Flood hazard assessment in Bangladesh using NOAA AVHRR data with geographical information system

Flood hazard assessment in Bangladesh using NOAAAVHRR data with geographical information system

Md. Monirul Islam and Kimiteru Sado*Department of Civil Engineering, Kitami Institute of Technology, 165 Koen-cho, Kitami 090-8507, Japan

Abstract:Flooded area and ¯ood hazard assessment in Bangladesh were evaluated using remote sensing (RS) data withthe geographical information system (GIS). Image data from the National Oceanographic and Atmospheric

Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR) were used to analyseBangladesh's historical ¯ood event of 1988, which sets a hundred-year record for the inundated areas, withsevere damage occurring throughout this region. Cloudy skies often occur during big ¯oods, and visible andinfrared rays observed by the NOAA AVHRR cannot penetrate cloud cover. We therefore proposed a new

algorithm for classifying cloud covered pixels into water or non-water categories. In order to estimate the¯ooded area accurately, the images taken during the ¯ood and dry season which were classi®ed into water andnon-water categories were superimposed. A new ¯ood hazard map, created from the land cover category and

elevation height data which were ranked by ¯ood a�ected frequency, can be used to evaluate ¯ood risk usingonly RS data and elevation height data. Copyright # 2000 John Wiley & Sons, Ltd.

KEY WORDS NOAA AVHRR data; cloud covered pixels; land cover classi®cations; GIS; ¯ood hazard

assessment; ¯ood hazard map

INTRODUCTION

Flooding occurs very frequently in Bangladesh. The whole area of Bangladesh is an immense river basincrisscrossed by a network of 230 rivers, including three major rivers: the Ganges, Brahmaputra andMeghna.Floods are one of the most devastating environmental hazards in Bangladesh. Normal seasonal ¯oodsincrease the fertility of the surrounding lands and provide the required moisture for crop cultivation, butsevere ¯oods damage the country and result in the untold su�ering of people, who may lose property,livestock, crops or even their lives. The ¯ood of 1988 set a hundred-year record and surpassed the previous¯ood extent record of 1987 (Rasid and Pramanik, 1990). Four types of ¯ood are generally recognized inBangladesh: (1) tidal ¯ood; coastal area ¯ooding by tidal and storm surges, (2) rainfall ¯ood; excessiverainfall with high stage in outfall rivers, (3) river ¯ood; overbank ¯ooding caused by river discharges, and(4) ¯ash ¯ood; a sharp rise followed by a relatively rapid recession in the northeastern and eastern hill basinof Bangladesh. All four types occurred in Bangladesh in 1988 (Rahman, 1996; Bangladesh Government andUNDP, 1989). Bangladesh is surrounded on three sides by India, which lies to the east, north, and west, andshares a small part of its southeastern border with Myanmar. To the south lies the Bay of Bengal. Thecountry contains the lowest riparian zones of the Ganges, Brahmaputra andMeghna rivers. The runo� fromtheir combined 1.72 million km2 catchment area passes through Bangladesh, even though only 7% of thisarea lies within Bangladesh borders. The unprecedented ¯ood of 1988 was caused by an unusual

CCC 0885±6087/2000/030605±16$17�50 Received 26 November 1998Copyright # 2000 John Wiley & Sons, Ltd. Accepted 25 February 1999

HYDROLOGICAL PROCESSESHydrol. Process. 14, 605±620 (2000)

*Correspondence to: Dr K. Sado, Department of Civil Engineering, Kitami Institute of Technology, 165 Koen-cho, Kitami, Hokkaido090-8507, Japan. E-mail: [email protected]

synchronization of the peak ¯ood ¯ow of the Ganges (72 300 m3 sÿ1), Brahmaputra (99 500 m3 sÿ1) andMeghna (19 000 m3 sÿ1) (Matin and Hussain, 1988; Rasid and Pramanik, 1990). The combined peak ¯ow of190 000 m3 sÿ1 greatly exceeded the capacity of the rivers' bank full ¯ow limit, which was estimated to be75 000 m3 sÿ1 (Rasid and Paul, 1987). The country's ¯ooding conditions are in¯uenced by the hydrologicalcondition of the three rivers and their upper catchment areas which lie in India and the Himalayas.

Bangladesh is a very densely populated country with a current population of nearly 120 million. Seventypercent of its citizens are directly or indirectly involved with agriculture. More and more land is beingcultivated as the population increases, and low lying areas and even river beds are being used for agriculturalpurposes during the dry season. Poor people are therefore forced to live on cultivated, ¯ood-prone lands, andbecause farmers must stay near their ®elds, they build their houses in these highly susceptible areas.Bangladesh is fully dependent on agricultural goods, so the Government of Bangladesh is very conscious ofthe ¯ood problem.

The most popular method of ¯ood mitigation in Bangladesh is the construction of embankments or levees.These structural measures provide protection up to a certain level and the failure of structural measures werereported during the ¯ood. A ¯ood forecasting and warning system began in 1972 (Ministry of ForeignA�airs, Japan, 1989), and the system has subsequently been expanded and enhanced but is still insu�cient.After another big event of the 1998 ¯ood, experts from di�erent ®elds recommended the need for ¯oodhazard and ¯ood maps for the historical events, because of its importance for developing an e�ective ¯oodmanagement system (Nazneen and Yasmin, 1998; Nishat, 1998). Although the technical expertise forreduction of ¯ood levels is available, the adequate resources and appropriate institutional set-up to deal withall levels of the hazard are still lacking in Bangladesh. Therefore we think that our derived hazard map willprovide helpful information.

The historic ¯ood of 1988 was monitored by Space Research and Remote Sensing Organization(SPARRSO) in Bangladesh using NOAA AVHRR data transmitted from the NOAA series of US meteoro-logical satellites (Rasid and Pramanik, 1990). Flooded areas were only visually interpreted by SPARRSOand some visual interpretation results for ¯ood damage in 1988 were published (Rasid and Pramanik, 1990;Rahman et al., 1991). Some results were also published in a local journal (Choudhury, 1989) and a report(Bangladesh Government and UNDP, 1989). This study presents some techniques for estimating the extentof ¯ood damaged areas and ¯ood hazard assessment using NOAA AVHRR data with GIS. Satellite imagerywith digital elevation data were appropriate for this study because ¯oods in the basin of the Ganges,Brahmaputra and Meghna rivers are characterized by extremely large magnitudes, high frequencies, longduration and extensive devastation.

DATA ACQUISITION AND PREPARATION

SPARRSO in Bangladesh obtained NOAA data to investigate the ¯oods and associated hydrologicalconditions that occurred in Bangladesh in 1984, 1987, and 1988. The main data sources for this study wereNOAA AVHRR images taken during the ¯ood of 1988, land use information and an elevation model. Thefull scenes of NOAA AVHRR covering the whole area of Bangladesh and parts of neighbouring countrieswere considered for this study. Vector data for the boundary of Bangladesh was prepared using an existinggeographic map. After geometric correction, sub-scenes corresponding to the whole area of Bangladesh wereextracted from the full scenes using a vector layer for the country. Geometric correction was carried out untilthe root mean square errors resulted in less than one pixel. Both the length of one pixel and of one line forNOAA AVHRR images represented 1.1 km on the ground surface. This is the ground resolution for NOAAAVHRR data.

One NOAA AVHRR image for the dry season (20 January 1988) and three images for the ¯ood (18 and24 September and 8 October 1988) were analysed. These four images were taken under di�erent vegetation,land cover and water content conditions. The dry season image was free of clouds, and the land covercategory and Normalized Di�erence Vegetation Index (NDVI) were calculated from this image. Land cover

Copyright # 2000 John Wiley & Sons, Ltd. Hydrol. Process., Vol. 14, 605±620 (2000)

606 M. M. ISLAM AND K. SADO

categories were estimated on the basis of land use maps produced by SPARRSO and ERIM (EnvironmentalResearch Institute of Michigan). Digital elevation data were obtained from Surface Water Modelling Centre,Bangladesh. The mesh size of these data is 300 m� 300 m and elevation height is of the order of millimetres.

LAND COVER CLASSIFICATION

Land cover classi®cation of dry season image

Land cover mapping is one of the most important and typical applications of remote sensing data. Landcover corresponds to the physical conditions of the ground surface while land use re¯ects human activitiessuch as the use of the land for their own purposes. Land cover classi®cation was carried out using supervisedland cover classi®cation (parallelpiped with maximum likelihood as overlap and unclassi®ed rule). Trainingareas were adopted through visual interpretation of the dry season image by referring to existing land usemaps produced by SPARRSO in association with ERIM using Landsat TM data. There are generally somedi�erences due to spatial resolution between derived land cover classi®cation by Landsat and NOAA data(Portmann, 1997). We used the results of land cover classi®cation by NOAA AVHRR data as one of the GIScomponents of our study. It is easy to obtain the image with less cloud amount from the NOAA satellitebecause of its short return period, and NOAA data is very useful for ¯ood monitoring over a large area dueto its wide swath width compared with other satellites (Ochi et al., 1991; Sado and Islam, 1997). The landcover classi®cation results obtained through NOAA AVHRR image data of 20 January 1988, are shown inTable 1.

Consideration of cloud covered pixels

Cloudy skies are expected during big ¯oods, and NOAA AVHRR can't observe a cloud covered groundsurface, so the presence of clouds over the damaged areas after an event limits the usefulness of this data anddi�culty arises with the interpretation of whether a given area beneath the clouds is dry or covered by water.We employed low cloud covered images (cloud coverage for the images taken 18 and 24 September and8 October were 15.00%, 15.40% and 5.51%, respectively) and an algorithm was developed to interpret thesecloud covered pixels. The ¯owchart of the algorithm is shown in Figure 1.

After di�erentiating all of the pixels into three categories (land, water and cloud) using land coverclassi®cation, this algorithm was used to estimate what lay beneath the cloud covered pixels. The shadow of acloud was included with the cloud covered area. Cloud covered pixels for the ¯ood images which hadrepresented water in the dry season image were interpreted as water, but the remaining cloud covered pixels

Table I. Major land cover classi®cation

Category Occupied pixels Area (%)

1. Cultivated land with scattered settlements(mixed aus, deep water & transplanted aman, jute, rabi crops)

26,363 22.20

2. Boro rice ®eld (low level crops by surface water) 18,791 15.833. Cultivated low land with scattered settlements

(aus, transplanted aman-fallow/rabi crops)22,166 18.67

4. Dry fallows area 12,153 10.245. Mixed cropped area with scattered settlements

(aus/jute-transplanted aman-rabi crops on ridge)12,392 10.44

6. Mangrove area 4,791 4.037. High land with mixed forest 4,432 3.738. High land with scattered settlements 8,137 6.859. Saline area, cultivated 993 0.8410. Water in dry season (20 January, 1988) 8,518 7.17

Total 118,736 100.00


FLOOD HAZARD ASSESSMENT 607

Figure 1. Flowchart for recovery of cloud covered pixels



were divided into two types through the use of digital elevation data. A ¯occulus of cloud covered pixels,which was surrounded by non-water pixels was considered to be non-water, while cloud covered ¯occulussurrounded by water pixels was compared with the neighboring water pixels. If the elevation of a cloudcovered pixel of interest (candidate pixel) was less than the land surface elevation of a neighboring waterpixel plus water depth, then the candidate pixel was considered to be a water pixel (it was assumed that the¯ow of water from one pixel to another pixel occurs across the common perimeter only), otherwise it wasconsidered to be a non-water pixel. For cloud covered pixels initially determined to be non-water in the same¯occulus, however, these procedures were repeated to compare them with neighboring water pixels in alldirections until the decision was ®nalized for the ¯occulus. On the other hand, pixels that were initiallydetermined to be water were not allowed to be changed to non-water pixel status. This procedure was usedfor all the ¯occuli in the image. Digital elevation and water depth data are necessary to apply this procedure.

After the recovery of cloud covered pixels using above procedure, water and non-water areas wereincreased by 1.74% and 13.26%, 6.06% and 9.34% and 0.75% and 4.76% for the images of 18 September,24 September and 8 October 1988, respectively. We could not use the water depth in this study becausewater depth data were not available. Results would be more accurate if the water depth could also be used.Some studies and mathematical modeling have been done in Bangladesh to estimate ¯ood water depth(Paudyal, 1996).

ESTIMATION OF FLOODED AREA FOR THE 1988 FLOOD IN BANGLADESH

NOAAAVHRR data is very useful for monitoring large surface phenomena, such as ¯oods, which involve abasic analysis of the spectral characteristics of major land cover types (Tappan et al., 1983). Investigation of¯ood extent using NOAA data has been done in some studies of Bangladesh (Rasid and Pramanik, 1990;Rahman et al., 1991; Islam and Sado, 1998a, 1998b) and some other countries (Barton and Bathols, 1989;Gale and Bainbridge, 1990). From these studies it is realized that the NOAA data is more applicable, becauseof the short return period and the wide coverage area, to monitor major ¯oods occurring in areas of poordrainage such as ¯ood plains with low gradients, slower runo� rates, longer periods of inundation andpotentially large covered areas.

The ¯ood of 1988, which covered most of the ¯ood plain areas, lasted for a long period. The ¯oodoccurred due to heavy rainfall in Bangladesh and outside the country. The ¯ood in the Brahmaputra Valleyin the upper catchment areas of India was particularly severe (Bhattacharyya, 1997), and had a tremendousa�ect on the lower catchment area of Bangladesh. Flash water that came from the northeastern foothillsoutside the country and heavy rainfall inside the country caused ¯ash ¯ooding in the northeastern part ofBangladesh. The Ganges and the Meghna also received monsoon rainfall water from their upper catchmentareas, which are comparatively lower than that of the Brahmaputra. Flooding started from the northern partof the country and moved south, with heavy rainfall inundating the ¯ood plain areas along the river due tothe synchronized peak ¯ow of the three major rivers, which set a new hundred-year record.

Di�culty arises with the interpretation of the ¯ood extent from a single image for long duration ¯oods.Normally ¯oods do not occur at the same time in the three major river basins of Bangladesh, although fordamage assessment and design purposes, maximum inundated areas are needed, and that information maybe gathered from the images acquired closer to peak ¯ooding. We used only data from three images availableat the time of ¯ooding due to the lack of images with low cloud coverage. In addition, because this eventoccurred ten years ago, there are no more data available at SPARRSO in Bangladesh or NOAA in USA. Onthe other hand, a Japanese inquiry commission investigated the severe ¯ood disaster of 1987 from the aspectof meteorology, geomorphology, hydrology, river engineering and sociology in order to be able to contributeto the planning of structural and non-structural countermeasures (Muramoto, 1988; Oya, 1993).

Four RS image data from 18 and 24 September, 8 October and 20 January 1988 were used to estimate the¯ood e�ects. The ®rst one displayed comparatively large inundated area because it was fairly close to thecombined peak ¯ow of the three rivers and the last one showed water ¯ow in the dry season. Each image for a



particular date during the ¯ood permitted mapping of all water bodies present at the time of dataacquisition, so the ®rst three images represent the ¯ood extent for their respective times. The e�ect of rainfall,river, ¯ash and tidal ¯ood appeared in each of the ®rst three images. Each image yields 118 736 pixels on thedisplay monitor and each pixel covers 1.1 km� 1.1 km on ground surface after geometric correction.

Superimposing the ¯ood and dry season images

To di�erentiate between water and non-water, unsupervised land cover classi®cation (ISO-DATAclustering) and supervised land cover classi®cation were performed for the four images. Initially all imageswere categorized into sixty classes by both supervised and unsupervised classi®cation, then the sixtycategories were divided into three categories: land, water and cloud, and ®nally two categories: water andnon-water, after the interpretation of cloud covered pixels. To estimate the inundated area and the extent ofdamage, a model was constructed using a spatial analysis tool of ERDAS Imagine software (MODELMAKER). A ¯owchart of the model is shown in Figure 2. The model has two parts. In the ®rst part, land,¯ooded and normal water areas were di�erentiated and in the second part, ¯ood hazard areas were computedusing the land cover categories, NDVI and elevation data with the output from the ®rst part. Existing waterbodies in the dry season (rivers, lakes, ponds, etc.) are considered as normal water. In Figure 2, input data areclassi®ed images for water and non-water. The ®rst function of the calculation is a recoding function todistinguish non-water from water pixels by assigning 1 for non-water and 2 for water. Output ®les arerecorded ®les for water and non-water. The second function which is a conditional statement composed of

Figure 2. Flowchart for the model to estimate the ¯ooded area and di�erentiate ¯ood damaged area using GIS



three conditions (Figure 2) was used to identify land, ¯ooded and normal water areas, and the second output®le is a three classes raster image ®le. In the second part of the model a function was used to combine thisthree classes image ®le with NDVI, land cover category and elevation ®les for the assessment of ¯ood hazard.

Inundated and land areas estimated by ISO-DATA clustering for the three days during the ¯ood (18 and24 September and 8 October1988) are shown in Table II. The columns of Table II show the results for theone day of the dry season while the rows show results for the three di�erent days of the ¯ood season. Theelements representing water and non-water areas both during the ¯ood and dry season are considered to benormal water in the dry season and land area during the ¯ood, respectively. An element that represents awater area during the ¯ood but a non-water area during the dry season is considered to be an inundated area,and an element that represents a non-water area during the ¯ood but a water area during the dry season isconsidered to be an error pixel. It was assumed that all river courses remained unchanging from 20 January1988 to the beginning of the 1988 ¯ood. If some e�ective channel transitions occurred during this period thenthe above mentioned three categories could be judged on the basis of the date of the channel transitions.Flooded area percentage can be estimated by

Flooded area percentage � a

a � b� 100% �1�

where a � inundated area, b � land area during the ¯ood. Using this equation, the estimated ¯ooded areafor 18 September, 24 September and 8 October 1988 are 36.63%, 33.38% and 28.21%, respectively. Thesevalues were increased by 1.62%, 5.78% and 0.67%, respectively, due to the consideration of cloud coveredinundated pixels. The ¯ooding condition of the three days are shown in Figure 3. The ¯ood progress for the

Table II. Estimated ¯ooded area percentage for three di�erent dates (Sept. 18, Sept. 24 and Oct. 8, 1988) calculated byISO-DATA clustering

(a) Flooded area (%) for 18 September 1988

20 January 1988

18 Sept. 1988 Water Non-water Total

Water 3.65 34.68 38.33

Non-water 1.67 60.00 61.67

Total 5.32 94.68 100.00

(b) Flooded area (%) for 24 September 1988

20 January 1988


Water 3.76 31.61 35.37

Non-water 1.56 63.07 64.63

Total 5.32 94.68 100.00

(c) Flooded area (%) for 8 October 1988

20 January 1988

8 Oct. 1988 Water Non-water Total

Water 3.61 26.71 30.32

Non-water 1.71 67.97 69.68

Total 5.32 94.68 100.00



three dates are shown in Table III, which was calculated by comparing the three images. In Table III(a), theelement of the ®rst row and ®rst column, 28.79%, shows the water area percentage for both 18 and24 September 1988, which represents the common inundated areas for both times (1 and 2 in Figure 3), while

Figure 3. Schematic concept of ¯ood event frequency analysis by three images (each circle represents the water area for respective dateduring the ¯ood and rectangle represents the whole area of Bangladesh)

Table III. Time series of ¯ood progress area percentage

(a) 24 September 1988 and 18 September 1988

24 Sept. 1988


Water 28.79 9.54 38.33

Non-water 6.58 55.09 61.67

Total 35.37 64.63 100.00

(b) 8 October 1988 and 24 September 1988

8 Oct. 1988


Water 22.59 12.78 35.37

Non-water 7.73 56.90 64.63

Total 30.32 69.68 100.00

(c) 8 October 1988 and 18 September 1988

8 Oct. 1988


Water 23.67 14.66 38.33

Non-water 6.65 55.02 61.67

Total 30.32 69.68 100.00



Plate 1. Flooded area, normal water and land area estimated from ISO-DATAclustering for combined three NOAA images of 18 and 24 September and

8 October 1988

Plate 2. Flood hazard map (El � elevation; LC � land cover category;HR � hazard rank)

Copyrig

ht#

2000JohnWiley

&Sons,Ltd.

Hydrol.Process.,

Vol.14(2000)

the element of the second column and ®rst row, 9.54%, indicates the areas covered by water on 18 Septemberbut non-water on 24 September, representing the ¯ood recovered area during that interval (4 and 5 inFigure 3). The element of the ®rst column and second row, 6.58%, represents newly inundated areas betweenthe interval of the above mentioned dates (3 and 6 in Figure 3). The element of the second column andsecond row, 55.09%, represents non-¯ooded areas for both dates (7 and 8 in Figure 3). Tables III(b) and (c)represent the same time variation results for their respective dates. Time series of newly ¯ooded and ¯oodrecovered areas can therefore be understood from Table III. In the same way, by time to time or day to day,the routine measurement for the ¯ood can be performed and updated by analyzing the RS images for thefollowing days of the ¯ood.

After classifying the data into two categories (water and non-water) the three ¯ood images weresuperimposed to produce an image of the maximum ¯ooded area which consisted of two categories: waterand non-water. This image was superimposed again with the dry season image, which also consisted of twocategories. This ®nal image was classi®ed into three categories: normal water, ¯ooded and non-¯ooded area,by using three basic conditions. These basic conditions and the details of the procedure are shown inFigure 2. The results of ¯ooded areas for di�erent land cover classi®cation are shown in Table IV. Floodedareas were estimated by using ISO-DATA clustering as unsupervised land cover classi®cation and supervisedland cover classi®cation. Parallelpiped, maximum likelihood, Mahalanobis distance and minimum distancewere considered in supervised land cover classi®cation. In parallelpiped, overlap rule and unclassi®ed rulewere considered as maximum likelihood, Mahalanobis distance and minimum distance. Parallelpiped withmaximum likelihood as an overlap rule and unclassi®ed rule shows maximum ¯ooded area, 50.20%. TheMahalanobis distance method shows minimum ¯ooded area of 46.82%. It is shown that the ¯ooded area forthe combined image calculated by di�erent methods produced similar results ranging from 47% to 50%. Thedistribution of ¯ooded area calculated by di�erent methods show similar results along the three main riversbut little di�erences appear in the northwest upper part, southwest lower part, southeast lower part and seashore areas of Bangladesh. The combined ¯ood a�ected area which was estimated by ISO-DATA clusteringis shown in Plate 1. It seems that the high ¯ood risk areas are along the riverbanks of the ¯ood plain andextend from the riverbanks to the cultivated land and pastoral area where farmers are normally cultivatingand building their habitats.

Examining the results of the ¯ooded area

We examined the result of the ¯ooded area for the 24 September image using the results published by Rasidand Pramanik (1990). They estimated the ¯ooded area to be 31% by visual interpretation of the whole image,but they did not consider the cloud-covered water pixels and also did not use the dry season image. The

Table IV. Estimated ¯ooded area percentage by using di�erent methods forcombined three images

Method Flooded area (%)

1. Unsupervised classi®cation ISO-DATA clustering 47.38

2. (a) Supervised classi®cationParallelpipedOverlap rule and Unclassi®ed rule(I) Maximum likelihood 50.20(II) Mahalanobis distance 47.27(III) Minimum distance 49.51

2. (b) Supervised classi®cation(I) Maximum likelihood 49.72(II) Mahalanobis distance 46.82(III) Minimum distance 50.01



equation for ¯ooded area percentage was not described clearly in their paper, and sowe thought they used therevised form of equation (1); adding normal water to both the numerator and the denominator of equation(1). Using the same equation as theirs, our initial estimation for the same image was 30.86% withoutconsideration of cloud covered water pixels, but our ®nal estimation is 36.92% with consideration of cloudcovered water pixels. Comparison between the two ¯ood images shows that the ¯ooded area is nearly the sameexcept for the ¯ooded area under the cloud cover. Regarding the image for 18 September and 8 October, thereare no available published results. Rasid and Pramanik (1990) also estimated the ¯ooded area for the imagewhich was taken on 15 September to be 43% by visual interpretation, but our estimated results for18 September indicate it was 38.26% without consideration of cloud covered water pixels and 40.00% withconsideration of cloud covered water pixels. There are no image results in map form in their publication tocompare images for 15 September. Although we have image data of 15 September, this data was severelydamaged and therefore we could not use it. There is another ¯ooded map based on the image of 15 and24 September, which was published in the report of Bangladesh ¯ood policy study (Bangladesh Governmentand UNDP, 1989). We compared our ®nal estimated ¯ood a�ected map (Plate 1) with this result, andconcluded that the damaged area of the Brahmaputra river basin and the northeastern part are nearly thesame except for small di�erences in the middle and lower parts of Bangladesh. Our estimated results showmore inundated areas than their estimations, because our estimated damage area is based on three imageswhile they used only two. Since there are no available accurate ground truth data and digital analysed resultsregarding the studied three ¯ood images, it is di�cult to clarify how much the accuracy of the ¯ooded areaestimation has been improved in this study. However, this study provides a more reasonable procedure for¯ooded area estimation due to consideration of cloud covered inundated and normal water areas.

FLOOD HAZARD ASSESSMENT

Flood hazard as seen through vegetation cover

NDVI is normally used as a quantitative measure for the estimation of vegetation vigour. Development ofurbanization and conditions of leaf area index (LAI) can be measured by NDVI and they have in¯uence onruno� conditions. Development of urbanization decreases the value of NDVI but increases the direct runo�which increases discharge and decreases the lag time from the midpoint of e�ective rainfall duration to peakdischarge. The decreasing of LAI means the increasing of the direct runo� due to decreasing of interceptionloss. In this study the dry season NOAA AVHRR image (20 January 1988) was used to estimate the NDVI.Land cover classi®cation can be conducted only for the dry season, because Bangladesh faces ¯oods everyyear and normal monsoon ¯oods extend about 30% of its area (Rahman, 1996). Therefore to evaluate theseasonal variation of vegetation in Bangladesh through NDVI is di�cult due to the inundation. In additionwe have no information regarding study of the seasonal variation of vegetation. The equation used for thecalculation of NDVI is as follows:

NDVI � IR ÿ R

IR � R�2�

where IR � DN value of near infrared band and R � DN value of the red band in the visible zone. In thecase of NOAA AVHRR, band-1 and band-2 are the red and infrared band, respectively. Theoretically, thevalue of NDVI ranges from ÿ1 to �1. Four categories were established by Kumar (1998) for land useclassi®cation. We used these categories of NDVI in the present study. These categories are (i) NDVI5 0,(ii) 0 to 0.075, (iii) 0.075 to 0.1 and (iv) NDVI4 0�1. Table V shows the di�erent dense vegetated areasduring the dry season and the inundation percentage of the same area for combined three images. It isunderstood that the areas with low dense vegetation during the dry season were damaged more than the highdense vegetation areas of the dry season. Low dense vegetation areas are normally cultivated low land andcultivated land with scattered settlements. It is therefore recognized that cultivated low land and cultivated



land with scattered settlements were severely damaged compared with high land with scattered settlements.However, inundation has many bene®cial e�ects for the cultivated land such as recharging ground water,increasing the soil moisture, making the land more fertile and cleaning up wastes. After the normal ¯oodfarmers cultivate Boro rice (December to May) in low land (Ministry of Foreign A�airs Japan, 1989). Thenormal ¯oods are expected by the people of Bangladesh.

Flood hazard assessment through land cover categories and elevation height

Derived land cover classi®cation from the dry season image was combined with the ¯ood season imagethrough digital elevation data. Then, ¯ood hazard analysis on the basis of land cover category and elevationheight was performed, which is shown in Table VI. Of course land cover categories re¯ect the vegetationcover which can be measured by NDVI. Table VI represents the ¯ooded area percentage and ¯oodoccurrence rate on the basis of elevation height and land cover category. The description of the categories areshown in detail in Table I. Elevation data are available except for the main three rivers' stream, mangrovearea and some portion of Chittagong hill tract. The elevation of the southeastern part (nearby Chittagonghill tract and Chittagong hill tract) is comparatively higher than the ¯ood plain areas. Elevation extremes forBangladesh are as follows: the lowest point is 0 m along the Bay of Bengal while the highest point is 957 m inReng Tlang. Some lower-lying areas did not look ¯ooded while some higher elevation areas were ¯oodedbecause of ¯ash ¯oods in the northeastern part of the country due to the heavy rainfall inside and outside ofthe country, and also because the Brahmaputra river over¯owed its banks after a huge amount of rainfall inIndian and Himalayan regions. Table VI shows that maximum ¯ooded area (12.23%) corresponds to theelement of category 3 and elevation 0±4 m, then followed by the element of category 2 and elevation 4±8 m,and so on, similarly maximum ¯ood occurrence rate (100.00%) corresponds to the element of category 9 andelevation 12±16 m, then followed by the element of category 9 and elevation 4±8 m, and so on. Hazardranking was considered for the marginal distribution on the basis of total ¯ood a�ected areas for eachcategory and elevation interval. Most ¯ooding (26.43% of ¯ooded areas) occurred at elevation 4 to 8 m, andwas assigned the highest hazard rank of 9, followed by 0 to 4 m. Similar hazard rankings were assigned forthe categories from 9 to 1 with respect to the intensity of the highest to the lowest ¯ood a�ected areas.

Table VI shows that the categories 1, 2 and 3 with the elevation from 0 to 12 m are high risk areas, whichmeans agricultural low lands are ®rst to be inundated. The physiography of Bangladesh consists of twodepression areas: Sylhet trough in the northeast and Faridpur-Khulna trough in the southeast, alluvial ¯atand deltaic irrigated land, and the high elevated land of Chittagong hill, Maduhupur Jungle and BarindTerrace. The depression, alluvial ¯at and delta land include most of the cultivated land of the country whichare specially characterized by poor drainage systems due to the low gradient of the three major rivers andtheir tributaries. Moreover most of the rivers carry a large amount of sediment during the ¯ood which alsoreduces the capacity of the channels. Therefore these cultivated areas can be easily inundated and the ¯oodslast for a long period.

The use of the above described data allows one to estimate ¯ood hazard areas for each land cover categoryand elevation interval, which in turn allows for the estimation of crop and structural damage providedproperty information is available.

Table V. Flood damaged area percentage based on NDVI (Veg.: vegetation, Scatt.: scattered)

Dry season (20 January 1988) During ¯ood (combined three images)

NDVI Label Water Non-water Total

50 No Veg. 58.18 41.82 100.000 � 0�075 Scatt. Veg. 47.06 52.94 100.000�075 � 0�1 Sparse Veg. 24.94 75.06 100.00

40�1 Dense Veg. 13.87 86.13 100.00



Table VI. Flood area percentage and ¯ood occurrence rate for combined three images corresponding to the combination of land cover category andelevation, and their hazard rank (Cat.: land cover category, FP: ¯ood percentage, OR: ¯ood occurrence rate and HR: hazard rank)

Elevation interval (m)

0 � 4 4 � 8 8 � 12 12 � 16 16 � 20 20 � 40 40 � 60 60 � 80 480 No data Total FP HR

Cat. 1 FR 2.99 7.88 10.87 4.86 1.79 0.28 0.02 0.00 0.00 2.72 31.31 9OR 48.71 72.36 78.91 63.31 69.91 20.00 16.67 0.00 0.00 34.84 62.15

Cat. 2 FP 7.55 11.07 2.70 0.43 0.05 0.02 0.00 0.00 0.00 2.50 24.32 8OR 67.27 79.74 67.61 63.33 22.22 11.11 0.00 0.00 0.00 44.00 67.64

Cat. 3 FP 12.23 5.29 1.72 0.79 0.21 0.07 0.00 0.00 0.00 2.54 22.85 7OR 47.66 71.04 71.70 66.04 69.23 37.50 0.00 0.00 0.00 49.56 54.01

Cat. 4 FP 0.09 1.29 1.07 1.47 1.11 2.02 0.05 0.00 0.00 0.95 8.05 6OR 44.44 45.24 46.53 42.48 28.00 25.87 28.57 0.00 0.00 38.53 34.70

Cat. 5 FP 0.29 0.52 0.82 1.04 0.50 2.25 0.70 0.05 0.05 0.59 6.81 5OR 65.00 21.70 29.27 36.22 62.86 26.54 18.24 6.90 50.00 44.07 28.64

Cat. 6 FP 0.39 0.11 0.18 0.11 0.02 0.02 0.00 0.00 0.00 1.11 1.94 4OR 32.08 26.32 36.36 26.32 50.00 25.00 0.00 0.00 0.00 17.31 21.29

Cat. 7 FP 0.00 0.07 0.05 0.09 0.07 0.16 0.05 0.05 0.00 1.16 1.70 3OR 0.00 25.00 25.00 28.57 42.86 30.43 4.00 4.44 0.00 1.05 10.95

Cat. 8 FP 0.02 0.00 0.07 0.27 0.23 0.39 0.36 0.02 0.07 0.18 1.61 2OR 33.33 0.00 75.00 57.14 14.29 28.81 12.90 2.50 27.27 8.51 19.03

Cat. 9 FP 0.89 0.29 0.09 0.03 0.00 0.00 0.00 0.00 0.00 0.11 1.41 1OR 70.91 92.86 80.00 100.00 0.00 0.00 0.00 0.00 0.00 55.56 74.70

Total FP 24.45 26.42 17.57 9.09 3.98 5.21 1.18 0.12 0.12 11.86 100.00

HR 8 9 7 6 4 5 3 2 1 0

Copyrig

ht#

2000JohnWiley

&Sons,Ltd.

Hydrol.Process.,

Vol.14,605±620(2000)

616

M.M.ISLAM

AND

K.SADO

Flood hazard assessment through a ranking matrix

Now it is well recognized in Bangladesh that ¯ood distress is caused not only by peak ¯ood levels but alsoby the duration above danger levels. Bangladesh grows rice: Aman, Aus and Boro on 26.22 million acres(Ministry of Foreign A�airs Japan, 1989). Paddies can't tolerate submergence for a long period of time and¯oods have caused damage to both the paddy seedlings and planted areas in the country. The seeds of rabicrops (vegetables, mustard, wheat, potatoes, etc.) can't be sowed on time in the case of a long duration of¯ood. The most damaging aspect of the ¯oods has been the destruction of peoples' livelihood due tointerruption of transportation and communications, submergence of houses, etc. The ¯ood victims who takerefuge in ¯ood shelter camps want to know how long they have to su�er. Slowly receding ¯ood water doesnot allow them to go back to their houses and does not bring them back to the pre-¯ood situation, whichconsequently creates problems. Therefore, the ¯ood a�ected frequency based on duration was adopted as aparameter to calculate a hazard score, besides this ¯ood depth and ¯ood water velocity could be used as aparameter. It is very di�cult to install the instruments and measure ¯ood water depth and velocity in manyplaces during a severe ¯ood. It is also di�cult to estimate the ¯ood water depth and velocity using NOAAdata, but ¯ood extent and duration can be calculated from NOAA data.

Classi®cation of ¯ood a�ected frequency is shown in Table VII. The three hypotheses can be surmized forthe newly ¯ooded areas on 18 and 24 September and 8 October 1988. The hypotheses concerning thepresence of water in the image taken on 24 September, but its absence on 18 September (3 and 6 in Figure 3),are that either it rained again between 18 and 24 September, or that water came from the previously ¯oodedareas and spread out over the new areas, or that both phenomena occurred simultaneously. The abovehypotheses can be adopted for water presence in the image of 8 October but absence on 24 September (4 and7 in Figure 3) or 18 September (3 and 7 in Figure 3). The hypotheses regarding the presence of commonwater (1 in Figure 3) in the three images are that either ¯ooded water did not drain or that areas ¯oodedagain after recovering between the two intervals of the three images. The common inundated area thatappeared in all three images was considered to be a high hazard area, the common inundated area thatappeared in two images was considered to be a medium hazard area, an inundated area that appeared in asingle image was considered to be a low hazard area and an area not appearing inundated in any image wasconsidered to be a non hazard area. The damage ranking was decided as class 4, class 3, class 2 and class 1 forthe high, medium, low and non hazard area, respectively. Using this database with land cover categories andelevations, two ¯ood hazard maps were prepared for the land cover category and the elevation. However weconsidered the land cover category and elevation had an interactive e�ect on ¯ood risk assessment, and newhazard ranks were decided based on a weighted score for the evaluation of ®nal ¯ood hazard assessment foreach pixel. A weighted score was calculated as follows:

Weighted score � 0�0 � class1 � 1�0 � class2 � 3�0 � class3 � 5�0 � class4 �3�

Table VII. Classi®cation of ¯ood a�ected frequency (W: water, NW: non-water)

18 Sept. 1988 24 Sept. 1988 8 Oct. 1988 Area (%) Class

W W W 20.42 4W W NW 8.37 3NW W W 2.17 3W NW W 3.25 3W NW NW 6.29 2NW W NW 4.40 2NW NW W 4.48 2NW NW NW 50.62 1

Total 100.00



In order to quantify the ¯ood hazard, the four rankings were obtained from the weighted score. The acquiredarea percentage of each class number for the land cover category and elevation interval area shown inTable VIII with their score, point and hazard rank. The higher the score, the higher the risk that ¯ooddamage may occur. Points for each category and elevation were estimated on the basis of linear interpolationbetween 0 and 100. The points 0 and 100 corresponded to the lowest (0) and the highest score. Hazard ranksfrom 1 to 4 were ®xed by the corresponding value of the point, (0) to 25 for hazard rank 1, 25 to 50 for 2, 50 to75 for 3, and 75 to 100 for 4.

Two thematic maps, ¯ood hazard maps created on the basis of land cover classi®cation and elevation, wereused to evaluate ¯ood risk. The ®nal ¯ood hazard map shown in Plate 2 was evaluated by the ranking matrixon a multiplication mode (shown in Plate 2) suggested by Ochi et al. (1991). Zone-0 in Plate 2 indicates thenormal water. Some non-¯ooded cultivated low land areas are designated as high hazard zones due to a highhazard rank for their low elevation and a high hazard rank for category 3. In general, our results agree withRahman et al. (1991) predictions for higher (Sylhet trough, ¯ood plains, etc.) and lower (Chittagong hill,Barind tract, etc.) levels, but for middle zones (Khulna-Faridpur trough, tidal ¯ood plain, etc.) there is noagreement. They used the ¯ood depth that was categorized into three categories and then the classi®ed ¯ooddepths were overlaid onto physiography and geological maps to calculate hazard scores. In this studyclassi®ed ¯ood durations were overlaid onto land cover classi®cations and elevation height maps. The ¯oodextent area has signi®cant in¯uence through the ¯ood depth and the ¯ood duration on the ¯ood hazardscore. This study shows a larger ¯ood extent area compared with Rahman et al. (1991) results. It is becausethree images (18 and 24 September and 8 October 1988) were used in this study for the estimation of the¯ooded area, but they used only one image taken on 8 October 1988, which shows a comparatively smallinundated area. It is considered that the di�erences between the results are mainly a result of the di�erencesof estimated ¯ood extent.

Table VIII. Score and hazard rank for land cover category and elevation (Cat.: land cover category, HR: hazard rank)

(a) For land cover category

Class 1 Class 2 Class 3 Class 4 Score Point HR

Cat. 1 38.00 17.02 22.15 22.83 197.62 79.25 4Cat. 2 32.93 13.91 15.17 37.99 249.37 100.00 4Cat. 3 47.08 17.89 12.91 22.12 167.22 67.06 3Cat. 4 65.30 14.76 11.44 8.50 91.58 36.72 2Cat. 5 71.46 17.72 7.28 3.54 57.26 22.96 1Cat. 6 80.15 11.17 3.72 4.96 47.13 18.90 1Cat. 7 80.97 10.72 5.90 2.41 40.47 16.23 1Cat. 8 89.49 6.28 3.80 0.44 19.88 7.97 1Cat. 9 28.57 26.19 17.86 27.38 216.67 86.89 3

(b) For elevation height

Class 1 Class 2 Class 3 Class 4 Score Point HR

0 � 4 46.67 19.73 12.63 20.97 162.47 64.97 34 � 8 31.14 14.99 17.14 36.73 250.06 100.00 48 � 12 32.33 18.00 22.70 26.97 220.95 88.36 412 � 16 45.00 18.05 18.65 18.29 165.45 66.16 316 � 20 47.92 15.05 22.68 14.35 154.84 61.92 320 � 40 72.11 16.74 8.37 2.68 55.25 22.09 140 � 60 85.60 13.02 1.38 0.00 17.16 6.86 160 � 80 96.52 3.48 0.00 0.00 3.48 1.39 1480 73.33 26.67 0.00 0.00 26.67 10.66 1No data 64.67 8.54 9.12 17.76 124.25 49.69 2



CONCLUSIONS

NOAA data can be used to monitor a catastrophic ¯ood occurring over large areas of poor drainage such as¯ood plains with low gradients and long periods of inundation. However the presence of cloud limits theNOAA data applicability. In this study, we proposed an algorithm to recover the cloud covered pixel for¯ood study, then ¯ood hazard assessment was carried out through NDVI, land cover category and elevationheight. The ¯ooded areas were increased by 1.62%, 5.78% and 0.67% after the recovery of cloud coveredpixels for the images of 18 and 24 September and 8 October 1988, respectively. Flood hazard assessmentthrough NDVI showed that ¯ood magnitude was higher for lower dense vegetated areas of the dry season,which correspond to cultivated low land and cultivated land with scattered settlements where high yieldingvarieties of crops are produced. Using NOAA data with digital elevation data, ¯ood risk assessment wasundertaken on the basis of each land cover category and elevation height interval. The study shows that avast area of Bangladesh was inundated during the ¯ood and the rivers were much wider than during the dryseason. Estimated combined ¯ooded area of 18 and 24 September and 8 October 1988 ranges from 47% to50%, while 68% of the inundated area lies between 0 and 12 m in elevation. Our results also indicate thatland cover categories 1 (cultivated land with scattered settlements), 2 (Boro rice ®eld) and 3 (cultivated lowland with scattered settlements) were highly damaged followed by 4 (dry fallows area), 5 (mixed cropped areawith scattered settlements) and so on. Flood and risk areas were identi®ed and a new ¯ood hazard map wascreated on the basis of the ranking matrix in a multiplication mode which was calculated using the databaseof land cover and elevation. This type of ¯ood hazard assessment can be performed using only RS data andthe digital elevation model.

The result described in this paper demonstrate that this type of study provides helpful information for¯ood control planning. We hope that post-event historical observation can be used to limit future damage byupdating and understanding preventative measures through the application and modi®cation of the presentstudy's results. This study will contribute to ¯ood risk management and help in the prevention of ¯oods inBangladesh.

REFERENCES

Bangladesh Government and UNDP. 1989. Bangladesh Flood Policy Study. Bangladesh Government: Dhaka; 6±22.Barton IJ, Bathols JM. 1989. Monitoring ¯oods with AVHRR. Remote Sensing and the Environment 30: 89±94.Bhattacharyya NN. 1997. Floods of the Brahmaputra river in India.Water International 22: 222±229, J. International Water Resources

Association.Choudhury AM. 1989. Flood Ð 1988 as seen by satellites. Bangladesh Quest 1: 52±54.Gale SJ, Bainbridge S. 1990. The ¯ood in eastern Australia. Nature 345: 767.Islam MM, Sado K. 1998a. Application of remote sensing techniques with a geographic information system to study ¯ood hazard in

Bangladesh. Proceedings of the 27th International Symposium on Remote Sensing of Environment: Tromso, Norway, 745±748.Islam MM, Sado K. 1998b. Assessment of the ¯ooded areas in collaboration with land cover classi®cation by using NOAA AVHRR

data with a digital elevation model Ð a case study in Bangladesh. Proceedings of the 3rd Korea-Japan Bilateral Symposium on WaterResources and Environmental Research: Seoul, Korea, 132±137.

Kumar N. 1998. Vegetation canopy cover in a metropolitan setup: a study of Jawaharlal Nehru University (JNU) and surroundings,South Delhi, India, 1987±1997. Proceedings of the 27th International Symposium on Remote Sensing of Environment: Tromso,Norway, 526±529.

Matin MA, Hussain MA. 1988. Hydrological Aspects of the 1988 Flood. Proceedings of the One Day Seminar on Floods in Bangladesh:Dhaka, Bangladesh, 1±10.

Ministry of Foreign A�airs, Japan. 1989. Report on Survey of Flood Control Planning in Bangladesh. Ministry of Foreign A�airs: Japan.Muramoto Y. 1988. Investigation of the ¯ood disaster caused by heavy rainfall in Bangladesh during the 1987 monsoon season (in

Japanese). Report of Scienti®c Research for Natural Diaster, No. B-62-5, Ministry of Education, Science and Culture, Tokyo.Nazneen S, Yasmin L. 1998. Coping with Floods. Report of Workshop on 1998 ¯ood, Center for Alternatives: Bangladesh.Nishat A. 1998. A discussion on Flood Management in Bangladesh. Special workshop on 1998 ¯ood. Nagorik Durjog-Mokabela

Uddogh: Bangladesh: The 1988 Deluge-Developing Coping Capacities.Ochi S, Rahman NM, Kakiuchi H. 1991. A study on ¯ood risk evaluation in Bangladesh using remote sensing and GIS (in Japanese).

Journal of the Japan Society of Photogrammetry and Remote Sensing 30: 34±38.Oya M. 1993. Fluvial Geography (in Japanese). Kokonshoin Inc.: Tokyo.Paudyal GN. 1996. An integrated GIS-numerical modeling system for advanced ¯ood management. Proceedings of the International

Conference on Water Resources & Environment Research: Towards the 21st Century. Kyoto University: Japan, 555±562.



Portmann TF. 1997. Hydrological runo� modeling by the use of remote sensing data with reference to the 1993±1994 and 1995 ¯oodsin the Rhine River catchment. Hydrological Processes 11: 1377±1392.

Rahman LM. 1996. Present situation and future issues regarding river and hydrological data base in Bangladesh. Proceedings of theSecond Experts Conference on River Information Systems. Sapporo: Japan, 31±38.

Rahman NM, Ochi S, Murai S, Hashimoto T, Kakiuchi H. 1991. Flood risk mapping in Bangladesh Ð ¯ood disaster managementusing remote sensing and GIS. In Application of Remote Sensing in Asia and Oceania Ð Environmental Change Monitoring. AsianAssociation on Remote Sensing: Tokyo, Japan; 263±268.

Rasid H, Paul BK. 1987. Flood problem in Bangladesh: is there an indigenous solution. Environmental Management 11: 155±173.Rasid H, Pramanik M. 1990. Visual interpretation of satellite imagery for monitoring ¯oods in Bangladesh. Environmental

Management 14: 815±821.Sado K, Islam MM. 1997. Satellite remote sensing data analysis for ¯ooded area and weather study: case study of Dhaka city,

Bangladesh. Annual Journal of Hydraulic Engineering, JSCE 14: 945±950.Tappan G, Horvath NC, Doraiswamy PC, Engman T, Goss DW. 1983. Use of NOAA-n satellites for land/water discrimination and

¯ood monitoring. Technical Report No. EW-L3-04394; JSC-18594. NASA: Houston.



Flood hazard assessment in Bangladesh using NOAA AVHRR data with geographical information system

Documents

Transcript of Flood hazard assessment in Bangladesh using NOAA AVHRR data with geographical information system