Full text of DOI:10.1080/01431160118510

This article was downloaded by: [Dalhousie University]
On: 21 November 2012, At: 07:51
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number:
1072954 Registered office: Mortimer House, 37-41 Mortimer Street,
London W1T 3JH, UK
International Journal of
Remote Sensing
Publication details, including instructions for
authors and subscription information:
http://www.tandfonline.com/loi/tres20
Forest fire analysis with
remote sensing data
F. Sunar & C. Özkan
Version of record first published: 25 Nov 2010.
To cite this article: F. Sunar & C. Özkan (2001): Forest fire analysis with remote
sensing data, International Journal of Remote Sensing, 22:12, 2265-2277
To link to this article: http://dx.doi.org/10.1080/01431160118510
PLEASE SCROLL DOWN FOR ARTICLE
Full terms and conditions of use: http://www.tandfonline.com/page/
terms-and-conditions
This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any
representation that the contents will be complete or accurate or up to
date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not
be liable for any loss, actions, claims, proceedings, demand, or costs or
damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.

int. j. remote sensing, 2001, vol. 22, no. 12, 2265–2277
Forest Ž re analysis with remote sensing data
F. SUNAR
Istanbul Technical University, Civil Engineering Faculty, Remote Sensing
/
Division, Maslak 80626, Istanbul Turkey
¨
and C. OZKAN
Erciyes University, Geodesy and Photogrammetry Engineering Department,
/
Kayseri Turkey
(Received 21 May 1999; in Ž nal form 14 March 2000)
Abstract. Forest Ž res cause major damage to the environment, human health
and property, and endanger life. Fires can be monitored and analysed over large
areas in a timely and cost-eVective manner by using satellite sensor imagery in
combination with spatial analysis as provided by Geographical Information
Systems (GIS). In this study, the forest area damage caused by a large Ž re which
Ï
occurred in the Marmaris, province of Mugla in July 1996 was analysed using
satellite sensor images. Digital image processing methods, such as spectral proŽ le
analysis, vegetation indices and multispectral classiŽ cation, were applied to the
satellite sensor images acquired before and after the forest Ž re. Besides the conven-
tional maximum likelihood classiŽ cation algorithm, a multilayer feed-forward
neural network architecture was also used for comparison and evaluation of its
eVectiveness. A GIS database was constructed from the raster (satellite sensor
data), vector (the forest type and topographical maps) and ancillary data (meteoro-
logical data). The GIS is being used to develop an information and decision
support system to monitor and predict forest Ž re activity, and to enhance Ž re
management eYciency. This study highlights the deŽ ciencies in the current
approach to Ž re management and emphasizes the need for an improved method
along the lines outlined.
1. Introduction
Forest Ž res, whether natural or initiated by man, have an important in uence
on the environment, human health (often leading to fatalities) and property. In terms
of annual damage, Turkey is one of the most susceptible countries in the
Mediterranean region (Ž gure 1). According to the National Forest Service records
(State Forest Directorate 1998), the causes of Ž res are various but are mainly, more
than 97% of the time, due to human activity while the rest, less than 3%, are caused
by natural phenomena, such as lightning. Human related Ž res are mainly grouped
as intentional and unintentional Ž res. The unintentional Ž res are mostly due to
carelessness caused by hunters or picnickers, whereas the intentional Ž res are mostly
related to forest or shrub elimination, for transformation of the land for agricultural
purposes, or generation of pastures.
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online © 2001 Taylor & Francis Ltd
http://www.tandf.co.uk/journals

¨
F. Sunar and C. Ozkan
2266
Figure 1. Forest Ž res and the most susceptible regions in Turkey (Forest Fire Statistics 1998).
The information needed for Ž re management includes Ž re risk mapping, Ž re
detection and monitoring, damage assessment and planning post-Ž re recovery.
EVective management requires the use of new techniques using remote sensing and
Geographical Information Systems (GIS); these must, however, produce results in a
timely and cost-eVective manner (Fox and Stuart 1994). In Ž re damage assessment,
satellite sensor imagery in conjunction with ground observations is the main source
of information from which pre-Ž re vegetation condition can be determined. Pre-Ž re
condition is important both in order to assess potential Ž re risk and to make rational
judgements on the precautionary measures required to avoid potential soil erosion
which topographic and meteorological conditions may promote on newly burnt land
(Chuvieco and Martin 1994).
One of the largest forest Ž res in Turkey in recent years occurred in the Marmaris,
province of MugÏla in July 1996. The eVect of this Ž re was analysed using satellite
sensor images acquired before and after the forest Ž re. A GIS was established to
monitor and predict the extent of the damage, and to enhance Ž re management
eYciency.
2. Study area
The Ž re occurred in the area bounded by the MugÏla Forestry Chief Directorate,
the Marmaris Forestry Chief Directorate, and the Cetibeli Forestry Chief Directorate
¸
OYce (42 W 35 N–40 W 33 N), located on the southern coast of Turkey (Ž gure 2).
°
°
°
°
The Ž re started in the northern part of the area on 27 July 1996, spreading to the
forests of the Hisaronu and Marmaris Forestry Chief Directorate, and continued
burning for 4 days.
¨
¨
The topography of the area is broken and steep commonly having slopes steeper
than 5–12%. Elevations range from 50 m to the ridge tops at 600 m. The average
August temperature is 42 C and the annual average precipitation is 1257 mm. The
°
study area is covered with forest dominated by Red Pine (Pinus brutia).
3. Data collection
EVective decision-support systems are needed to monitor, and predict Ž re activity
and enhance sustainable Ž re management. To develop and use such a system
requires a computerized forest Ž re information system supported by remotely sensed
information. In this study, inputs to the system were obtained from the following
sources:
(i) Raster data: one Landsat TM (Thematic Mapper) image acquired before the

Forest Ž re analysis with remote sensing data
2267
Figure 2. The study area.
Ž re (24 July 1996), and two images, IRS-1C LISS (Linear Imaging Self-Scanner) and
SPOT HRV (High Resolution Visible) Panchromatic, acquired after the Ž re
(22 August 1996 and 15 November 1996, respectively) were used in the analysis. Pre-
Ž re condition was assessed from the Landsat TM imagery, whereas post-Ž re condition
and the areal extent of the burned area were assessed from the IRS-1C data. SPOT
panchromatic data was only used for visual analysis due to its broad spectral
resolution.
(ii) Vector data consists of map data. The three national topographic maps
(1:25 000 scale) which covered the area were digitized at 50 m contour intervals.
Thematic forestry maps (1:25 000 scale), prepared by the MugÏla Forestry Directorate
OYce, which showed previously burned areas, were digitized. On these forest maps
the burnt areas were classed as productive, disturbed and aVorested.
(iii) Ancillary data: meteorological data (mean and maximum wind velocity, wind
direction, mean relative humidity (%), mean air temperature (dry thermometer C)
°
and total precipitation (mm)) were obtained from the meteorological station located
in the study area. The data were selected with respect to the acquisition dates of the
satellite sensor data, i.e. between July and November 1996.
4. Application
There are several diVerent ways to detect Ž re damaged forested areas using digital
image processing: spectral proŽ le analysis, vegetation indices and classiŽ cation.
4.1. Spectral proŽ le analysis
All surface features have spectral response patterns. If these patterns, or spectra,
are unique then the features can clearly be identiŽ ed from spectral information using

¨
F. Sunar and C. Ozkan
2268
remotely sensed data. Although it may not be feasible to deŽ ne unique signatures
for features with similar spectral characteristics (such as between some vegetation
types), distinctive spectral patterns for vegetated versus non-vegetated areas, and
also the existence of diseased or stressed vegetation, can often be recognized. With
this in mind, four identiŽ able points on the images were selected arbitrarily and their
spectral re ectance proŽ les were extracted from the pre- and post-Ž re imagery
(Ž gure 3). These points were then evaluated with, and compared to, the reference
information contained in the forestry maps.
4.2. Vegetation indices
These are quantitative measures, based on vegetation spectral properties, that
attempt to measure biomass or vegetative vigour. The simplest form of vegetation
index is simply a ratio between two digital values from independent spectral bands.
High values of the vegetation index identify pixels covered by substantial proportions
of healthy vegetation (Campbell 1987). Vegetation indices can be classiŽ ed into two
broad categories: the ratio indices and orthogonal indices. The Ž rst category, used
in this study, are based on ratios of the digital values in the red (R) and near-infrared
(NIR) spectral bands, chosen because of the inverse relationship between the
re ectance of vegetation in these regions. In this study, the Normalized DiVerence
Vegetation Index (NDVI NIR R/NIR R), the most commonly used ratio trans-
=
-
formation for vegetation studies, was applied. The normalization of the NDVI
+
reduces the eVects of variations caused by topography.
The two satellite sensor images, taken before and after the Ž re, were co-registered
with a normal registration procedure and ratioed (IRS NIR/Landsat NIR and
IRS NDVI/Landsat NDVI) to assess the changes occurring between the two dates.
4.3. ClassiŽ cation
In this study, two diVerent classiŽ cation methods were applied to classify Ž re
damaged forest areas: conventional classiŽ cation and neural network classiŽ cation.
4.3.1. Conventional classiŽ cation
The classiŽ cation process is described as the identiŽ cation of the pattern associ-
ated with each pixel position in an image in terms of characteristics of the objects
or materials at the corresponding point on the Earth’s surface (Mather 1987). The
problem of allocating pixels to their most likely class can be approached in one of
two ways. The Ž rst one, unsupervised classiŽ cation, can be deŽ ned as the identiŽ ca-
tion of natural groups, or structures, within multispectral data whereas the latter,
supervised classiŽ cation, can be deŽ ned informally as the process of using samples
of known identity (Campbell 1987).
In this study, the classiŽ cation method was based on an interactive Iterative Self-
Organizing DATA (ISODATA) algorithm and was used to determine the initial
classes. The resulting classes were examined both spectrally and spatially with ground
information and merged, split or discarded and reclassiŽ ed as appropriate. The
statistics from these classes were than used as input to a maximum likelihood
classiŽ cation which basically allocates pixels to the class to which they have the
highest probability of belonging. This approach assumes that the real classes to
discriminate are represented in the image data as normal distributions around class
mean values. The form of each normal distribution is determined by mean and
variance/covariance of the class. Allocation of pixels to classes involves using class

Forest Ž re analysis with remote sensing data
2269
distribution statistics as a probability density function for the class which indicates
the likelihood of a particular pixel value belonging to the class (Swain and Davis
1978, Harrison and Jupp 1990).
4.3.2. Neural network classiŽ cation
A neural network, applied to a number of image classiŽ cation problems recently,
is a directed graph consisting of neurons or nodes arranged in layers with intercon-
necting links (Haykin 1994). These structures represent a system composed of many
simple processing elements operating in parallel whose function is determined by
network structure, connection, weights and node function (Hara et al. 1994). The
main characteristics of neural networks are (i) they have an intrinsic ability to
generalize, (ii) they make weaker a priori assumptions about the statistical distribu-
tion of the classes in the dataset than a parametric Bayes classiŽ er, and (iii) they are
capable of forming highly nonlinear decision boundaries in the feature space (Bock
1996 ).
Various neural network architectures and algorithms have been developed and
used extensively, e.g. in classiŽ cation, forecasting and modelling. There exist two
primary types of neural network architectures: supervised and unsupervised. The
supervised method, used in this study, is a process achieved by the presentation of
the input data to the neural network and then performing a comparison of the actual
known output together with the predicted output. The unsupervised method is a
process when the network is able to discover statistical regularities in its input space
and automatically develops diVerent modes of behaviour to represent diVerent classes
¨
of inputs (Ozkan and Sunar 1999, BerberogÏlu and Curran 1998).
5. Results
Since Ž re-aVected green areas are stressed and nutritionally deŽ cient environ-
ments for plant growth, the spectral properties of these areas, measured through
time from remotely sensed data, will permit the burnt vegetation canopy to be
identiŽ ed and assessed (Sunar and Taberner 1998). As may be seen in Ž gure 3(a), a
characteristic canopy re ectance curve for selected four points in the Landsat TM
image shows relatively low values in the red and blue regions in the visible spectrum
with a peak in the green spectral band due to absorption of blue and red light by
chlorophyll and other pigments, whereas there is high re ection in the infrared region
due to internal leaf structure (Lillesand and Kiefer 1994). However, the curve has
diVerent characteristics in the IRS-1C image due to forest Ž re burning (Ž gure 3(b)).
The diVerences observed in these two graphs made it possible to analyse the vegeta-
tion cover destruction due to forest Ž re and to interpret the classiŽ cation results
more accurately.
Since Ž re removes variable amounts of green biomass from a plant canopy, and
the amount removed depends upon the severity of the Ž re, there is a good correlation
between the vegetation indices and Ž re severity. In this study, the NDVI was applied
to Landsat TM (pre-Ž re, Ž gure 4(a)) and IRS-1C (post-Ž re, Ž gure 4(b)) images and
scaled from 0 to 255 for image display. As may be seen in Ž gure 4(a), vegetated areas
generally yield high values because of their relatively high near-infrared re ectance
and low visible re ectance, whereas the heavily burned areas (black) were easily
distinguished from unburned areas (white) in Ž gure 4(b). The degree of change that
has taken place between the two dates in the study area can be readily delineated
in the IRS NIR/Landsat NIR ratio image (Ž gure 4(c)).

¨
F. Sunar and C. Ozkan
2270

Forest Ž re analysis with remote sensing data
2271
The classes obtained from the classiŽ cation procedures corresponding to the
burned forest areas were identiŽ ed by examining them both spectrally and spatially
(Ž gure 5(a)). The major ground cover types isolated in the study area were (i) burned
forest area; (ii) forest area; (iii) urban areas; (iv) bare soil; (v) rocky area, and (vi) sea.
The classiŽ ed image agreed with the thematic map prepared by the regional Forestry
Chief Directorate in general, but it was found that there were some diVerences in
the forest structures in the southern and northern parts (Ž gure 5(a), (b)). This was
especially notable in the southern parts—the vigorous forest structure (shown in
pink in the thematic map (Ž gure 5(b)) had been aVected, depicted by a decrease in
¨
vegetation cover with time (Ozkan 1998).
A multilayer feed-forward perceptron model, which consists of an input layer,
two hidden layers of processing nodes and an output layer, was applied to IRS-1C
image data. The network was trained with a back-propagation learning algorithm
comprising a forward and a backward phase through the neural-network structure
and a 0.4 learning momentum. The network architecture consists of two input units
(IRS-1C near-infrared and red bands, respectively), Ž ve Ž rst hidden units, 15 second
hidden units, and six output units representing land cover classes. The training
dataset used comprised 864 pixels with diVerent number of pixels per land cover
class (259, 217, 57, 219, 50 and 67 pixels for burned forest area, forest area, urban
areas, bare soil, rocky area and sea, respectively) and the network was trained cycles
of 2000 with a learning rate of 0.3. The transfer function used in this study was
sigmoid. The neural network classiŽ cation result of IRS-1C image was given in
Ž gure 6.
The classiŽ ed IRS-1C images obtained each classiŽ cation methods were registered
to the national Universal Transverse Mercator (UTM) map projection system to
quantify and delineate the burned areas. According to the records of the regional
Forestry Chief Directorate, the area burned was 7094 ha, this is much higher than
that calculated from both conventional methods (6290 ha) and artiŽ cial neural net-
work algorithm (6294 ha). There are three main reasons for this discrepancy:
(1) The thematic map prepared by the regional Forestry Chief Directorate delin-
eated the burned areas using conventional methods: the boundaries of the
burnt forest areas were approximately delineated on the forestry plans by
inspecting the damage from helicopter or by Ž eld checking. The areal extent
was calculated using planimetric methods.
(2) Some of the aVorested regions, shown with red circles in the classiŽ ed image
(Ž gure 5(c)), were not classiŽ ed as burnt though they were marked as burnt
on the thematic map. It is thought that this is a result of the generalization
of the thematic map prepared after the Ž re.
(3) Remote sensing enables more objective, deŽ nitive and unbiased decisions to
be made from a synoptic viewpoint with capabilities to detect and monitor
vegetation biomass in more detail than approximate ground estimates.
6. Forest Ž re management system
Assessments of the impact of human activities on the environment requires
qualitative and quantitative information on the spatial distribution and dynamics of
natural resources. Fire, the most important disruptive factor in forestry in the
Mediterranean, often leads to rapid loss of soil fertility, to erosion, to desertiŽ cation,
and ultimately to poorer human habitability and survival. GIS integration with

¨
F. Sunar and C. Ozkan
2272
(a)
(b)
(c)
Figure 4. NDVI image of the (a) Landsat TM image and (b) IRS-1C LISS image. (c) The
/
ratio image of IRS NIR Landsat NIR.

Forest Ž re analysis with remote sensing data
2273
Burned forest area
Forest area
Bare soil
Rocky area
Urban
Sea
(a)
Productive forest
Afforested area
Disturbed area
Unburned forest
3
0
3
6 Kilometers
(b)
(c)
Figure 5. (a) ClassiŽ ed image of the Landsat TM (left) and the IRS-1C (right). (b) Thematic
map of the burned area prepared by the Forestry Chief Directorate. (c) Con icts
between the classiŽ ed IRS-1C image and the thematic map. Red circles demonstrate
unburned aVorested areas.

¨
F. Sunar and C. Ozkan
2274
Figure 6. Neural network supervised classiŽ cation result of IRS-1C image.
multitemporal remotely sensed data is necessary, and achievable, to protect and
manage forests before, during, and after Ž res (Chuvieco and Congalton 1989, Klaver
et al. 1997).
The reasons why a computerized forest Ž re information system used in conjunc-
tion with remotely sensed imagery is required can be outlined as follows:
1. To identify the forest Ž re risk areas by taking into account the actual and
predicted meteorological conditions and the vegetation condition before the
Ž re to enable the necessary control and precautionary measures to be
instigated.
2. To supply other types of crucial information, such as the location of human
settlements, water resources, road networks, etc., for sustaining Ž re manage-
ment and controlling Ž re-Ž ghting during the Ž re.
3. To retrieve information on the vegetation condition after the Ž re in order to
assess forest Ž re damage and to evaluate the post-Ž re situation with and, in
particular, in relation to soil and topography in order to minimize erosion
and control recovery.
In this study, the Forest Fire Information system was established with
ArcView V.3.0. The thematic map derived from the Forestry Chief Directorate

Forest Ž re analysis with remote sensing data
2275

¨
F. Sunar and C. Ozkan
2276
was overlayed on the pre- and post-Ž re satellite sensor images to determine and
compare the inter-relationships between Ž re eVects, the types of vegetation aVected,
and the areal extents (Ž gure 7). Meteorological data was only used as an input to this
system, but will be intended to use for preparation of the forest Ž re risk map in
conjunction with the weather forecast and the vegetation condition in future. All entries
in the dataset relating to the study area were registered and referenced to a national
UTM projection system. For the raster datasets, the geometric registration process was
carried out using standard techniques with 75 ground control points, a Ž rst-degree
polynomial equation, and resampling using cubic convolution with a
registration accuracy.
0.5 pixel
±
Future improvements will include the use of any additional information such as
short-term risk prediction, i.e. distance to roads, distance to human settlements and
to land-use interfaces together with digital terrain model.
7. Conclusions
In recent years, GIS has emerged as a powerful tool in conjunction with multitem-
poral remote sensing data for studying and monitoring the impact of many environ-
mental factors. Many government agencies now use a wide range of computer-based
information systems to handle geographic aspects of their business, particularly
mapping, by integrating spatial data collected from diVerent sources, and in dif-
ferent formats, into GISs. In this way, it is possible to address the critical questions
of how economic, meteorological, hydrological and other processes interact with
geographically disposed natural resources like forests.
With its rich biological diversity, the Turkish forests (covering 26% of the country)
are very important, and forest Ž res are the most important of the factors threatening
them. Today, the concept of sustainable management of forest has changed consider-
ably, and advanced forest management systems play an important role in organizing,
maintaining, accessing and reporting on all aspects of forest information required to
support management planning functions such as control, scheduling and monitoring
resources for the future. Forestry in Turkey has to change quickly to adapt to this
new concept, however, there is still a signiŽ cant gap between what is possible, in
terms of knowledge and management from the new techniques such as remote
sensing, and traditional forestry techniques. Reducing this gap is only possible if it
can be shown that digital remote sensing techniques are at least as accurate and
economically viable as current operational tools, and that GIS can play a decisive
role in the measurement and assessment stages of a forest management strategy. An
integrated management system could have an important role in the planning, review
and improvement of current methods, primarily through better modelling.
Over recent years, technologies using GIS and remote sensing have gained
momentum in Turkey. Many of these GIS/remote sensing applications, however, are
not fully integrated with organizational business processes nor are they being used
to their greatest potential in resource analysis. Success will also require a government
initiative in data management and in the integration of data  ows within and between
agencies. As shown in this paper, the analysis of the Ž re eVects using satellite sensor
images and GIS uncovered signiŽ cant inadequacies in the results of areal extent and
type of forest destruction achieved by the Forestry Chief Directorate using conven-
tional and neural network classiŽ cation algorithms. It was also seen that the result
from the neural network classiŽ cation is quite harmonious with that calculated using
conventional remote sensing methods. The results presented here suggest that a

Forest Ž re analysis with remote sensing data
2277
co-operative forest management programme that monitors, assesses and reports on
the long-term status, changes and trends in forest ecosystem health, based on the
new technologies, should be initiated, developed and incorporated within the regional
forest services/public sector.
Acknowledgments
We wish to thank Mr Engin Akgoz and the GAF company for IRS-1C data
¨
supply and support. We also thank Mrs AysÉe Demirel for supplying the thematic
forest maps. We would like to thank Dr M. Taberner, Bristol University, England,
for help and valuable suggestions.
References
Ï
Berberoglu, S., and Curran, P. J., 1998, Optimising the remote sensing of Mediterranean
land cover, Proceedings of the 24th Annual Conference and Exhibition of the RSS, T he
University of Greenwich, 9–11 September, UK (Nottingham: Remote Sensing Society),
pp. 105–106.
Bock, S., 1996, A region-based approach to land-use classiŽ cation of remotely-sensed image
data using artiŽ cial neural networks. International Archives of Photogrammetry and
Remote Sensing, Vol. XXXI, Part B7, Vienna.
Campbell, J. B., 1987, Introduction to Remote Sensing, 2nd edn (New York: The Guildford
Press).
Chuvieco, E., and Congalton R. G., 1989, Application of remote sensing and geographic
information systems to forest Ž re hazard mapping. Remote Sensing of Environment, 29,
147–159.
Chuvieco, E., and Martin, M. P., 1994, A simple method for Ž re growth mapping using
AVHRR channel 3 data. International Journal of Remote Sensing, 15, 3141–3146.
Fox, L. III, and Stuart, J. D., 1994, Detecting changes in forest condition following wildŽ re,
using image processing and GIS. ASPRS T echnical Papers: 1994 ASPRS-ACSM
Annual Convention, American Society of Photogrammetry and Remote Sensing, Reno,
Nevada, April 1994 (Maryland: American Society for Photogrammetry and Remote
Sensing), pp. 197–206.
Hara, Y., Atkins, R., Yueh, S., Shin, R., and King, J., 1994, application of neural networks
to radar image classiŽ cation. IEEE T ransactions on Geoscience and Remote Sensing,
32, 100–109.
Harrison, B. A., and Jupp, D. L. B., 1990, Introduction to image processing. CSIRO, Australia.
Haykin, S., 1994, Neural Networks (New York: Macmillan).
Klaver, J. M., Klaver, R. W., and Burgan, R. E., 1997, Using GIS to assess forest Ž re hazard
in the Mediterranean region of the United States, 1997 ESRI Users Conference, San
Diego, CA, 8–11 July,
//
/
/
/
/
/
http: www.esri.com base common userconf proc97 HOME. HTM.
Mather, P. M., 1987, Computer Processing of Remotely-Sensed Images, An Introduction
(Chichester: Wiley).
¨
Ozkan, C., 1998, Forest Ž re analysis with remote sensing data. MSc Thesis, ITU, Istanbul,
Turkey {in Turkish}.
Ozkan, C., and Sunar, F., 1999, The use and eVectiveness of artiŽ cial neural networks in
¨
forest Ž re classiŽ cation. Proceedings of the RSS 99 Earth Observation, From Data to
Information, University of Wales, 8–10 September, UK (Nottingham: Remote Sensing
Society), pp. 767–772.
State Forest Directorate, 1998, Forest Ž re statistics, Turkey, HYPERLINK
//
//
http: www.ogm.gov.tr : www.ogm.gov.tr
Ï
Sunar, F., and Taberner, M., 1998, An evaluation of forest burning in the Kapidag Peninsula,
Turkey, using multitemporal satellite data. Proceedings of the 27th International
Symposium on Remote Sensing of Environment—Information for Sustainability,
8–12 June, T romso, Norway (Baltimore, MD: International Center for Remote Sensing
of Environment), pp. 648–651.
Swain, P. H., and Davis, M. S., 1978, Remote Sensing: T he Quantitative Approach (New York:
McGraw Hill).