Postflood Damage Evaluation Using Landsat TM and ETM+ Data ...
236
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 44, NO. 1, JANUARY 2006
Postflood Damage Evaluation Using Landsat TM and ETM+ Data Integrated With DEM Marco Gianinetto, Paolo Villa, and Giovanmaria Lechi
Abstract—In recent decades, radar and optical satellite imagery have been used for evaluating flooding extent. In this paper, a straightforward technique based on the sequential use of the spectral-temporal principal component analysis, logical filtering, and image segmentation integrated with the digital elevation model was developed as a decisional support tool for the allocations of the resource destined for the flooded areas. The mapping technique was first applied to the catastrophic event that occurred in the Piemonte Region (Italy) in November 1994, which was the worst event of the past century for that region, with 44 casualities and over 2000 homeless. Next, it was applied to the Obion/Forked Deer inundation that occurred in Tennessee (U.S.) between November and December 2001, in which heavy damage to the infrastructure was reported. Two Landsat-5 Thematic Mapper (path 194, row 28/29) and two Landsat-7 Enhanced Thematic Mapper Plus (path 23, row 35) images were processed, two of them collected before and two after the events. The method proposed proved to be an effective approach for evaluating flood extent and for assessing the damage produced by the flooding. An overall accuracy of 85.6%, a user accuracy of 87.5%, and a producer accuracy of 97.5% were achieved, and an agreement of 83% between ground measures and remotely sensed data in the estimation of flood water volumes was also achieved on a regional scale. Index Terms—Algorithms, image processing, inundation maps, Landsat.
I. INTRODUCTION
I
N recent decades, remotely sensed data collected by radar and optical satellites have been used for flood extent evaluation over regions characterized by different climate, land cover conditions, and topography [1]. Aerial photos proved to be an efficient tool for mapping and can be considered as concurrent terms of results and costs over small and medium flooded areas. However, the advantages of using satellite remote sensing in flood mapping are the easy availability of the data (especially for the optical data where large archives are often available), the lower cost for a large extent, and the robustness and the efficiency of the data processing techniques developed. Because of their cloud penetration and their all-weather acquisition capability, spaceborne radar systems have been widely used as real-time mapping systems. For example, the National
Manuscript received February 8, 2005; revised July 15, 2005. This work was performed under a research framework supported in part by the Italian Ministry for University and Research (COFIN 2003) under Contract “Tecnologie innovative per la previsione, il controllo e la mitigazione dell’impatto delle emergenze ambientali.” The authors are with the Remote Sensing Laboratory of the Ingegneria Idraulica, Ambientale, Infrastrutture Viarie, Rilevamento (DIIAR) (Hydraulics, Environment) Department, Politecnico di Milano University, 20133 Milan, Italy (e-mail: [email protected]). Digital Object Identifier 10.1109/TGRS.2005.859952
Aeronautics and Space Administration’s Shuttle Imaging Radar Mission B (SIR-B) synthetic aperture radar (SAR) data were used to estimate the flooding damage caused by monsoon in Asia [2], and SIR-C/X-SAR data were used to study the inundation patterns in Brazil [3] and Mexico [4]. The Japanese Earth Resource Satellite 1 (JERS-1) SAR data were used to derive the inundation extent of the Amazon rainforest [5], [6]. The European Remote Sensing Satellite 1 (ERS-1) SAR data were used to study the relationship between discharge and inundated areas in Alaska and British Columbia [7]. Regarding the use of the optical sensors, the cloud cover always associated with floods restricts the availability of data during the event. For this reason, all optical sensors have been mainly used as postflooding mapping systems to derive precise maps of the inundated areas. In the last decades, studies have been carried out using the Système Probatoire pour l’Observation de la Terre Multispectral (SPOT XS) data collected over Europe [8] and Asia [9]. The Landsat Multispectral Scanner (MSS) data collected over Africa [10], and the Landsat Thematic Mapper (TM) data collected over North America [11]–[15] and over Australia [16]–[18] were also used for postflooding damage assessment. Also meteorological and low-resolution satellites have been used for mapping the inundated areas, such as the National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR) data collected over North America [19] and over Australia [20], or the Moderate Resolution Imaging Spectroradiometer (MODIS) data collected over the U.S. [21]. Finally, every kind of satellite sensor was used for monitoring the heavy damage produced by the tsunami that struck the far east Asia at the end of 2004. In this paper, a straightforward method for mapping the inundated areas and for estimating flood water volumes through the use of optical Landsat multispectral data integrated with topography data is presented. The methodology is based on the sequential use of the principal component analysis (PCA), in its spectral-temporal implementation, logical filtering (LF), and image segmentation (IS), and it was developed as a decisional support tool for the allocations of the resource destined to the reconstruction of the inundated municipalities. Two test sites—Piemonte (Italy) and Tennessee (U.S.)—have been used for testing the algorithm described. First, the study areas and the extent of damage due to the floods are described. Then, a discussion of the Landsat TM and Enhanced Thematic Mapper Plus (ETM+) data and of the digital elevation models (DEMs) used is provided. Next, the flood mapping methodology is detailed and explained, and finally results are presented and discussed.
0196-2892/$20.00 © 2006 IEEE
GIANINETTO et al.: POSTFLOOD DAMAGE EVALUATION USING LANDSAT TM AND ETM+ DATA
Fig. 1.
Site location map for the Piemonte study area.
II. STUDY AREAS A. Tanaro Flood of November 1994 Accordingly to the semicircle disposition of the mountainous chains of the western Alps, the Piemonte Region (northern Italy) has a radial arranged hydrographic net divided into the two drain systems of the outflows of the Po River and of its tributary, the Tanaro River. The volume of the water at Pieve del Cairo section (Pavia, Italy) of the Po River, just over the Piemonte Region border, is annually about 14 500 000 000 m , corresponding to a mean annual discharge of 460 m s of which a mean discharge of 129 m s is attributable to the Tanaro River at Montecastello section (Alessandria, Italy). In the first ten days of November 1994 the majority part of the Piemonte Region was struck by an intense low-pressure weather system. In the days of Saturday 4 and Sunday 5 a serious and diffused hydrogeological instability affected a densely inhabited part of the regional territory, causing serious damage to the infrastructure and heavy human losses. From November 4–7, the integrated precipitations on nearly all the Piemonte Region reached record values: more than 100 mm of rainfall on the whole Piemonte Region and more than 200 mm in the upper and in the middle part of the Tanaro basin, with a maximum hourly intensity of 55 mm/h. The resulting flood was the worst event in the last century for the Piemonte Region. In particular, the Tanaro River and its tributaries of the mountainous and hilly areas among the Ligurian Alps and the Langhe reliefs have been heavily involved. All the inhabited centers placed along the Tanaro, the Belbo, and the Bormida di Millesimo valleys have been inundated with very serious and extended damage. Only for the passage of the flood level of the Tanaro River, about 30 have been the recorded victims, for a total of 44 casualities and over 2000 people being homeless [22]. Fig. 1 shows a site location map of the study area. B. Obion and Forked Deer Flood of December 2001 Tennessee’s topography is among the most varied in the U.S., ranging from wide, swampy river valleys in the west, to mountains in the east, with rolling hill country, karstic
237
Fig. 2. Site location map for the Tennesse study area.
terrain, plateaus, mountains, deep gorges, and other features in between. The West Tennessee Plain physiographic province includes three major watersheds: Obion/Forked Deer River, Hatchie River, and the Memphis basin. In Tennessee, the Obion/Forked Deer and Hatchie Rivers encompass 4412 and 1877 square miles, respectively. Topography is characterized as gently rolling, interrupted by small streams and drainage divides. Some gullied topography has developed, and wetlands are common [23]. In 2001, from November 28 to 29 a strong storm system brought record rainfall to the midsouth U.S., with rainfall from 200–300 mm throughout the area. In Memphis, the Wednesday rainfall was 149 mm, becoming the sixth wettest day on record in 129 years. In addition, the 24-h rainfall of 174 mm from noon Wednesday to noon Thursday was the highest 24-h total recorded in the last 50 years. This resulted in extensive flash flooding. The Obion and the Forked Deer rivers overflowed between November 28 and December 7. Many homes, buildings, and roads were inundated, and heavy damage was reported. Fig. 2 shows a site location map of the study area. III. DATA A. Tanaro Dataset For the preflood analysis, the closest cloud-free Landsat TM data of the Tanaro basin were collected on October 16, 1994, three weeks before the heavy rainfall. The Tanaro River reached the peak flood stage on November 6, 1994. Due to the severe cloud coverage during the flooding, the November 2, November 18, December 4, and December 20, 1994 images were not suitable for processing, so, for the postflood analysis the January 5, 1995 Landsat TM image was used (Fig. 3). A digital elevation model of the Tanaro basin was available from the Cartographic Office of the Piemonte Region in ASCII format. The DEM, derived from a photogrammetric survey of the region, had a 50 m 50 m spatial resolution and was arranged in the Roma40 Italian geodetic datum and expressed in the Gauss–Boaga cartographic coordinate system. Descriptive statistics for the DEM area used in this study included min.
238
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 44, NO. 1, JANUARY 2006
Fig. 3. Landsat TM data (path 194, row 28/29) taken over the Tanaro basin (Piemonte, Italy). (a) Preflood TM image collected on October 16, 1994 (4-3-2 band composition). (b) Postflood TM image collected on January 5, 1995 (4-3-2 band composition).
Fig. 4. Landsat ETM+ data (path 23, row 35) taken over the Obion/Forked Deer basin (Tennesse). Data were collected from the Global Land Cover Facility Program (University of Maryland). (a) Preflood ETM+ image collected on November 14, 2001 (4-3-2- band composition). (b) Postflood ETM+ image collected on December 24, 2001 (4-3-2- band composition).
m, mean m, max. m, and standard devim. The DEM was first converted into a raster ation format (for further analysis), then was converted to the UTMWGS84 system, and finally resampled to match the Landsat TM resolution.
For this reason, our study focused on the use of Landsat TM and ETM+ data taken some weeks after the inundations. To identify all the areas that were flooded and to estimate the mean volume of water and sediments carried by the floodwaters, a straightforward four-step algorithm was used.
B. Obion/Forked Deer Dataset The Obion/Forked Deer flood was studied using the Landsat ETM+ data collected from the Global Land Cover Facility Program (University of Maryland). For the preflood analysis, the closest cloud-free Landsat ETM+ data to the Obion/Forked Deer overflow were collected on November 14, 2001, two weeks before the rainfall. For the postflood analysis, the December 24, 2001 Landsat ETM+ image was used (Fig. 4). The DEM of the Obion/Forked Deer basin was available from the U.S. Geological Survey’s (USGS) web site in the Spatial Data Transfer Standard (SDTS) format. The original DEM had a 10 m 10 m cell resolution, with an elevation interval of 0.10 m. Descriptive statistics for the DEM area used in this study inm, mean m, max. m, and cluded min. m. The DEM was converted from standard deviation the NAD27 to the UTM-WGS84 reference system and spatially resampled to match the Landsat ETM+ resolution.
1) In the first step, preflood and postflood Landsat data were first georeferenced to the UTM-WGS84 reference system. For the Tanaro dataset, both the Landsat TM images taken on October 16, 1994 and on January 5, 1995 were rectified to the UTM-WGS84 F32N system, using as reference an already georeferenced Landsat ETM+ image taken on April 24, 2003 (path 194, row 28/29) and supplied by the Italian National Research Council. For the Obion/Forked Deer dataset, the Landsat ETM+ images collected from the Global Land Cover Facility Program (University of Maryland) were already georeferenced to the UTM-WGS84 F15N system. Subsequently, preflood and postflood images were merged together into a synthetic 12-band image, with the first six spectral bands corresponding to the preflood data and the latter six spectral bands corresponding to the postflood data. On the 12-band synthetic image, the spectral-temporal principal component analysis was
GIANINETTO et al.: POSTFLOOD DAMAGE EVALUATION USING LANDSAT TM AND ETM+ DATA
239
Fig. 6. Tanaro River. Areas classified as flooded (plain) superimposed to the map of the flood provided by the Piemonte Region Office (cross-hatching). The classification algorithm allowed to establish an overall accuracy of 85.6%, a user accuracy of 87.5%, and a producer accuracy of 97.5%. TABLE I TANARO DATASET. CONFUSION MATRIX AND ACCURACY ASSESSMENT BASED ON AVAILABLE GROUND MEASUREMENTS SUPPLIED BY THE CARTOGRAPHIC OFFICE OF THE PIEMONTE REGION
Fig. 5. Flood maps of the Tanaro basin derived from the Landsat data processing. (a) Intermediate flood map obtained after the spectral-temporal PCA analysis and logical filtering. (b) Final flood map obtained after the image segmentation and clumping steps.
performed [24]. The PCA involves a mathematical procedure that transforms a number of correlated variables into a smaller number of uncorrelated variables, and in its spectral-temporal implementation, it is able to emphasize the area of change in an image pair [25]–[27]. 2) In the second step, from the DEM of the hydrographical basin were derived the maps of slopes [28]–[30] and the maps of the river flows, with the same resolution of the Landsat data. Using the maps of slopes, the algorithm classified as flooded areas those with the normalized principal component (PC) values and slope values lower than threshold limits. Depending of the relative percentage area of change in the preflood and the postflood image couple, the most significant PC was identified in the higher order PCs (for the Tennessee flood case study) or in the lower order PCs (for the Tanaro flood case study). To solve the misclassification that raised with the exclusion of the course of the rivers from the flooding classification maps, due to the fact that the shores have elevated slopes, the flow of the rivers were also added.
Thresholds were selected in the range of nearly the 50% of the cumulative histogram for the normalized PC and in the range of 30% to 40% values PC . In particular, the following for slope values values were used for the Tanaro dataset: (corresponding to 48% of the cua) PC mulative histogram data for PC6); while the following values were b) used for the Obion/Forked Deer dataset: (corresponding to 58% of the cuc) PC mulative histogram data for PC2); . d) 3) In the third step, image segmentation with eight connectivity neighbors was applied [31], [32] to obtain the final flood map (Fig. 5). Image segmentation partitions the classified image into regions of connected pixels that are contained in the same class. Both for the Tanaro and the Tennesse study areas, we found an optimum threshold for the image segmentation stage in regions with minimum number of 10 000 pixels, followed by a clump operation using a 10 10 pixel kernel dimension. The clump operator was used to cluster adjacent similar classified areas together, producing an homogenization of the data. 4) As a final step, flood intensity was estimated with a simplified empirical model. The height reached by the water along the boundary of the flooded areas is described by the intersection of the inundation contours (derived from the inundation maps) and the topography (derived from
240
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 44, NO. 1, JANUARY 2006
TABLE II TANARO FLOOD. AVERAGE WATER HEIGHT AND WATER VOLUME COMPUTED FOR EVERY CITY COUNCIL AND COMPARED WITH IN SITU MEASURES OF 1994 SUPPLIED BY THE CARTOGRAPHIC OFFICE OF THE PIEMONTE REGION
the DEM). The flood volumes (both water and sediments) were determined by computing the volume included between the water surface and the ground level. Regarding the water height computation, some ground control points (GCPs) belonging to the edges of the flooded areas were selected and, through a second-order linear regression, it was estimated a surface fitting the water heights data with a correlation coefficient of 95%. Using this surface and the ground elevation derived from the DEM, cross sections were computed along transects and, consequently, flood volumes were estimated. Raster data were converted into vectors and municipality and district damage were derived superimposing a georeferenced administrative limit layer. The average water height was computed for every city council and results were compared with some in situ measures available from the Cartographic Office of the Piemonte Region. For assessing the classification accuracy, when available (only for the Tanaro flood case study), in situ measures of hydrometric levels have been compared with flood maps and flood volumes estimated from Landsat data, while when not available (for the Tennessee flood case study), results have been visually compared with the inundation maps comparing flood limits measured from MODIS and NOAA AVHRR data at the Dartmouth Floods Observatory [21]. IV. RESULTS AND DISCUSSION A. Tanaro Flood Because of the relative low percentage of change identified in the Tanaro dataset (11.6% of change versus 88.4% of no change), the areas inundated were well correlated with the lower order PCs [24]. In particular, for the Tanaro dataset PC6 was used for the data processing. The accuracy evaluation for the identified flooded areas in the Landsat images was assessed by using available maps of the
flood provided by the Piemonte Region Office (Fig. 6). Using these maps as ground truth, the classification algorithm proposed allowed to establish an overall accuracy (OA) of 85.6%, a user accuracy (UA) of 87.5%, and a producer accuracy (PA) of 97.5% (Table I). Concerning the estimation of the flood volumes from the Landsat imagery, for the Tanaro basin we estimated a mean water height of 1.42 m and a total water volume of 118 283 229 m (Table II). Via comparison to water heights observed in 1994 of 1.72 m and to an estimated total water volume of 142 727 715 m , an accordance of 83% emerges between ground data and remotely sensed data. The map of the flood supplied from the Piemonte Region Office shown as comparison in Fig. 6 and the survey carried out for evaluating the water height reached in each city council were created with traditional surveys techniques (aerial photogrammetry interpretation integrated with in situ observations). With reference to the flood mapping, the map derived from the Landsat data processing is nearly the same as the official map, except for the valley in the northwest part which was not mapped in the Piemonte Region data. Regarding the flood volume estimation (jointly water and sediments), again it is to point out that the official water height reached in each municipality was evaluated in an empiric way (e.g., measuring the trace of the water level reached on the buildings), as practice. Even if on a municipality scale our study, which tried to use a simplified approach for the flood volume estimation, might show local discrepancies between water heights and water volumes derived from satellite imagery and in situ measures (e.g., Rocca d’Arazzo, Azzano d’Asti, Isola d’Asti), it is interesting to observe that on a regional scale data can be compared. In fact, globally (in mean value) data are comparable and the order of magnitude are correctly detected. As final product, a map of the damage level of the Tanaro Basin (Fig. 7) was produced, based on the topography and on
GIANINETTO et al.: POSTFLOOD DAMAGE EVALUATION USING LANDSAT TM AND ETM+ DATA
Fig. 7. Flood damage evaluation for the Tanaro basin obtained from the image processing of Landsat data.
241
Fig. 9. Flood damage evaluation for the south Forked Deer river obtained from the image processing of Landsat data.
inundation map derived at the Dartmouth Floods Observatory [21] shown in Fig. 10. The two maps are in a good accordance, even if the Dartmouth Floods Observatory’s underestimated the flooded areas because they processed the satellites data of December 2, 2001, and, considering that the rains were continuous until on December 7, the MODIS/AVHRR inundation maps produced were only a partial identification of the real flooding. Moreover, the lower spatial resolution of the MODIS and NOAA AVHRR data, compared to Landsat TM/ETM+ data, concurred in the underestimation of local damaged areas. V. CONCLUSION Fig. 8. Flood map of the Obion/Forked Deer basin derived from the Landsat data processing. Final flood map after image segmentation and data clumping.
the water volume carried by the flooding. This map, describing not only the extent but also the intensity of the flooding, can be used to identify the damage level on a municipality scale and can also be used as a decision support tool for the allocation of the financing destined for the postflooding reconstruction. B. Obion/Forked Deer Flood To test the robustness of the methodology proposed, the processing scheme was also applied to the Obion/Forked Deer flooding that occurred in Tennessee between November and December 2001. In this case, due to the higher percentage of change identified in the Tennessee dataset (44.4% of change versus 55.6% of no change), the areas inundated were well correlated with the higher order PCs [24]. In particular, for the Obion/Forked Deer dataset PC2 was used for the data processing. The final map is shown in Fig. 8. Because no ground control measures were available for that location, for a rough accuracy evaluation the map of inundated areas produced (Fig. 9) was visually compared with the flood
This paper described a straightforward method that used Landsat TM and ETM+ optical imagery for mapping flood extent, estimating the volume of water and sediments carried by the floodwaters, and evaluating the damage associated with floodwaters. This method is based on a four-step algorithm: 1) spectral-temporal principal component analysis; 2) logical filtering with DEM data; 3) image segmentation; and 4) flood intensity estimation. For the analysis, satellite images should be taken one as close as possible before the event, and another as close as possible after the flooding. Optical sensors show serious limitations in postflooding data acquisition, due to the bad weather conditions and cloud coverage always associated with flood events. Our study, however, showed that inundated areas and damage can be precisely mapped even using data taken some weeks after the floods. For the Tanaro event, the comparison with available official flood maps supplied by the Cartographic Office of the Piemonte Region assessed an overall accuracy of 85.6%, a user accuracy of 87.5%, and a producer accuracy of 97.5%. Regarding the volume of water and sediments estimation, the comparison with in situ measures assessed an accordance of 83% between ground measures and remotely sensed data on a regional scale. On a
242
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 44, NO. 1, JANUARY 2006
Fig. 10. Inundation map comparing flood limits derived from MODIS and NOAA AVHRR data on December 2, 2001 at the Dartmouth Floods Observatory (Image courtesy of E. Anderson, Dartmouth Floods Observatory). The rectangle shows the area studied with Landsat data.
municipality scale, however, there might be a lower accuracy both due to insufficient local estimation and to the uncertainty of the ground measures accuracy used for comparison. For the Obion/Forked Deer flood, no ground data were available, so a rough accuracy evaluation was performed by visually comparing the flood map produced with the methodology described with the inundation map obtained from MODIS and NOAA AVHRR satellite data at the Dartmouth Floods Observatory. A good accordance was found. As derived products, maps of the damage level of the Tanaro and of the Obion/Forked Deer basins were obtained, based on the topography and on the estimated water height reached during the flooding. ACKNOWLEDGMENT The authors wish to thank the National Aeronautics and Space Administration for supporting the Global Land Cover Facility Program from which were collected the Obion/Forked Deer Landsat ETM+ data (path 23, row 35) taken on November 14, 2001 and December 24, 2001. Thanks also to the Italian National Research Council (CNR-IREA, Milan) for supplying the Landsat ETM+ image taken over Piemonte (Italy) on April 24, 2003 (path 194, row 28/29) that was used for georeferencing the 1994 Piemonte Landsat TM images. The DEM of the Tanaro basin and the in situ water height measures for the 1994 flood were supplied by the Cartographic Office of the Piemonte Region. The DEM of the Obion/Forked Deer basin
was collected from the U.S. Geological Survey’s web site. Special thanks to E. Anderson (Dartmouth Floods Observatory) for supplying the inundation map comparing flood limits measured from MODIS and AVHRR satellite data on December 2, 2001 used for comparison. REFERENCES [1] P. A. Brivio, R. Colombo, M. Maggi, and R. Tomasoni, “Integration of remote sensing data and GIS for accurate mapping of flooded areas,” Int. J. Remote Sens., vol. 23, no. 3, pp. 429–441, 2002. [2] M. L. Imhoff, C. Vermillon, M. H. Story, A. M. Choudhury, and A. Gafoor, “Monsoon flood boundary delineation and damage assessment using spaceborne imaging radar and Landsat data,” Photogramm. Eng. Remote Sens., vol. 4, pp. 405–413, 1987. [3] L. L. Hess, J. M. Melack, S. Filoso, and Y. Wang, “Delineation of inundated area and vegetation along the Amazon floodplain with the SIR-C synthetic aperture radar,” IEEE Trans. Geosci. Remote Sens., vol. 33, no. 4, pp. 896–904, Jul. 1995. [4] K. O. Pope, E. Rejmankova, J. F. Paris, and R. Woodruff, “Detecting seasonal flooding cycles in marshes of the Yucatan peninsula with SIR-C polarimetric radar imagery,” Remote Sens. Environ., vol. 59, pp. 157–166, 1997. [5] J. M. Melack and Y. Wang, “Delineation of flooded area and flooded vegetation in Balbina reservoir (Amazonas, Brazil) with synthetic aperture radar,” Verhandlung. Int. Vereinigung Limnol., vol. 26, pp. 2374–2377, 1998. [6] F. P. Miranda, L. E. N. Fonseca, and J. R. Carr, “Semivariogram textural classification of JERS-1 (Fuyo-1) SAR data obtained over a flooded area of the Amazon rainforest,” Int. J. Remote Sens., vol. 19, pp. 549–556, 1998. [7] L. C. Smith, L. Bryan, and A. L. Bloom, “Estimation of discharge from three braided rivers using synthetic aperture radar satellite imagery: Potential application to ungauged basins,” Water Resources Res., vol. 32, pp. 2021–2034, 1996.
GIANINETTO et al.: POSTFLOOD DAMAGE EVALUATION USING LANDSAT TM AND ETM+ DATA
[8] M. Robin and J. Le Rhum, “L’inondation de février 1988 dans la marais de monts (Vendée) à partir d’une image SPOT-1 multibande,” PhotoInterprétation, vol. 1, no. 2, pp. 9–16, 1989. [9] F. Blasco, M. F. Bellan, and M. U. Chaudhury, “Estimating the extent of floods in Bangladesh using SPOT data,” Remote Sens. Environ., vol. 39, pp. 167–178, 1992. [10] P. A. Brivio, J. Gregoire, and E. Zilioli, “The detection of hydrological indicators in the study of Niger River regime by means of Landsat imageries,” ITC J., vol. 3, pp. 191–199, 1984. [11] R. Lougeay, P. Baumann, and M. D. Nellis, “Two digital approaches for calculating the area of regions affected by the great American flood of 1993,” Geocarto Int., vol. 9, pp. 53–59, 1994. [12] “Flooding from Hurricane Floyd,” Dartmouth Observatory, Dartmouth College, Hanover, NH, DFO-1999-076, 1999. [13] J. D. Colby, K. A. Mulcahy, and Y. Wang, “Modeling flooding extent from Hurricane Floyd in the coastal plains of North Carolina,” Environ. Hazard, vol. 2, pp. 157–168, 2000. [14] Y. Wang, J. D. Colby, and K. A. Mulcahy, “An efficient method for mapping flood extent in costal floodplain using Landsat TM and DEM data,” Int. J. Remote Sens., vol. 23, no. 18, pp. 3681–3696, 2002. [15] Y. Wang, “Using Landsat 7 TM data acquired days after a flood event to delineate the maximum flood extent on a coastal floodplain,” Int. J. Remote Sens., vol. 25, no. 5, pp. 959–974, 2004. [16] D. Green, M. Shaikh, M. Maini, H. Cross, and J. Slaven, Assessment of Environmental Flow Needs for the Lower Darling River. Sydney, Australia: Dept. Land and Water Conservation, 1998. [17] I. Overton, B. Newman, B. Erdmann, N. Sykora, and S. Slegers, “Modeling floodplain inundation under natural and regulated flows in the lower River Murray,” in Proc. 2nd Austral. Stream Management Conf., Adelaide, Australia, 1999, pp. 493–498. [18] P. Frazier, K. Page, J. Louis, S. Briggs, and A. I. Robertson, “Relating wetland inundation to river flow using Landsat TM data,” Int. J. Remote Sens., vol. 24, no. 19, pp. 3755–3770, 2003. [19] C. P. Berg, M. Matson, and D. R. Wiesnet, Assessing the Red River of the North 1978 Flooding From NOAA Satellite Data. Satellite Hydrology, M. Deutsch, D. R. Wiesnet, and A. Rango, Eds. Minneapolis, MN: Amer. Water Resources Assoc., 1981, pp. 309–315. [20] I. J. Barton and J. M. Bathols, “Monitoring floods with AVHRR,” Remote Sens. Environ., vol. 30, pp. 89–94, 1989. [21] E. Anderson, “MODIS 250 image map series of western Tennessee with inundation maps,” Dartmouth Flood Observatory, Hanover, NH, 2002. [Online]. Available: http://www.dartmouth.edu~floods/images/2001-147b.html. [22] F. Luino, “The flood and landslide event of November 4–6 1994 in Piedmont Region (Northwestern Italy): Causes and related effects in Tanaro Valley,” in Proc. 22nd General Assembly Eur. Geophysical Society, vol. 24, Vienna, Austria, Nov. 4–6, 1999, pp. 123–129. [23] DOA. (1989) NPS management program document. Dept. Agricult., Ellington Agricult. Center, Nashville, TN. [Online]. Available: http://www.state.tn.us/agriculture/nps/npsdoc. [24] R. A. Schowengerdt, Remote Sensing-Models and Methods for Image Processing, 2nd ed. Orlando, FL: Academic, 1997. [25] S. E. Ingebritsen and R. J. P. Lyon, “Principal components analysis of multitemporal image pairs,” Int. J. Remote Sens., vol. 6, pp. 687–696, 1985. [26] T. Fung and E. LeDrew, “Application of principal component analysis to change detection,” Photogramm. Eng. Remote Sens., vol. 53, no. 12, pp. 1649–1658, 1987. [27] J. A. Richards, Remote Sensing Digital Image Analysis-An Introduction, 2nd ed. New York: Springer-Verlag, 1993. [28] J. R. Carter, “The effect of data precision on the calculation of slope and aspect using gridded DEM’s,” Cartographica, vol. 29, no. 1, pp. 22–34, 1992. [29] C. Huang, C. Gascuel-Odoux, and S. Cros-Cayot, “Hillslope topographic and hydrologic effects on overland flow and erosion,” Catena, vol. 46, no. 2–3, pp. 177–188, 2002. [30] M. F. Hutchinson and J. C. Gallant, Digital Elevation Models and Representation of Terrain Slope. Terrain Analysis, Principles and Applications. New York: Wiley, pp. 29–50.
243
[31] R. E. Roger, “Principal components transform with simple, automatic noise adjustment,” Int. J. Remote Sens., vol. 17, no. 14, pp. 2719–2727, 1996. [32] R. C. Gonzales and R. E. Woods, Digital Image Processing, 2nd ed. Upper Saddler River, NJ: Prentice-Hall, 2002.
Marco Gianinetto received the laurea (M.S.) degree in environmental engineering from the Politecnico di Milano Technical University, Milan, Italy, in 2000. He is currently pursuing the Ph.D. degree in geomatics at the Politecnico di Milano Technical University. Since 2000, he has been with the Remote Sensing Laboratory (RSL), DIIAR Department, Politecnico di Milano Technical University. He has been designated Contract Professor at the Politecnico di Milano Technical University, Polo Regionale di Lecco, Lecco, Italy, in 2004, and currently teaches remote sensing and GIS. His main research activity is in the area of hyperspectral remote sensing and high-resolution satellite data processing. In these fields, he conducts research within several national projects. He is author (or coauthor) of more than 30 scientific publications, including book chapters, journal articles, conference proceedings, and technical reports. He is a referee for the International Journal of Remote Sensing, Remote Sensing of Environment, Photogrammetric Engineering and Remote Sensing, and the Italian Journal of Remote Sensing. In 2003, he was appointed a member of the scientific committee of the Italian Remote Sensing Journal. Mr. Gianinetto is a member of the International Society for Photogrammetry and Remote Sensing (ISPRS) and an active member of the ISPRS WG’s III/1, VII/6, and VIII/2. He is a member of the Italian Association for Remote Sensing (AIT).
Paolo Villa received the laurea (M.S.) degree in environmental engineering from the Politecnico di Milano University, Milan, Italy, in 2004. He is currently pursuing the Ph.D. degree in geomatics at the Politecnico di Milano University, Milan. Since 2004, he has been with the Remote Sensing Laboratory of the Ingegneria Idraulica, Ambientale, Infrastrutture Viarie, Rilevamento (DIIAR) (Hydraulics, Environment) Department, Politecnico di Milano University, and his main research activity is in the area of change detection with midresolution satellite data.
Giovanmaria Lechi received the laurea (M.S.) degree in electrical engineering from the Politecnico di Milano Technical University, Milano, Italy, in 1970. He began his career in 1970 with the Italian National Research Council, attending in remote sensing. He has been appointed director of the Lithosphere Geophysics Dept. of the Italian National Research Council (CNR). From 1987 he is Associate Professor of Remote Sensing at the Politecnico di Milano Technical University, Milano, Italy. He is author (or coauthor) of more than 100 scientific publications, including books, book chapters, journals and conference proceedings. Mr. Lechi he was appointed a member of the UE Commission for the development of a cadastre for the Italian olive tree plantings in 1975. He is a founding member of the Italian Association for Remote Sensing (AIT), and from 1996 to 2003, he has been also President of the Italian Association for Remote Sensing. He is member and arbitrator of the Italian Topographic and Photogrammetric Society.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 44, NO. 1, JANUARY 2006
Postflood Damage Evaluation Using Landsat TM and ETM+ Data Integrated With DEM Marco Gianinetto, Paolo Villa, and Giovanmaria Lechi
Abstract—In recent decades, radar and optical satellite imagery have been used for evaluating flooding extent. In this paper, a straightforward technique based on the sequential use of the spectral-temporal principal component analysis, logical filtering, and image segmentation integrated with the digital elevation model was developed as a decisional support tool for the allocations of the resource destined for the flooded areas. The mapping technique was first applied to the catastrophic event that occurred in the Piemonte Region (Italy) in November 1994, which was the worst event of the past century for that region, with 44 casualities and over 2000 homeless. Next, it was applied to the Obion/Forked Deer inundation that occurred in Tennessee (U.S.) between November and December 2001, in which heavy damage to the infrastructure was reported. Two Landsat-5 Thematic Mapper (path 194, row 28/29) and two Landsat-7 Enhanced Thematic Mapper Plus (path 23, row 35) images were processed, two of them collected before and two after the events. The method proposed proved to be an effective approach for evaluating flood extent and for assessing the damage produced by the flooding. An overall accuracy of 85.6%, a user accuracy of 87.5%, and a producer accuracy of 97.5% were achieved, and an agreement of 83% between ground measures and remotely sensed data in the estimation of flood water volumes was also achieved on a regional scale. Index Terms—Algorithms, image processing, inundation maps, Landsat.
I. INTRODUCTION
I
N recent decades, remotely sensed data collected by radar and optical satellites have been used for flood extent evaluation over regions characterized by different climate, land cover conditions, and topography [1]. Aerial photos proved to be an efficient tool for mapping and can be considered as concurrent terms of results and costs over small and medium flooded areas. However, the advantages of using satellite remote sensing in flood mapping are the easy availability of the data (especially for the optical data where large archives are often available), the lower cost for a large extent, and the robustness and the efficiency of the data processing techniques developed. Because of their cloud penetration and their all-weather acquisition capability, spaceborne radar systems have been widely used as real-time mapping systems. For example, the National
Manuscript received February 8, 2005; revised July 15, 2005. This work was performed under a research framework supported in part by the Italian Ministry for University and Research (COFIN 2003) under Contract “Tecnologie innovative per la previsione, il controllo e la mitigazione dell’impatto delle emergenze ambientali.” The authors are with the Remote Sensing Laboratory of the Ingegneria Idraulica, Ambientale, Infrastrutture Viarie, Rilevamento (DIIAR) (Hydraulics, Environment) Department, Politecnico di Milano University, 20133 Milan, Italy (e-mail: [email protected]). Digital Object Identifier 10.1109/TGRS.2005.859952
Aeronautics and Space Administration’s Shuttle Imaging Radar Mission B (SIR-B) synthetic aperture radar (SAR) data were used to estimate the flooding damage caused by monsoon in Asia [2], and SIR-C/X-SAR data were used to study the inundation patterns in Brazil [3] and Mexico [4]. The Japanese Earth Resource Satellite 1 (JERS-1) SAR data were used to derive the inundation extent of the Amazon rainforest [5], [6]. The European Remote Sensing Satellite 1 (ERS-1) SAR data were used to study the relationship between discharge and inundated areas in Alaska and British Columbia [7]. Regarding the use of the optical sensors, the cloud cover always associated with floods restricts the availability of data during the event. For this reason, all optical sensors have been mainly used as postflooding mapping systems to derive precise maps of the inundated areas. In the last decades, studies have been carried out using the Système Probatoire pour l’Observation de la Terre Multispectral (SPOT XS) data collected over Europe [8] and Asia [9]. The Landsat Multispectral Scanner (MSS) data collected over Africa [10], and the Landsat Thematic Mapper (TM) data collected over North America [11]–[15] and over Australia [16]–[18] were also used for postflooding damage assessment. Also meteorological and low-resolution satellites have been used for mapping the inundated areas, such as the National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR) data collected over North America [19] and over Australia [20], or the Moderate Resolution Imaging Spectroradiometer (MODIS) data collected over the U.S. [21]. Finally, every kind of satellite sensor was used for monitoring the heavy damage produced by the tsunami that struck the far east Asia at the end of 2004. In this paper, a straightforward method for mapping the inundated areas and for estimating flood water volumes through the use of optical Landsat multispectral data integrated with topography data is presented. The methodology is based on the sequential use of the principal component analysis (PCA), in its spectral-temporal implementation, logical filtering (LF), and image segmentation (IS), and it was developed as a decisional support tool for the allocations of the resource destined to the reconstruction of the inundated municipalities. Two test sites—Piemonte (Italy) and Tennessee (U.S.)—have been used for testing the algorithm described. First, the study areas and the extent of damage due to the floods are described. Then, a discussion of the Landsat TM and Enhanced Thematic Mapper Plus (ETM+) data and of the digital elevation models (DEMs) used is provided. Next, the flood mapping methodology is detailed and explained, and finally results are presented and discussed.
0196-2892/$20.00 © 2006 IEEE
GIANINETTO et al.: POSTFLOOD DAMAGE EVALUATION USING LANDSAT TM AND ETM+ DATA
Fig. 1.
Site location map for the Piemonte study area.
II. STUDY AREAS A. Tanaro Flood of November 1994 Accordingly to the semicircle disposition of the mountainous chains of the western Alps, the Piemonte Region (northern Italy) has a radial arranged hydrographic net divided into the two drain systems of the outflows of the Po River and of its tributary, the Tanaro River. The volume of the water at Pieve del Cairo section (Pavia, Italy) of the Po River, just over the Piemonte Region border, is annually about 14 500 000 000 m , corresponding to a mean annual discharge of 460 m s of which a mean discharge of 129 m s is attributable to the Tanaro River at Montecastello section (Alessandria, Italy). In the first ten days of November 1994 the majority part of the Piemonte Region was struck by an intense low-pressure weather system. In the days of Saturday 4 and Sunday 5 a serious and diffused hydrogeological instability affected a densely inhabited part of the regional territory, causing serious damage to the infrastructure and heavy human losses. From November 4–7, the integrated precipitations on nearly all the Piemonte Region reached record values: more than 100 mm of rainfall on the whole Piemonte Region and more than 200 mm in the upper and in the middle part of the Tanaro basin, with a maximum hourly intensity of 55 mm/h. The resulting flood was the worst event in the last century for the Piemonte Region. In particular, the Tanaro River and its tributaries of the mountainous and hilly areas among the Ligurian Alps and the Langhe reliefs have been heavily involved. All the inhabited centers placed along the Tanaro, the Belbo, and the Bormida di Millesimo valleys have been inundated with very serious and extended damage. Only for the passage of the flood level of the Tanaro River, about 30 have been the recorded victims, for a total of 44 casualities and over 2000 people being homeless [22]. Fig. 1 shows a site location map of the study area. B. Obion and Forked Deer Flood of December 2001 Tennessee’s topography is among the most varied in the U.S., ranging from wide, swampy river valleys in the west, to mountains in the east, with rolling hill country, karstic
237
Fig. 2. Site location map for the Tennesse study area.
terrain, plateaus, mountains, deep gorges, and other features in between. The West Tennessee Plain physiographic province includes three major watersheds: Obion/Forked Deer River, Hatchie River, and the Memphis basin. In Tennessee, the Obion/Forked Deer and Hatchie Rivers encompass 4412 and 1877 square miles, respectively. Topography is characterized as gently rolling, interrupted by small streams and drainage divides. Some gullied topography has developed, and wetlands are common [23]. In 2001, from November 28 to 29 a strong storm system brought record rainfall to the midsouth U.S., with rainfall from 200–300 mm throughout the area. In Memphis, the Wednesday rainfall was 149 mm, becoming the sixth wettest day on record in 129 years. In addition, the 24-h rainfall of 174 mm from noon Wednesday to noon Thursday was the highest 24-h total recorded in the last 50 years. This resulted in extensive flash flooding. The Obion and the Forked Deer rivers overflowed between November 28 and December 7. Many homes, buildings, and roads were inundated, and heavy damage was reported. Fig. 2 shows a site location map of the study area. III. DATA A. Tanaro Dataset For the preflood analysis, the closest cloud-free Landsat TM data of the Tanaro basin were collected on October 16, 1994, three weeks before the heavy rainfall. The Tanaro River reached the peak flood stage on November 6, 1994. Due to the severe cloud coverage during the flooding, the November 2, November 18, December 4, and December 20, 1994 images were not suitable for processing, so, for the postflood analysis the January 5, 1995 Landsat TM image was used (Fig. 3). A digital elevation model of the Tanaro basin was available from the Cartographic Office of the Piemonte Region in ASCII format. The DEM, derived from a photogrammetric survey of the region, had a 50 m 50 m spatial resolution and was arranged in the Roma40 Italian geodetic datum and expressed in the Gauss–Boaga cartographic coordinate system. Descriptive statistics for the DEM area used in this study included min.
238
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 44, NO. 1, JANUARY 2006
Fig. 3. Landsat TM data (path 194, row 28/29) taken over the Tanaro basin (Piemonte, Italy). (a) Preflood TM image collected on October 16, 1994 (4-3-2 band composition). (b) Postflood TM image collected on January 5, 1995 (4-3-2 band composition).
Fig. 4. Landsat ETM+ data (path 23, row 35) taken over the Obion/Forked Deer basin (Tennesse). Data were collected from the Global Land Cover Facility Program (University of Maryland). (a) Preflood ETM+ image collected on November 14, 2001 (4-3-2- band composition). (b) Postflood ETM+ image collected on December 24, 2001 (4-3-2- band composition).
m, mean m, max. m, and standard devim. The DEM was first converted into a raster ation format (for further analysis), then was converted to the UTMWGS84 system, and finally resampled to match the Landsat TM resolution.
For this reason, our study focused on the use of Landsat TM and ETM+ data taken some weeks after the inundations. To identify all the areas that were flooded and to estimate the mean volume of water and sediments carried by the floodwaters, a straightforward four-step algorithm was used.
B. Obion/Forked Deer Dataset The Obion/Forked Deer flood was studied using the Landsat ETM+ data collected from the Global Land Cover Facility Program (University of Maryland). For the preflood analysis, the closest cloud-free Landsat ETM+ data to the Obion/Forked Deer overflow were collected on November 14, 2001, two weeks before the rainfall. For the postflood analysis, the December 24, 2001 Landsat ETM+ image was used (Fig. 4). The DEM of the Obion/Forked Deer basin was available from the U.S. Geological Survey’s (USGS) web site in the Spatial Data Transfer Standard (SDTS) format. The original DEM had a 10 m 10 m cell resolution, with an elevation interval of 0.10 m. Descriptive statistics for the DEM area used in this study inm, mean m, max. m, and cluded min. m. The DEM was converted from standard deviation the NAD27 to the UTM-WGS84 reference system and spatially resampled to match the Landsat ETM+ resolution.
1) In the first step, preflood and postflood Landsat data were first georeferenced to the UTM-WGS84 reference system. For the Tanaro dataset, both the Landsat TM images taken on October 16, 1994 and on January 5, 1995 were rectified to the UTM-WGS84 F32N system, using as reference an already georeferenced Landsat ETM+ image taken on April 24, 2003 (path 194, row 28/29) and supplied by the Italian National Research Council. For the Obion/Forked Deer dataset, the Landsat ETM+ images collected from the Global Land Cover Facility Program (University of Maryland) were already georeferenced to the UTM-WGS84 F15N system. Subsequently, preflood and postflood images were merged together into a synthetic 12-band image, with the first six spectral bands corresponding to the preflood data and the latter six spectral bands corresponding to the postflood data. On the 12-band synthetic image, the spectral-temporal principal component analysis was
GIANINETTO et al.: POSTFLOOD DAMAGE EVALUATION USING LANDSAT TM AND ETM+ DATA
239
Fig. 6. Tanaro River. Areas classified as flooded (plain) superimposed to the map of the flood provided by the Piemonte Region Office (cross-hatching). The classification algorithm allowed to establish an overall accuracy of 85.6%, a user accuracy of 87.5%, and a producer accuracy of 97.5%. TABLE I TANARO DATASET. CONFUSION MATRIX AND ACCURACY ASSESSMENT BASED ON AVAILABLE GROUND MEASUREMENTS SUPPLIED BY THE CARTOGRAPHIC OFFICE OF THE PIEMONTE REGION
Fig. 5. Flood maps of the Tanaro basin derived from the Landsat data processing. (a) Intermediate flood map obtained after the spectral-temporal PCA analysis and logical filtering. (b) Final flood map obtained after the image segmentation and clumping steps.
performed [24]. The PCA involves a mathematical procedure that transforms a number of correlated variables into a smaller number of uncorrelated variables, and in its spectral-temporal implementation, it is able to emphasize the area of change in an image pair [25]–[27]. 2) In the second step, from the DEM of the hydrographical basin were derived the maps of slopes [28]–[30] and the maps of the river flows, with the same resolution of the Landsat data. Using the maps of slopes, the algorithm classified as flooded areas those with the normalized principal component (PC) values and slope values lower than threshold limits. Depending of the relative percentage area of change in the preflood and the postflood image couple, the most significant PC was identified in the higher order PCs (for the Tennessee flood case study) or in the lower order PCs (for the Tanaro flood case study). To solve the misclassification that raised with the exclusion of the course of the rivers from the flooding classification maps, due to the fact that the shores have elevated slopes, the flow of the rivers were also added.
Thresholds were selected in the range of nearly the 50% of the cumulative histogram for the normalized PC and in the range of 30% to 40% values PC . In particular, the following for slope values values were used for the Tanaro dataset: (corresponding to 48% of the cua) PC mulative histogram data for PC6); while the following values were b) used for the Obion/Forked Deer dataset: (corresponding to 58% of the cuc) PC mulative histogram data for PC2); . d) 3) In the third step, image segmentation with eight connectivity neighbors was applied [31], [32] to obtain the final flood map (Fig. 5). Image segmentation partitions the classified image into regions of connected pixels that are contained in the same class. Both for the Tanaro and the Tennesse study areas, we found an optimum threshold for the image segmentation stage in regions with minimum number of 10 000 pixels, followed by a clump operation using a 10 10 pixel kernel dimension. The clump operator was used to cluster adjacent similar classified areas together, producing an homogenization of the data. 4) As a final step, flood intensity was estimated with a simplified empirical model. The height reached by the water along the boundary of the flooded areas is described by the intersection of the inundation contours (derived from the inundation maps) and the topography (derived from
240
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 44, NO. 1, JANUARY 2006
TABLE II TANARO FLOOD. AVERAGE WATER HEIGHT AND WATER VOLUME COMPUTED FOR EVERY CITY COUNCIL AND COMPARED WITH IN SITU MEASURES OF 1994 SUPPLIED BY THE CARTOGRAPHIC OFFICE OF THE PIEMONTE REGION
the DEM). The flood volumes (both water and sediments) were determined by computing the volume included between the water surface and the ground level. Regarding the water height computation, some ground control points (GCPs) belonging to the edges of the flooded areas were selected and, through a second-order linear regression, it was estimated a surface fitting the water heights data with a correlation coefficient of 95%. Using this surface and the ground elevation derived from the DEM, cross sections were computed along transects and, consequently, flood volumes were estimated. Raster data were converted into vectors and municipality and district damage were derived superimposing a georeferenced administrative limit layer. The average water height was computed for every city council and results were compared with some in situ measures available from the Cartographic Office of the Piemonte Region. For assessing the classification accuracy, when available (only for the Tanaro flood case study), in situ measures of hydrometric levels have been compared with flood maps and flood volumes estimated from Landsat data, while when not available (for the Tennessee flood case study), results have been visually compared with the inundation maps comparing flood limits measured from MODIS and NOAA AVHRR data at the Dartmouth Floods Observatory [21]. IV. RESULTS AND DISCUSSION A. Tanaro Flood Because of the relative low percentage of change identified in the Tanaro dataset (11.6% of change versus 88.4% of no change), the areas inundated were well correlated with the lower order PCs [24]. In particular, for the Tanaro dataset PC6 was used for the data processing. The accuracy evaluation for the identified flooded areas in the Landsat images was assessed by using available maps of the
flood provided by the Piemonte Region Office (Fig. 6). Using these maps as ground truth, the classification algorithm proposed allowed to establish an overall accuracy (OA) of 85.6%, a user accuracy (UA) of 87.5%, and a producer accuracy (PA) of 97.5% (Table I). Concerning the estimation of the flood volumes from the Landsat imagery, for the Tanaro basin we estimated a mean water height of 1.42 m and a total water volume of 118 283 229 m (Table II). Via comparison to water heights observed in 1994 of 1.72 m and to an estimated total water volume of 142 727 715 m , an accordance of 83% emerges between ground data and remotely sensed data. The map of the flood supplied from the Piemonte Region Office shown as comparison in Fig. 6 and the survey carried out for evaluating the water height reached in each city council were created with traditional surveys techniques (aerial photogrammetry interpretation integrated with in situ observations). With reference to the flood mapping, the map derived from the Landsat data processing is nearly the same as the official map, except for the valley in the northwest part which was not mapped in the Piemonte Region data. Regarding the flood volume estimation (jointly water and sediments), again it is to point out that the official water height reached in each municipality was evaluated in an empiric way (e.g., measuring the trace of the water level reached on the buildings), as practice. Even if on a municipality scale our study, which tried to use a simplified approach for the flood volume estimation, might show local discrepancies between water heights and water volumes derived from satellite imagery and in situ measures (e.g., Rocca d’Arazzo, Azzano d’Asti, Isola d’Asti), it is interesting to observe that on a regional scale data can be compared. In fact, globally (in mean value) data are comparable and the order of magnitude are correctly detected. As final product, a map of the damage level of the Tanaro Basin (Fig. 7) was produced, based on the topography and on
GIANINETTO et al.: POSTFLOOD DAMAGE EVALUATION USING LANDSAT TM AND ETM+ DATA
Fig. 7. Flood damage evaluation for the Tanaro basin obtained from the image processing of Landsat data.
241
Fig. 9. Flood damage evaluation for the south Forked Deer river obtained from the image processing of Landsat data.
inundation map derived at the Dartmouth Floods Observatory [21] shown in Fig. 10. The two maps are in a good accordance, even if the Dartmouth Floods Observatory’s underestimated the flooded areas because they processed the satellites data of December 2, 2001, and, considering that the rains were continuous until on December 7, the MODIS/AVHRR inundation maps produced were only a partial identification of the real flooding. Moreover, the lower spatial resolution of the MODIS and NOAA AVHRR data, compared to Landsat TM/ETM+ data, concurred in the underestimation of local damaged areas. V. CONCLUSION Fig. 8. Flood map of the Obion/Forked Deer basin derived from the Landsat data processing. Final flood map after image segmentation and data clumping.
the water volume carried by the flooding. This map, describing not only the extent but also the intensity of the flooding, can be used to identify the damage level on a municipality scale and can also be used as a decision support tool for the allocation of the financing destined for the postflooding reconstruction. B. Obion/Forked Deer Flood To test the robustness of the methodology proposed, the processing scheme was also applied to the Obion/Forked Deer flooding that occurred in Tennessee between November and December 2001. In this case, due to the higher percentage of change identified in the Tennessee dataset (44.4% of change versus 55.6% of no change), the areas inundated were well correlated with the higher order PCs [24]. In particular, for the Obion/Forked Deer dataset PC2 was used for the data processing. The final map is shown in Fig. 8. Because no ground control measures were available for that location, for a rough accuracy evaluation the map of inundated areas produced (Fig. 9) was visually compared with the flood
This paper described a straightforward method that used Landsat TM and ETM+ optical imagery for mapping flood extent, estimating the volume of water and sediments carried by the floodwaters, and evaluating the damage associated with floodwaters. This method is based on a four-step algorithm: 1) spectral-temporal principal component analysis; 2) logical filtering with DEM data; 3) image segmentation; and 4) flood intensity estimation. For the analysis, satellite images should be taken one as close as possible before the event, and another as close as possible after the flooding. Optical sensors show serious limitations in postflooding data acquisition, due to the bad weather conditions and cloud coverage always associated with flood events. Our study, however, showed that inundated areas and damage can be precisely mapped even using data taken some weeks after the floods. For the Tanaro event, the comparison with available official flood maps supplied by the Cartographic Office of the Piemonte Region assessed an overall accuracy of 85.6%, a user accuracy of 87.5%, and a producer accuracy of 97.5%. Regarding the volume of water and sediments estimation, the comparison with in situ measures assessed an accordance of 83% between ground measures and remotely sensed data on a regional scale. On a
242
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 44, NO. 1, JANUARY 2006
Fig. 10. Inundation map comparing flood limits derived from MODIS and NOAA AVHRR data on December 2, 2001 at the Dartmouth Floods Observatory (Image courtesy of E. Anderson, Dartmouth Floods Observatory). The rectangle shows the area studied with Landsat data.
municipality scale, however, there might be a lower accuracy both due to insufficient local estimation and to the uncertainty of the ground measures accuracy used for comparison. For the Obion/Forked Deer flood, no ground data were available, so a rough accuracy evaluation was performed by visually comparing the flood map produced with the methodology described with the inundation map obtained from MODIS and NOAA AVHRR satellite data at the Dartmouth Floods Observatory. A good accordance was found. As derived products, maps of the damage level of the Tanaro and of the Obion/Forked Deer basins were obtained, based on the topography and on the estimated water height reached during the flooding. ACKNOWLEDGMENT The authors wish to thank the National Aeronautics and Space Administration for supporting the Global Land Cover Facility Program from which were collected the Obion/Forked Deer Landsat ETM+ data (path 23, row 35) taken on November 14, 2001 and December 24, 2001. Thanks also to the Italian National Research Council (CNR-IREA, Milan) for supplying the Landsat ETM+ image taken over Piemonte (Italy) on April 24, 2003 (path 194, row 28/29) that was used for georeferencing the 1994 Piemonte Landsat TM images. The DEM of the Tanaro basin and the in situ water height measures for the 1994 flood were supplied by the Cartographic Office of the Piemonte Region. The DEM of the Obion/Forked Deer basin
was collected from the U.S. Geological Survey’s web site. Special thanks to E. Anderson (Dartmouth Floods Observatory) for supplying the inundation map comparing flood limits measured from MODIS and AVHRR satellite data on December 2, 2001 used for comparison. REFERENCES [1] P. A. Brivio, R. Colombo, M. Maggi, and R. Tomasoni, “Integration of remote sensing data and GIS for accurate mapping of flooded areas,” Int. J. Remote Sens., vol. 23, no. 3, pp. 429–441, 2002. [2] M. L. Imhoff, C. Vermillon, M. H. Story, A. M. Choudhury, and A. Gafoor, “Monsoon flood boundary delineation and damage assessment using spaceborne imaging radar and Landsat data,” Photogramm. Eng. Remote Sens., vol. 4, pp. 405–413, 1987. [3] L. L. Hess, J. M. Melack, S. Filoso, and Y. Wang, “Delineation of inundated area and vegetation along the Amazon floodplain with the SIR-C synthetic aperture radar,” IEEE Trans. Geosci. Remote Sens., vol. 33, no. 4, pp. 896–904, Jul. 1995. [4] K. O. Pope, E. Rejmankova, J. F. Paris, and R. Woodruff, “Detecting seasonal flooding cycles in marshes of the Yucatan peninsula with SIR-C polarimetric radar imagery,” Remote Sens. Environ., vol. 59, pp. 157–166, 1997. [5] J. M. Melack and Y. Wang, “Delineation of flooded area and flooded vegetation in Balbina reservoir (Amazonas, Brazil) with synthetic aperture radar,” Verhandlung. Int. Vereinigung Limnol., vol. 26, pp. 2374–2377, 1998. [6] F. P. Miranda, L. E. N. Fonseca, and J. R. Carr, “Semivariogram textural classification of JERS-1 (Fuyo-1) SAR data obtained over a flooded area of the Amazon rainforest,” Int. J. Remote Sens., vol. 19, pp. 549–556, 1998. [7] L. C. Smith, L. Bryan, and A. L. Bloom, “Estimation of discharge from three braided rivers using synthetic aperture radar satellite imagery: Potential application to ungauged basins,” Water Resources Res., vol. 32, pp. 2021–2034, 1996.
GIANINETTO et al.: POSTFLOOD DAMAGE EVALUATION USING LANDSAT TM AND ETM+ DATA
[8] M. Robin and J. Le Rhum, “L’inondation de février 1988 dans la marais de monts (Vendée) à partir d’une image SPOT-1 multibande,” PhotoInterprétation, vol. 1, no. 2, pp. 9–16, 1989. [9] F. Blasco, M. F. Bellan, and M. U. Chaudhury, “Estimating the extent of floods in Bangladesh using SPOT data,” Remote Sens. Environ., vol. 39, pp. 167–178, 1992. [10] P. A. Brivio, J. Gregoire, and E. Zilioli, “The detection of hydrological indicators in the study of Niger River regime by means of Landsat imageries,” ITC J., vol. 3, pp. 191–199, 1984. [11] R. Lougeay, P. Baumann, and M. D. Nellis, “Two digital approaches for calculating the area of regions affected by the great American flood of 1993,” Geocarto Int., vol. 9, pp. 53–59, 1994. [12] “Flooding from Hurricane Floyd,” Dartmouth Observatory, Dartmouth College, Hanover, NH, DFO-1999-076, 1999. [13] J. D. Colby, K. A. Mulcahy, and Y. Wang, “Modeling flooding extent from Hurricane Floyd in the coastal plains of North Carolina,” Environ. Hazard, vol. 2, pp. 157–168, 2000. [14] Y. Wang, J. D. Colby, and K. A. Mulcahy, “An efficient method for mapping flood extent in costal floodplain using Landsat TM and DEM data,” Int. J. Remote Sens., vol. 23, no. 18, pp. 3681–3696, 2002. [15] Y. Wang, “Using Landsat 7 TM data acquired days after a flood event to delineate the maximum flood extent on a coastal floodplain,” Int. J. Remote Sens., vol. 25, no. 5, pp. 959–974, 2004. [16] D. Green, M. Shaikh, M. Maini, H. Cross, and J. Slaven, Assessment of Environmental Flow Needs for the Lower Darling River. Sydney, Australia: Dept. Land and Water Conservation, 1998. [17] I. Overton, B. Newman, B. Erdmann, N. Sykora, and S. Slegers, “Modeling floodplain inundation under natural and regulated flows in the lower River Murray,” in Proc. 2nd Austral. Stream Management Conf., Adelaide, Australia, 1999, pp. 493–498. [18] P. Frazier, K. Page, J. Louis, S. Briggs, and A. I. Robertson, “Relating wetland inundation to river flow using Landsat TM data,” Int. J. Remote Sens., vol. 24, no. 19, pp. 3755–3770, 2003. [19] C. P. Berg, M. Matson, and D. R. Wiesnet, Assessing the Red River of the North 1978 Flooding From NOAA Satellite Data. Satellite Hydrology, M. Deutsch, D. R. Wiesnet, and A. Rango, Eds. Minneapolis, MN: Amer. Water Resources Assoc., 1981, pp. 309–315. [20] I. J. Barton and J. M. Bathols, “Monitoring floods with AVHRR,” Remote Sens. Environ., vol. 30, pp. 89–94, 1989. [21] E. Anderson, “MODIS 250 image map series of western Tennessee with inundation maps,” Dartmouth Flood Observatory, Hanover, NH, 2002. [Online]. Available: http://www.dartmouth.edu~floods/images/2001-147b.html. [22] F. Luino, “The flood and landslide event of November 4–6 1994 in Piedmont Region (Northwestern Italy): Causes and related effects in Tanaro Valley,” in Proc. 22nd General Assembly Eur. Geophysical Society, vol. 24, Vienna, Austria, Nov. 4–6, 1999, pp. 123–129. [23] DOA. (1989) NPS management program document. Dept. Agricult., Ellington Agricult. Center, Nashville, TN. [Online]. Available: http://www.state.tn.us/agriculture/nps/npsdoc. [24] R. A. Schowengerdt, Remote Sensing-Models and Methods for Image Processing, 2nd ed. Orlando, FL: Academic, 1997. [25] S. E. Ingebritsen and R. J. P. Lyon, “Principal components analysis of multitemporal image pairs,” Int. J. Remote Sens., vol. 6, pp. 687–696, 1985. [26] T. Fung and E. LeDrew, “Application of principal component analysis to change detection,” Photogramm. Eng. Remote Sens., vol. 53, no. 12, pp. 1649–1658, 1987. [27] J. A. Richards, Remote Sensing Digital Image Analysis-An Introduction, 2nd ed. New York: Springer-Verlag, 1993. [28] J. R. Carter, “The effect of data precision on the calculation of slope and aspect using gridded DEM’s,” Cartographica, vol. 29, no. 1, pp. 22–34, 1992. [29] C. Huang, C. Gascuel-Odoux, and S. Cros-Cayot, “Hillslope topographic and hydrologic effects on overland flow and erosion,” Catena, vol. 46, no. 2–3, pp. 177–188, 2002. [30] M. F. Hutchinson and J. C. Gallant, Digital Elevation Models and Representation of Terrain Slope. Terrain Analysis, Principles and Applications. New York: Wiley, pp. 29–50.
243
[31] R. E. Roger, “Principal components transform with simple, automatic noise adjustment,” Int. J. Remote Sens., vol. 17, no. 14, pp. 2719–2727, 1996. [32] R. C. Gonzales and R. E. Woods, Digital Image Processing, 2nd ed. Upper Saddler River, NJ: Prentice-Hall, 2002.
Marco Gianinetto received the laurea (M.S.) degree in environmental engineering from the Politecnico di Milano Technical University, Milan, Italy, in 2000. He is currently pursuing the Ph.D. degree in geomatics at the Politecnico di Milano Technical University. Since 2000, he has been with the Remote Sensing Laboratory (RSL), DIIAR Department, Politecnico di Milano Technical University. He has been designated Contract Professor at the Politecnico di Milano Technical University, Polo Regionale di Lecco, Lecco, Italy, in 2004, and currently teaches remote sensing and GIS. His main research activity is in the area of hyperspectral remote sensing and high-resolution satellite data processing. In these fields, he conducts research within several national projects. He is author (or coauthor) of more than 30 scientific publications, including book chapters, journal articles, conference proceedings, and technical reports. He is a referee for the International Journal of Remote Sensing, Remote Sensing of Environment, Photogrammetric Engineering and Remote Sensing, and the Italian Journal of Remote Sensing. In 2003, he was appointed a member of the scientific committee of the Italian Remote Sensing Journal. Mr. Gianinetto is a member of the International Society for Photogrammetry and Remote Sensing (ISPRS) and an active member of the ISPRS WG’s III/1, VII/6, and VIII/2. He is a member of the Italian Association for Remote Sensing (AIT).
Paolo Villa received the laurea (M.S.) degree in environmental engineering from the Politecnico di Milano University, Milan, Italy, in 2004. He is currently pursuing the Ph.D. degree in geomatics at the Politecnico di Milano University, Milan. Since 2004, he has been with the Remote Sensing Laboratory of the Ingegneria Idraulica, Ambientale, Infrastrutture Viarie, Rilevamento (DIIAR) (Hydraulics, Environment) Department, Politecnico di Milano University, and his main research activity is in the area of change detection with midresolution satellite data.
Giovanmaria Lechi received the laurea (M.S.) degree in electrical engineering from the Politecnico di Milano Technical University, Milano, Italy, in 1970. He began his career in 1970 with the Italian National Research Council, attending in remote sensing. He has been appointed director of the Lithosphere Geophysics Dept. of the Italian National Research Council (CNR). From 1987 he is Associate Professor of Remote Sensing at the Politecnico di Milano Technical University, Milano, Italy. He is author (or coauthor) of more than 100 scientific publications, including books, book chapters, journals and conference proceedings. Mr. Lechi he was appointed a member of the UE Commission for the development of a cadastre for the Italian olive tree plantings in 1975. He is a founding member of the Italian Association for Remote Sensing (AIT), and from 1996 to 2003, he has been also President of the Italian Association for Remote Sensing. He is member and arbitrator of the Italian Topographic and Photogrammetric Society.
