Patrono, Andrea 
  
whole set of mitigation measurements. Erosional problems may be exacerbated by subsequent overgrazing of burnt 
areas. 
1.2 The Study Areas 
A set of 16 test sites heavily affected by fires in the last decade was requested to be investigated by the Western European 
Union Satellite Centre. The analysed sites were principally located in the regions of Thessaloniki, Halkidiki, Attiki, 
Lakonia, Ilia and Arcadia in Greece and the regions of Gelibolu and Marmaris in Turkey. The analysed fire seasons cov- 
ered a time range of 7 years, from 1992 to 1998; the available LANDSAT TM data set covered a time frame of 9 years 
from 1991 to 1999. The acquisition philosophy adopted for each fire, was to acquire a spring image pre-fire, an image 
immediately post-fire and all possible spring images post-fire (trying to match as much as possible the dates of acquisi- 
tion). In the individual areas, the regions affected by fire varied from 1,000 to 8,300 ha. 
2 METHODS 
2.1 Burnt Areas Identification 
Measurement of the surface area of vegetation burnt during fire events is a fundamental input required by environmental 
managers for estimating quantities of biomass burnt and for studying land change processes such as deforestation, soil 
erosion, etc. Spectral characteristics of burnt surfaces are dependent on the type of vegetation cover affected by fire, the 
soil characteristics and the time elapsed since burning occurred, leading to a wide range of spectral responses. Specifi- 
cally, the spectral signature of a burnt area is influenced by two different aspects: the spectral characteristics of the com- 
bustion products and the spectral change due to the partial or complete removal of the pre-existing vegetation (Maselli et 
al., 1996). 
The TM sensor on board LANDSAT 4 and 5 provides useful information on the characteristics of burnt surfaces. The 
TM bands widths are optimised for vegetation discrimination. TM band 4 (near-infrared) is particularly useful to deter- 
mine vegetation types, health, and biomass content. The methods used in this study are illustrated below; they are based 
on the use of spectral response of burnt areas at a given wavelength and on vegetation indices (see 2.3). 
The methods used are the following: (a) the TM band 4 post-fire imagery, when stretched and filtered, delineates the loss 
of vegetation vigour, e.g. caused by the fire, generally highlighted with darker (gray) tones; (b) TM band 7, band 4 and 
band 1 (of post-fire imagery) in RGB colour composites provide in dark red tones a clear overview of the extent of fires; 
(c) multi-temporal RGB colour composites detect spectral changes; pre-fire and post-fire images are used to detect fire- 
related alterations (if available, three TM band 4 images, pre-fire, immediately post-fire and a third from a following 
year, may serve this purpose); NDVIs may be an excellent substitute for the use of TM band 4; (d) principal component 
transformations, a technique designed to analyse information content and distribution in multi-spectral data, were applied 
to images produced by merging pre- and post-fire imagery. The resulting principal components are oriented within the 
data such that areas that show high correlation, i.e. areas with similar DN values on each date, appear in the first compo- 
nent; areas that show low correlation, i.e. the areas that have been affected by fire, appear in the lower order components. 
Results of (a), (b), (c) and (d) had been enhanced combining them through logical and mathematical operation of mask- 
ing and subtracting. Due to varying environmental conditions each method may have differing results and effectiveness, 
however, by combining them, they always fulfilled the objective of detecting all the fire affected sites, analysed in this 
study. 
2.2 Land Cover Classification - 
Once areas affected by fire had been identified, the following objective of the study was to analyse, on pre-fire imagery, 
what had been burnt. This was necessary to study a posteriori the relation between the original vegetation coverage and 
the natural re-growth rates. Multi-spectral data were used to perform the classification using the spectral pattern present 
within the data for each pixel was used as numerical basis for categorization. First several classes of different combina- 
tions of digital values, based on their inherent spectral reflectance and emittance properties were recognized. Representa- 
tive training samples had then to be identified and a numerical description of the spectral attributes of each land cover 
type has to be developed with the help of available land cover maps. Each pixel in the image data set was then catego- 
rized into the land cover class it most closely resembles using the Maximum Likelihood algorithm. 
LANDSAT TM band 3, band 4 and band 5 (of the pre-fire images) had been selected in the study areas for the supervised 
  
1132 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.