Burned area mapping on the Mediterranean island of Thasos using low, medium-high and very high spatial resolution satellite data Khaldoun Nasri Rishmawi and Ioannis Zois Gitas Mediterranean Agronomic Institute of Chania (MAICh), Alsylio Agrokepiou, P.O. Box 85, GR 73100, Chania, Greece Ioannis@maich.gr Abstract The aim of this work was to investigate the potential of using low (AVHRR), medium-high (LANDSAT TM) and very high (IKONOS) spatial resolution data to accurately map the burned areas on the Mediterranean island of Thasos. In order to exploit best each sensor s capabilities, sensor specific image processing methods were employed. Results from the application of the methods indicated that each sensor type had both advantages and disadvantages. Although the AVHRR sensor was unable to detect such details as patches of non-burned vegetation within the fire perimeter, it was the only sensor of the three that provided daily coverage. LANDSAT TM and IKONOS, however, were shown to be effective in providing more accurate and detailed information about the burned areas studied. A further advantage of the IKONOS sensor was its ability to differentiate between surface and canopy burns. 1. Introduction It has been well established in the literature that natural fires are an integral part of many terrestrial ecosystems such as boreal forests, temperate forests, Mediterranean ecosystems, savannas and grasslands among others. However, from the 1960s until today, the general trend in the number of fires and surface burnings in the European Mediterranean areas has increased exponentially (Moreno et. al. 1998). This increase is mainly due to: (a) changes in traditional land uses, the consequence of which is higher fuel accumulation, and (b) global climatic warming (Gitas 1999). When a forested area is damaged by fire, detailed and current information concerning the location and extent of the burned areas and the level of fire damage is important to assess economic losses and ecological effects, to monitor land use and land cover changes, and to model the atmospheric and climatic impacts of biomass burning (Caetano et. al. 1994, Pereira et. al. 1997). Moreover, accurate assessments aid in evaluating the effectiveness of measures taken to rehabilitate the fire damaged area, and allow forest managers to identify and target areas for intensive or special restoration (Jakubauskas 1988, Jakubauskas et. al. 1990) thus avoiding longterm site degradation. To date, most of the National Forest Services in Mediterranean Europe do not provide cartographic representation of the burned areas (Chuvieco 1997). In some cases, only a general estimation of the burned area is provided and the fire perimeter is not available. This lack of information results in a poor understanding of the spatial consequences of fires. The methods conventionally employed to produce maps of burned areas and to estimate the damage caused by fire (i.e. degree of vegetation damage and timber loss) in Greece are based on extensive field visits to the burned area and visual observations of the fire s effects. A rough map of the area is drawn, an autopsy statement and a fire report are prepared and the collected data is then plotted on topographic maps (Gitas 1999). The main problems related to these conventional methods are: RSPS2001 Proceedings Poster Session 408
information provided is often qualitative, concerning only the fire perimeter; there is a lack of information about species affected and the severity of damage; information is not available even months after the fire (Martin et. al. 1994) with the result that vegetation recovery cannot be assessed. Furthermore, the lack of vegetation recovery may constitute a severe soil erosion hazard (Isaacson et. al. 1982). Small burns can be mapped and assessed by conventional methods, but when a large fire occurs, such methods are not applicable (Gitas 1994). According to Chuvieco (1997), satellite remote sensing has supplied a suitable alternative to conventional techniques for monitoring burned areas. To this end, several studies have investigated the suitability of low and mediumhigh spatial resolution satellite imagery to produce accurate burned area estimates. Moreover, studies comparing the use of low and medium-high resolution data have also been carried out (Martin et. al. 1994, Fraser et. al. 2001, Sa et. al. 2001). To date, many comparative studies have showed that low spatial resolution satellite imagery, such as that provided by AVHRR or SPOT VEGETATION, produce an overestimation, or a positive bias, of the burned areas (Eva and Lambin 1998, Fraser et. al. 2000, Fraser et. al. 2001), whereas medium-high resolution satellite data such as that provided by LANDSAT TM, produce more accurate fire assessments. These studies revealed that the positive bias associated with utilizing coarse resolution satellite data might be due to spatial aggregation effects where unburned patches smaller than the sensor resolution were mapped as burned (Eva and Lambin 1998). Although LANDSAT TM produces more accurate burned area estimates at detailed cartographic scales, several problems relating to its utility in burned vegetation mapping and in burn severity assessment remain. These problems can be categorized as follows: confusion in the discrimination between burned areas and shaded unburned areas (Tanaka et al. 1983, Milne 1986, Chuvieco and Congalton 1988, Parnot 1988, Pereira 1992, Caetano et al. 1994, Lombrana 1995, Pereira et. al. 1997, Gitas 1999); confusion between slightly burned and unburned sparse vegetative classes (Chuvieco and Congalton 1998, Gitas 1999); difficulties in separating burned vegetation from other non-vegetation categories and especially water bodies (Tanaka et. al. 1983, Ponzoni et al. 1986, Chuvieco and Congalton 1988, Parnot 1988, Pereira and Setzer 1993, Lombrana 1995, Siljestrom and Moreno 1995, Silva 1996), urban areas (Tanaka et al. 1983, Chuvieco and Congalton 1988, Lombrana 1995, Caetano et. al. 1996, Silva 1996) and bare soil (Parnot 1988, Pereira and Setzer 1993, Siljestrom and Moreno 1995). Although a considerable body of research has focused on the mapping of burned areas using AVHRR and LANDSAT data, it seems that the potential of the existing satellite data has not been fully exploited. As reported by Justice et al. (1993), the development of methodologies capable of producing more accurate burned area estimates from remotely sensed data remains an active topic of research at geographic scales ranging from local to global. On 24 th September 1999, the IKONOS satellite, the world s first commercial high resolution Earth imaging satellite, was launched from Vandenberg Air Force Base, California, USA. The spatial resolution of an IKONOS image is 82 cm in nadir viewing, resulting in images that are exactly the same as those of an aerial photo. According to Tanaka and Sugimura (2001), it seems that IKONOS images are opening up a new field of remote sensing. One major application in which IKONOS images are expected to bring new insight is the mapping of burned areas. RSPS2001 Proceedings Poster Session 409
The aim of this work was to investigate the potential of using low (AVHRR), medium-high (LANDSAT TM) and very high (IKONOS) spatial resolution data to accurately map burned areas on the Mediterranean island of Thasos. The specific objectives were: to investigate ways of improving existing methods employed in mapping the burned areas using AVHRR and LANDSAT TM satellite images; to investigate the ability of IKONOS satellite images in burned area mapping; and to clarify the advantages and disadvantages of each of the three sensors in burned area mapping. 2. Study Area And Dataset The study area is located on the island of Thasos, in the prefecture of Kavala, Macedonia, Greece. The area extends from 24 o 30 to 24 o 48 East and 40 o 33 to 40 o 49 North. Its total area is 39,000 ha, while its perimeter is approximately 104 km. The elevation ranges from sea level to 1217 m and terrain slopes range from 0 to 76 degrees. The climate of Thasos is Mediterranean, characterised by hot, dry and sunny summers and a cool winter, where precipitation occurs mainly in short, intense showers during the spring and autumn. Pinus brutia dominates the forests of Thasos, with Pinus nigra being the co-dominant forest species. Three large fires between 1984 and 1989 resulted in the loss of about 20,000 ha of Pinus brutia and Pinus nigra forest, on area representing more than half the size of Thasos. In July 2000, a medium-sized fire burned 185 ha of Pinus brutia forest; the fire was a mixed (Surface and Crown) fire. Today, the Pinus brutia forest has a spatial extent of about 2000 ha, mainly in the northern and eastern parts of the island. This study area is ideal for addressing the proposed objectives because: (1) the complex relief characteristics guarantee the existence of shaded unburned areas usually confused with burned areas, (2) the availability of cartographic presentations of the burned parameter which, in turn, allows a quantitative validation of remotely sensed data products, and (3) the land cover types in the study area are representative of most of the landcover types affected by fires in Greece. The data set used in this study consists of: (1) two NOAA/AVHRR Local Area Coverage products: one pre-fire (11 August, 1989) and one post-fire (20 August, 1989); (2) one post-fire LANDSAT TM image captured on the 19 th September 1989; (3) one post-fire IKONOS very high resolution image captured on the 17 th July 2000; (4) fire perimeter data published by the Greek Forest Service; (5) a digital elevation model (DEM) with 10 m pixels generated from a 1:5000 contour maps; (6) ground truth points of scorched areas burned in the fire of 2000 collected using a global positioning system and (7) planimetric maps. 3. Data Preprocessing Prior to the mapping of the burned areas, the satellite images were pre-processed. Pre-processing of the satellite images included their atmospheric and geometric correction. Topographic correction was further employed on the LANDSAT TM image. 3.1. Atmospheric Correction The objective of atmospheric correction is to reduce pixel Brightness Value (BV) variation caused by atmospheric attenuation so that variation in pixel BVs between images can be related RSPS2001 Proceedings Poster Session 410
to actual changes in surface conditions. In the case of burned area mapping atmospheric correction is necessary whenever more than one image is used to map the burned area and to bolster the potential of extending the classification algorithms produced in this study into an operational system of mapping Pine forest fires in the Mediterranean. An absolute atmospheric correction method developed by Richter (1997) for sensors with a small swath angle was employed on the LANDSAT and the IKONOS images. The method, which employs the MORTRAN radiation propagation model, first measures the optical depth (ground visibility) and then proceeds in the calculation of the ground reflectance for each pixel. In order to render the AVHRR images radiometrically comparable, relative radiometric normalization was employed on the broad swath angle sensor, where the post-fire image was regressed to the pre-fire image, resulting in a reduction in the variation of pixel BVs caused by atmospheric attenuation. 3.2 Geometric Correction Geometric correction is essential in order to render the images and the auxiliary data sets geographically comparable. As the DEM and the fire perimeter were referenced to the HAGS 1970 coordinate system, a 1:50,000 topographic map (HAGS 1970) was used to geometrically correct the satellite images. Since the transformation accuracy depends on the even distribution of the Ground Control Points over the image, an effort was made to space the points as widely as possible across the study areas. Characteristic features on the images, such as the coastline and road intersections, the latter used only with LANDSAT and IKONOS images, provided excellent locations for the selection of data file coordinates. Using a first order polynomial transformation, an overall root mean square error (RMS error ) of 0.23 was attained for the IKONOS image and an RMS error of less than 0.01 for AVHRR images. The LANDSAT image was orthorectified using a LANDSAT polynomial model, which accounts for relief displacement. The modification in the LANDSAT polynomial model is that it also takes into account terrain elevation, local earth curvature, distance from nadir, and flying height above datum to get a polynomial transformation between the image and ground coordinates. It was found that the registration accuracy of the orthorectified LANDSAT image is higher than that obtained using a simple polynomial transformation. As topographic normalization will be further applied to the LANDSAT image, miss-registration between the LANDSAT image and the DEM can lead to unfavourable results. 3.3 Topographic Correction Several studies concerned with burned area mapping using Landsat TM reported confusion between burned areas and topographically shadowed areas (Tanaka et. al. 1983, Milne 1986, Chuvieco and Congalton 1988, Parnot 1988, Pereira 1992, Caetano et. al. 1994, Lombrana 1995, Pereira et. al. 1997). In order to minimize the confusion, topographic correction was applied to the LANDSAT TM. Here, the goal of topographic correction is to remove topographically induced illumination variation so that two objects having the same reflectance properties show the same brightness value in the image, despite their different orientation to the suns position. The topographic correction model is derived from the algorithm LH = LT*cose / [ (cosθ) ^k. (cose) ^k ] (1) where: L H = Normalized brightness values L T = Observed brightness values RSPS2001 Proceedings Poster Session 411
cos θ = cosine of the sun incidence angle in relation to the normal on the pixel cos e = cosine of the exitance angle, or slope angle k = Minnaert constant. The Minnaert constant was empirically derived by: (i) logarithmically linearizing Equation 1, (ii) obtaining a sufficiently large sampling size of pixels located on moderate to steep, east and west facing slopes (L T Values) and on horizontal slopes (L H Values) and (iii) estimating the value of the Minnaert constant using linear regression analysis. A Minnaert constant value ranges from 0 to 1 and is a measure of the extent to which a surface is Lambertian. Table 1 shows the values of the Minnaert constant for each band, as well as their respective goodness of fit as obtained from the statistics of the regression analysis. Table 1. The Empirically derived Minnaert constant (K), Pearson correlation (R) and goodness of fit (R^2 ) values per band for the Landsat TM scene. Landsat 89 K R R^2 TM Band 1 0.651 0.811 0.606 TM Band 2 0.685 0.752 0.566 TM Band 3 0.681 0.733 0.537 TM Band 4 0.439 0.526 0.277 TM Band 5 0.544 0.722 0.521 TM Band 7 0.58 0.519 0.269 It is interesting to note that R^2 values for bands 4 and 7 were less than those observed for the other bands, an indication that infrared bands are more severely affected by the topographic effect. However, the result of topographic normalization was an overall reduction in the illumination variation induced by topography. 4. Methods In this study, the burned area in the year 1989 was mapped using AVHRR and LANDSAT TM data. Multitemporal thresholding of the Normalized Difference Vegetation Index was used with the radiometrically normalized AVHRR data, whereas binary logistic regression was used with the topographically normalized LANDSAT TM image. For the burn scar of the 2000 fire, a post-fire IKONOS image was obtained. Here, a method that involves the classification of discriminant scores was used to map the burned area. 4.1 Burned Area Mapping Using AVHRR Images The two radiometrically corrected local area coverage AVHRR images were used to map the burned area. The AVHRR images were used to produce NDVI images corresponding to the preand post-fire dates. The two NDVI layers were differenced. Burned areas were identified as those exceeding a given NDVI decrease threshold. In order to decide upon the optimum threshold boundary between burned and unburned pixels, a variety of restrictive thresholds ranging from low to high were tested and compared with the burned area perimeter provided by the National Forest Service. 4.2 Burned Area Mapping Using LANDSAT TM Images RSPS2001 Proceedings Poster Session 412
In this study, binary logistic regression modelling was used to classify the LANDSAT image pixels into two categories, burned and non-burned areas. The output of the binary logistic equation formed a realistic probability surface with a minimum of 0 and a maximum of 1. The pixels with a probability value less than 0.5 were classified as burned, and those greater than the value 0.5 unburned. To ensure the successful development of the logistic regression models, the sampling criteria provided by Koutsias and Karteris (1998) were followed: first, the sampling areas for the burned and non-burned cases were accurately located on the satellite image; second, the sampling size regarding both cases was about the same to avoid bias in the sampling process; and third, a satisfactory absolute sampling size of 11,900 pixels was obtained to represent all the spectral variability occurring on the satellite image within and outside the canopy burned area. The non-burned pixels represented water, urban fabric, bare surfaces, vegetation, and shade. Burned areas were assigned the value Zero, whereas non-burned areas were assigned the value one to create the dependent variable in the modelling process. Two logistic regression equations were structured: the first using the original spectral channels of the TM image as the explanatory variables, and the second using the topographically normalized data. Following the construction of the two equations, the performance of each was evaluated by calculating the percentage of correct classified observations (burned, unburned and overall). The model with the better performance was applied to the entire satellite data to map the burned areas. 4.3 Burned Area Mapping Using IKONOS Images It has been noticed that spectral confusion between burned areas and the other landcover categories such as water, bare soil and shade can result from utilizing IKONOS data that lacks the spectral depth provided by LANDSAT TM in the MIR range of the spectrum. Furthermore, it has also been noticed that a new source of spectral confusion can generate from the spectral similarity between the signatures of surface burned areas and those of shadowed understory vegetation. Actually, spectral confusion results from the spectral overlap in the multidimensional space between the signatures of the different landcover types. In order to reduce the spectral confusion between the classes, a two-step approach was followed: (1) the selection of the data layers that potentially provides the best discrimination between classes; the data layers can be the original spectral channels and/or those arising from different transformations such as IHS and NDVI, and (2) the application of Canonical Discriminant Analysis, which maximizes the between landcover type variance relative to the within landcover type variance. Therefore, in order to reduce the confusion between burned areas and the other land cover categories; the following procedure was followed: first, a comparison of the spectral signatures between the burned areas (canopy and surface) and the other land use/cover categories was made using the Euclidean Distance separability index; second, all possible RGB color composites consisting of the original spectral channels of IKONOS sensor were transformed to IHS color models. The hue components of the IHS models were visually evaluated for their ability to discriminate between burned areas and the other landcover types; RSPS2001 Proceedings Poster Session 413
third, an NDVI layer was produced and evaluated for its ability to discriminate between the burned classes and the other landcover types present on the image; fourth, the spectral bands and the data layers arising from different transformations that potentially provide the best discrimination between the desired classes were selected; fifth, sample areas representative of each land cover type were located on the selected data layers and assigned a unique numerical value. The procedure resulted in a sample of cases for which the landcover type is known; sixth, a set of linear discriminant functions was generated from the sample of cases. The functions were then applied to the selected layers to produce discriminant scores; and finally, a supervised Minimum distance classification was conducted using the discriminant scores. 4.4 Accuracy Assessment The burned area perimeter produced by the Greek National Forest Service was used in the assessment. The burn perimeter delineates the boundary without any reference to burn severity. In addition to the burn perimeter, a field visit was carried out to the scorched area in the year 2000 to collect real time GPS data, where areas with different levels of fire damage were geographically located and photographed. Burned area maps were assessed for their accuracy using the fire perimeter. However, the ability of the IKONOS sensor to differentiate between different levels of burns was assessed only by using real time GPS data. The classification accuracy assessment was based on a sampling of approximately 256 pixel comparisons between each classified image and a combination of real time GPS data, and the National Forest Service fire perimeter. Given the complexity of digital classification, particular attention was given to assess the reliability of the results and an error matrix was determined. The error matrix was found to be a very effective tool to represent the accuracies of each category, together with both the errors of inclusion and the errors of exclusion always present in a classification (Congalton, 1991). 5. Results and Discussion 5.1 Burned Area Mapping Using AVHRR Images The advantage of using radiometrically normalized satellite images is that changes in pixel brightness values can now be related to actual changes in surface conditions. In this case, the changes in NDVI values can be related directly to vegetation loss. Now, the critical step in the process to produce the burned versus unburned binary mask was the choice of the threshold boundary. It was observed that using a high restrictive threshold with an NDVI decrease increment greater than 0.20 resulted in an underestimation of the burned area. In this case, the AVHRR burned area product underestimated the actual burned area by 35%. However, it is interesting to note that no false alarms were detected. The use of a less restrictive threshold with an NDVI decrease increment greater than 0.15 was found to better match the field estimates. However, the use of a less restrictive threshold produced a number of false alarms. The classification error matrix further showed the commission and emission errors in the classification (Table 2). It was observed that the users accuracy for the burned area class is considerably low; this is in fact due to: (1) commission errors arising mainly from false detections within 1 km distance off the seashore, the main RSPS2001 Proceedings Poster Session 414
reason for which is miss-registration between the two images and (2) from aggregation effects, mainly attributable to the coarse resolution of the sensor. Table 3. Accuracy assessment Reference Data Classified Reference Classified Number Producers Users Data NB B Totals Totals Correct Accuracy Accuracy Non-Burned Areas (NB) 115 10 125 125 115 85.2% 92.0% Burned Areas (B) 20 105 125 125 105 91.3% 84.0% Column Total 135 115 250 250 220 Overall Classification Accuracy = 88.0% 5.2 Burned Area Mapping Using LANDSAT TM Images In order to investigate the effect of topographic normalization on the ability of the LANDSAT TM sensor to produce more accurate burned area estimates, two logistic regression models were structured. The first utilized the original reflectance channels as the independent explanatory measurements, whereas the second utilized the topographically normalized channels. The prediction ability of the models is depicted in Table 4. Table 4. The classification results of the logistic regression models based on the comparisons of observed versus predicted observations Percentage of correct Classified Observations Statistics of Performance % Unburned % Burned % Overall Nagelkerke R Square Original Reflectance Channels 90.8 92.5 91.7 0.797 Topographically Normalized Channels 97.6 97.2 97.4 0.921 The prediction ability of the model increased from 91.7% to 97.4% when the topographically normalized channels were used as the independent explanatory variables. The mathematical formulation of the better performing logistic regression model was built into the ERDAS graphical model maker and applied to the entire topographically normalized TM image. This resulted in a new continuous data layer with a minimum of 0 and a maximum of 1. All pixels with a value greater than 0.5 were classified as burned, and those with a smaller value than 0.5 unburned. The results of the binary logistic classification and the accuracy assessment showed that burned areas could be accurately mapped with LANDSAT TM data. The comparison of the classification with the burned area perimeter using 256 points is reported in Table 5. The classification error matrix shows the commission and emission errors in the classification. Of the 128 pixels predicted by the logistic model to be burned, only three pixels were erroneously classified as burned (commission error). The emission error was almost equal to the commission RSPS2001 Proceedings Poster Session 415
error where 4 pixels were erroneously classified as unburned. The overall classification accuracy was estimated to be 97.3%. The landcover types confused with the burned areas were bare/low vegetated areas and urban fabrics. It is interesting to note that due to topographic normalization, the confusion between burned and shaded areas was eliminated. Table 5. Accuracy assessment Reference Data Reference Classified Totals Totals Classified Data NB B Number Correct Producers Accuracy Users Accuracy Non-Burned Areas (NB) 124 4 127 128 124 97.6% 96.9% Burned Areas (B) 3 125 129 128 125 96.9% 97.6% Column Total 127 129 256 256 249 Overall Classification Accuracy = 97.3% 5.3 Burned Area Mapping Using IKONOS Images The results of the Euclidean distance seperability index (Table 6) show that the NIR and Red bands are the most useful in discriminating between the burned classes and the other landcover categories present on the image. Nevertheless, the spectral distances between canopy burned areas and surface burned areas, water and shaded forest understory indicate that spectral confusion can arise between these landcover types. Providentially, the visual evaluation of the IHS models showed that the hue component of the NIR-Red-Blue color model transformation clearly discriminates between the canopy and surface burned areas. Moreover, the same hue component reduces the confusion between canopy burned area and shaded forest understory. Therefore, the NIR and Red spectral channels, the hue component and the NDVI were selected to be used in the analysis. The latter was selected for its ability to discriminate between canopy burned areas and seawater. Table 6. Euclidean-Distance separability index between canopy burned areas and the other landcover types present on the image for each band Canopy Burn Band 1 Band 2 Band 3 Band 4 Surface Burned Areas 11 23 37 4 Pine Forest 20 15 20 186 Shade Forest Understory 50 70 74 31 Water 2 25 67 81 Bare/ Low Vegetated Soils 42 84 102 112 Bare Rocks 166 312 390 492 Average 48.5 88.2 115 151 Four linear discriminant functions were structured. The functions were then applied to the entire dataset to produce discriminant scores. The discriminant scores further maximized the distances between the mean values of the burned classes and the other landcover types. It was RSPS2001 Proceedings Poster Session 416
evident that the area of internal overlap between the values of the landcover types was minimized. This was achieved by the statistical decision rule, inherit in CDA, of maximizing the between group variance relative to the within group variance. Supervised classification using the minimum distance classifier was applied to the discriminant scores. The results of the classification produced a relatively high accuracy in mapping the burned areas and in distinguishing between canopy and surface burned classes. A comparison of the classification with the burned area perimeter and the GCPs using 256 points is reported in Table 7. The overall classification accuracy was estimated to be 92.5%. The classification produced good results in discriminating between canopy burned areas and surface burned areas where no confusion between the two classes was detected. However, the confusion between surface burned areas and shaded understory vegetation was not eliminated. This can be attributed to the fact that some pixels representative of surface burned areas are shadowed. Overall, the inclusion of the hue component and the NDVI and the application of CDA to produce discriminant scores were crucial for the attainment of reliable results. Reference Data Table 7. Accuracy Assessment Reference Totals Classified Totals Number Correct Producers Accuracy Users Accuracy Classified Data CB SB S W V BR BS Canopy Burned Area (CB) 37 0 0 0 0 0 1 37 38 37 100.00% 97.37% Surface Burned Area (SB) 0 48 3 0 0 0 0 60 51 48 80.00% 94.12% Shaded Understory (S) 0 12 11 0 3 0 0 14 23 11 78.57% 47.83% Water (W) 0 0 0 73 0 0 0 73 73 73 100.00% 100.00% Vegetation (V) 0 1 0 0 72 2 0 75 75 72 96.00% 96.00% Bare Rocks (BR) 0 0 0 0 0 12 0 14 12 12 85.71% 100.00% Bare/Low Veg. Soil (BS) 0 0 0 0 0 0 6 7 6 6 85.71% 100.00% Column Total 37 61 14 73 75 14 7 280 278 259 Overall Classification Accuracy = 92.5% 6. Conclusions Accurate information concerning fire size, fire severity and the spatial distribution of burned and unburned patches within the fire perimeter is important to understand the effects of forest fires. Satellite remote sensing provides the necessary means for gathering information about the burned area in a timely/cost effective manner and at a range of spatial resolutions from coarse to high. In this work, low (AVHRR), medium-high (LANDSAT TM) and very high (IKONOS) spatial resolution sensors were used to map the burned areas on the Mediterranean island of Thasos. RSPS2001 Proceedings Poster Session 417
It was found that the burned areas could be accurately mapped using AVHRR, LANDSAT TM and IKONOS data. However, the accuracies of the burned area maps derived from the LANDSAT TM and IKONOS data were higher than the accuracy obtained using AVHRR data. The relative low accuracy of 88% obtained when using AVHRR data was due to: (1) the inability of the sensor to detect patches of unburned vegetation within the fire perimeter, and (2) the confusion between the burned areas and the seashore. In relation to the specific objectives of this work, the major findings can be summarised as follows: Although binary logistic regression can be used with LANDSAT TM data to accurately map the burned areas, the accuracy of the method can still be improved by topographically correcting the image. Topographic correction, followed by its classification using binary logistic regression produced very accurate burned area maps. When AVHRR data were used, the application of relative radiometric correction resulted in a reduction in the variation of pixel brightness values caused by external factors. This, indeed, proved useful in burned area mapping where changes in reflectance values between the pre- and post-fire images were related to the actual changes caused by fire. IKONOS data were shown to be effective in providing more detailed information about the burned area. Small unburned patches, otherwise undetected when using LANDSAT TM, were accurately mapped. It was also possible to discriminate with high accuracy between unburned areas, surface burned areas, and canopy burned areas. Two main problems arose with the use of IKONOS images in burned area mapping, namely, the introduction of noise due to shadows of individual trees; and the spectral confusion that results from the inability of the sensor to record electromagnetic radiation in the middle infra red part of the spectrum. These problems were overcome by employing Canonical Discriminant Analysis on data layers that best separated the signatures of the dominant landcover types. This method led to a significant improvement in the accuracy of the burned area map, but it is time-consuming and requires greater efforts. Each sensor seems to have both advantages and disadvantages. Although detailed burned area maps can be produced using high spatial resolution data, the time and cost of processing can restrict their use in large-scale operations. Another closely related issue is the spatial coverage of each sensor. The spatial coverage of AVHRR data is global, while that of LANDSAT TM is regional and that of IKONOS is local. Therefore, very high-resolution data that cover a large spatial extent can be very difficult to collect and to process. 7. References Caetano, O., Mertes, L., and Pereira, J., 1994, Using spectral mixture analysis for fire severity mapping, in: 2nd Int Conf Forest Fire Research, coimbra, pp. 667-677. Chuvieco, E., 1997, Foreword, in: A review of remote sensing methods for the study of large wildland fires (E. Chuvieco, ed.), Departamento de Geografía, Universidad de Alcalá, Alcalá de Henares, pp. 3-5. Chuvieco Salinero, E., 1989, Multitemporal analysis of Thematic Mapper images. Applications to forest fire mapping and inventory in a Mediterranean environment, in: Proceedings of a workshop "Earthnet pilot Project on Landsat Thematic Mapper Applications", Frascati, pp. 279-285. Chuvieco Salinero, E., and Congalton, R., 1998, Mapping and inventory of forest fire from digital processing of TM data, Geocarto Internationl,4, 41-53. Chuvieco Salinero, E., 1999, Remote sensing of large wildfires in the European Mediterranean RSPS2001 Proceedings Poster Session 418
Basin, Springer, New York. Congalton, R., 1991, A review of assessing the accuracy of classifications of remotely sensed data, Remote sensing of Environment,52, 229-241. Eva, H., and Lambin, E., 1998, Remote sensing of biomass burning in tropical regions: sampling issues and multisensor approach, Remote sensing of Environment,16, 975-982. Fraser, R., Li, Z., and Cihlar, J., 2000, Hotspot and NDVI differencing Synergy (HANDS): A new technique for burned area mapping over boreal forests, Remote sensing of Environment,74, 362-376. Fraser, R., Hall, R., and Landry, R., 2001, Burnt area mapping across Canada's boreal forest zone using SPOT VEGETATION calibrated with Landsat TM imagery, in: 3rd International Workshop on Remote Sensing and GIS applications to Forest Fire Management, New Methods and Sensors (E. Chuvieco, and M. Martin, eds.), Paris, pp. 133-137. Gitas, I.Z., 1994, Geographical Information Systems and Remote Sensing in forest management, Department of Geography, University of Cambridge, M.Phil. Thesis, p.96. Gitas, I.Z., 1999, Geographical Information Systems and Remote Sensing in mapping and monitoring fire-altered forest landscapes, Department of Geography, University of Cambridge, Ph.D. Dissertation, p.23. Isaacson, D., Smith, H., and Alexander, C., 1982, Erosion hazard reduction in a wildfire damaged area, in: Remote Sensing for Resource Management (C. Johannsen, and J. Sanders, eds.), Soil Conservation Society of America, Ankeny, pp. 179-190. Jakubauskas, M., 1988, Postfire vegetation change detection using LANDSAT MSS and TM data, in: School of Graduate Studies, Indiana State University, Terre Haute, Indiana, pp. 99. Jakubauskas, M., Lulla, K., and Mausel, P., 1990, Assessment of vegetation change in a firealtered forest landscape, Photogrametric Engineering and Remote Sensing,56, 371-377. Justice, C., Malingreau, J.-P., and Setzer, A., 1993, Satellite remote sensing of fires: potential and limitation, in: Fire in the environment: The ecological atmospheric and climatic importance of vegetation fires (P. Crutzen, and G. JG, eds.), John Wiley and Sons, New York, pp. 77-88. Koutsias, N., Karteris, M., and Chuvieco, E., 1988, The use of intensity-hue-saturation transformation of Landsat-5 Thematic Mapper data for burned land mapping, Photogrametric Engineering and Remote Sensing,19, 3499-3514. Koutsias, N., and Karteris, M., 1996, Logistic regression modelling of Thematic Mapper data for burnt area mapping, in: Unpublished Manuscript, Aristotelian University of Thessaloniki, Dept. of Forestry and Natural Environment, Lab. of Forest Mangement and remote Sensing. Koutsias, N., and Karteris, M., 1998, Logistic regression modelling of multtemporal Thematic Mapper data for burned area mapping, International Journal of Remote Sensing, 19, 3499-3514. Koutsias, N., and Karteris, M., 2000, Burned area mapping using logistic regression modelling of a single post-fire Landsat 5 Thematic Mapper image, International Journal of Remote Sensing, 21, 673-687. Lombrana, M., 1995, Monitoring of burnt forest areas with remote sensing data. A study in North-East Spain using Landsat TM and SPOT XS data. Technical note I.95.80, in: Institute for Remote Sensing Applications, Joint research Center, Ispra. Lopez, G., and Cazsellas, V., 1991, Mapping burns and and natural reforestration using thematic Mapper data, Geocarto Internationl,6, 31-37. Martin, M., Viedma, D., and Chuvieco Salinero, E., 1994, High versus low resolution satellite RSPS2001 Proceedings Poster Session 419
images to estimate burned areas in large forest fires, in: 2nd Int Conf Forest Fire Research, coimbra, pp. 653-663. Milne, A., 1986, The use of remote sensing in mapping and monitoring vegetational change assossiated with bushfire events in eastern Australia, Geocarto Internationl,1, 25-35. Moreno, J. M., and Oechel, W. C., 1994, The Role of fire in Mediterranean-type ecosystems, in: Ecological studies ; v. 107, Springer-Verlag, New York, pp. xiii, 201. Moreno, J. M., and Oechel, W. C., 1995, Global change and Mediterranean-type ecosystems, in: Ecological studies ; v. 117, Springer-Verlag, New York, pp. vii, 527. Moreno, J. M., 1998, Large Fires, Backhuys Publishers, Leiden. Navaro, R., Salas, F., Navaro, C., Fernandez, P., and Gonzalez, M., 1997, Evaluacion de danos producidos por incendio y regeneracion posterior de la vegetacion. Aplicacion de imagenes Landsat TM a su caracterizacion y seguimiento, Internal Report, Consejeria de Medio Ambiente, Junta de Andalucia, Sevilla. Parnot, J., 1988, Inventaire des feux de brouse au Burkina Faso saison seche 1986-1987, in: 22nd International Symposium on Remote Sensing of the Environment, ERIM, Ann Arbor, Michigan, Abidjan, Cote d' Ivoire. Pereira, J. M., 1992, Burned area mapping with conventional and selective principal component analysis, Finisterra,27, 61-76. Pereira, J., and Setzer, A., 1993a, Spectral Characteristics of deforestration fires in NOAA/AVHRR images, International Journal of Remote Sensing,14, 583-597. Pereira, J., Chuvieco, E., Beaudion, A., and Desbois, N., 1997, Remote sensing of burned areas, in: A review of remote sensing methods for the study of large wildland fires (E. Chuvieco Salinero, ed.), Universidad de Alcala, Alcala de Henares, pp. 127-183. Ponzoni, F., Lee, D., and Filho, P., 1986, Assessment of area burnt and vegetation recovery at Brazilia National Park, using Landsat TM data, in: IV Symposio Brazileiro de Sensoriamento Remoto, INPE, Sao Jose Dos Campos, pp. 615-621. Richter, R., 1997, ATCOR2 for ERDAS IMAGINE- Atmospheric correction and haze removal, in: Users Manual, Version 1.6, Geosystems GmbH, Germany. Sa, A., Silva, J., Pereira, J., and Vasconcelos, M., 2001, Burned area detection in the Miombo of Northern Mozabique using MODIS and Landsat Data, in: 3rd International Workshop on Remote Sensing and GIS applications to Forest Fire Management, New Methods and Sensors (E. Chuvieco, and M. Martin, eds.), Paris, pp. 156-159. Siljestrom, p., and Moreno, A., 1995, Monitoring Burnt area by principal component analysis of multitemporal TM data, International Journal of Remote Sensing,16, 1577-1587. Silva, J., 1996, Comparing the vegetation indices NDVI and VI7 for burnt area mapping with Landsat 5 TM imagery, Instituto Superior de Agronomia, Universidad Tecnica de Lisboa, Lisbon. Tanaka, S., Kimura, H., and Suga, Y., 1983, Preperation of 1:25000 Landsat map for assessment of burnt areas in Etajima Island, International Journal of Remote Sensing,4, 17-31. Tanaka, S., and Sugimura, T., 2001, A new frontier of remote sensing from IKONOS images, International Journal of Remote Sensing,22, 1-5. Viedma, O., Melia, J., Segarra, D., and Carcia-Haro, J., 1997, Modeling the rate of ecosystem recovery after fires by using Landsat TM data, Remote sensing of Environment,61, 383-398. RSPS2001 Proceedings Poster Session 420