




































In ternationa l
Scholars
Journa ls

 

African Journal of Environmental Economics and Management ISSN 2375-0707 Vol. 8 (3), pp. 001-012, March, 
2020. Available online at www.internationalscholarsjournals.org © International Scholars Journals 

 

Author(s) retain the copyright of this article. 
 
 
 

Full Length Research Paper 

 

Evaluation of three classifiers in mapping forest stand 

types using medium resolution imagery: A case study 

in the Offinso Forest District, Ghana 

 
Baatuuwie, N. B and Ir. Louise Van Leeuwen* 

 
1
Faculty of Renewable Natural Resources, University for Development Studies, P. O. Box TL 1882, Tamale, Ghana. 

2
International Institute for Geo-Information Science and Earth Observation, Enschede, The Netherlands. 

 
Accepted 12 December, 2019 

 
Loss of Ghana’s natural forests has been counteracted by plantation establishment and development. 
As at 2003, Ghana had a plantation area of about 97,000 hectares comprising different tree species. 
With the rapid expansion of plantations in the country, it is anticipated that major managerial 
challenges will arise due to insignificant technical personnel for monitoring and management. The 
application of GIS and remote sensing will be a powerful intervention and tool for monitoring and 
managing these forest resources in the area. The aim of this study was to determine a suitable method 
of mapping the different forest stand types using medium resolution images in the study area. Three 
classifiers were examined for their suitability in mapping the different forest stand types in the area 
(maximum likelihood, spectral angle mapper and decision tree). The results showed that using 
maximum likelihood classifier and ASTER imagery, different forest stand types can be accurately 
mapped with an overall accuracy of 88.50%. 

 
Key words: Geo-referenced, classification, algorithm, spectral, mapper. 

 
INTRODUCTION 
 
Ghana’s total land area of 238,500 km

2
 is made up of 

savanna (56%); closed forest (35%) and the remaining 
represents other land cover. All the vegetation types in 
Ghana, except for those comprising the savanna, are 
considered tropical forests and located in the southern belt 
of the country (Blay et al., 2007).  
These forests play very vital roles in supporting the livelihood of 
the inhabitants of Ghana in various ways. Trading in non-timber 
forest products (NTFPs: chewing sticks, pestles, canes, nuts, 
fruits, bush meat, fodder, artifacts and medicine extracted from 
the forest) are economically important within all areas of the 
high forest zone (Blay, 2004) . The forest also serves as haven 
for numerous species of flora and fauna, and generally helps 
maintain the biological diversity of the area (Abeney and 
Owusu, 1999). It also contains many genetic materials of known 

and yet to be discovered species of importance (Blay, 2004). 
  

 
 
 
*Corresponding author. E-mail: taawiebernard@yahoo.com. 

 

 
 

Regardless of the substantial contribution of the forest 
resources to the economy, degradation of Ghana’s forest 
resources is assuming an alarming proportion. It is 
estimated that about 60% of the reserved forests are 
degraded (TED, 1997; FAO, 2000). It was reported that 26% 
of the forest cover was lost between 1990 and 2005 
(Mongabay.com, 2006).  

Like many other tropical countries, the loss of Ghana’s 
natural forests has been counteracted by comprehensive 
reform programmes in the forestry sector (MLF, 2004). Key 
among the reforms was the Government’s initiative in 
establishing plantation in the country. Plantation forestry can 
be used as a tool not only to halt forest degradation but also 
to catalyze important native forest flora restoration after long 
period of anthropogenic and non-anthropogenic 
disturbances (Lemenih and Teketay, 2004).  

Ghana by 2003 had a plantation area of about 97,000 

ha (MLF, 2004). With the rapid expansion of plantations 

in Ghana, it is anticipated that major managerial 



 
 
 

 

challenges will arise due to insignificant technical 
personnel for monitoring and management. The 
application of Geographic Information System (GIS) and 
remote sensing will be a powerful intervention and tool for 
monitoring and managing these forest resources in the 
area. 

GIS and remote sensing have been utilized in many 
regions over the past few decades for both mapping and 
monitoring forest resources. Maintenance of ecosystem 
integrity and protection of forest diversity are essential 
processes to achieve sustainable forest management 
(Jensen and Everett, 1994). To achieve these multiple 
objectives, remote sensing techniques have the capability 
of providing multi-scale, spectral, spatial, and temporal 
information in a consistent format to help forest managers 
meet their management objectives (Huggins, 2006).  

It was acknowledged at the environmental conferences 
in Rio de Janeiro and Kyoto that satellite imagery will 
offer the most promising and is most likely the only 
feasible way for a detailed mapping and monitoring of 
forests resources over large geographical areas per unit 
time (De Carvalho et al., 2004). Remote sensing data and 
other supplementary data that is being produced on 
regular basis allow researchers to study natural forest 
processes, such as forest succession, encroachment and 
regeneration, which were previously limited to small 
geographical areas and often times impracticable 
(Huggins, 2006). Thus remote sensing may be 
considered the only viable option for monitoring forest 
succession in a timely and cost effective manner (Song 
and Woodcock, 2002).  

Extracting this information from the data is however 

accomplished by image classification (Lillesand and 
Kiefer, 2000). There are several techniques and 

variations exist between these techniques for extracting 
thematic information from an image (Huggins, 2006). 

Each technique is superior over the other depending on 
how accurately it identifies and maps the resource/feature, its 

cost and time effectiveness (Franklin, 2001). Mapping 
accuracy will depend on the spectral characteristics of the 

resource/feature to be mapped. In this study, the different 
forest cover types in the study area are mapped.  

The forest stand types in the study area differ from the 
standpoint of tree species composition. However there 
are no definite and explicit boundaries between these 
forests stand types and other vegetation types in the 
area. In a nutshell, these vegetation types are seen as 
the same forest cover type on the satellite data due to 
their similar species composition. There is need to 
identify an effective classification/mapping procedure that 
will discriminate between these forest cover types with 
least cost and human intervention (Franklin, 2001), if 
reliable estimates of forest cover are to be made 
frequently over large geographical areas over time for 
strategic forestry planning and decision making (Dees et 
al., 2003). This study was therefore conducted to 

 
 
 
 

 

determine a suitable method for mapping forest stand 
types in the study area using medium resolution images. 
Three image classification algorithms were then con-

sidered and used for this research: Maximum likelihood 
classifier, spectral angle mapper and decision tree 
classifier. 

 

MATERIALS AND METHODS 
 
Description of the study area 
 
The study was conducted in the Afram Headwaters Forest Reserve 
(Figure 1) located in the Offinso District of Ashanti Region of 
Ghana. The Reserve lies between longitude 1º 32’W and 1º 48W 
and 6º 45’N and 7º 25’N. It covers 20,100 hectares, comprising both 
natural and plantation forests. The reserve is fringed by eight 
villages, with their inhabitants depending heavily on the reserve for 
sustenance. Each of the communities has a population size of not 
less than 600 people (Ghana Statistical Service, 2000). Due to 
heavy dependency of this populace around the reserve, it has 
consequently subjected the forest reserve to various forms of 
human activities leading to its fragmentation and degradation. 
Consequently, there are ongoing restoration programmes in the 
reserve through plantation establishment by the government and 
private developers. Plantation establishment methods employed are 
mainly monoculture of exotic tree species such as Cedrela odorata 
(Cedrela) and Tectona grandis (Teak) and mixed stands of local 
tree species. 

 

Land covers data collection 
 
An iPAQ (handheld computers) equipped with an internal Global 
Positioning System receiver was used to capture coordinates of 
pre-determined land cover samples which were used as training set 
for image classification and testing data for the accuracy 
assessment of the classification. It is recommended that 50 points 
be collected per land cover for image classification and accuracy 
assessment (Curran, 1985). However due to time constrain, 219 
points were collected for all land cover types. 

 

Image processing and classification 
 
A geo-referenced and atmospherically corrected ASTER February 
24, 2008 image was obtained from ITC geo-database. The image 
was re-projected from geographic coordinate system (WGS84) to 
the Ghana local coordinate system in Erdas image processing 
software.  

The ASTER image was classified using the three classifiers: 
Maximum likelihood, spectral angle mapper and decision tree. The 
main aim was to assess the performance of the three classification 
algorithms using medium resolution data acquired from the ASTER 
sensor in the study area. Their performance was based on the 
ability to identify different forest stand types accurately based on 
ground truth data (Mohd, 2007) with minimum time and cost 
(Franklin, 2001) . Overall classification accuracies as well as the 
Kappa statistics were calculated for each of the classifiers used and 
their results compared. 

 

The maximum likelihood classifiers 
 
Supervised classification using maximum likelihood algorithm was 



   
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 

Figure 1. Study area showing Afram headwaters forest reserve on Aster image. 
 
 
 
applied for classification of the image. Supervised classification 
technique using maximum likelihood algorithm is the most 
commonly and widely used method for land cover classification (Jia 
and Richards, 2006). In Australia, the maximum likelihood classifier 
was effectively used to map different forest stand types with high 
precision (Chia et al., 2005). Stated below are steps involved in the 
procedure for achieving image classification using the maximum 
likelihood classifier: 
 
Training sample selection for classification: Training samples 
were collected on the Aster image to train the classifier. One 
hundred and six (106) training sets were used for the classification. 
During analysis, five land use/cover classes were considered: Build-
up/bare (15 training sets), teak plantation forest (15 training sets), 
new plantations (24 training sets), natural forest (26 training sets) 
and mixed/degraded forest (26 training sets). The quality of a 
classification output is dependent on how training samples were 
collected (Franklin, 2001). To ensure high accuracy, the selection of 
the training samples was aided by feature space in bands 2 and 3 
which gives an idea of the homogeneity and spectral separability 

 
 
 
 
between the forest classes. 
 
Assessing the normality of the training sample: Normal 
distribution of the training samples is the basic requirement of the 
maximum likelihood classification. Therefore, training samples were 
subjected to normality test using the histogram plot control curve. 
For a normal distribution, the histogram shows bell shaped curve. 
Outliers are often located in the tail ends of the curve. The 
histograms of the training samples of the various classes in bands 
1, 2 and 3 are shown in Figures 2 to 6. Means of the classes differ 
from band to band. Nevertheless the data distribution shows the 
same trend in the different bands; thus, majority of the training 
pixels have values close to the mean and only a few are outliers. 
This form of distribution makes it suitable for the maximum 
likelihood classification. 

 

The spectral angle mapper (SAM) 
 
As with all supervised classifications, SAM requires that the user 



  
 
 

 

 Teak      Teak     Teak   
 

                 
 

Band number  1    Band number  2    Band number 3  
 

    

Histogram 
           

Histogram 
 

        

Histogram 
     

 

             
 

               
 

                 
 

                 
 

 
 
 
 
 
 
 
 
 
 
 
 
 

 
Figure 2. Histogram of class teak plantation.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 

Figure 3. Histogram of class build-up/bare.  
 
 
 

 

Band number 1    Band number 2    Band number 3   
 

  

Histogram 
   

Histogram  

 

 Histogram 
 

      
 

           
 

 
 
 
 
 
 
 
 
 
 
 
 

 
Figure 4. Histogram of class mixed/degraded forest. 



   
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Figure 5. Histogram of class natural forest. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
Figure 6. Histogram of class new plantations.  

 

 
defines the training classes and training spectra for the 
classification. The mean measurement vector for a training class is 
the vector representation of its mean spectrum. The creation of 
training classes can be accomplished in one of several ways in 
ENVI. One way is to define a set of regions/area of interest (ROIs) 
for pixels in the image for which you have ground truths. The mean 
measurement vector for a given ROI training class is simply the 
mean measurement vector for all the pixels in that class. A second 
way of defining training classes is to import spectra from spectral 
libraries. For this study, the former method was used to create the 
class training set. ROIs here designate the spectral signatures of 
the five classes (bare, mixed/degraded forest, teak forest, natural 
forest and new plantations). In all, 77 ROIs were used to train the 
classifier for the classification. 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Decision tree 
 
The decision tree is made up of a series of binary decisions that are 
used to categorized pixels into classes. The decisions can be based 
on any characteristic of the dataset that is available.  

For current research, image DN values were extracted and used 
to create the decision tree. Based on the ground truth, regions of 
interest of the various land use/cover were obtained and the 
spectral values extracted for inputs in the decision tree. With the 
exception of the build-up/bare, maximum and minimum DN values 
of bands 2 and 3 of the image were used to develop the 
expressions for the forest stand types (Table 1). Band one was 
used to develop an expression for the build-up/bare. Even though 
the decision tree classifier is not commonly employed within 



 
 
 

 
Table 1. Input expressions for the decision tree. 

 
Classes Expression    

Border (B1 EQ 0) AND (B2 EQ 0) AND (B3 EQ 0)  

Build-up/bare (B1 GE 165) AND (B1LE 169)   

Teak forest (B2 GE 104) AND (B2 LE 118) AND (B3 GE 77) AND (B3 LE 92) 

New plantation (B2 GE 104) AND (B2 LE 110) AND (B3 GE 93) AND (B3 LE 95) 

Degraded forest (B2 GE 98) AND (B2 LE 105) AND (B3 GE 96) AND (B3 LE 101) 

mixed forest (B2 GE 99) AND (B2 LE 106) AND (B3 GE 96) AND (B3 LE 102) 

Natural forest (B2 GE 94) AND (B2 LE 112) AND (B3 GE 84) AND (B3 LE 99)   
B= Band, GE= greater than, EQ= equal to, LE= less than.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
Figure 7. Decision tree scheme. 

 

 

classical remote sensing, it has been shown to produce a higher 
level of classification accuracy in mapping features than that of the 
traditional method of maximum likelihood supervised classification 
(Pal and Mather, 2003).  

Figure 7 shows the decision tree scheme that was developed for 
the image classification. Each decision node is named for the class 
it is intended to classify and if a pixel meets the criteria at that node 
it is assigned to that class. If it does not meet up the criteria, then it 
moves down the tree to the next decision node. Within the 
classification image, the colour represented within the “yes” result of 
a decision node is the actual colour that the class will be 
represented (Huggins, 2006). 

 

Accuracy assessment of classification 
 
To evaluate the performance of a classifier by the use of the 
methods mentioned in the preceding chapter, it requires that a 
randomly selected set of test samples (pixels) for each land cover 
class is used for computing the classification accuracy (Richards, 
1993). In this study, accuracy assessment was performed with 113 
test samples. The test samples were carefully chosen in such a way 
that they were equally distributed geographically with the training 
data set. The classified images were then crossed with the test data 

 
 

 
to generate error matrices for the calculation of the various 
accuracies (producer’s, user’s accuracy, class mapping accuracy 
for each class and the overall accuracy) as well as the kappa 
statistics. A significance test of difference between the classification 
methods was performed between the confusion matrices using the 
Z test with = 0.05. Classification outputs were also assessed 
qualitatively by visual examination of the classified maps in relation 
to the field knowledge. This is to ascertain whether the classification 
outputs matches reality. 

 

RESULTS 

 

The maximum likelihood classification output 
 

Figure 8a shows the classification output of the maximum 
likelihood classifier. Brown locations in the map represent 
build- up/bare areas. The rest are different forest stand 
types in the area. By ocular examination, it is quite clear 
that build-up/bare area covers smaller proportion of the 
area while teak plantations and natural forest cover equal 
proportion of the area. By quantitative analysis, the 



  
 
 

 

Maximum likelihood classification output  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
Build-up/bare 

 
Figure 8a. Output of maximum likelihood classification. 

 

 

classification identified 5639.80 hectares of the total area 

covered by teak plantations which is the highest, while 
build-up/bare covered the least area of 280.34 hectares 

(Figure 8b). 
 

 

Spectral angle mapper (SAM) 
 

Figure 9a shows classification output of the spectral 
angle mapper (SAM). The colour designations match 
those used for the maximum likelihood classification. 
Unlike the maximum likelihood classifier, SAM identified 
natural forest to cover the largest proportion of 6963.33 
hectares whiles mixed/degraded forest was identified to 
cover the least of 517.16 hectares (Figure 9b). Thus, the 
estimated area covered by mixed/degraded forest is three 
times lower than that estimated by the maximum 
likelihood classifier. It is also clear from visual 
examination that the SAM also estimated large area 
covered by build-up/bare than that of the maximum 
likelihood classifier. 

 
 

 

Decision tree classification output 
 

The map displayed in Figure 10a is the output of the 
decision tree classifier. It estimated a large area covered 
by natural forest. The area covered by this forest stand 
type is 8269.49 hectares. Build-up/bare like that of the 
maximum likelihood classifier was detected the least 
covering an area of 1048.46 hectares (Figure 10b). 
However the area covered by build-up/bare estimated by 
decision tree is still higher than that estimated by the 
maximum likelihood classifier. Compared with the 
previous two classifiers (ML and SAM), the decision tree 
classifier estimated the least area covered by teak 
plantations by 4362.91 hectares. Nevertheless, it 
estimated the largest area covered by natural forest. 
 

 

Comparison of classification accuracies of the three 

methods 
 
Table 2 shows the summary of the overall accuracy and 



  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
Figure 8b. Area representation of the different land use/cover classes. 

 
 
 

Spectral angle mapper classification output  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 

Build- up/bare 
 

Figure 9a. Classification output of spectral angle mapper. 



   
 
 
 
 

A
re

a
s
 i
n
 h

e
c
ta

re
s
 

 
 
 
 
 
 
 

 

Land use/cover type 

 
Figure 9b. Area representation of the different land use/cover classification.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
Figure 10a. Output of decision tree classification. 

 

 

kappa statistics of the three classifiers. The Maximum 
likelihood classifier had the highest classification 
accuracy of 88.50% while the spectral angle mapper had 
the lowest of 69.03%. The maximum likelihood classifier 
again had the highest kappa value of 0.85 and the 
spectral angle mapper had the lowest value of 0.62. The 
significance of this difference was tested using the Z-test. 
The Z-test showed that the maximum likelihood classifier 

 
 

 

is significantly higher than the spectral angle mapper (Z-
test: Z= 3.5, P= 0.0004). There is no significant difference 
between the accuracies of the maximum likelihood and 
the decision tree classifier (Z-test: Z= 1.46, P= 0.144). 
Decision tree and spectral angle mapper differ 
significantly in the accuracies at (Z-test: Z= 2.10, P= 
0.029)  

Apart from the overall accuracy and the kappa 



  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
Figure 10b. Area representation of the different land use/cover classes. 

 

 
Table 2. Summary of accuracy results of the three classifiers. 

 
Classifier Overall accuracy (%) Kappa coefficient 

Maximum likelihood 88.50 0.85 

Decision tree 81.74 0.77 

Spectral angle mapper 69.03 0.62  
 

 
Table 3. Comparison of user’s and producer’s accuracies of the three classifiers. 

 
Class  Producer©s accuracy (%)   User’s accuracy (%)  

 ML DT SAM ML DT SAM 

Teak forest 89.66 79.31 86.21 83.87 85.19 83.33 

Build-up/bare 76.92 84.62 92.31 100 91.67 75 

Mixed/degraded 100 94.29 37.14 87.5 89.19 100 

Natural forest 94.12 94.12 94.12 100 72.73 42.11 

New plantations 68.42 52.38 63.16 81.25 78.57 75  
 

 

statistics, producer’s and user’s accuracy were also 
examined to determine the accuracy of the different 
classification methods. User’s and producers accuracies 
are necessary for this study as the main aim is to map the 
forest categories. The user’s and producer’s accuracies 
when compared (Table 3) revealed that the maximum 
likelihood has a maximum producer accuracy of 100% for 
mixed forest with user accuracy of 87.5% , decision tree 
recorded 94.29% as a maximum producer accuracy for 
mixed/degraded forest with user’s accuracy of 89.19%. 
Spectral angle mapper recorded its maximum producer 
accuracy of 94.12% for natural forest with it counterpart 
user’s accuracy of 42.11%. The least producer’s 
accuracy of maximum likelihood, decision tree 

 
 

 

classification and the spectral angle mapper are; 68.42, 
52.38 and 37.14% respectively (Table 3). The least user’s 
accuracies for the three classifications methods included: 
maximum likelihood (81.25), decision tree classification 
(72.73) and spectral angle mapper (42.11). Based on the 
confusion matrix, it was also clear that some classes 
were confused with others. This may be due to the 
spectral heterogeneity of these features in the landscape 
of the study area and thereby become a hindrance to 
excellent mapping. In addition, out of the 113 reference 
pixels, Maximum likelihood classifier accurately classified 
100 pixels whiles spectral angle mapper and decision 
tree classifiers accurately classified 75 and 94 pixels 
respectively. 



  
 
 

 

DISCUSSION 

 

Comparison of image classification methods 

 

The forest cover mapping using image classification will 
be incomplete if the classification accuracies are not 
known. Thus accuracy assessment is a fundamental 
principle in assuring the quality of thematic maps for their 
intended application (Stehman and Czaplewski, 1998). 
This principle determines how useful and applicable a 
particular classification method is (Huggins, 2006; 
Campbell, 2002). It is also the surest way to compare 
different classification methods. However, cost-
effectiveness and technical uncomplicatedness are also 
key factors determining the usability and applicability of 
the method (Franklin, 2001). Methods that are less 
expensive and technically simple will always be opted if 
the classification accuracies are not significantly different 
(Congalton and Green, 1999). All these factors however 
will depend on the purpose and financial value of the 
resources (Hussin and Atmopawiro, 2004).  

Comparison of the accuracies between the three 
classifiers indicated that the maximum likelihood classifier 
has the highest classification accuracy of 88.50%, 
followed by the decision tree with classification accuracy 
of 81.74%, and the spectral angle mapper recorded the 
lowest of 69.03%. An overall accuracy of 85% is 
considered standard for land cover mapping (Campbell, 
2002). With reference to this overall accuracy, only the 
maximum likelihood classification has met this standard 
while the other two classifiers (decision tree and spectral 
angle mapper) have overall accuracies lower than this 
standard value recommended by Campbell (2002). This 
means that the maximum likelihood classification 
technique is a better method in mapping the forest cover 
types in the study area than both spectral angle mapper 
and decision tree classifiers when using medium 
resolution images.  

Further comparison of the mapping accuracies of the 

three classifiers by their kappa statistics (KHAT) showed 
that the maximum likelihood classifier still performed 
better in mapping the forest cover types in the study area 
than the two classifiers with a kappa value of 0.85 while 
the decision tree recorded 0.77 and the spectral angle 
mapper had the lowest value of 0.62.  

Apart from the overall accuracy and the kappa 
statistics, class-wise mapping accuracy measurement 
such as producer’s accuracy and user’s accuracy also 
revealed that the maximum likelihood classification 
method did better in mapping the different forest cover 
types in the study area than the other two methods. 
Spectral angle mapper relatively has poorer performance 
in mapping various categories of forest cover. This result 
is not in divergence with that found by Mohd (2007) who 
used the same methods in mapping Malayan tropical 
forest. 

 
 

 

The variation in the classification accuracies however 
might have been due to the method of training the 
classifier. This is because all the three methods have 
different approaches in selecting the training samples for 
the classification and this may have an influence on the 
classification accuracies. According to Lillesand and 
Kiefer (2000), the classification accuracies of any land 
cover maps will depend largely on the quality of the 
training set used. So any variation in training sample 
selection is likely to influence the classification results.  

The superior performance of the maximum likelihood 
classifier over the others may be due to the thorough 
examination of the spectral signatures separability of the 
different forest stand types in the area using the spectral 
feature space curve. Difficulty in determining the precise 
DN values for the various forests stand types due to the 
spectral heterogeneity of their reflectance may have 
caused the decision tree classifier to have performed 
below standard. The performance of SAM however 
showed that more than the direction of a vector is 
required in order to separate the tropical forest stand 
types, which are spectrally difficult to separate.  

In terms of technical simplicity and accessibility, the 
maximum likelihood classifier has been used for many 
years and is one of the inbuilt extensions of many GIS 
and RS based software such as ILWIS, ENVI, and 
ERDAS etc. Spectral angle mapper and decision tree 
classifiers are restricted to specific image analysis 
software such as ENVI. Execution of maximum likelihood 
classification technique is also straight forward and 
simple. Thus its implementation is less laborious and 
requires relatively low technical know-how before exe-
cution. On the other hand, Spectral angle mapper and 
decision tree techniques are more complex, time 
consuming and require expert knowledge in signature 
derivation and decision rule setting before final classifica-
tion is attempted. The maximum likelihood classification 
method even though the oldest and traditional method, 
still may be considered to be the suitable method for 
mapping forest stand types in the study area than the 
other two methods especially when medium resolution 
images are employed. 
 

 

Conclusions and recommendations 

 

The classification accuracy of the three classifiers was 
used as the determining factor for identifying the suitable 
mapping method. From the finding of this work, it may be 
concluded that the maximum likelihood classification 
method is a suitable method in mapping the different 
forest stand types in the area in comparison with the 
other two methods.  

In a nutshell, we may conclude that using maximum 

likelihood classifier and medium resolution imagery 

(ASTER), different forest stand types can be accurately 



 
 
 

 

mapped with an overall accuracy of 88.50% in the study 
area.  

With regard to mapping forest stand types in the study 
area, the maximum likelihood classification method is 
recommended because of its high accuracy, technically 
simplicity and being available in many remote sensing 
data analysis softwares. 
 

 
REFERENCES 
 
Abeney EA, Owusu JGK (1999). Workshop for media personnel on 

forestry and wildlife reporting proceedings, Institute of Renewable 
Natural Resources, UST, Ghana, p. 95. 

Blay D, Appiah M, Damnyag L, Dwomoh KF, Ari PLO (2007). Involving 
local farmers in rehabilitation of degraded tropical forests: some 
lessons from Ghana. 

Blay D (2004). Rehabilitation of degraded forests through the 
collaboration of local communities in the Dormaa District of Brong 
Ahafo Region of Ghana, pp. 31–35. 

Campbell JB (2002). Introduction to remote sensing. The Guilford Press 
Chia J, Caccetta PA, Furby SL, Wallace JF (2005). Mathematical and 
Information Sciences, Private Bag 5 PO Wembley Western Australia 

6913.  
Congalton RG, Green K (1999). Assessing the accuracy of remotely 

sensed data: principles and practices. Boca Raton: Lewis Publishers.  
Curran PJ (1985). Principles of Remote Sensing. Longman Scientific 

and Technical Group, Essex, England.  
De Carvalho LMT, Clevers JGPW, Skidmore AK, Jong SM (2004). 

Selection of imagery data and classifiers for mapping Brazilian semi-
deciduous Atlantic forests. 

Dees M, Asan U, Yesil A (2003). Ideas and options for a national forest 
inventory for Turkey. Advances in forest inventory for sustainable 
forest management and biodiversity. Kluver Academic Publishers, 
For. Sci. pp. 375-395.  

FAO (2000). Global forest resources assessment main report. FAO 
Forestry Paper, pp. 140,479.  

Franklin SE (2001). Remote Sensing for Sustainable Forest 
Management. CRC Press, p. 7. 

Ghana Statistical Service (2000). Summary of the 2000 Population 
Census Report. Accra; Ghana: Ghana Statistical Service. 

 
 
 
 

 
Huggins NJ (2006). Remote sensing classification methods for 

monitoring eastern red cedar. MSc. Thesis, Missouri State University. 
Hussin A, Atmopawiro VP (2004). Sub-pixel and maximum likelihood 

classification of Landsat ETM+ images for detecting illegal logging 
and mapping tropical rain forest cover types in Berau, East 
Kalimantan, Indonesia. The International Institute for Geoinformation 
Science and Earth Observation (ITC), Enschede, Netherlands.  

Jia X, Richards JA (2006). Remote sensing digital image analysis, an 
introduction, Taylor and Industrial Arts.  

Jensen ME, Everett R (1994). An overview of ecosystems management 
principles. Ecosystem Management Principles and Applications, 
Eastside Forest Ecosystem Health Assessment, USDA Forest 
Service, Pacific Northwest Research Station, General Technical 
Report. PNW-GTR-318.  

Lillesand TM, Kiefer RW (2000). Remote sensing and image 
interpretation (4th ed.). New York: John Wiley and Sons. 

Lemenih M, Teketay D (2004b). Restoration of native forest flora under 
plantation forests established on abandoned degraded agricultural 
sites in Ethiopia. Environ. Manage. 

MLF (2004). Criteria and Indicators for Sustainable Management of 
Natural Tropical Forests. Ghana.  

Mohd SHZ (2007). Remote Sensing & GIS Programme, Dept. of Civil 
Eng., UPM, 43400 Selangor, Malaysia.  

Mongabay.com (2006). Forest protection could earn tens of millions for 
Ghana http://news.mongabay.com/2006/1106-ghana.html.  

Pal M, Mather PM (2003). An assessment of effectiveness of decision 
tree methods for land cover classification, Remote Sens. Environ., 
86: 554-565 

Richards JA (1993). Remote sensing digital image analysis: an 
introduction. Springer-Verlag, Berlin. 

Song C, Woodcock CE (2002). The spatial manifestation of forest 
succession in optical imagery. The potential of multi-resolution 
imagery, Remote Sensing Environ., 82: 271-284. 

Stehman SV, Czaplewski LR (1998). Design and analysis of thematic 
map accuracy assessment: fundamental principles. Remote Sensing 
Environ., 64: 331-344. 

TED (1997). Ghana Forest Loss. Case Studies 

http://www.american.edu/TED/ghana. htm 


