Combining spectral mixture analysis and object-based classification for fire severity mapping

This study shows an accurate and fast methodology in order to evaluate fire severity classes of large forest fires. A single Landsat Enhanced Thematic Mapper multispectral image was utilized with the aim of mapping fire severity classes (high, moderate and low) using a combined-approach based in a spectral mixing model and object-based image analysis. A large wildfire in the Northwest of Spain was used to test the model. Fraction images obtained by Landsat unmixing were used as input data in the object-based image analysis. A multilevel segmentation and a classification were carried out by using membership functions. This method was compared with other simpler in order to evaluate the suitability to distinguish between the three fire severity classes above mentioned. McNemar’s test was used to evaluate the statistical significance of the difference between approaches tested in this study. The combined approach achieved the highest accuracy reaching 97.32% and kappa index of agreement of 95.96% and improving accuracy of individual classes.


Introduction
Large forest fires are becoming more frequent in Mediterranean areas due to climatic factors and changes in lifestyles and economic conditions.They are one of the most important causes of environmental alteration and land degradation in the Mediterranean Basin, because of the post-fire exposure of bare soil to rainfall (Leone and Lovreglio, 2005).The main consequences of fire on vegetation depend largely on fire severity.In this study, the term fire severity is defined as the conditions resulting from fire, which can be described by the degree of mortality in above ground vegetation (Ryan and Noste, 1985;Patterson and Yool, 1998;Morgan et al., 2001;Rogan and Franklin, 2001;Key and Benson, 2002;Miller and Yool, 2002;Van Wagtendonk et al., 2004;Doerr et al., 2006).Fire severity maps may complement ecosystem management, providing foresters with baseline data on fire severity and extent required for fire management.Data from these maps may be used to identify areas that have experienced differing fire severity, to plan and monitor restoration and recovery activities, to provide a method for updating current vegetation maps and information for future pre-fire planning (Patterson and Yool, 1998;Brewer et al., 2005).It is important therefore, to dispose of techniques to efficiently evaluate fire effects in burned areas.
Considering the extremely broad spatial expansion and often limited access to areas affected by fires, satellite remote sensing provides an important means of gathering information about a burned area in a timely and consistent manner (Rogan and Yool, 2001).Optical satellite imagery from Landsat Enhanced Thematic Mapper (ETM+) has been chosen for this work because the midinfrared reflectance of vegetation is strongly related to important vegetation canopy characteristics relative to fire effects.It was decided to employ only one post-fire image as it is considered of great interest to find quick and affordable methodology for obtaining fire severity maps avoiding the use of pre-fire images.In doing this, money and time would be saved in terms of obtaining, correcting and normalising images.Landsat missions such as Multi Spectral Scanner (MSS), Thematic Mapper (TM) and ETM+ have been widely used for mapping fire severity (Ryan and Noste, 1985;Milne, 1986;Chuvieco and Congalton, 1988;White et al., 1996;Key and Benson, 1999;Key et al., 2002;Key and Benson, 2004;Roldán-Zamarrón et al., 2006;De Santis and Chuvieco, 2007;González-Alonso et al., 2007;Miller and Thode, 2007;Wimberly and Reilly, 2007;Hoy et al., 2008;Verbyla et al., 2008;Norton et al., 2009).
A lack of spectral contrast is partly responsible for the classical errors related to post-fire classifications of burned areas (Koutsias et al., 1999): confusion of burned areas with dark land covers (water, dark forests), confusion between slightly burned and sparsely unburned vegetation (problem of the mixed pixels), difficulties in discriminating severity of burning, and confusion between burned vegetation and non-vegetated categories, such as urban areas.To minimize these prob-lems, it has usually been necessary to combine diverse remote sensing systems and a variety of image processing techniques (Justice et al., 1993).The range of methods dealing with level-of-damage mapping using post-fire satellite data includes, among others: (1) vegetation indices (White et al., 1996;Key and Benson, 1999;Key et al., 2002;Díaz-Delgado et al., 2003;Chafer et al., 2004;Van Wagtendonk et al., 2004;Epting et al., 2005), (2) linear transformation techniques such as principal components (PC) analysis and Kauth-Thomas transform (KT) (Patterson and Yool, 1998), (3) spectral unmixing (Roldán-Zamarrón et al., 2006), etc.
Among the large number of techniques applied for the characterization of burned areas, only a few have quantitatively compared their accuracies (Chuvieco and Congalton, 1988;Koutsias et al., 1999), offering little information about the potential and limitations of each technique.To address this issue, this study focuses on a quantitative comparison of four approaches for mapping fire severity using a case of study of a large fire that burned in Northwest Spain, in 1998.We are particularly interested in finding synergies combining both a subpixel-based approach such as Spectral Mixture Analysis (SMA) and an Object-based Analysis Image (OBIA).SMA approach has been widely used due to its ability to cope better with the problem of the mixed pixel and minimize the effects of topography on satellite data (Caetano et al., 1994;Caetano, 1995;Caetano et al., 1996;Cochrane and Souza, 1998;Rogan and Franklin, 2001;Rogan et al., 2002).SMA has the potential of producing results that are directly related to post-fire land management (Caetano et al., 1994;Cochrane and Souza, 1998;Roldán-Zamarrón et al., 2006).In the case of post-fire assessment, the potential of spectral unmixing relies on the sub-pixel analysis of the materials of a burned area and it has been considered advantageous over vegetation index-based methods, due to its improved capability to distinguish burns from other bare or sparsely vegetated areas (Caetano et al., 1996;Díaz-Delgado et al., 2001).Despite Object-based classifications are increasingly being used to.In comparison with pixels, image objects carry much more useful information and, therefore, can be characterised by far more properties, such as form, texture, neighbourhood or context, than pure spectral or spectral-derivative information (Baatz and Shäpe, 1999).Object-based classification models have been developed and applied on Landsat TM (Mitri and Gitas, 2002;Mitri and Gitas, 2004a;Mitri and Gitas, 2004b), NOAA-AVHRR images (Gitas et al., 2004), and IKONOS images (Mitri and Gitas, 2006;Mitri and Gitas, 2008) resulting in the accurate mapping of burned areas in the Mediterranean areas.
The main objective of our research is to demonstrate the superior accuracy obtained using a combined approach (SMA plus OBIA) for fire severity mapping with medium-resolution remote sensing image than the obtained by a more traditional approaches.

Study site description
The study site, 'Tabuyo del Monte', is located in the Sierra del Teleno, in Northern Spain (figure 1).It is a small mountain chain in the South-East (SE) of León province (Spain) with SE aspect, a maximum slope of 11% and elevation ranges from sea level between 850 to 2,100 m.
The climate is Mediterranean with an average annual rainfall between 650 and 900 mm and two or three months of dryness in the summer time.Soil in this area is very sandy and acidic (pH=5.5)(Calvo et al., 1998).Currently vegetation is a large natural Pinus pinaster Ait.community covering 11,500 ha.Spanish Vegetation Map shows that into the fire scar roughly the 78% was covered by pineland, 18% by shrubs and 4% by Pyrenean oak.Fires have occurred frequently in this community, generally affecting small areas and mostly caused by dry spring-summer storms.However, in September 1998 there was a large fire presumably caused by a military manoeuvre, which burned more than 3,000 ha during four days (between September 13 and 17).This fire is the object of this study.

Remotely sensed data
No Landsat cloud-free scenes close to the wildfire date were found so the first scene available corresponded on September, 16 1999.Van Wagtendonk et al. (2004) used also one year after the fire occurred postfire Landsat ETM+ for fire detection.Key (2005) pointed out that extended assessment (EA), (it occurs during the first growing season after fire) may provide more complete representation of actual fire effects.It captures first-order effects that include survivorship and delayed mortality of vegetation present before fire.The former is detected by regrowth from roots and stems of vegetation that burns but remains viable (McCarron and Knapp, 2003;Safford, 2004).Most other first-order effects, such as char, scorch and fuel consumption, are expected to persist until the next growing season, with two exceptions.Areas prone to surface erosion from wind or precipitation may show a decrease in ash cover and an increase of newly exposed mineral soil.Also, canopy foliage that is heat scorched or dies from girdling may drop to ground litter over the interval before EA.Since such effects are more or less complementary in regards to severity assessment, these delayed responses are not expected to significantly alter the remotely sensed magnitude of change detected between initial and extended assessment.In addition, is complete, so the extent of perimeters and distribution of severity represents final conditions.
Preprocessing of remotely sensed images is a preparatory phase that, in principle, improves image quality for further analyses.In this study only a geometric correction was performed.Atmospheric correction was not necessary since only one post-fire image was used to map the fire severity into the fire scar and it was cloud free.In addition, there is a likelihood that uneven implementation of corrections would not necessarily provide a better representation of the mixing space of the SMA model (Elmore et al., 2000).
For the geometric correction a set of 22 Ground Control Points (GCP's), selected using the National Topographic Map (Instituto Geográfico Nacional, IGN) at 1:50,000 (UTM 30 T European 1950 mean), and a 25 (1) low: areas where shrubs to 2 m burned and no or partial canopy scorched.(2) moderate: areas where shrubs incinerated and canopy scorched.(3) high: areas where shrubs incinerated and canopy completely burned and apparently dead, even though some plants may still be able to sprout.

Data analysis
The development of the main proposed methodology involved two cascaded image analysis techniques: linear spectral mixture (SMA) and object-based image analysis (OBIA).
Image objects were extracted from the fraction images (obtained from SMA algorithm) in the segmentation procedure prior to classification (4 th approach).In order to emphasize the benefits achievable using the adopted approach they were provided quantitative evaluations and comparisons with other approaches (1 st , 2 nd and 3 rd ) (figure 2).
(1) First approach: data analysis for Pixel-Based Method (ETM+ISODATA) The first approach is a pixel-based image unsupervised classification by Iterative Self-Organizing Data Analysis Technique (ISODATA) (Sunar and Özkan, 2001;Miller and Yool, 2002) to ETM+ image.ISODA-TA is clustering algorithm that compares the radiometric value of each pixel with predefined number of cluster attractors and shifts the cluster mean values in a way that the majority of the pixels belongs to a cluster.In this case, we interacted with the procedure at the beginning indicating the number of the predefined cluster to be created and the iterations to be carried out and at the end, where it decides which class represents which surface objects and merges or rejects the classes with nonrealistic representatives.We masked the satellite image with the official fire perimeter polygon in order to estimate the fire severity categories in the fire scar.
(2) Second approach: data analysis for Subpixel-Based Method (SMA+ISODATA) Because the spatial resolution of Landsat ETM+ imagery is 30 by 30 m, the materials in a given picture element (pixel) are rarely represented by a single physical component.Therefore, in the first stage of this approach (figure 3), a linear spectral model was used which is based on the assumption that the image spectra m-grid size digital elevation model (DEM) were used.A first-order polynomial warp function was applied and a nearest neighbour resampling protocol was then used to preserve original pixel values (Jensen, 1996;Lillesand and Kiefer, 2000).The root mean square error (RMSE) for the transformation was less than 1 pixel.In general, road intersections were used as GCP's since it was possible to locate them in the image.An illumination correction was performed with the C-correction (Teillet et al., 1982).This lessens the effects of shadows that may occur due to elevation variations in the landscape.

Field data
In addition to the optical satellite image data, field data were collected in the autumn of 1999 from 72 random plots for fire severity in the Tabuyo burned area (figure 1).Two different datasets were used: one for training and developing the classification rules, and another one for assessing the accuracy of the classification.Separate and independent data were used for training and for accuracy assessment.Despite this time lag, sufficient material remained in the field (scorched leaves on branches and the ground, char on the tree trunks, etc.) for an adequate, qualitative estimate of the degree of severity.Resprouting green leaves did not interfere with these observations.
The field survey plots were sized, with an average area of 0.78 hectare (100 m diameter).Random sampling plots location was correlated with the ETM+ location using a global positioning system.The plots were randoming located within pre-selected large areas with homogenous fire severity levels and low slope gradients by interpreting a 0.7 m-pixel post-fire colour aerial photograph in order to locate in the fire scar representative fire severity categories (scale 1:25,000, digital images orthorectified, mosaicked and examined on-screen in a GIS, captured on October 1998).
Classification of each field plot was determined by visual inspection, based on the observed majority fire severity class within each plot.Three possible fire severity categories were defined according to the degree of scorching vegetation (figure 2).We considered a high-severity, moderate-severity and low-severity as the used by other researchers (e.g., Jakubauskas et al., 1990;Turner et al., 1994;DeBano et al., 1998;Patterson and Yool, 1998;Brown and Smith, 2000;Rogan and Yool, 2001;Arno and Fiedler, 2005).The different fire severity classes were defined as follows: are formed by a linear combination of n pure spectra, such that: where DN b is the digital number in band b, DN i,b is the digital number for endmember i, in band b, F i the fraction of endmember i, and ε b is the residual error for each band.
The most common approach is to assume linear unmixing (Shimabukuro et al., 1991), although non-linear mixing can occur (Adams et al., 1993;Roberts et al., 1993).Smith et al. (2005) tested the most appropriate mixing model to use (linear or non-linear) in fire severity estimation.Whether the optical mixing was lin-ear or non-linear was largely controlled by the size of the particles present in the ash.
Endmember selection is the most important step in SMA.It determines how accurately the mixture model can represent the spectra.The endmember selection must accommodate the dimensionality of the mixing space.It involves determination of the number of endmembers and the methods to select these endmembers.Possible endmembers, however, are restricted to the number of bands the image data plus one (Hill, 1993;Small, 2004).The Landsat ETM+ sensor has sufficiently low noise that the inherent dimensionality of spectrally diverse images is generally equal to the full six dimensions.We limited this analysis to bands 3, 4, 5, and 7 as White et al. (1996) did when they tried to map  members spectra, and their spectral response was visually verified using local knowledge (Goodwin et al., 2005).
Usually shade could be included either implicitly (fractions sum to 1 or less) or explicitly as an endmember (fractions sum to 1).In our case it was included implicitly (the following equation, ΣFi = 1.0, was not included into the equation system of the unmixing model; unconstrained solution).
The least-squares solution is the method most often used for solving the linear mixture model (Smith et al., 1990;Shimabukuro and Smith, 1991;García-Haro et al., 1996) due to its simplicity and ease of implementation.As the results from the unconstrained solution do not reflect the true abundance fractions of endmembers then the root-mean-square error (RMSE) was used to assess the fit of the model (Adams et al., 1993;Roberts et al., 1998a) and it is shown in equation ( 2), where m is the number of bands.
The ISODATA classifier was used to classify fraction image into fire severity categories: high, moderate and low (Sunar and Özkan, 2001;Miller and Yool, 2002;Roldán-Zamarrón et al., 2006).Fraction image was masked by fire perimeter polygon before performing unsupervised classification.

(a) Image segmentation
Segmentation is a prerequisite to object-based classification which is the subdivision of an image into separated regions or objects by gathering together many pixels in fire severity.The definition of appropriate spectral endmembers may be either done using reference endmember from spectral libraries or from the image itself (image endmember).As appropriate reference endmembers were not available for the study site, an approach to extract pure pixels from the image was applied to retrieve image endmembers.For most SMA applications, image endmembers are utilized because they can be easily obtained and can represent spectra measured at the same scale as the image data (Roberts et al., 1998a).A minimum noise fraction (MNF) technique (essentially two cascaded principal components transformations) was used to determine the inherent dimensionality of image data, to segregate noise in the data, and to reduce the computational requirements for subsequent processing (Boardman and Kruse, 1994).The data space could be divided into two parts: one part associated with large eigenvalues and coherent eigenimages, and a complementary part with near-unity eigenvalues and noisedominated images.By using only the coherent portions, the noise was separated from the data, thus improving spectral processing results (ENVI, 2000).It was possible to run an inverse MNF transform using a spectral subset to include only the good bands, or smoothing the noisy bands before the inverse.Separating purer from more mixed pixels reduced the number of pixels to be analyzed for endmember determination and made separation and identification of endmembers easier.
Four new MNF transformed bands were then analysed to find the most spectrally pure (extreme) pixels in the image using a pixel purity index (PPI) classifier.The PPI image was the result of several thousand iterations of the PPI algorithm.The higher values indicated pixels that were nearer to the corners of the n-dimensional data cloud, and were thus relatively purer than pixels with lower values.After the purer pixels were identified in the n-dimensional scatter plot, an inverse-MNF transform was applied to obtain the end-  ( ) certain way.In comparison to pixels, image objects carry much more useful information and, therefore, can be characterized by far more properties (such as form, texture, neighbourhood or context) than pure spectral or spectral-derivative information (Baatz and Schäpe, 1999).
The segmentation used in this study was a bottom up region-merging process, starting with one-pixel objects.Throughout the segmentation procedure, the whole image was segmented and image objects were generated based upon several criteria of homogeneity in colour and shape (compactness and smoothness).In a subsequent step smaller image objects (Level 1 or fine scale to define fire severity categories in the fire scar) were merged into bigger ones (Level 2 or coarse scale to define the object boundaries of fire scar).The scale parameter was set to 5 and 20 at level 1 and 2, respectively.The composition of homogeneity criterion was set as follows: colour 0.8 and shape 0.2.For the shape criterion, smoothness was 0.1 and compactness was 0.9.
This process is called multiresolution segmentation, which was used to construct a hierarchical network of image object that simultaneously represented image information in different spatial resolutions (level 1 and 2).

(b) Object-based classification
The classification of the image objects was performed by using membership functions based on fuzzy theory combined with user-defined rules.A membership function ranges from 0 to 1 for each object's feature values with regard to the object's assigned class (Navulur, 2007).Spectral, shape, and statistical characteristics as well as relationships between linked levels of the image objects can be used in the rule base to combine objects into meaningful classes (Benz et al., 2004).The fuzzy sets were defined by membership functions that identify those values of a feature that are regarded as typical, less typical, or not typical of a class.
In our combined approach, an object-based image classification was performed using fraction images obtained in the second approach of this study as input of the model.
Burned vegetation fraction image performed better result in order to fit the image objects on both first and second segmentation levels.The scale parameter was set to 1 and 5 at level 1 and 2, respectively.The composition of homogeneity criterion was set as follows: colour 0.9 and shape 0.1.For the shape criterion, smoothness was 0.2 and compactness was 0.8.As it was done at the third approach, a fuzzy set was defined by membership functions that identified those values of a feature that were regarded as typical, less typical, or not typical of a class.
Classification accuracy was evaluated using ground referenced data.To ensure independence, no training data were used for the validation.Ground referenced data in this context means having been derived from a presumably more accurate data source than the thematic map, in this case from ground visits.The same set of ground data was used in the assessment of the accuracy of the thematic maps obtained by different classifiers in order to compare their suitability in fire severity mapping.
The accuracy assessment was based on confusion matrices, Overall Accuracy (OA), Producer's Accuracy (PA), User's Accuracy (UA) and Kappa Index of Agreement (KIA) statistic (Congalton, 1991).Error matrices were formed with data from thematic map and ground data (Congalton and Green, 1999).McNemar's test was selected to determine significant differences among classifications.Foody (2004) stated that for dependent samples, the statistical significance of the difference between two proportions might be evaluated using McNemar's test.It is a non-parametric test that is based upon confusion matrixes that are 2 by 2 in dimension.The attention is focused on the binary distinction between correct and incorrect class allocations.The McNemar test is based on the standardized normal test statistic (3) in which f ij indicates the frequency of ground data lying in confusion matrix element i, j. f 12 and f 21 are the number of pixels that with one method were correctly classified, while with the other one were incorrectly classified.
Spectral unmixing was performed using the four endmembers derived from the image data leading to 4-fraction images and a root-mean-square error (RMSE) image.The RMSE was calculated for all image pixels.Therefore, the error image was used to assess whether the endmembers were properly selected and whether the number of selected endmembers was sufficient.The value of the RMSE must be lower than the level of noise in the system, in order to guarantee the viability of the results.Landsat ETM+ signal-to-noise value is approximately 2 DN.The unmixing model results were evaluated as proposed by Adams et al., 1995.First, we evaluated the RMSE image.Our final model showed low RMSE (<2 DN values).Typically, a reasonable mixing model results in an overall RMS-threshold-error of 2.5 DN values for an image (Roberts et al., 1998a).Next, fraction images were evaluated and interpreted in terms of field context and spatial distribution.In this study, final fractions were allowed to be negative or superpositive (Román-Cuesta et al., 2005).
Fraction images derived from different combinations of image endmembers were evaluated with visual interpretation and the error extent and distribution in the error fraction image.The criteria used to identify the best suitable fraction images were based on: (1) high-quality fraction images in the fire scar, and (2) relatively low errors in the fire scar.The best results of the spectral mixture analysis for the Tabuyo fire scar are shown in figure 6.
Bright values in these images indicated areas of high fractional abundance for the endmember in question.Bright values on the RMS error image indicated areas that were poorly modelled by the least squares algorithm (values were greater than the 2.5 threshold).A cross-check with the Spanish vegetation map (1:50,000) revealed that these areas with a high RMS error were represented by crops at the time of image acquisition.

General results by approach
The different image processing methods employed and the classification techniques applied with either ISODATA or OBIA yielded varied results (figure 4).
When the unsupervised classification (ISODATA) was applied directly over the satellite image (first approach), the best results for mapping fire severity were reached using bands 3-5 and band 7, 5 iterations and forcing to 3 clusters.This approach could be considered as the best cost-effective method since no image processing technique was applied to the digital number data but it carried out the worst results among the tested approaches.
Regarding second approach, it was obtained a finally four endmembers dataset formed by: soil, two kinds of vegetation (veg1 and veg2) and burned vegetation endmember.The final vegetation 1 (veg 1) endmember was extracted from canopy of pine stands (Pinus pinaster Ait.), the vegetation 2 (veg 2) was mainly derived from canopy of Quercus pyrenaica Willd., while the soil endmember was located on agricultural areas and the burned vegetation endmember was extracted from the fire scar (figure 5).The rest of the image presented a random error distribution.Given the fact that these areas fell outside of the fire scar, we assumed that the endmembers chosen had produced robust and representative image fractions of burned areas, soil and two kinds of vegetation.In this method, fire severity map was carried out by means of applying the ISODATA classifier to burned vegetation fraction image (figure 4).
For the third approach, an object-based analysis was carried out to the ETM+ image.Visual interpretation of different image segmentation results showed that it was extremely beneficial to use band 4 (NIR) and band 7 (SWIR) since they are related with wildfire reflectance values.Official fire perimeter polygon was used as thematic layer in the segmentation in order to a better delineation of fire scar boundaries (figure 7).
Two different levels of image objects representing different scales were created: a fine scale to capture fire severity categories and a coarser scale to define the burned area.Classification at level 2 included the following classes: not burned and possibly burned.This level provided a context to detect the burned area in the image and it was used as super-object information for level 1. Features based on object spectral information (image DN) as well as object contextual information, such as relation to super-objects were used in the classification.The features based on object spectral information were: brightness and ratio B5/ratio B7 to level 2 and the Normalized Burn Ratio (NBR=B4-B7/B4+B7).Existence of super-objects was used as contextual feature.The object NBR was calculated from the NBR values of all n pixels forming an image object.Membership functions were adapted for each chosen classification feature.Aerial photos and field notes were used to help interpret the satellite image and select burn thresholds.
For the fourth approach (combined-based approach using SMA and OBIA) two different levels of image objects representing different scales were created also with the aim of capturing fire severity categories.Same class hierarchy as developed in third approach was also adopted (figure 7).Fraction images were used to extract features that were not well distinguishable in the multispectral image.Fuzzy membership functions, which are the knowledge-based part of the classification methodology in eCognition® software, were used to apply fuzzy range to the selected features (which separates a class from other classes, fire severity classes for instance).Because it is necessary to choose the feature to which to apply the membership function, it was explored the feature space to determine which feature(s) best separate the problem classes (fire severity classes).Features based on object information (from fraction image-abundance values) as well as object contextual information, such as neighbourhood and relation to super-objects were used in the classification (figure 8).The features mean burned vegetation, mean difference to scene burned vegetation and mean soil were useful at differentiating possibly burned and not burned from each other in level 2. At level 1, features such as existence of super-objects and mean burned vegetation were the best in separating fire severity classes (low, moderate and high) from each other.Membership functions were adapted for each chosen classification feature by interactively funding the lower and upper limits of the fuzzy ranges on segmentation level 1 for the classes Low, Moderate and High fire severity (figure 9).

Fire severity class areas
The area of fire severity categories varied among the fourth approaches analyzed in this study.Total areas of each approach depend on the combined ability of the classifier and the potential of each technique to separate

Accuracy assessments
The classification accuracy between sites visited on the ground and the final fire severity classification displayed different results (table 2).To ensure independence, no training data were used for validation.The remote sensing fire severity values at pixel-level of the plots were then compared to the field survey (100 mdiameter) fire severity classes (table 4).
Regarding the KIA (Kappa Index of Agreement) statistics, the highest values corroborated the overall accuracy (OA) results, with object-based approaches (3 and 4) showing the highest accuracies, in opposite relation with the pixel-based approaches.As happen in overall accuracy, KIA values displayed better values when fraction images were introduced.
A revision of significant differences among KIA statistics based on equation (3) was calculated and showed in table 3. It was apparent that the large differences in accuracy observed between the classifications expressed in terms of proportions of correctly allocated pixels were statistically significant at the 0.1% level of significance.This led to the conclusion that the classification accuracy derived from the four approaches was distinctively different, and the advantage of the approach 4 over the rest of approaches was significant.
An error matrix for each approach was also produced.Based on the User's, Producer's and KIA per class accuracy, individual class accuracies revealed diverse differences among methodologies (table 4).

Discussion
Confusion among classes occurred for all approaches even though classes were supposedly quite different in their spectral responses.For the approaches that used ETM+ data instead of fraction imagery as input data for distinguishing among classes, the moderate and low severity classes were confused in the visible and NIR band, but displayed well in the SWIR range.In this regard, White el al. (1996) reported Landsat TM band 7 (SWIR) data is useful for distinguishing among different burn severity classes.Regarding approaches using fraction images, they displayed high severity areas characterized by a large amount of burned vegetation and low amount of vegetation and soil, the opposite trend for the low severity areas and an intermediate trend for the moderate severity areas (figure 10).The incorporation of the fraction image into the classification procedure increased the accuracy for both subpixel-and objectbased approaches.By employing fraction image the overall accuracy of fire severity classification was improved by 18.30% in subpixel-based and by 12.65% in combined-based approaches (Table 2).Several authors have reported that using fraction images in a classification produces a higher accuracy than results produced by classifying the single sensor bands (Smith et al., 1990;Settle and Drake, 1993;Caetano et al., 1994;Ustin et al., 1996;Huguenin et al., 1997;Cochrane and Souza, 1998;Settle and Campbell, 1998;Aguiar et al., 1999;Elmore et al., 2000;Riaño et al., 2002;Theseira et al., 2002).
The methods that used ISODATA as classifier performed more moderate and low categories, whereas the OBIA methods (object-based classification), which classified more high and moderate classes.Approach 1 (pixel-based) displayed the highest values for the low severity class, being almost two times larger than the object-based approaches, and lowest values for the high severity class.This was because the ISODATA clustering algorithm only had a one-dimensional space to separate pixels into classes and the means of the classes  therefore tended to be uniformly distributed along the one-dimensional space (Miller and Yool, 2002).The better results obtained by approaches that included object-based image analysis into its process are mainly due to the ability of a context-based classification to reduce the speckle in the classification.Obviously, the object-based classification, which first extracts homogeneous regions and then classifies them, avoids the annoying salt-and-pepper effect of the spatially-fine classification results, which is typical of pixel-based analysis.Besides, combined-based approach (SMA/OBIA) performed better results at individual class level.It dealt satisfactory with the problems of classes confusion (burned vegetation and non-vegeted, slighty burned and unburned vegetation) using the information contained into the fraction imagery to the object-based classification and, it had an advantage over the rest of approaches tested by supplying the opportunity to combine contextual and subpixel information (contribution of each surface material in each mixed pixel) into classification which enhanced the accuracy.KIA per class, PA and UA reached high accuracy values, indicating that confusion between problematic classes such as moderate and high were minimized.This implies that introducing fraction image and object-based classification may helpful for improving separability between classes (Table 4).
The proposed method shows potential for further applications such as land cover changes, mining activities, etc.Despite future studies will examine alternative approaches to analyzing images of different resolutions.

Conclusions
Fire severity mapping is an important step by providing operational information for post-fire restoration.This paper investigated the utilization of satellite image and image processing techniques to derive fire severity information.Different approaches were tested in order to obtain accurate fire severity maps.Results showed that fraction images generated by unmixing of a Landsat ETM+ post-fire image can be used as input in an object-based classification improving the result accuracy.Results complement the findings of a small number of previous studies that support the use of SMA in mapping fire severity due to its ability to produce fractions representative of subpixel components directly related to fire severity.The accuracy of fire severity categories was better combining SMA and OBIA than for the rest of approach tested.McNemar's test was used to evaluate the statistical significance of the difference between the four methods.The difference in accuracy expressed in terms of proportions of correctly allocated pixels was statistically significant at the 0.1% level, which meant that the thematic mapping result using the combinedapproach (SMA/OBIA) achieved a much higher accuracy than the rest of approaches.

Figure 1 .
Figure 1.Location of study area in Spain; and location of field samples in the fire scar.

Figure 4 .
Figure 4. Final classified images obtained by means of ISO-DATA (approaches 1 and 2) and Object-based classification (approaches 3 and 4).Colours corresponding to each class are indicated in the legend.

Figure 5 .
Figure 5. Image endmembers used in the spectral mixing model, expressed in image radiance (DN) for the 3, 4, 5 and 7 reflective ETM+ bands.

Figure 7 .
Figure 7.A section of the study area showing the segmentation results on level 1.(a) Approach 3rd.(b) Approach 4th.(c) Class hierarchy created for both approach 3rd and 4th.

Figure 8 .
Figure8.Membership functions of level 2 and 1 for fourth approach.

Figure 9 .
Figure 9. Membership functions of level 1 for fourth approach.

10.
Mean values for each fire severity class (high, moderate and low), for approaches using ETM+ data (left) and for those using fraction imagery (right).Left graphic bars represent the utilized bands: B3, B4, B5 and B7.Right graphic bars represent the different endmembers: vegetation 1, vegetation 2, burned vegetation, soil and error term (RMS).Each sub-section in the right graphic shows the three selected classes: high, moderate and low severity.

Table 1 .
Severity area estimated for each approach

Table 2 .
Overall accuracies and KIA statistics for each considered approach