Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 7-Day Trial for You or Your Team.

Learn More →

An algorithm for rapid flood inundation mapping from optical data using a reflectance differencing technique

An algorithm for rapid flood inundation mapping from optical data using a reflectance... Introduction Floods are major natural disasters that affect many regions around the world year after year, causing loss of lives, damaging economies and human health. More than one third of the world's land area is flood prone, affecting about 82% of the world's population (Dilley et al ., ). According to EM‐DAT ( ), about 3 billion people in more than 110 countries are affected by catastrophic flooding. Between 1980 and 2011, about 212 460 deaths were associated with floods worldwide. Destructive floods are common in tropics, particularly in Asia (Kundzewicz et al ., ). About 3881 flood disasters associated with 2651 different rivers were recorded in the Dartmouth Flood Observatory catalogue (DFO, ) between 1985 and 2011. Remote Sensing technologies allow a quick mapping of large regions and may provide valuable support in flood crisis management and effective early warning for disasters. It is used for flood mapping (Jain et al ., ; Sakamoto et al ., ; Xiao et al ., ), flood monitoring (Sakamoto et al ., ; Jain et al ., ), flood damage assessment (Dewan et al ., ), flood forecasting (Patro et al ., ; Van Der Knijff et al ., ), validating numerical inundation models and rainfall run‐off analyses (Machado and Ahmad, ; Patro et al ., ). Flood prediction and flood risk analysis are important components of global flood mitigation efforts (Sakamoto et al ., ; Jain et al ., ; Koutroulis et al ., ). Nonstructural measures (flood forecasting, flood warning, etc.) are equally important for mitigation of flood hazards as structural measures (dams, dikes, etc.) (Jain et al ., ; Sharma et al ., ). Floods are currently monitored systematically only in a few countries (Kugler and De Groeve, ; De Groeve and Riva, ). At a global scale, space charters such as the International Charter Space and Major Disasters ( http://www.disasterscharter.org/ ); Global Monitoring for Environment and Security ( http://www.emergencyresponse.eu/gmes/en/ref/home.html ); Sentinel Asia ( http://dmss.tksc.jaxa.jp/ ) and SERVIR provide emergency response services to access satellite data in the event of natural or man‐made disasters. For any catastrophic flood event, the nominated users of such charters can acquire satellite imagery and related data products free of charge for further analyses. However, this requires significant preprocessing; requests for such products are often delayed due to hindrance in data acquisition, processing, and analysis of flood extent and duration. In order to minimise the response time, there is the need for a generic approach to flood inundation mapping using satellite‐based data. There are many methods that use multiresolution satellite data both spatial and temporal resolution for mapping flood inundation and analysing damage induced by floods, such as visual interpretation of satellite images (Jensen, ), multispectral image classification (Sharma et al ., ), band rationing (Jain et al ., ), contextual multitemporal classification and object‐based classification (Cleve et al ., ). Of these methods, multispectral image classification is the most common. Sheng and Gong ( ) proposed a ratio of Channel 2 (0.725–1.0 mm) to Channel 1 (0.58–0.68 mm) of National Oceanic and Atmospheric Administration (NOAA)/Advanced Very High Resolution Radiometer (AVHRR) data to enable discrimination between water and land surfaces. Xiao et al . ( ) evaluated Normalized Difference Water Index (NDWI) calculated from near‐infrared and shortwave infrared (SWIR) data acquired by a Système Pour l'Observation de la Terre/vegetation sensor. Moderate Resolution Imaging Spectroradiometer (MODIS) data with the band rationing technique were used to detect temporal changes in extent of annual flooding (Sakamoto et al ., ; Xiao et al ., ). Hence for rapid flood mapping products, it is very relevant and important to develop a suitable band rationing technique for detecting surface water changes. The objective of this study is to assess the potential of optical remote sensing data like Earth Observatory‐1 Advanced Land Imager (EO‐1 ALI) and Landsat Thematic Mapper (TM) to identify the extent of flood inundation with a composite index based on spectral differences. To validate such an index, free globally available satellite data sets were used. The algorithm developed for ALI and Landsat images can be applied quickly, once the satellite images are acquired to prepare rapid flood inundation information products. Data used and sensor description The list of satellite images used in the implementation and validation of the algorithm is given in Table . All the satellite images were made available freely from Earth Explorer of the United States Geological Survey (USGS). The NASA New Millennium Program's EO‐1 satellite was successfully launched on 21 November 2000, and the Landsat‐5 TM was launched on 1 March 1984. There are three primary instruments on the EO‐1 spacecraft: the multispectral ALI, the hyperspectral Hyperion sensor and the Linear Etalon Imaging Spectrometer Array Atmospheric Corrector. Table summarises the essential spatial and spectral characteristics of the ALI and Landsat instruments. ALI satellite data are made available for rapid mapping activities at times of major disasters within 4–8 h after its acquisition. It is also used in environmental and natural resource mapping. The ALI is a 10‐band multispectral system with multiple linear arrays embedded in a single sensor chip assembly (Pearlman et al ., ). These bands have been designed to mimic six Landsat bands with three additional bands covering 433–453 nm, 845–895 nm and 1200–1300 nm (Table ). List of satellite data used in index development and web link for individual flood description Sl Sensors Location Country Date Flood event Flood site description Cloud cover 1 Zambezi River in Southern Africa South Africa 2009‐04‐19 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=38282 0 to 9% 2 Flooding reaches Vicksburg, Mississippi USA 2011‐05‐10 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=50568 20% to 29% 3 Flooding near Hamburg, Iowa USA 2011‐07‐17 During http://earthobservatory.nasa.gov/IOTD/view.php?id=51380 0 to 9% 4 Flooding around Manchhar Lake, Pakistan Pakistan 2010‐09‐18 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=45828 10% to 19% 5 Flooding on the Nzoia River, Western Kenya Kenya 2008‐11‐13 During http://earthobservatory.nasa.gov/IOTD/view.php?id=35977 10% to 19% 6 EO‐1 ALI Flooding near Manaus, Brazil Brazil 2009‐06‐29 During http://earthobservatory.nasa.gov/IOTD/view.php?id=39359 50% to 59% 7 Monsoon Floods in Northeast India Nepal, India 2008‐09‐08 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=20396 60% to 69% 8 Monsoon Floods in Northeast India Nepal, India 2004‐11‐10 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=20394 0 to 9% 9 Flooding along Ganges River, India India 2007‐08‐22 During 40% to 49% 10 Deadly Flooding in Namibia Namibia 2010‐05‐08 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=37711 0 to 9% 11 Flooding near the Betsiboka River, Madagascar Madagascar 2010‐03‐23 During http://earthobservatory.nasa.gov/IOTD/view.php?id=43240 0 to 9% 12 Pakistan 2010‐09‐04 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=45808 0 to 9% 13 Pakistan 2010‐09‐04 During 0 to 9% 14 Pakistan 2010‐02‐24 Before 0 to 9% 15 Landsat TM Flooding in Pakistan Pakistan 2010‐02‐24 Before 0 to 9% 16 Pakistan 2010‐12‐25 After 0 to 9% 17 Pakistan 2010‐12‐25 After 0 to 9% Satellite sensor description for both the E arth O bservatory‐1 A dvanced L and I mager ( ALI ) and L andsat T hematic M apper ( TM ) sensors Parameters ALI Landsat‐5 TM Spectral range 0.4–2.4 μ m 0.4–2.4 μ m Spatial resolution 30 m 30 m Swath width 37 km 185 km Spectral resolution Variable Variable Spectral coverage Discrete Discrete Panchromatic‐band resolution 10 m 15 m Number of bands 10 7 Spectral bands range Pan 0.48–0.69 0.52–0.90 MS‐1' 0.433–0.453 MS‐1 0.45–0.515 0.452–0.518 MS‐2 0.525–0.605 0.528–0.609 MS‐3 0.63–0.69 0.626–0.693 MS‐4 0.775–0.805 0.776–0.904 MS‐4' 0.845–0.89 MS‐5' 1.2–1.3 MS‐5 1.55–1.75 1.567–1.784 MS‐7 2.08–2.35 2.097–2.349 MS‐6 10.45–12.42 MS, multispectral. The ALI has a 30‐m resolution for multispectral pixels and 10‐m resolution for panchromatic pixels. The instrument swath width can represent one 37 km by 100 km land area per image. The TM has six multispectral visible near‐infrared (VNIR) and SWIR bands, one panchromatic band and one thermal band with spatial resolutions of 30 m for six VNIR/SWIR bands, 60 m for one thermal band and 15 m for one panchromatic band. The EO‐1 sensors are ground pointing capable, allowing shorter revisit times than the Landsat's 16‐day cycle. The instrument swath width achieves one 185 km by 185 km land area per image. In this study, I used only the multispectral bands for development of the algorithm. Background for algorithm development To date, more than 40 multispectral remote sensing‐based indices have been developed and used to monitor water and vegetation properties (Yan et al ., ). Among these, Normalised Difference Vegetation Index (NDVI) is the most widely used vegetation index (Tucker, ; Teillet et al ., ); however, it is sensitive to atmospheric aerosols and soil background. In mapping floods, studies have used a combination of several indices, including enhanced vegetation index (EVI), land surface water index and NDWI, to generate results that are better than that obtained by only using NDVI (Xiao et al ., , ; Sakamoto et al ., ; Yan et al ., ). In this study, 13 sets of globally available satellite data showing floods and postflood events were used to implement the flood mapping algorithm (Table ). All the satellite data sets, which are radiometric and geometrically corrected, were provided by the USGS Earth Resources Observation and Science Center. For mapping inundated areas, the following indices were used in the process of algorithm development. N ⁢ D V I = ρ n i r − ρ r e d ρ n i r + ρ r e d where: ρ red and ρ nir represent the surface reflectance values of the red and NIR bands, respectively. Similarly, the NDWI (McFeeters, ) was calculated, based on the green and NIR bands: N ⁢ D W I = ρ g r e e n − ρ n i r ρ g r e e n + ρ n i r where: ρ green and ρ nir are the surface reflectance of the green and NIR bands, respectively, for both the sensors. There are various definitions of NDWI that combine different pairs of bands (normally of ALI, TM or ETM), typically and originally including green and NIR (McFeeters, ), NIR and SWIR (Gao, ; Chen et al ., ), and red band and middle infrared (Xiao et al ., ). Wavelengths for both ALI and Landsat sensors are given in detail in Table . A number of water and vegetation indices have been developed that use the fuzzy logic and criteria‐selection approach in flood inundation mapping (Brakenridge et al ., ; Sakamoto et al ., ; Xiao et al ., ; Islam et al ., ). Recent development of flood indices includes those that were specific for satellite sensors such as NOAA/AVHRR data (Jain et al ., ), MODIS (Sakamoto et al ., ; Islam et al ., ) and Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager (Proud et al ., ). These indices, even though different in formulae, were developed using the reflectance characteristics of the Earth's surface in the red, NIR and SIR, and therefore they are primarily correlated with flood‐related pixels. There were only a few indices available as a simple but practical approach to estimate inundated areas that help in the preparation of rapid response flood detection. N ormalized D ifference S urface W ater I ndex ( NDSWI ) approach Using the capabilities of indices from NDVI and NDWI, an improved algorithm was created using EO‐1 ALI and Landsat TM images. The algorithm is similar to the EVI in terms of band approach but differs largely in the usability of wavelength spectrum. It is referred to as the NDSWI. The first task in the process of indices development was to find a way of mapping inundation and estimating the floods. This was accomplished by using a simple index, the NDSWI, derived from the SWIR (2.08–2.35 μ m), NIR (0.775–0.805 μ m) and green (0.525–0.605 μ m) bands. Calculation of at‐sensor spectral radiance is the fundamental step in converting image data from multiple sensors and platforms into a physically meaningful common radiometric scale. Radiometric calibration of the TM and ALI sensors involves rescaling the raw digital numbers transmitted from the satellite to calibrated digital numbers, which have the same radiometric scaling for all scenes processed on the ground for a specific period. A reduction in scene‐to‐scene variability was achieved by converting the at‐sensor spectral radiance to exoatmospheric Top‐of‐Atmosphere reflectance. Necessary values in calculation of solar exoatmospheric spectral irradiances (ESUN λ ) for the TM and ALI sensors were obtained from Chander et al . ( ) using the Thuillier solar spectrum (Thuillier et al ., ). In areas covered by clouds and shadows, it is more difficult to identify water bodies. As an optical sensor, Landsat and ALI cannot penetrate thick clouds. No method can eliminate cloud contamination under such circumstances. In case of thin clouds, the sensor does receive some signal from the underlying surface, even though mixed with cloud signal. Moreover, the spectral characteristics of water and land are so different that it is possible to distinguish water from land under thin cloud covers. The ratio image is immune to thin cloud contamination, but the original channels are vulnerable. Sheng et al . ( ) demonstrated the effectiveness of the ratio image in reducing cloud effect on water body recognition during floods. Development of NDSWI was based on the relationship between water content in vegetation canopy and characteristics of the water surface as an open water body and soil moisture. Selection of these wavelengths maximises the reflectance properties of water. That is, it (1) maximises the typical reflectance of water features by using green wavelengths; (2) minimises the low reflectance of NIR by water features and (3) maximises the high reflectance of NIR by terrestrial vegetation and soil features. SWIR is highly sensitive to moisture content from both water bodies and water contained in plant grids. Considering the fact that the SWIR and green bands are sensitive to water content, it was reasoned that, as vegetation dries, the SWIR and green band reflectance should increase and therefore the NDSWI would enhance the water‐related pixels distinctly. The corrected reflectance data were then spectrally enhanced through the NDSWI algorithm as follows: N ⁢ D S W I = G * ρ s w i r − ρ n i r ρ s w i r + c 1 ρ n i r − c 2 ρ g r e e n where: ρ swir , ρ nir and ρ green are the reflectance values of ALI (9, 5, 3) and Landsat (6, 4, 2) bands, respectively, G is a given gain factor, and c 1 and c 2 are applied to adjust absorption and backscattering coefficient. Several iterations were performed to adjust absorption and backscattering coefficient to determine surface‐water pixels. The coefficients used in this study were different for both the sensors as ALI has two SWIR bands compared with Landsat, which has a single SWIR band. So, the coefficients of ALI were c 1 = 6.4, c 2 = 6.3 and G = 4.5, while for Landsat the values were c 1 = 9.8, c 2 = 6.3 and G = 4.5. Results and discussion Results indicate that a NDSWI derived from satellite images (Table ) and high‐resolution images obtained from ALOS/PALSAR data showed good agreement and potential to determine inundated areas and water turbidity. Spectral reflectance of surface water bodies, vegetation and bare soil is shown in Figure . The figure depicts the response of the ALI sensor to varying land surface types across all its spectral channels, providing information on the spatial pattern of flood inundation. Spectral reflectance of water and vegetated areas using E arth O bservatory‐1 A dvanced L and I mager ( EO ‐1 ALI ) sensor. NIR, near‐infrared; SWIR, shortwave infrared. The results explain that SWIR is more effective than NIR in detecting inundated areas because of the presence of turbid water. The distinction of water properties between SWIR and NIR has helped in the development of the NDSWI algorithm. Although spectral signatures of water are quite distinct from other land uses like vegetation, built‐up area and soil surface, identification of water pixels at the water/soil interface is very difficult and depends on the interpretative ability of the analyst. Deep water bodies (e.g. reservoirs) have quite distinct and clear representation in the imagery (Figure ). However, very shallow water/turbid water (e.g. flood inundation) can be mistaken for soil, while saturated soil can be mistaken for water pixels. Also, it is possible that a pixel may represent mixed conditions (some parts as water and other parts as soil) only at the soil/water interface. It is difficult to differentiate between suspended sediment and shallow water by focusing only on the tone and colour of a pixel. The development of an algorithm in combination with reflectance could suitably identify inundated pixels and concentration of turbidity. Spatial comparison of indices using E arth O bservatory‐1 A dvanced L and I mager ( EO ‐1 ALI ) for the K osi river embankment in N epal and I ndia. (a) ALI satellite data R ed, G reen, B lue ( RGB ) = shortwave infrared ( SWIR ), near‐infrared ( NIR ), Green ( G ); (b) N ormalized D ifference V egetation I ndex ( NDVI ); (c) N ormalized D ifference W ater I ndex ( NDWI ) and (d) N ormalized D ifference S urface W ater I ndex ( NDSWI ). Spatial distribution of surface‐water pixels identified using the two indices, NDVI and NDWI for the case of Kosi floods of Bihar (India) in 2008, is given in Figure . In the case of NDVI, the values obtained vary in the range between −1.0 and +1.0, with vegetated areas typically having values greater than zero, with negative values indicating clouds and water, positive values near zero indicating bare soil. On the other hand, with NDWI, water features have positive values, while vegetation and soil usually have zero or negative values. However, delineation based only on the surface‐water pixels is not achieved from either of the two indices. By applying a threshold method, both the water and nonwater pixels can be segregated. Results from the NDSWI, which is a reflectance differencing approach (Figure ), were able to determine only the surface‐water pixels. Even though the NDSWI formula has an output of −1.0 to 1.0, the NDSWI process scales this output to 0 to 255, thereby creating a one‐band 8‐bit output image. Additionally, this index successfully masks all cloud and shadow areas that are not eliminated in NDVI and NDWI. Mapped inundated areas were validated using the training samples acquired through ALOS Palsar satellite imagery obtained during the floods in Pakistan – July 2010 (Figure ). Fifty random samples were generated for flood and nonflood pixels for before‐, during‐ and after‐flood maps and an accuracy assessment was generated between the classified map and reference data (Table ). Overall, kappa statistics for flood maps were found to be 0.94, 0.90 and 0.96, respectively to assess the accuracies of the classification maps. Image showing before flooding from G oogle E arth and inundation extent shown in black using A dvanced L and O bservation S atellite ( ALOS ) of P hased A rray type L ‐band S ynthetic A perture R adar ( PALSAR ) top bottom; N ormalized D ifference S urface W ater I ndex ( NDSWI ) applied using L andsat T hematic M apper ( TM ) sensors and principal component ( PC ) analysis showing deep and shallow flood depth for the case of floods in P akistan in 2010. RGB , R ed, G reen, B lue. Indices comparison for different classes and accuracy assessment for the three different flood categories compared between classified map and reference data NDVI NDWI River −0.028 0.611 Land 0.596 0.361 Land‐water 0.156 0.312 Flood 0.007 0.558 Sand 0.159 0.036 NDVI, Normalised Difference Vegetation Index; NDWI, Normalized Difference Water Index. Before floods Individual category accuracies (%) Flood Nonflood Producer User Overall Kappa index 96.08 98.00 95.00 0.94 100 97 Error matrix Classified data Reference data Flood Nonflood Flood 48 2 Nonflood 1 49 During floods Individual category accuracies (%) Producer User Overall Kappa index Flood 94.12 96.00 95.00 0.90 Nonflood 100 98.85 Error matrix Classified data Reference data Flood Nonflood Flood 48 2 Nonflood 3 47 After floods Individual category accuracies (%) Producer User Overall Kappa index Flood 96.15 100.00 98.00 0.96 Nonflood 100 99.21 Error matrix Classified data Reference data Flood Nonflood Flood 48 2 Nonflood 0 50 With the highest level of accuracy achieved in the three different stages of flood maps, the index has been tested for selected cases studies (Figure ) using ALI and TM images which are available free of charge to users of Earth Explorer of the USGS (Table ). N ormalized D ifference S urface W ater I ndex applied using E arth O bservatory‐1 A dvanced L and I mager ( ALI ). (a) flood along the M ississippi river in USA on 10 M ay 2011; (b) monsoon floods along the G anges river in I ndia on 22 A ugust 2007; (c) river embankment along K osi river in N epal and I ndia on 8 S eptember 2008; (d) flooding near the B etsiboka R iver, M adagascar on 23 M arch 2010; (e) deadly flooding in N amibia on 8 M ay 2010. NDSWI and the applicability of flood turbidity Using the NDSWI, it is possible to detect floodwater turbidity (Figure ). The negative values explain the high concentration of turbidity, with the image being depicted in blue. For positive values, the image is seen in light shades of green, showing medium to low concentration of turbidity. It is observed that the increase in turbidity is caused not by mud erosion, but by the high concentration of mud in the inflow of water during floods. However, without ground observation on turbidity measurement, it is difficult to assess the accuracy of the turbidity level of floodwaters. NDSWI will be helpful in assessing the turbidity level and will serve as an indicator of flood depth, which is considered as the most important indicator of flood hazard (Wadge et al ., ; Townsend and Walsh, ; Islam and Sado, ). An increased flood depth is associated with high discharge, which is a determining factor in flood‐induced destruction of life and property. During the floods in Pakistan in July, 2010, the depth of water was as high as 1.5 to 3 m and there was up to half‐a‐meter of mud in some areas (Pakistan Meteorological Department, ). It is very difficult to determine flood depth from remotely sensed imagery, but an indirect method exists to classify a flooded area into different flood‐depth zones. Except for the blue band, all optical bands have very high correlation with the turbidity and sediment concentration of the water. Deeper water has more turbidity than shallower waters because of its high velocity (Islam and Sado, ). Owing to a higher velocity and discharge, the sediment concentration and consequent turbidity in the River Indus is much higher than the other smaller rivers, like the Kabul and Swat. This factor also contributes to deep inundation along the Indus river bank. Interband correlation is a major impediment in analysing multispectral data. Principal component (PC) transformation is performed to overcome this problem. This image‐processing technique makes the bands less correlated and reduces the dimensionality of the original data set (Lillesand and Kiefer, ). To enhance contrast and facilitate flood turbidity levels, a PC transformation has been applied over bands 2, 3, 4, 5 and 7 of the Landsat TM data (4 September 2010) obtained during the floods in Pakistan in 2010 (Figure ). The first three components capture 96.97% of the total variation that was selected for further analysis. The other components were excluded as their signal‐to‐noise ratio was expected to be very high. Kunte and Wagle ( ) and Sanyal and Lu ( ) attempted to classify the depth of water in the Gulf of Kutch and River Ganga in India, and reported that the PC2 band is particularly sensitive to the concentration of suspended sediments and that they could therefore be effectively used for a broad classification of water depth. To enhance the amount of information, different combinations of the three PC bands into RGB (Red, Green, Blue) were tried to create a false colour composite (FCC), and it was found that PC2, PC1 and PC3 (RGB) generate the best FCC. Figure clearly represents at least two turbidity zones in the shades of bright red and maroon. The general trend of tonal variation reveals that highly turbid water exists at the core of the flooded zone, with the sediment concentration of the water decreasing gradually towards the interface between water and land. After referring to Islam and Sado ( ), we have assumed that highly turbid water is associated with high velocity. During a flood, highly turbid water flows at high velocity through the heavily inundated zone. Highly turbid/deep water appears in bright red and shallow water with less turbidity is presumed to appear in maroon tint. The major rivers of the region, namely, the Indus, Jhelhum and Chenab, fall in the red zone, which further confirms our visual comparison with the principal component analysis method (Figure ). Thus, the NDSWI index has good potential in both mapping inundated areas as well as detecting floodwater turbidity, but there is no real validation carried out in the present study. NDSWI in flood damage analysis After identifying surface water extent using the NDSWI approach, flood‐affected areas were identified for the floods in Pakistan in 2010 (Figure ). A simple comparison of the areas of water bodies before, during and after floods can derive the area of flooded regions. The total area of water bodies as well as the flood‐affected area were calculated and are given in Table for each available cloud‐free data. Thus, the water‐related feature is best identified using the NDSWI approach. Flood‐inundated area mapped using E arth O bservatory‐1 A dvanced L and I mager image for three different periods (before, during and after floods) for the case of floods in P akistan in 2010. TM , T hematic M apper. Indices comparison for different classes Sl. Taluk Tehsil area km 2 Before flood (June) During flood (July–August) After flood (December) 1 Bhiria 665.68 2.86 45.62 6.88 2 Dadu 819.50 5.49 185.98 11.16 3 Daulat Pur 1071.74 55.00 215.56 29.92 4 Dokri 802.78 10.39 192.45 26.40 5 Gambat 571.40 5.34 154.93 11.52 6 Garhi Khairo 732.95 0.06 454.50 60.05 7 Garhi Vasin 931.57 6.54 79.08 18.21 8 Hala 740.52 38.85 297.16 26.72 9 Johi 3592.48 66.65 163.72 103.30 10 Kambar Ali Khan 2293.74 13.35 447.51 64.84 11 Kandiaro 917.82 11.96 287.55 30.74 12 Khairpur Nathan Shah 2610.55 4.76 386.70 65.63 13 Khanpur 673.30 11.64 363.70 23.10 14 Kingri 566.68 3.85 87.39 11.57 15 Kotri 3630.83 27.33 339.77 25.78 16 Lakhi 382.95 1.92 57.60 15.86 17 Larkana 514.39 4.41 20.76 10.38 18 Matiari 712.03 7.67 92.85 10.33 19 Mehar 993.08 3.24 173.87 29.58 20 Miro Khan 774.19 0.15 104.57 0.98 21 Moro 760.73 8.45 295.76 27.15 22 Naushahro Feroze 692.48 2.33 53.87 7.46 23 Pano Aqil 899.14 19.43 282.31 27.80 24 Ratodero 590.26 3.04 34.08 9.06 25 Sakrand 874.73 88.93 303.68 35.13 26 Sehwan 2338.74 155.59 294.34 201.15 27 Shahdad Kot 1480.37 2.29 863.70 29.04 28 Shikarpur 575.61 0.10 292.86 26.27 29 Sobho Dero 508.51 2.07 122.09 6.43 30 Sukkur 324.27 5.60 79.39 12.29 31 Usta Mohammad 987.86 0.11 520.96 9.56 32 Warah 1059.03 51.66 392.43 72.49 From Table , it can be inferred that 30–40% of the area is affected by floods in the months of July and August. The flood‐affected areas have been overlaid on the district of the three provinces (Sindh, Punjab and Balochistan) of Pakistan. Then, using these overlaid figures, the flood‐affected districts are shown at administrative level (e.g. taluk). From Table , it can be seen that flood‐affected areas (approx. more than 50% of the taluk area) are more for taluk of Garhi Khairo, Shahdad Kot, Shikarpur, Usta Mohammad and Khanpur. Results explain that the NDSWI approach of flood mapping using satellite data and GIS would help in coordinating the emergency response phase, and will also be an important source of information for decision‐makers. Conclusions The results of this study demonstrate the flood inundation mapping algorithm using reflectance differencing technique for the EO‐1 ALI and Landsat TM sensor. The analysis of NDSWI goes beyond the previous mapping algorithm that is primarily dependent on NDVI or NDWI as the spectral input. The NDSWI‐based approach produces the best results in mapping flood‐inundated areas when compared with the reference satellite data and ALOS/PALSAR images. Availability of ALI sensor during disaster situations and provision of water‐sensitive SWIR bands from a new generation of optical sensors has made it possible to detect and monitor floods rapidly. During the premonsoon season, when there is less wet area, the index produced better results. During the monsoon and postmonsoon seasons, when there is more wet area, the index produced good results. There are certain sources of error that are inherent to optical sensors, such as cloud contamination, topographic efforts and resolution limitation (both spatial and temporal). Development of NDSWI can be a useful input for assessing the progression of floodwaters and the severity of the flood situation. It is clear that the rapid mapping products aim to be quickly available and are easy to interpret by the end users. Thus, indices become a critical and universal tool for evaluating the physical attributes of land and water resources, and provide an opportunity to understand floodplains inundation dynamics. The products serve as valuable tools for the restoration and rehabilitation of rivers, whose environment has been altered by flooding events, and also for future analyses and decision‐making on the efficiency and environmental impact of the structural work. Further it can help in expediting the flood mapping process and timely dissemination of flood maps to the concerned disaster response agencies. Acknowledgements This research has been funded by the International Water Management Institute (IWMI) and the CGIAR Research Program on Climate Change, Agriculture and Food Security. I would like to acknowledge in particular the contributions of Dr Vladimir Smakhtin, Dr Pramod Aggarwal, Mr Rajah Ameer of IWMI and the United States Geological Survey for freely providing all EO‐1 and Landsat data sets. Also thanks to Japan Aerospace Exploration Agency for providing ALOSdata for the 2010 floods in Pakistan. I would like to specially thank to the anonymous reviewers for valuable comments and suggestions, which considerably improved the quality of the paper. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Flood Risk Management Wiley

An algorithm for rapid flood inundation mapping from optical data using a reflectance differencing technique

Amarnath, G.
Journal of Flood Risk Management , Volume 7 (3) – Sep 1, 2014
Download PDF
12 pages

Loading next page...
 
/lp/wiley/an-algorithm-for-rapid-flood-inundation-mapping-from-optical-data-pfP300sSTY

References (38)

Publisher
Wiley
Copyright
Copyright © 2014 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
eISSN
1753-318X
DOI
10.1111/jfr3.12045
Publisher site
See Article on Publisher Site

Abstract

Introduction Floods are major natural disasters that affect many regions around the world year after year, causing loss of lives, damaging economies and human health. More than one third of the world's land area is flood prone, affecting about 82% of the world's population (Dilley et al ., ). According to EM‐DAT ( ), about 3 billion people in more than 110 countries are affected by catastrophic flooding. Between 1980 and 2011, about 212 460 deaths were associated with floods worldwide. Destructive floods are common in tropics, particularly in Asia (Kundzewicz et al ., ). About 3881 flood disasters associated with 2651 different rivers were recorded in the Dartmouth Flood Observatory catalogue (DFO, ) between 1985 and 2011. Remote Sensing technologies allow a quick mapping of large regions and may provide valuable support in flood crisis management and effective early warning for disasters. It is used for flood mapping (Jain et al ., ; Sakamoto et al ., ; Xiao et al ., ), flood monitoring (Sakamoto et al ., ; Jain et al ., ), flood damage assessment (Dewan et al ., ), flood forecasting (Patro et al ., ; Van Der Knijff et al ., ), validating numerical inundation models and rainfall run‐off analyses (Machado and Ahmad, ; Patro et al ., ). Flood prediction and flood risk analysis are important components of global flood mitigation efforts (Sakamoto et al ., ; Jain et al ., ; Koutroulis et al ., ). Nonstructural measures (flood forecasting, flood warning, etc.) are equally important for mitigation of flood hazards as structural measures (dams, dikes, etc.) (Jain et al ., ; Sharma et al ., ). Floods are currently monitored systematically only in a few countries (Kugler and De Groeve, ; De Groeve and Riva, ). At a global scale, space charters such as the International Charter Space and Major Disasters ( http://www.disasterscharter.org/ ); Global Monitoring for Environment and Security ( http://www.emergencyresponse.eu/gmes/en/ref/home.html ); Sentinel Asia ( http://dmss.tksc.jaxa.jp/ ) and SERVIR provide emergency response services to access satellite data in the event of natural or man‐made disasters. For any catastrophic flood event, the nominated users of such charters can acquire satellite imagery and related data products free of charge for further analyses. However, this requires significant preprocessing; requests for such products are often delayed due to hindrance in data acquisition, processing, and analysis of flood extent and duration. In order to minimise the response time, there is the need for a generic approach to flood inundation mapping using satellite‐based data. There are many methods that use multiresolution satellite data both spatial and temporal resolution for mapping flood inundation and analysing damage induced by floods, such as visual interpretation of satellite images (Jensen, ), multispectral image classification (Sharma et al ., ), band rationing (Jain et al ., ), contextual multitemporal classification and object‐based classification (Cleve et al ., ). Of these methods, multispectral image classification is the most common. Sheng and Gong ( ) proposed a ratio of Channel 2 (0.725–1.0 mm) to Channel 1 (0.58–0.68 mm) of National Oceanic and Atmospheric Administration (NOAA)/Advanced Very High Resolution Radiometer (AVHRR) data to enable discrimination between water and land surfaces. Xiao et al . ( ) evaluated Normalized Difference Water Index (NDWI) calculated from near‐infrared and shortwave infrared (SWIR) data acquired by a Système Pour l'Observation de la Terre/vegetation sensor. Moderate Resolution Imaging Spectroradiometer (MODIS) data with the band rationing technique were used to detect temporal changes in extent of annual flooding (Sakamoto et al ., ; Xiao et al ., ). Hence for rapid flood mapping products, it is very relevant and important to develop a suitable band rationing technique for detecting surface water changes. The objective of this study is to assess the potential of optical remote sensing data like Earth Observatory‐1 Advanced Land Imager (EO‐1 ALI) and Landsat Thematic Mapper (TM) to identify the extent of flood inundation with a composite index based on spectral differences. To validate such an index, free globally available satellite data sets were used. The algorithm developed for ALI and Landsat images can be applied quickly, once the satellite images are acquired to prepare rapid flood inundation information products. Data used and sensor description The list of satellite images used in the implementation and validation of the algorithm is given in Table . All the satellite images were made available freely from Earth Explorer of the United States Geological Survey (USGS). The NASA New Millennium Program's EO‐1 satellite was successfully launched on 21 November 2000, and the Landsat‐5 TM was launched on 1 March 1984. There are three primary instruments on the EO‐1 spacecraft: the multispectral ALI, the hyperspectral Hyperion sensor and the Linear Etalon Imaging Spectrometer Array Atmospheric Corrector. Table summarises the essential spatial and spectral characteristics of the ALI and Landsat instruments. ALI satellite data are made available for rapid mapping activities at times of major disasters within 4–8 h after its acquisition. It is also used in environmental and natural resource mapping. The ALI is a 10‐band multispectral system with multiple linear arrays embedded in a single sensor chip assembly (Pearlman et al ., ). These bands have been designed to mimic six Landsat bands with three additional bands covering 433–453 nm, 845–895 nm and 1200–1300 nm (Table ). List of satellite data used in index development and web link for individual flood description Sl Sensors Location Country Date Flood event Flood site description Cloud cover 1 Zambezi River in Southern Africa South Africa 2009‐04‐19 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=38282 0 to 9% 2 Flooding reaches Vicksburg, Mississippi USA 2011‐05‐10 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=50568 20% to 29% 3 Flooding near Hamburg, Iowa USA 2011‐07‐17 During http://earthobservatory.nasa.gov/IOTD/view.php?id=51380 0 to 9% 4 Flooding around Manchhar Lake, Pakistan Pakistan 2010‐09‐18 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=45828 10% to 19% 5 Flooding on the Nzoia River, Western Kenya Kenya 2008‐11‐13 During http://earthobservatory.nasa.gov/IOTD/view.php?id=35977 10% to 19% 6 EO‐1 ALI Flooding near Manaus, Brazil Brazil 2009‐06‐29 During http://earthobservatory.nasa.gov/IOTD/view.php?id=39359 50% to 59% 7 Monsoon Floods in Northeast India Nepal, India 2008‐09‐08 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=20396 60% to 69% 8 Monsoon Floods in Northeast India Nepal, India 2004‐11‐10 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=20394 0 to 9% 9 Flooding along Ganges River, India India 2007‐08‐22 During 40% to 49% 10 Deadly Flooding in Namibia Namibia 2010‐05‐08 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=37711 0 to 9% 11 Flooding near the Betsiboka River, Madagascar Madagascar 2010‐03‐23 During http://earthobservatory.nasa.gov/IOTD/view.php?id=43240 0 to 9% 12 Pakistan 2010‐09‐04 During http://earthobservatory.nasa.gov/NaturalHazards/view.php?id=45808 0 to 9% 13 Pakistan 2010‐09‐04 During 0 to 9% 14 Pakistan 2010‐02‐24 Before 0 to 9% 15 Landsat TM Flooding in Pakistan Pakistan 2010‐02‐24 Before 0 to 9% 16 Pakistan 2010‐12‐25 After 0 to 9% 17 Pakistan 2010‐12‐25 After 0 to 9% Satellite sensor description for both the E arth O bservatory‐1 A dvanced L and I mager ( ALI ) and L andsat T hematic M apper ( TM ) sensors Parameters ALI Landsat‐5 TM Spectral range 0.4–2.4 μ m 0.4–2.4 μ m Spatial resolution 30 m 30 m Swath width 37 km 185 km Spectral resolution Variable Variable Spectral coverage Discrete Discrete Panchromatic‐band resolution 10 m 15 m Number of bands 10 7 Spectral bands range Pan 0.48–0.69 0.52–0.90 MS‐1' 0.433–0.453 MS‐1 0.45–0.515 0.452–0.518 MS‐2 0.525–0.605 0.528–0.609 MS‐3 0.63–0.69 0.626–0.693 MS‐4 0.775–0.805 0.776–0.904 MS‐4' 0.845–0.89 MS‐5' 1.2–1.3 MS‐5 1.55–1.75 1.567–1.784 MS‐7 2.08–2.35 2.097–2.349 MS‐6 10.45–12.42 MS, multispectral. The ALI has a 30‐m resolution for multispectral pixels and 10‐m resolution for panchromatic pixels. The instrument swath width can represent one 37 km by 100 km land area per image. The TM has six multispectral visible near‐infrared (VNIR) and SWIR bands, one panchromatic band and one thermal band with spatial resolutions of 30 m for six VNIR/SWIR bands, 60 m for one thermal band and 15 m for one panchromatic band. The EO‐1 sensors are ground pointing capable, allowing shorter revisit times than the Landsat's 16‐day cycle. The instrument swath width achieves one 185 km by 185 km land area per image. In this study, I used only the multispectral bands for development of the algorithm. Background for algorithm development To date, more than 40 multispectral remote sensing‐based indices have been developed and used to monitor water and vegetation properties (Yan et al ., ). Among these, Normalised Difference Vegetation Index (NDVI) is the most widely used vegetation index (Tucker, ; Teillet et al ., ); however, it is sensitive to atmospheric aerosols and soil background. In mapping floods, studies have used a combination of several indices, including enhanced vegetation index (EVI), land surface water index and NDWI, to generate results that are better than that obtained by only using NDVI (Xiao et al ., , ; Sakamoto et al ., ; Yan et al ., ). In this study, 13 sets of globally available satellite data showing floods and postflood events were used to implement the flood mapping algorithm (Table ). All the satellite data sets, which are radiometric and geometrically corrected, were provided by the USGS Earth Resources Observation and Science Center. For mapping inundated areas, the following indices were used in the process of algorithm development. N ⁢ D V I = ρ n i r − ρ r e d ρ n i r + ρ r e d where: ρ red and ρ nir represent the surface reflectance values of the red and NIR bands, respectively. Similarly, the NDWI (McFeeters, ) was calculated, based on the green and NIR bands: N ⁢ D W I = ρ g r e e n − ρ n i r ρ g r e e n + ρ n i r where: ρ green and ρ nir are the surface reflectance of the green and NIR bands, respectively, for both the sensors. There are various definitions of NDWI that combine different pairs of bands (normally of ALI, TM or ETM), typically and originally including green and NIR (McFeeters, ), NIR and SWIR (Gao, ; Chen et al ., ), and red band and middle infrared (Xiao et al ., ). Wavelengths for both ALI and Landsat sensors are given in detail in Table . A number of water and vegetation indices have been developed that use the fuzzy logic and criteria‐selection approach in flood inundation mapping (Brakenridge et al ., ; Sakamoto et al ., ; Xiao et al ., ; Islam et al ., ). Recent development of flood indices includes those that were specific for satellite sensors such as NOAA/AVHRR data (Jain et al ., ), MODIS (Sakamoto et al ., ; Islam et al ., ) and Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager (Proud et al ., ). These indices, even though different in formulae, were developed using the reflectance characteristics of the Earth's surface in the red, NIR and SIR, and therefore they are primarily correlated with flood‐related pixels. There were only a few indices available as a simple but practical approach to estimate inundated areas that help in the preparation of rapid response flood detection. N ormalized D ifference S urface W ater I ndex ( NDSWI ) approach Using the capabilities of indices from NDVI and NDWI, an improved algorithm was created using EO‐1 ALI and Landsat TM images. The algorithm is similar to the EVI in terms of band approach but differs largely in the usability of wavelength spectrum. It is referred to as the NDSWI. The first task in the process of indices development was to find a way of mapping inundation and estimating the floods. This was accomplished by using a simple index, the NDSWI, derived from the SWIR (2.08–2.35 μ m), NIR (0.775–0.805 μ m) and green (0.525–0.605 μ m) bands. Calculation of at‐sensor spectral radiance is the fundamental step in converting image data from multiple sensors and platforms into a physically meaningful common radiometric scale. Radiometric calibration of the TM and ALI sensors involves rescaling the raw digital numbers transmitted from the satellite to calibrated digital numbers, which have the same radiometric scaling for all scenes processed on the ground for a specific period. A reduction in scene‐to‐scene variability was achieved by converting the at‐sensor spectral radiance to exoatmospheric Top‐of‐Atmosphere reflectance. Necessary values in calculation of solar exoatmospheric spectral irradiances (ESUN λ ) for the TM and ALI sensors were obtained from Chander et al . ( ) using the Thuillier solar spectrum (Thuillier et al ., ). In areas covered by clouds and shadows, it is more difficult to identify water bodies. As an optical sensor, Landsat and ALI cannot penetrate thick clouds. No method can eliminate cloud contamination under such circumstances. In case of thin clouds, the sensor does receive some signal from the underlying surface, even though mixed with cloud signal. Moreover, the spectral characteristics of water and land are so different that it is possible to distinguish water from land under thin cloud covers. The ratio image is immune to thin cloud contamination, but the original channels are vulnerable. Sheng et al . ( ) demonstrated the effectiveness of the ratio image in reducing cloud effect on water body recognition during floods. Development of NDSWI was based on the relationship between water content in vegetation canopy and characteristics of the water surface as an open water body and soil moisture. Selection of these wavelengths maximises the reflectance properties of water. That is, it (1) maximises the typical reflectance of water features by using green wavelengths; (2) minimises the low reflectance of NIR by water features and (3) maximises the high reflectance of NIR by terrestrial vegetation and soil features. SWIR is highly sensitive to moisture content from both water bodies and water contained in plant grids. Considering the fact that the SWIR and green bands are sensitive to water content, it was reasoned that, as vegetation dries, the SWIR and green band reflectance should increase and therefore the NDSWI would enhance the water‐related pixels distinctly. The corrected reflectance data were then spectrally enhanced through the NDSWI algorithm as follows: N ⁢ D S W I = G * ρ s w i r − ρ n i r ρ s w i r + c 1 ρ n i r − c 2 ρ g r e e n where: ρ swir , ρ nir and ρ green are the reflectance values of ALI (9, 5, 3) and Landsat (6, 4, 2) bands, respectively, G is a given gain factor, and c 1 and c 2 are applied to adjust absorption and backscattering coefficient. Several iterations were performed to adjust absorption and backscattering coefficient to determine surface‐water pixels. The coefficients used in this study were different for both the sensors as ALI has two SWIR bands compared with Landsat, which has a single SWIR band. So, the coefficients of ALI were c 1 = 6.4, c 2 = 6.3 and G = 4.5, while for Landsat the values were c 1 = 9.8, c 2 = 6.3 and G = 4.5. Results and discussion Results indicate that a NDSWI derived from satellite images (Table ) and high‐resolution images obtained from ALOS/PALSAR data showed good agreement and potential to determine inundated areas and water turbidity. Spectral reflectance of surface water bodies, vegetation and bare soil is shown in Figure . The figure depicts the response of the ALI sensor to varying land surface types across all its spectral channels, providing information on the spatial pattern of flood inundation. Spectral reflectance of water and vegetated areas using E arth O bservatory‐1 A dvanced L and I mager ( EO ‐1 ALI ) sensor. NIR, near‐infrared; SWIR, shortwave infrared. The results explain that SWIR is more effective than NIR in detecting inundated areas because of the presence of turbid water. The distinction of water properties between SWIR and NIR has helped in the development of the NDSWI algorithm. Although spectral signatures of water are quite distinct from other land uses like vegetation, built‐up area and soil surface, identification of water pixels at the water/soil interface is very difficult and depends on the interpretative ability of the analyst. Deep water bodies (e.g. reservoirs) have quite distinct and clear representation in the imagery (Figure ). However, very shallow water/turbid water (e.g. flood inundation) can be mistaken for soil, while saturated soil can be mistaken for water pixels. Also, it is possible that a pixel may represent mixed conditions (some parts as water and other parts as soil) only at the soil/water interface. It is difficult to differentiate between suspended sediment and shallow water by focusing only on the tone and colour of a pixel. The development of an algorithm in combination with reflectance could suitably identify inundated pixels and concentration of turbidity. Spatial comparison of indices using E arth O bservatory‐1 A dvanced L and I mager ( EO ‐1 ALI ) for the K osi river embankment in N epal and I ndia. (a) ALI satellite data R ed, G reen, B lue ( RGB ) = shortwave infrared ( SWIR ), near‐infrared ( NIR ), Green ( G ); (b) N ormalized D ifference V egetation I ndex ( NDVI ); (c) N ormalized D ifference W ater I ndex ( NDWI ) and (d) N ormalized D ifference S urface W ater I ndex ( NDSWI ). Spatial distribution of surface‐water pixels identified using the two indices, NDVI and NDWI for the case of Kosi floods of Bihar (India) in 2008, is given in Figure . In the case of NDVI, the values obtained vary in the range between −1.0 and +1.0, with vegetated areas typically having values greater than zero, with negative values indicating clouds and water, positive values near zero indicating bare soil. On the other hand, with NDWI, water features have positive values, while vegetation and soil usually have zero or negative values. However, delineation based only on the surface‐water pixels is not achieved from either of the two indices. By applying a threshold method, both the water and nonwater pixels can be segregated. Results from the NDSWI, which is a reflectance differencing approach (Figure ), were able to determine only the surface‐water pixels. Even though the NDSWI formula has an output of −1.0 to 1.0, the NDSWI process scales this output to 0 to 255, thereby creating a one‐band 8‐bit output image. Additionally, this index successfully masks all cloud and shadow areas that are not eliminated in NDVI and NDWI. Mapped inundated areas were validated using the training samples acquired through ALOS Palsar satellite imagery obtained during the floods in Pakistan – July 2010 (Figure ). Fifty random samples were generated for flood and nonflood pixels for before‐, during‐ and after‐flood maps and an accuracy assessment was generated between the classified map and reference data (Table ). Overall, kappa statistics for flood maps were found to be 0.94, 0.90 and 0.96, respectively to assess the accuracies of the classification maps. Image showing before flooding from G oogle E arth and inundation extent shown in black using A dvanced L and O bservation S atellite ( ALOS ) of P hased A rray type L ‐band S ynthetic A perture R adar ( PALSAR ) top bottom; N ormalized D ifference S urface W ater I ndex ( NDSWI ) applied using L andsat T hematic M apper ( TM ) sensors and principal component ( PC ) analysis showing deep and shallow flood depth for the case of floods in P akistan in 2010. RGB , R ed, G reen, B lue. Indices comparison for different classes and accuracy assessment for the three different flood categories compared between classified map and reference data NDVI NDWI River −0.028 0.611 Land 0.596 0.361 Land‐water 0.156 0.312 Flood 0.007 0.558 Sand 0.159 0.036 NDVI, Normalised Difference Vegetation Index; NDWI, Normalized Difference Water Index. Before floods Individual category accuracies (%) Flood Nonflood Producer User Overall Kappa index 96.08 98.00 95.00 0.94 100 97 Error matrix Classified data Reference data Flood Nonflood Flood 48 2 Nonflood 1 49 During floods Individual category accuracies (%) Producer User Overall Kappa index Flood 94.12 96.00 95.00 0.90 Nonflood 100 98.85 Error matrix Classified data Reference data Flood Nonflood Flood 48 2 Nonflood 3 47 After floods Individual category accuracies (%) Producer User Overall Kappa index Flood 96.15 100.00 98.00 0.96 Nonflood 100 99.21 Error matrix Classified data Reference data Flood Nonflood Flood 48 2 Nonflood 0 50 With the highest level of accuracy achieved in the three different stages of flood maps, the index has been tested for selected cases studies (Figure ) using ALI and TM images which are available free of charge to users of Earth Explorer of the USGS (Table ). N ormalized D ifference S urface W ater I ndex applied using E arth O bservatory‐1 A dvanced L and I mager ( ALI ). (a) flood along the M ississippi river in USA on 10 M ay 2011; (b) monsoon floods along the G anges river in I ndia on 22 A ugust 2007; (c) river embankment along K osi river in N epal and I ndia on 8 S eptember 2008; (d) flooding near the B etsiboka R iver, M adagascar on 23 M arch 2010; (e) deadly flooding in N amibia on 8 M ay 2010. NDSWI and the applicability of flood turbidity Using the NDSWI, it is possible to detect floodwater turbidity (Figure ). The negative values explain the high concentration of turbidity, with the image being depicted in blue. For positive values, the image is seen in light shades of green, showing medium to low concentration of turbidity. It is observed that the increase in turbidity is caused not by mud erosion, but by the high concentration of mud in the inflow of water during floods. However, without ground observation on turbidity measurement, it is difficult to assess the accuracy of the turbidity level of floodwaters. NDSWI will be helpful in assessing the turbidity level and will serve as an indicator of flood depth, which is considered as the most important indicator of flood hazard (Wadge et al ., ; Townsend and Walsh, ; Islam and Sado, ). An increased flood depth is associated with high discharge, which is a determining factor in flood‐induced destruction of life and property. During the floods in Pakistan in July, 2010, the depth of water was as high as 1.5 to 3 m and there was up to half‐a‐meter of mud in some areas (Pakistan Meteorological Department, ). It is very difficult to determine flood depth from remotely sensed imagery, but an indirect method exists to classify a flooded area into different flood‐depth zones. Except for the blue band, all optical bands have very high correlation with the turbidity and sediment concentration of the water. Deeper water has more turbidity than shallower waters because of its high velocity (Islam and Sado, ). Owing to a higher velocity and discharge, the sediment concentration and consequent turbidity in the River Indus is much higher than the other smaller rivers, like the Kabul and Swat. This factor also contributes to deep inundation along the Indus river bank. Interband correlation is a major impediment in analysing multispectral data. Principal component (PC) transformation is performed to overcome this problem. This image‐processing technique makes the bands less correlated and reduces the dimensionality of the original data set (Lillesand and Kiefer, ). To enhance contrast and facilitate flood turbidity levels, a PC transformation has been applied over bands 2, 3, 4, 5 and 7 of the Landsat TM data (4 September 2010) obtained during the floods in Pakistan in 2010 (Figure ). The first three components capture 96.97% of the total variation that was selected for further analysis. The other components were excluded as their signal‐to‐noise ratio was expected to be very high. Kunte and Wagle ( ) and Sanyal and Lu ( ) attempted to classify the depth of water in the Gulf of Kutch and River Ganga in India, and reported that the PC2 band is particularly sensitive to the concentration of suspended sediments and that they could therefore be effectively used for a broad classification of water depth. To enhance the amount of information, different combinations of the three PC bands into RGB (Red, Green, Blue) were tried to create a false colour composite (FCC), and it was found that PC2, PC1 and PC3 (RGB) generate the best FCC. Figure clearly represents at least two turbidity zones in the shades of bright red and maroon. The general trend of tonal variation reveals that highly turbid water exists at the core of the flooded zone, with the sediment concentration of the water decreasing gradually towards the interface between water and land. After referring to Islam and Sado ( ), we have assumed that highly turbid water is associated with high velocity. During a flood, highly turbid water flows at high velocity through the heavily inundated zone. Highly turbid/deep water appears in bright red and shallow water with less turbidity is presumed to appear in maroon tint. The major rivers of the region, namely, the Indus, Jhelhum and Chenab, fall in the red zone, which further confirms our visual comparison with the principal component analysis method (Figure ). Thus, the NDSWI index has good potential in both mapping inundated areas as well as detecting floodwater turbidity, but there is no real validation carried out in the present study. NDSWI in flood damage analysis After identifying surface water extent using the NDSWI approach, flood‐affected areas were identified for the floods in Pakistan in 2010 (Figure ). A simple comparison of the areas of water bodies before, during and after floods can derive the area of flooded regions. The total area of water bodies as well as the flood‐affected area were calculated and are given in Table for each available cloud‐free data. Thus, the water‐related feature is best identified using the NDSWI approach. Flood‐inundated area mapped using E arth O bservatory‐1 A dvanced L and I mager image for three different periods (before, during and after floods) for the case of floods in P akistan in 2010. TM , T hematic M apper. Indices comparison for different classes Sl. Taluk Tehsil area km 2 Before flood (June) During flood (July–August) After flood (December) 1 Bhiria 665.68 2.86 45.62 6.88 2 Dadu 819.50 5.49 185.98 11.16 3 Daulat Pur 1071.74 55.00 215.56 29.92 4 Dokri 802.78 10.39 192.45 26.40 5 Gambat 571.40 5.34 154.93 11.52 6 Garhi Khairo 732.95 0.06 454.50 60.05 7 Garhi Vasin 931.57 6.54 79.08 18.21 8 Hala 740.52 38.85 297.16 26.72 9 Johi 3592.48 66.65 163.72 103.30 10 Kambar Ali Khan 2293.74 13.35 447.51 64.84 11 Kandiaro 917.82 11.96 287.55 30.74 12 Khairpur Nathan Shah 2610.55 4.76 386.70 65.63 13 Khanpur 673.30 11.64 363.70 23.10 14 Kingri 566.68 3.85 87.39 11.57 15 Kotri 3630.83 27.33 339.77 25.78 16 Lakhi 382.95 1.92 57.60 15.86 17 Larkana 514.39 4.41 20.76 10.38 18 Matiari 712.03 7.67 92.85 10.33 19 Mehar 993.08 3.24 173.87 29.58 20 Miro Khan 774.19 0.15 104.57 0.98 21 Moro 760.73 8.45 295.76 27.15 22 Naushahro Feroze 692.48 2.33 53.87 7.46 23 Pano Aqil 899.14 19.43 282.31 27.80 24 Ratodero 590.26 3.04 34.08 9.06 25 Sakrand 874.73 88.93 303.68 35.13 26 Sehwan 2338.74 155.59 294.34 201.15 27 Shahdad Kot 1480.37 2.29 863.70 29.04 28 Shikarpur 575.61 0.10 292.86 26.27 29 Sobho Dero 508.51 2.07 122.09 6.43 30 Sukkur 324.27 5.60 79.39 12.29 31 Usta Mohammad 987.86 0.11 520.96 9.56 32 Warah 1059.03 51.66 392.43 72.49 From Table , it can be inferred that 30–40% of the area is affected by floods in the months of July and August. The flood‐affected areas have been overlaid on the district of the three provinces (Sindh, Punjab and Balochistan) of Pakistan. Then, using these overlaid figures, the flood‐affected districts are shown at administrative level (e.g. taluk). From Table , it can be seen that flood‐affected areas (approx. more than 50% of the taluk area) are more for taluk of Garhi Khairo, Shahdad Kot, Shikarpur, Usta Mohammad and Khanpur. Results explain that the NDSWI approach of flood mapping using satellite data and GIS would help in coordinating the emergency response phase, and will also be an important source of information for decision‐makers. Conclusions The results of this study demonstrate the flood inundation mapping algorithm using reflectance differencing technique for the EO‐1 ALI and Landsat TM sensor. The analysis of NDSWI goes beyond the previous mapping algorithm that is primarily dependent on NDVI or NDWI as the spectral input. The NDSWI‐based approach produces the best results in mapping flood‐inundated areas when compared with the reference satellite data and ALOS/PALSAR images. Availability of ALI sensor during disaster situations and provision of water‐sensitive SWIR bands from a new generation of optical sensors has made it possible to detect and monitor floods rapidly. During the premonsoon season, when there is less wet area, the index produced better results. During the monsoon and postmonsoon seasons, when there is more wet area, the index produced good results. There are certain sources of error that are inherent to optical sensors, such as cloud contamination, topographic efforts and resolution limitation (both spatial and temporal). Development of NDSWI can be a useful input for assessing the progression of floodwaters and the severity of the flood situation. It is clear that the rapid mapping products aim to be quickly available and are easy to interpret by the end users. Thus, indices become a critical and universal tool for evaluating the physical attributes of land and water resources, and provide an opportunity to understand floodplains inundation dynamics. The products serve as valuable tools for the restoration and rehabilitation of rivers, whose environment has been altered by flooding events, and also for future analyses and decision‐making on the efficiency and environmental impact of the structural work. Further it can help in expediting the flood mapping process and timely dissemination of flood maps to the concerned disaster response agencies. Acknowledgements This research has been funded by the International Water Management Institute (IWMI) and the CGIAR Research Program on Climate Change, Agriculture and Food Security. I would like to acknowledge in particular the contributions of Dr Vladimir Smakhtin, Dr Pramod Aggarwal, Mr Rajah Ameer of IWMI and the United States Geological Survey for freely providing all EO‐1 and Landsat data sets. Also thanks to Japan Aerospace Exploration Agency for providing ALOSdata for the 2010 floods in Pakistan. I would like to specially thank to the anonymous reviewers for valuable comments and suggestions, which considerably improved the quality of the paper.

Journal

Journal of Flood Risk Management – Wiley

Published: Sep 1, 2014

Keywords: ; ; ;

There are no references for this article.