and techniques to handle these multi-dimensional data sets
and accurately extract the required information. These new
methods and techniques include the use of GEographic
Object-Based Image Analysis (
GEOBIA
) where the informa-
tion necessary to interpret an image is represented in image
objects and not represented in a single pixel (Blaschke, 2010).
Methodology of Work
Study Area and Dataset Description
The study area is the Country of Lebanon (total surface of
10,452 km
2
) in the Mediterranean Region (Figure 1). Lebanon
is divided into four distinct physiographic regions: the coastal
plain, the Lebanon mountain range, the Beqaa Valley, and
the Anti-Lebanon mountains. The Lebanon mountain range
where most of Lebanon’s forests are present is carved by nar-
row and deep gorges. The mountain range rises steeply from
the Mediterranean coast to mountains reaching 3,088 meters
above sea level.
An assessment of the land cover classes in Lebanon
showed that burnable lands (forests and other wooded land)
cover 24.5 percent of the Lebanese territory (
MOE/UNDP/ECODIT
2011). The major forest species widespread in Lebanon are
Quercus calliprinos, Quercus infectoria, Quercus cerris var.
pseudo cerris, Juniperus excelsa, Cedrus libani, Abies silicica,
Pinus pinea, Pinus halepensis, Pinus brutia,
and
Cupressus
sempervirens
.
Lebanon’s climate is characterized by dry summers extend-
ing from June to November (Salloum and Mitri, 2013) with
average daytime temperatures above 30°C, and little rain with
around 90 percent of the total annual precipitation falling
between November and March.
Like other Mediterranean countries, fire occurrence and
extent of burned areas in Lebanon occur during the dry sea-
son and are driven by the interactions between fuel accumu-
lation, seasonal drought, temperature, and precipitation (
MOE/
UNDP/ECODIT
2011; Salloum and Mitri, 2013) although most
wildland fires in Lebanon are started by human activities
(
FAO
, 2013).
Geospatial biophysical and climatic data in additional to
other ancillary data were employed. These included the 1998
land-cover/land-use map of Lebanon which was produced
for the Lebanese Ministry of Agriculture (
MoA
) in 2002 using
Landsat and Indian Remote Sensing (
IRS
) satellite images
acquired in 1998. This was the only official map providing a
detailed representation of the National land-cover/land-use.
However, changes in landlcover/landluse (i.e., conversions to
settlements) were identified using a recent database of urban
areas and a map of urban areas produced by Tragsatec in 2010
with the use of Ikonos satellite imagery (Tragsatec, 2012). Data
on protected natural reserves published by
MoA
was also used.
A geo-referenced Landsat
TM
imagery (acquired in 2012), and
a Digital Elevation Model (
DEM
) of 25 meter resolution and
an administrative map of Lebanon were also collected (
MOE/
UNDP/UOB
, 2013).
Datasets of monthly maximum temperature, monthly pre-
cipitation, and mean annual rainfall (1-km spatial resolution)
of current (1950 to 2000) conditions were extracted from the
“Worldclim” database for the calculation of the
KBDI
index.
These data were generated through interpolations of repre-
sentative observed data from major global climate databases
(Hijmans
et al.
, 2005).
Finally, the extents of burned areas between 1999 and 2012
were collected in shapefiles for use in the evaluation of the
classification results. These data were extracted from multi-
temporal Landsat
TM
imagery on a yearly basis from the year
1999 to 2012 (
MOE/UNDP/UOB
, 2013).
GEOBIA Framework
To map wildfire hazard, vulnerability and overall wildfire risk
we used the software ecognition
®
within a
GEOBIA
framework
(Mitri and Gitas, 2004) which utilizes image objects instead of
image pixels. More precisely,
GEOBIA
allows the integration of
a broad spectrum of different object features such as spectral,
shape, and texture (when using Very High Resolution imag-
ery) and contextual values, for image analysis. Synergizing all
these features is expected to address image analysis tasks that
have not been possible until present (Mouflis
et al.
, 2008).
Each mapping iteration required two steps within the
GEOBIA
framework, namely, segmentation and classification of seg-
mented images. First, the strategy behind image segmentation
was to create networked image objects. This means that dif-
ferent object levels can be analyzed in relation to each other.
Segmentations at the image pixel level were generated using
an average abstract scale of 10. The segmentation process
included the following weights: 10 percent for shape, 90 per-
cent for color, and 50 percent for compactness. The segmen-
tation parameters were determined empirically in order to
produce highly homogeneous objects in specific resolutions
and for specific purposes. As a result of conducting segmenta-
tion at two different levels, a hierarchy of image objects was
created. Every image object in this hierarchy is networked in a
manner that each image object ‘knows’ its context in relation
to other image objects. The sub-objects were created at the
pixel level of the Landsat
TM
image, while the super-objects
were created at the cadastral units’ level (i.e., villages) using
the administrative map of Lebanon.
The classification incorporated also contextual information
in addition to other features extracted from the digital values
of the images and from the attribute tables of the shapefiles.
The classification was based on fuzzy logic and consisted of a
combination of several conditions that had to be fulfilled for
an object at a specific level to be assigned to a class. The fuzzy
sets were defined by membership functions that identified
those values of a feature that were regarded as typical, less
typical, or not typical of a class, (i.e., high, low, or zero mem-
bership, respectively, of the fuzzy set).
Subsequent steps are then completed for mapping wildlife
hazard, vulnerability and overall wildfire risk, respectively
(Figure 2). A description of each subsequent step and its as-
sociated term is presented in the following sections.
Mapping Wildfire Hazard
Wildfire hazard mapping involved the classification of (a) fire
spread which was a function of fuel type, fuel combustibility,
and slope of terrain, and (b) fire potential index, which was
mainly derived from geospatial climatic data. Both were clas-
sified at the sub-object level.
The land-cover/land-use map of Lebanon and the Landsat
TM
imagery were used for image segmentation and classifi-
cation for fuel type mapping (Mitri
et al.
, 2012). The Pro-
metheus fuel type classification system (Lasaponara
et al.
,
2006) which is considered to be better adapted to the Mediter-
ranean ecosystem was used (Table 1) for classifying the image
objects into the different types of fuel. Consequently, the fuel
combustibility classification system was associated with the
different fuel type classes (Table 1). Fuel categorization in re-
lation to combustibility was based on a comparison with the
combustibility of the fuel models classification established by
Rothermel and adapted for the Spanish Forest System by the
Directorate General for Biodiversity (Tragsatec, 2012).
Slope steepness plays a significant role in fire spread and
behavior (Keane
et al.
, 2001; Millington, 2005). Accordingly,
fire spread involved classifying image objects based on the
contextual information of fuel combustibility and the slope
500
June 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
