against the incorporation of its use into land cover studies. In addition,
under some situations, the use of a coarser scale of data is desirable and can remove
the need for further image processing such as the application of filters as a post-
classification technique. Due to the current situation with Landsat 7, with its
improved spatial resolution this will hopefully lead to many more uses for this data
source and much wider incorporation of thermal band data into analyses of land
cover and land cover change.
It is evident that thermal remote sensing is potentially a powerful tool for
examining land cover and is particularly suited for use in tropical forests where
access is difficult. Clearly, knowledge of the spatial pattern of temperatures, which
relate to land cover, as well as to significant aspects of forest energy budgets and
vegetation patterns, will contribute to our understanding of the determinants of
these processes and patterns. This will enhance our ability to model land cover over
landscape scales (Luvall et al. 1990). This is essential within the research arena of
global change where we must better monitor and explain current patterns of land
cover in order to better determine potential future change. As such, for both
climate and vegetation based research, Landsat TM band 6 temperature data are an
increasingly important component in remote sensing research. Hopefully, research
such as this will help to negate some current discussions calling for the removal of
thermal band data collection on future Landsat missions. While this type of data is
expensive to collect and is currently underused by many researchers, its importance
is paramount in the LULCC and global change research arenas. Its continued
presence on the Landsat missions is thus imperative.

Acknowledgments
This research was supported by the National Science Foundation (NSF) (SBR-
9521918) as part of the ongoing research at the Center for the Study of Institutions,
Population, and Environmental Change (CIPEC) at Indiana University. Fieldwork
benefited from the invaluable research assistance of Julie Hanson. The author is
very grateful to Dr J. C. Randolph, Dr Elinor Ostrom, Dr Otto Doering, Dr V. J.
Meretsky and Michael Kohlhaas for their comments on earlier drafts; thanks Dr G.
Green for many helpful discussions; and, finally, thanks Dr H. Nagendra for final
comments and editing of the later versions of this manuscript.
 Landsat TM band 6 data for analysing land cover 705

References
ANDREFOUET, S., and ROUX, L., 1998, Characterisation of ecotones using membership
degrees computed with a fuzzy classifier. International Journal of Remote Sensing, 19,
3205–3213.
BASTIAANSSEN, W. G. M., MENENTI, M., FEDDES, R. A., and HOLTSLAG, A. A. M., 1998, A
remote sensing surface energy balance algorithm for land (SEBAL) 1. Formulation.
Journal of Hydrology, 212–213, 198–212.
BOYD, D. S., FOODY, G. M., CURRAN, P. J., LUCAS, R. M., and HONZAK, M., 1996, An
assessment of radiance in Landsat TM middle and thermal infrared wavebands for
the detection of tropical forest regeneration. International Journal of Remote Sensing,
17, 249–261.
CHAVEZ, P. S. JR., 1996, Image-based atmospheric corrections-revisited and improved.
Photogrammetric Engineering and Remote Sensing, 62, 1025–1036.
DEFRIES, R. S., HANSEN, M. C., and TOWNSHEND, J. R. G., 2000, Global continuous fields
of vegetation characteristics: a linear mixture model applied to multi-year 8km
AVHRR data. International Journal of Remote Sensing, 21, 1389–1414.
ESTÈVE, P., FONTÈS, J., and GASTELLU-ETCHEGORRY, J. P., 1998, Tropical dry ecosystems
modeling and monitoring from space. Ecological Modelling, 108, 175–188.
FOODY, G. M., 1996, Approaches for the production and evaluation of fuzzy land cover
classifications from remotely sensed data. International Journal of Remote Sensing, 17,
1317–1340.
FOODY, G. M., 2001, Accuracy of thematic maps derived from remote sensing. Accuracy
2000: Proceedings of the 4th International Symposium on Spatial Accuracy Assessment
in Natural Resources and Environmental Science, edited by G. B. M. Heuvelink and
M. J. P. M. Lemmens (Delft, Holland: Delft University Press), pp. 217–224.
FRANÇOIS, C., BOSSENO, R., VACHER, J. J., and SEGUIN, B., 1999, Frost risk mapping
derived from satellite and surface data over the Bolivian Altiplano. Agricultural and
Forest Meteorology, 95, 113–137.
GARSON, D. G., 2001, PA 765 Statnotes: an online textbook. http://www2.chass.ncsu.edu/
garson/pa765/statnote.htm.
HALL, F. G., BOTKIN, D. B., STREBEL, D. E., WOODS, K. D., and GOETZ, S. J., 1991, Large-scale
patterns of forest succession as determined by remote sensing. Ecology, 72, 628–640.
HANSEN, M. C., DEFRIES, R. S., TOWNSHEND, J. R. G., and SOHLBERG, R., 2000, Global
land cover classification at 1km spatial resolution using a classification tree approach.
International Journal of Remote Sensing, 21, 1331–1364.
HURTADO, E., VIDAL, A., and CASELLES, V., 1996, Comparison of two atmospheric
correction methods for Landsat TM thermal band. International Journal of Remote
Sensing, 17, 237–247.
JENSEN, J. R., 2000, Remote Sensing of the Environment: An Earth Resource Perspective (New
Jersey: Prentice Hall).
LAMBIN, E. F., 1997, Modelling and monitoring land-cover change processes in tropical
regions. Progress in Physical Geography, 21, 375–393.
LAMBIN, E. F., 1999, Monitoring forest degradation in tropical regions by remote sensing:
some methodological issues. Global Ecology and Biogeography, 8, 191–198.
LAMBIN, E. F., and EHRLICH, D., 1996, The surface temperature-vegetation index space for
land cover and land-cover change analysis. International Journal of Remote Sensing,
17, 463–487.
LI, Y., MAUSEL, P., WU, Y., MORAN, E., and BRONDIZIO, E., 1994, Discrimination between
advanced secondary succession and mature moist forest near Altamira, Brazil, using
Landsat TM data. Proceedings of the American Society for Photogrammetry and
Remote Sensing, pp. 350–364; URL: http://www.odyssey.maine.edu/gisweb/spatdb/
acsm/ac94040.html.
LO, C. P., QUATTROCHI, D. A., and LUVALL, J. C., 1997, Application of high-resolution
thermal infrared remote sensing and GIS to assess the urban heat island effect.
International Journal of Remote Sensing, 18, 287–304.
LUVALL, J. C., LIEBERMAN, D., LIEBERMAN, M., HARTSHORN, G. S., and PERALTA, R.,
1990, Estimation of tropical forest canopy temperatures, thermal response numbers,
and evapotranspiration using an aircraft-based thermal sensor. Photogrammetric
Engineering and Remote Sensing, 56, 1393–1401.
 706 Landsat TM band 6 data for analysing land cover

MAUSEL, P., WU, Y., LI, Y., MORAN, E., and BRONDIZIO, E., 1993, Spectral identification of
successional stages following deforestation in the Amazon. GeoCarto International, 4,
61–71.
MAUSER, W., and SCHÄDLICH, S., 1998, Modeling the spatial distribution of evapotran-
spiration on different scales using remote sensing data. Journal of Hydrology,
212–213, 250–267.
MCVICAR, T. R., and JUPP, D. L. B., 1999, Estimating one-time-of-day meteorological data
from standard daily data as inputs to thermal remote sensing based energy balance
models. Agricultural and Forest Meteorology, 96, 219–238.
MEYER, W. B., and TURNER, B. L. II., 1992, Human population growth and global land-use/
cover change. Annual Review of Ecology and Systematics, 23, 39–61.
NICHOL, J., 1995, Monitoring tropical rain forest microclimate. Photogrammetric Engineering
and Remote Sensing, 61, 1159–1165.
OKE, T. R., 1987, Boundary Layer Climates (New York: Methuen and Co.).
ROY, D. P., 1997, Investigation of the maximum normalized difference vegetation index
(NDVI) and the maximum surface temperature (Ts) AVHRR compositing
procedures for the extraction of NDVI and Ts over forest. International Journal of
Remote Sensing, 18, 2383–2401.
ROY, D. P., KENNEDY, P., and FOLVING, S., 1997, Combination of the normalized difference
vegetation index and surface temperature for regional scale European forest cover
mapping using AVHRR. International Journal of Remote Sensing, 18, 1189–1195.
SADER, S. A., HAYES, D. J., HEPINSTALL, J. A., COAN, M., and SOZA, C., 2001, Forest
change monitoring of a remote biosphere reserve. International Journal of Remote
Sensing, 22, 1937–1950.
SHORT, N., 1996, The warm Earth: thermal remote sensing. Code 935, Goddard Flight
Center, National Aeronatucis and Space Administration (NASA), USA.
SNYDER, W. C., WAN, Z., ZHANG, Y., and FENG, Y.-Z., 1998, Classification-based emissivity
for land surface temperature measurement from space. International Journal of
Remote Sensing, 19, 2753–2774.
SOBRINO, J. A., COLL, C., and CASELLES, V., 1991, Atmospheric correction for land surface
temperature using NOAA-11 AVHRR Channels 4 and 5. Remote Sensing of
Environment, 38, 19–34.
SOHN, Y., MORAN, E. F., and GURRI, F., 1999, Deforestation in North-Central Yucatan
(1985–1995): mapping secondary succession of forest and agricultural land use in
Sotuta using the cosine of the angle concept. Photogrammetric Engineering and
Remote Sensing, 65, 947–958.
TRISURAT, Y., EIUMNOH, A., MURAI, S., HUSSAIN, M. Z., and SHRESTHA, R. P., 2000,
Improvement of tropical vegetation mapping using a remote sensing technique: a case
of Khao Yai National Park, Thailand. International Journal of Remote Sensing, 21,
2031–2042.
TURNER, B. L., MOSS, R. H., and SKOLE, D. L., 1993, Relating land use and global land
cover change: a proposal for an IGBP-HDP core project. IGBP Report 24, HDP
Report 5, International Geosphere-Biosphere Programme, Stockholm, Sweden.
WHIGHAM, D. F., ZUGASTY TOWLE, P., CABRERA CANO, E., O’NEILL, J. O., and LEY, E.,
1990, The effect of annual variation in precipitation on growth and litter production
in a tropical dry forest in the Yucatan of Mexico. Tropical Ecology, 31, 23–34.
WITTICH, K.-P., 1997, Some simple relationships between land-surface emissivity, greenness
and the plant cover fraction for use in satellite remote sensing. International Journal
of Biometeorology, 41, 58–64.

View publication stats