CLASSIFICATION OF NATURAL BROAD-LEAVED EVERGREEN FORESTS
BASED ON MULTI-DATA FOR FOREST INVENTORY IN THE CENTRAL
HIGHLANDS OF VIETNAM

Thesis submitted in partial fulfillment of the requirements of
the degree Doctor rer. nat. of the
Faculty of Forest and Environmental Sciences,
Albert-Ludwigs-Universität
Freiburg im Breisgau, Germany by Nguyen Thi Thanh Huong Name of Dean: Prof. Dr. Heinz Rennenberg
Name of Supervisor: Prof Dr. Barbara Koch
Name of 2
nd

relation with spectral values. Inverse J-shaped distribution Meyer N-DBH was selected as
the basis to calculate mean characteristics of individual forest class. Finally, an improved
classification system of natural wooden forest was defined based on the result of image
classification and mean value of forest conditions of quantitative criteria along with
qualitative characteristics. Four band SPOT5, three Principle Component (PCs), and a
Normalized Difference Vegetation Index (NDVI) image, and the methods of regression,
kNN and geostatistics were used to predict the forest stand volume. The best result was
obtained by applying regression kriging method on SPOT5 image. In order to predict
potential risk for the forest at the study area, the factors which related to accessibility in
forest utilization were also analyzed. These factors were divided into four different impact
ii

levels for the two relevant classes for better forest management. A thematic map showing
the potentially vulnerable sites was developed for forest management and planning.
For the monitoring purposes as well as for sustainable forest management of the
Vietnamese forests, a combination of remote sensing data and field inventory to produce
suitable classified forest maps is a prerequisite. In combination with the appraisal of
potential risks to forest stands, this information is of extreme importance for forest
planning activities.
Das Untersuchungsgebiet liegt im zentralen Hochland von Vietnam. Dort kommen noch
verbreitet natürliche tropische Wälder vor. Als Fernerkundungsmittel wurde eine SPOT 5
Szene vom 5. Februar 2006 verwendet. Für die geometrischen und topographischen
Korrekturen des Satellitenbildes wurden Passpunkte (Ground Control Points) aus GPS
Messungen sowie ein Geländemodell aus digitalisierten Höhenschichtlinien eingesetzt. Für
die topographische Normalisierung wurden drei verschiedene Algorithmen getestet und
zwar Cosine, Minnaert und C-correction. Die besten Ergebnisse lieferte die C-correction,
weshalb diese Ergebnisse in die weiteren Verarbeitungsschritte einflossen. Die
vorbearbeiteten Satellitendaten wurden sowohl mit unüberwachter Klassifizierung, als
auch mit überwachten Klassifizierungsverfahren ausgewertet. Für letzteres und auch für
die Genauigkeitsüberprüfung der Ergebnisse wurden eine Reihe von Trainingsgebieten in
dem Untersuchungsgebiet festgelegt. Insgesamt 120 Plots wurden im Gelände angelegt,
verteilt auf die vier unterschiedenen, natürlichen Waldklassen sowie für vier weiteren
Landbedeckungsklassen. In den Plots wurden forstliche Parameter wie BHD, Baumhöhe,
Kronendurchmesser, Position der Bäume und deren relative Lage zu den Nachbarbäumen
gemessen, Kronenschluß und Bedeckungsgrad wurden geschätzt. Zur Berechnung der
Zusammenhänge zwischen den erhobenen forstlichen Variablen selbst sowie den
Variablen und den spektralen Fernerkundungsdaten wurden Regressionsanalysen
angewandt. Dabei wurden enge Zusammenhänge zwischen spektralen Eigenschaften mit
der Dichte des Kronendaches, dem mittleren Durchmesser, sowie der Dichte aller Bäume
über 35 cm Durchmesser gefunden. Für die Berechnung von charakteristischen spektralen
Eigenschaften der einzelnen Waldklassen wurde J-shaped Verteilung Meyer N-BHD zu
Grunde gelegt. Basierend auf den Ergebnissen der Satellitenbildklassifikation und den
berechneten Kennziffern wird in einem letzten Schritt ein neues, verbessertes
Klassifikationsschema für natürliche tropische Wälder in vorgeschlagen. Für die Schätzung
des Holzvorrates wurden die vier Kanäle der Spot 5 Szene, drei Principle Component
Berechnungen sowie der Normalized Difference Vegetation Index (NDVI) herangezogen.
iv

Dabei kamen verschiedene Methoden aus der Geostatistik, Regressionsanalyse und kNN

scholarship and research subsidy support by the DAAD (German Academic Exchange
Service). I am also greatly indebted to the OASIS for providing remote sensing images.
I would like to express my sincere appreciation to all of the friends and colleagues of Felis,
Freiburg University, Germany and Department of Forest Resource and Environment
Management, Tay Nguyen University, Vietnam for their kind assistance during my
research.
Last but not least, my deepest thanks go to my family for their patience, support and
encouragement throughout my many years of higher education. My parents have been an
endless source of love and understanding throughout my life. My husband, daughter,
brothers and sisters have always given me all their infinite love and best wishes. To all I
am grateful for their roles in my life.

Nguyễn Thị Thanh Hương
vi Table of Contents
Abstract i
Zusammenfassung iii
Acknowledgements v
List of Tables xi
Illustrations xiii
List of Abbreviation xv
1 INTRODUCTION……………………………………………………….1
1.1 Background
1
1.2 Problem analysis

3.3.3 Estimation of forest variable using remote sensing 33
3.4 GIS technique
36
3.5 Relation bet
ween relevant factors and forest status 37
3.6 Literature re
view 37
3.6.1 Classification of forest using remote sensing images 37
3.6.2 Prediction of forest parameters using remotely sensed data 38
4 DESCRIPTION OF RESEARCH AREA…………………………….43
4.1 Location
43
4.2 Topography 44
4.3 Geology an
d soil 44
4.4 Climate 44
4.5 Description
of forest types 45
4.6 Forest jurisd
iction 46
5 METHODOLOGY…………………………………………………… 48
5.1 Data, Software and equipment
49
5.1.1 Data availability 49
5.1.2 Software 51
viii

5.1.3 Equipment and tools 51
5.2 Developme
nt of a set of field data 52

6.1 Topographic correction 77
6.2 Statistical a
pproach with stand forest variables 78
6.2.1 Height – Diameter Relationship (H - DBH) 78
6.2.2 Volume equation for forest stands 80
6.3 Proposal for
a new forest classification system based on both quantitative and
qualitative criteria of the stand 80
6.3.1 Average characteristics of each forest status 80
6.3.2 Estimates of forest stand parameters 81
6.3.3 Qualitative description as a definition of the forest class 85
6.4 Forest/land cover
classification 86
6.4.1 Forest/Land cover map 86
6.4.2 Natural broad-leaved evergreen forest 88
6.4.3 Plantation forest 90
6.4.4 Bamboo 90
6.4.5 Shrub/Grass land 90
6.4.6 Other lands 91
6.5 Accuracy assessment
91
6.6 Relation of
directly surveyed data with spectral images 92
6.6.1 Correlation Analysis 92
6.6.2 Relationship model 96
6.7 Estimation of stand volume from spe
ctral image and field data survey 96
6.7.1 Regression method 96
6.7.2 KNN method 101
6.7.3 Geostatistical method with regression kriging 102

List of Tables

Table 2.1 Classification system of forest developed by Thai Van Trung (1978, 1999) 14
Table 2.2 Class definition for Vietnamese Classification of Evergreen Natural Wooden
Forest 15
Table 5.1 Characteristics of SPOT 5 image 50
Table 5.2 Description of land cover and their appearance on SPOT-5 image 61
Table 5.3 Interpretation of Kappa Values 64
Table 5.4 Relevant factor classes indirectly affect forest status 74
Table 6.1 The regression between digital value with incidence angle 77
Table 6.2 Result of homogeneity test 81
Table 6.3 The mean forest characteristics of forest status 82
Table 6.4 Density and basal area of mature trees 82
Table 6.5 Description of forest class 85
Table 6.6 The land cover area in ha and percentage 87
Table 6.7 Confusion matrix of the Maximum Likelihood classification 92
Table 6.8 Pearson’s correlation pairs of variables 93
Table 6.9 Pearson correlation matrix for forest variables and classified forest classes 95
Table 6.10 Pearson correlation matrix for the variables analyzed for stand volume
estimation 97
Table 6.11 Results of simple regression modeling for stand volume 98
Table 6.12 Results of simple regression modeling for logarithmic stand volume 98
Table 6.13 Error of the volume estimates using three different methods 105
Table 6.14 Risk levels based on relevant factors from poor forest 109
xii

Table 6.15 Risk levels based on relevant factors from agricultural land 112
Table 7.1 Correlation of spectral data and stand volume characteristics 117
Figure 4.3 Rate of the forest categories 46
Figure 5.1 Main approach 49
Figure 5.2 Equipment for stem diameter and tree height measurement
52
Figure 5.3 A common semivariogram form 70
Figure 5.4 A schematic example of regression-kriging: fitting a vertical cross-section with
assumed distribution of
an environmental variable in horizontal space 72
Figure 6.1 SPOT 5 False Colour Image 78
Figure 6.2 Scattergram of the relationship between tree height and breast height diameter
79
Figure 6.3 Inverse J-shaped distribution Meyer N-DBH 84
Figure 6.4 Land cover classification map 87
Figure 6.5 Rate of land cover classes 88
Figure 6.6 Volume map using SPOT 5 image and regression estimator 100
Figure 6.7 Volume map using SPOT 5 image and kNN method 101
Figure 6.8 Histogram of original and log transformed volume 102
Figure 6.9 Experimental variogram with fitted model, log (volume) 103
Figure 6.10 Volume map using SPOT 5 image and regression-kriging method 104
Figure 6.11 Accuracy assessment of standing volume estimations 106
xiv

Figure 6.12 Relation of poor forest area with (a) distance to stream, (b) elevation and (c)
slope level 107
Figure 6.13 Map of risk potential from poor forest status 108
Figure 6.14 Relation of agricultural land area with (a) distance to stream, (b) elevation and
(c) slope lev
el 110
Figure 6.15 Map of risk potential from agricultural land 111
Figure 7.1 Exponential negative relationship of stand volume and band 2 118

C Celsius
Ca.co Canopy cover
Veg.co Vegetation cover
CPU Central processing Unit
Cr.A Crown Area
DBH Diameter at Breast Height
DEM Digital Elevation Model
Dis. to NN tree Distance to nearest neighbor tree
DN Digital Number
e.g. for example
Eq. Equation
ETM+ Enhanced Thematic Mapper Plus
FD Forestry Department
FAO Food Agricultural Organization
GCP Ground Control Point
GoVN Government of Vietnam
GIS Geographic Information System
GPS Global Position System
H Tree height
HRS High Resolution Stereoscopic
ISODATA Iterative Self – Organization Data Analysis
kNN k Nearest Neighbor
LAI Leaf Area Index
MARD Ministry of Agricultural Rural Development
MLC Maximum likelihood classifier
MSS Multi Spectral Scanner
N Tree density
N
DBH35
Density of tree having DBH equal to 35cm and over
1

Chapter I
1 INTRODUCTION
1.1 Background
Forest mapping is one of the most important factors for organization of forest resource
inventory and monitoring. With forest mapping, a classification of the forest in
homogenious strata is carried out as a basis for sustainable forest management plans.
The main purpose of the classification of forest type is to support forest management
planning. The classification should be based on needs of potential resources, forest
characteristics with up-to-date information using the minimum time necessary and at a
reasonable cost. Therefore, the definition of a forest ecosystem and the relevant
characteristics vary with the resource managed and the issue under consideration (Wulder
and Franklin, 2007). The role of classification is to provide a set of criteria that bring a
certain degree of order to ecological community patterns. A classification system has been
developed to address a wide variety of spatial scale and purpose, therefore, a suitable
classification system is not only important for management planning but also for forestry
development strategy. The system can be directly interpreted from imagery and can be
associated with the ground information as well as the ancillary information with an
acceptable accuracy. According to Franklin (2001), the use of remote sensing in
classification is based on the fact that the differences on the ground between vegetation
types can be isolated or separated as differences in the image characteristics. When
different vegetation structures define the classes, and the latter correspond with
recognizable vegetation types on the ground, there is a good reason to believe that the
types can then be mapped with digital remote sensing data and methods (Merchant, 1981).
The goal of remote sensing in forestry is the provision, based on available or purposely
acquired remote sensing data, of information that foresters need to accomplish the various
activities that comprise sustainable forest management. The objective of classification is to

be of great value for understanding patterns of structure and composition, as well as being
important for defining appropriate management interventions. Characterizing the
disturbance regime typically involves assessing the severity, timing, and spatial
distribution of the different types of disturbance affecting the forest (Newton, 2007).
Consequently, an understanding of the natural and societal dynamics also has great
implications for forest management. Knowing the factors which are related to forest
accessibility e.g. topography, accessibility to forest area and so on, and how they impact
natural resources in indirect manner may contribute to better forest management planning.
Although the great progress in remote sensing and GIS in recent decades has provided
considerable opportunities for improving the effectiveness of forest management, an
application of remote sensing in forestry in Vietnam is limited due to several reasons,
3

namely in technique and resources as well. Meanwhile, characteristics of natural
rainforests are specific and complex, and always be put under change by forest land
utilization and social pressures. Correspondingly, the traditional method in forest inventory
cannot satisfy increasing information requirements in a sustainable forest management
system. For this reason, a multi-source approach of combining remote sensing, GIS and
field data may improve effectiveness of forest management planning at the location.
1.2 Problem analysis
The forestry issues in Continental Southeast Asia are probably the most complex in the
world (Stibig, 2003). As in many Southeast Asian countries, Vietnam has experienced the
highest rates of net forest cover decline. Estimates of the change in forest cover in Vietnam
in the last half century vary greatly in measurement. The forests have dramatically
decreased overall during the last 60 years, despite having slightly increased recently. The
forest cover declined from 43% of the country’s area to 30% in 1985, and 28% in 1995
(Lung, 2001; Maurand, 1943). Especially in the war time, large areas of mangrove and
rainforests of Vietnam were eradicated by toxic chemicals (Nakamura, 2007). The
summary tally is that 104,909 ha of mangrove forests and about 3,000,000 ha of inland
forests were eliminated. The total defoliated area amounted to 3,104,000 ha. Some

quality in terms of biophysical variables (i.e. diameter at breast height, stand height,
species composition, basal area, volume, tree density and age) are important. Knowing just
what forests are left, and where these are is a good starting point (Stibig et al., 2003). Thus,
a suitable classification of forests and mapping of forest cover to provide a comprehensive
and reliable view of the region’s forest cover is important to help forestry managers make
appropriate decisions in forest management and planning.
Several ecological classification systems have been developed for Vietnamese forests
during the past century in North and South Vietnam separately. In forest planning and
management, a common system, which came into force from the Vietnamese Government,
has been used since 1975 (see chapter II). This system was established for the whole
country without a region specific focus. The classification criteria of this system was
developed from basis of forest condition in the North region, then applied to the whole
country. As a consequence, this leads to not only difficulties in inventory but also in forest
management strategies.
Vietnam basically has a tropical climate as it lies inside the Northern Tropical Zone.
However, due to a varying monsoon season and complex terrain, the climate of Vietnam
differs with latitude and altitude (Lap, 1999; Thai, 1998). In other words, the climate
differs from the South to the North. Meanwhile, climate – hydrology is the decisive factor
group to physiognomy and structure of vegetation type. Forest types vary in general as a
function of environmental and climatic factors including temperature, rainfall, humidity,
seasonality - often governed by latitude and topography (Thai, 1978). As a result,
characteristics of vegetation vary not only in physiognomy but in structures from the North
5

to the South. Thus, to apply the classification system which was based on the forest
condition of the Northern Vietnam to the whole country is unsuitable, especially for the
forests distributed in the Central Highlands, because they have far different ecological
conditions from the North. In practice, this system has caused some issues while used in
inventory in the Central Highlands, namely:
- When the status is used for forest classification the factor of basal area or


cause of deforestation is due to highest population (people) growth rates in the world
(Laurance, 1999; Wright and Muller-Landau, 2006). Therefore, a successful forest strategy
requires not only a good understanding of the natural but also social conditions and their
relations to the strategy’s objectives. For this reason, assessment of accessibility to forest
resources is also one of the considerations under study.
The main problems as mentioned above are summarised in a diagram of cause and effect
relations, illustrated in Figure 1.1. The unsuitable forest classification system along with
drawback techniques in forest inventory are major causes leading to ineffective forest
management planning.

Figure 1.1 Problem Analysis Tree

7

Within limitation of time and scholarship as well as individual capacity, the three causes
which were rounded in diagram in Figure 1.2 were concerned in this study. The endeavors
to combine modern techniques (remote sensing and GIS) with terrestrial data were done to
improve these problems.

Inadequate Forest
Classification System in the
Central Highlands
Effects
Causes
Not popular RS and
GIS techniques
Unsuitable Forest
Classification System
Difference of natural forests of the