3.2. Image Classification of the 1987 Landsat TM and 2001 Landsat ETM Images
Ground reference data was obtained from more than 110 ground data points as signatures for each satellite image. The training points were equally distributed to each cover type with at least 10 points per cover type. For the supervised classification of the 1987 image, the stand type maps of 1980 and 1995 were combined to create ground signatures. Likewise, the stand type maps of 1995 and 2005 were combined to create ground signatures for the supervised classification of the 2000 image. These ground reference points were sampled on the cover type (stand) maps derived from the 2002 Ariel Photography and verified through ground measurements undertaken by the State Forest Management Teams in 2005. In order to classify cover types from the images, signatures were taken from the ground corrected stand type maps and adjusted based on the Transformed Vegetation Index, Principle Components Analysis-PCA and unsupervised classification image. Supervised maximum likelihood classification methods were employed in the analyses. Then the 1987 and 2000 images were checked for accuracy using ground data points that were not used in the original classification process together with other points of known condition, such as forest areas visually surveyed with binoculars, stand maps, urban areas and rock outcrops identified in the image. Equal Control Point methods were used in Erdas Imagine 9.0™ program with at least 30 points for each class [38
]. The accuracy assessment of image was checked for each image and accepted if the accuracy was higher than %80. After the accuracy assessment, all images were clumped, eliminated 3×1 pixels and vectorized in Erdas Imagine 9.0™ program. These coverages were pre-processed to eliminate areas less than 0.3 ha for spatial landscape analysis with Fragstats™.
The Landsat TM image (1987) was successfully classified for nine fuel types, but regeneration areas and open areas were classified with a lower accuracy than other classes (69%, Table 1
). However, this is generally acceptable as the overall classification accuracy is much higher (83%) with the Kappa statistics (Conditional Kappa for each Class) value of 0.812.
The Landsat ETM image (2000) was classified into 11 fuel type classes, which differed from Landsat TM because of the changing land cover classes over time, and better distinguished the mixed land cover class (ÇzMab3) with the kappa statistics value of 0.888. However, regeneration and open area classes were not distinguished successfully from each other (69%, Table 2
). Notwithstanding this, Landsat ETM classification is generally acceptable due to a higher overall classification accuracy of 86% and Kappa statistics (Conditional Kappa for each Class) value of 0.853.
3.3. Determination of fire risk and danger potential indices
Fire risk refers to the probability of ignition, as determined by the presence and activity of causative agents (i.e., man, lightning, etc) [39
]. The term “danger” refers to sum of constant and variable factors affecting the ignition, spread, and resistance to control, and subsequent fire damage [45
]. The risk and danger potential of wildland fires must be determined and mapped both spatially and temporally [39
]. Fire risk and danger potential (FRDP) maps are digital cartography of fire ignition and severity and are based on stand characteristics, topographic features and land use practices in a specific region. These maps are developed through incorporating satellite and surface observations in an index that correlates well with fire ignition and danger.
Fire risk potential index was determined based on land use attributes such as settlement areas and agricultural lands together with species composition, slope and insolation (Table 3
). Following steps were taken in the process. First, each variable class was assigned a fire risk rating (extreme, high, moderate or low) according to the risk potential of each class. Second, each fire risk class was rated on a scale from 1 to 5 (Table 3
). Third, all variables (layers) were then integrated through GIS using the equation generated. The equation used was of the form [46
where FRI is the relative numerical rating of fire risk; SC, species composition (5 classes); AL, proximity of agricultural lands to forest (4 classes); SA, proximity to settlement areas (4 classes); S, slope factor (4 classes); and IS, insolation factor (9 classes). The subscripts i
indicate subclasses determined by the fire risk potential. Finally, criterion-based analysis (Table 4
] was carried out to create fire risk maps showing different categories for 1987 and 2000 (Figure 2
Fire danger potential index was determined based on species composition, stages of stand development, stand crown closure and topographic features such as insolation and slope (Table 5
). Following steps were taken in the process. First, each variable class was assigned a fire danger rating (extreme, high, moderate or low) according to the danger potential of each class. Second, each fire danger class was rated on a scale from 1 to 5 (Table 5
). Third, all variables (layers) were then integrated through GIS using the equation generated. The equation used was of the form:
where FDI is the relative numerical rating of fire danger; SC, species composition (5 classes); CC, stand crown closure (5 classes); SD, stages of stand development (6 classes); S, slope factor (4 classes); and IS, insolation factor (9 classes). The subscripts i
indicate subclasses determined by the fire danger potential.Unlike the FRI, a different approach was employed here in developing the relationship used. Species composition was weighted (squared) and incorporated into the relationship as a multiplier so as to eliminate incorrect classification and obtain a wider range of values. This approach avoids the limitation of a simple additive model in which incorrect values could be obtained irrespective of species composition. Species composition is one of the most important factors affecting fire danger potential. Finally, criterion-based analysis (Table 6
) was carried out to create fire danger maps showing different categories for 1987 and 2000 (Figure 3
As a result, FRDP maps were obtained using seven factors, namely slope, insolation, stages of stand development, species composition, crown closure, proximity of agricultural lands to forest and distance from settlement areas (Tables 3
). The factors used in the calculation of the FRI and FDI were selected based on experience and relevant literature.
Slope does not necessarily have an effect on the probability of an ignition but has a strong effect on fire behavior [46
]. Forest stands on steeper slopes have greater fire danger. Slope was taken as the mean percent slope for each polygon (Tables 3
). Insolation was taken as the mean aspect for each polygon. Southern and southwestern exposures in the northern hemisphere have the greatest fire danger (Tables 3
). Stages of stand development are a measure of forest structure. The accumulation of crown and surface fuels increases with stand age and development [47
]. Forest structure and stand fuel characteristics (fuel loading and continuity) can dramatically change fire danger. The highest fire danger usually occurs in the pole/very young and young stages of stands due to lower crown base height and increased vertical fuel continuity [48
]. The stages of stand development considered in this study involved non-forested/newly planted, shrub/herb (1±15 years), pole/very young (15±30 years), young (31±60 years), mature (61±100 years), and old forest (101 years or greater) (Table 5
Relative species composition is an indicator of site conditions and directly affects flammability of the fuel complexes. Deciduous forest stands are considered to represent low fire danger areas, whereas coniferous (usually pine) stands are generally associated with high fire danger. Mediterranean shrubs are also known to have high fire danger potential due to the high flammability of the fuels and fast spread of fires in these vegetation types (Tables 3
). Crown closure provides an indicator of the ease with which fire can spread. The higher the crown closure, the more intense the fires burn [48
]. Crown closure was taken as the mean percent cover for each polygon (Table 5
). In addition, distances from settlement areas and proximity of agricultural lands to forest were used to provide an assessment of fire risk together with other factors. Both factors are a measure of the extent to which human activities contribute to fire risk (Table 3
). All variables included in the analyses were rated on a scale from 1 to 5.
As a last step, Fragstats™ [51
] was used to quantify landscape structure, fire risk and danger potential of Korudag Forest District for each of the land use classes. Forest stand attributes were classified to easily compare the accuracy of the satellite images and determine the changes in landscape structure at spatio-temporal scale (Table 7
Fragstats calculates a number of spatial metrics for each patch and cover class as well as for the entire landscape. Selected metrics were analyzed for the land use class for the study area in 1987 and 2000. The metrics were: class Percent of Landscape (PL), Number of patches (NP), Largest Patch Index (LPI), Mean Patch Size (MPS) and Area Weighted Mean Shape Index (AWMSI).