Rapid response flood detection using the MSG

International Journal of Applied Earth Observation and Geoinformation 13 (2011) 536–544

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and Geoinformation journal homepage: www.elsevier.com/locate/jag

Rapid response flood detection using the MSG geostationary satellite Simon Richard Proud ∗ , Rasmus Fensholt, Laura Vang Rasmussen, Inge Sandholt Department of Geography and Geology, University of Copenhagen, Øster Voldgade 10, DK-1350, Copenhagen, Denmark

a r t i c l e

i n f o

Article history: Received 10 March 2010 Accepted 6 February 2011 Keywords: Meteosat Second Generation Flooding BRDF Anisotropy SEVIRI

a b s t r a c t A novel technique for the detection of flooded land using satellite data is presented. This new method takes advantage of the high temporal resolution of the Spinning Enhanced Visible and InfraRed Imager (SEVIRI) aboard the Meteosat Second Generation (MSG) series of satellites to derive several parameters that describe the sensitivity of land surface reflectivity to variation in solar position throughout the day. Examination of these parameters can then yield information describing the nature of the surface being viewed, including the presence of water due to flooding, on a 3-day basis. An analysis of data gathered during the 2009 flooding events in West Africa shows that the presented method can detect floods of comparable size to the SEVIRI pixel resolution on a short timescale, making it a valuable tool for large scale flood mapping. © 2011 Elsevier B.V. All rights reserved.

1. Introduction Floods are the most frequent of natural disasters, affecting more than one billion people and killing in excess of 100,000 between 2002 and 2010 (EM-DAT, 2011a), with similar statistics being reported for the previous decade (Jonkman, 2005; Alcántara-Ayala, 2002). Despite this, few techniques exist for the rapid detection and monitoring of flooded land. Typically such techniques, where they do exist, are based on local knowledge, news reports and governmental information, such as the EM-DAT and Dartmouth Flood Observatory databases, or require in situ monitoring of water conditions using gauging stations placed at numerous intervals along a river’s course. The latter method produces good quality results but is expensive, which is a particular concern in the developing world, where the majority of people affected by such flooding events live. It is therefore important to develop global techniques for flood detection, particularly as it is predicted that climate change may lead to more frequent and more severe flooding in the future (Kleinen and Petschel-Held, 2007; McGranahan et al., 2007). Limited flood mapping from space has been achieved by examining changes in the Normalised Difference Vegetation or Water Indices (NDVI and NDWI respectively) due to the presence of water on the land surface (Sanyal and Lu, 2004; Jain et al., 2005; McFeeters, 1996). But NDVI is designed to monitor vegetation, and so is unsuitable for flood mapping if very sparse or dense vegetation is present (Beget and Di Bella, 2007). Recently the Global Disaster Alert and Coordination System (GDACS) has implemented a method to allow flood monitoring on a daily or bi-daily basis

∗ Corresponding author. Tel.: +45 35 32 25 84; fax: +45 35 32 25 01. E-mail address: [email protected] (S.R. Proud). 0303-2434/$ – see front matter © 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.jag.2011.02.002

from space using microwave sensors such as NASA’s AMSR-E (De Groeve et al., 2006; Brakenridge et al., 2007). This method provides a spatial resolution of 100 km2 , meaning that small floods are not visible. Countering this, its use of microwave radiation allows water to be visible despite cloud cover – something that is a severe problem when examining the surface within the Visible or Near InfraRed (VNIR) wavelengths. A recent study of flood detection using data from AMSR-E over Namibia produced positive results, with the majority of flood events being correctly identified and mapped (De Groeve, 2010). Additionally, satellite constellations such as COSMO-SkyMed allow for the analysis of flood events at high spatial resolution (Boni et al., 2008; Pierdicca et al., 2010; Hahmann et al., 2008). There has also been much work in integrating satellite measurements into hydrological models, including the measurement of precipitation and soil moisture (Sandholt et al., 2003a; Ottlé and Vidal-Madjar, 1994; Chen et al., 2005), but these typically focus on producing higher quality hydrological models rather than on the detection and mapping of flooded land. It has been shown (Sandholt et al., 2003b) that optical sensors such as AVHRR can be useful in the detection of flooded land, although care must be taken when employing such approaches to minimise the effects of vegetation and other surface features. With the advent of sensors such as the Spinning Enhanced Visible and InfraRed Imager (SEVIRI) aboard Meteosat Second Generation (MSG) that produce data every 15 min (Aminou, 2002), it is now possible to gain cloudfree VNIR observations of land surfaces much more rapidly than before. It has been shown that the land surface can be viewed on multiple occasions on a better than 3-day timescale with SEVIRI (Fensholt et al., 2007). SEVIRI records the top of atmosphere (ToA) reflectance in three VNIR spectral bands that are named channels 1, 2 and 3. These are centred on wavelengths of 635, 810 and 1640 nm respectively, providing a pixel spacing as good as 3 km/pixel.

S.R. Proud et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 536–544

Due to the dynamic and transient nature of flooding events, it is vital to examine them with an instrument capable of rapid data acquisition. The MSG series of satellites fulfils this requirement well, due to its frequent imaging and fixed position relative to the Earth’s surface. The Earth’s surface is not equally reflective under all illumination conditions. Consequently, as the Sun moves across the sky over the course of a day, the reflectance of the land surface will vary – in some cases by more than an order of magnitude (Coulson, 1966; Kriebel, 1978). The type of land cover present on the surface and the wavelength of light being used have a large effect upon the size of this diurnal variation. This study shows that analysing the variation in surface reflectance as a function of the sun’s position allows for examination of the properties of the land and, in particular, the ability to map areas that are flooded at a given time. As an example of this new technique a comparison is made to traditional flood detection methods for the severe floods that occurred during mid-2009 in West Africa. These floods resulted in widespread damage, particularly in the city of Ouagadougou, Burkina Faso, where the flooding affected more than 150,000 people. The severity of these floods makes them useful for testing the SEVIRI flood detection method. 2. Methodology 2.1. Modelling diurnal reflectance trends For most remote sensing applications, the variation in surface reflectance as a function of solar position reduces the accuracy of a data set and must therefore be minimised (Meyer et al., 1995). To facilitate this, various models have been produced that use a Bidirectional Reflectance Distribution Function (BRDF) to describe the reflectance variation as a function of the illumination and viewing conditions. One such model, originally designed for use with the MODerate resolution Imaging Spectroradiometer (MODIS), is known as the MODIS direct broadcast BRDF algorithm (Lucht et al., 2000; Schaaf et al., 2002). It utilises a BRDF that models the reflectance as a series of three mathematical expressions, known as kernels, that each describe a particular scattering method. The relative strength of each of the scattering modes is strongly dependant upon the type of land being observed and hence it is possible to gain information about the land surface by using the BRDF (Gao et al., 2003; Diner et al., 2005). Water, in particular, displays a very distinct set of scattering patterns and these can therefore be used to identify areas that are wholly or partially submerged. This BRDF algorithm combines three kernels in order to calculate the surface reflectance, R, at a wavelength for a given solar zenith angle , viewing zenith angle ϑ and relative azimuth angle : R(, ϑ, , ) = fiso ()Kiso + fvol ()Kvol (, ϑ, ) + fgeo ()Kgeo (, ϑ, )

(1)

The three kernel values are represented by K, with Kiso being equal to 1, whilst Kvol and Kgeo are functions only of the solar and viewing geometry and do not depend upon the land surface reflectance itself. The properties of the land surface are accounted for by the BRDF parameters, fiso , fvol and fgeo , that describe the surface reflectance as a function of the different scattering modes: isotropic, volumetric and geometric. The isotropic parameter measures the reflectance of the surface that is constant, no matter what the geometrical conditions, and can be thought of as the reflectance that would be measured if both the sun and sensor were nadir to the target. The volumetric parameter describes scattering within objects such as tree canopies. Finally, the geometric parameter represents the reflectance that can be modelled by scattering from a series of discrete surface objects, such as buildings. By examining

537

a time series of surface reflectance data and with knowledge of the Sun’s position, it is possible to invert this model and hence derive the values for each of the three parameters. Typically these parameter values are then used to normalise reflectances to a common set of viewing and illumination conditions, enabling the comparison of data gathered in different locations and at different times of year. However, within this study the parameter values themselves are used as the basis for the flood detection method – the actual reflectances are discarded. 2.2. Data used in this study To test the ability of the BRDF parameters to detect flooded land a data source that supplies a large number of images within a short space of time is required. Previously, the BRDF has been calculated for instruments such as MODIS that provide a relatively high spatial resolution but can only collect, at best, one or two images of a particular region each day (Justice et al., 1998). To calculate the BRDF a large number of observations is required, and thus data gathered over many days must be combined in order to successfully retrieve the BRDF parameter values. Typically 8 or 16 days are used to produce one BRDF (Schaaf et al., 2002), which results in a highly accurate set of parameters but due to the long timescale, transient events such as flooding or fire may not be detected. To overcome this, data from the MSG satellites can be used to generate the parameter values on a timelier basis. The high temporal resolution of SEVIRI means that up to 60 sunlit observations of an area can be recorded each day – more than enough to generate the BRDF parameters. However, tests showed that a 3-day acquisition time was required. Using a shorter acquisition time resulted in much of the image being obscured by cloud, whilst a 3-day period allowed the land surface itself to be examined with only a few unprocessed pixels due to cloud cover. This is still a substantial improvement on the MODIS 16 day timescale and should enable most flooding events to be detected. As the SEVIRI pixel resolution is at best 3 km/pixel, localised flooding may not be visible, although major flooding events will still be seen. At worst, no flooding events covering an area of less than 9 km2 will be detected as flooded, although depending upon the land cover for a pixel smaller floods may well be detected. Additionally, the spatial resolution decreases as pixels further from the subsatellite point are examined. In the extremities of Africa, such as Egypt and South Africa, resolution is closer to 5 km/pixel. During normal operation the SEVIRI views an area covering Africa, Europe and the Arabian Peninsula, but for the purposes of this study only data from a portion of this scene, known as the West Africa subset, was examined. This subset covers the area from 19◦ 34 N, 19◦ W to 4◦ 36 N, 8◦ 15 E and many severe floods have occurred within this subset over the past several years. Between 2002 and 2011 a total of almost 450,000 people were affected by 9 separate flooding events in Burkina Faso, with particularly severe floods occurring in July 2007, September 2009 and July 2010 (EMDAT, 2011b). Neighbouring countries have also been hit by large scale flooding. In the previous decade Mali experienced 11 events that affected a total of more than 180,000 people, whilst 7 large floods occurred in Niger – affecting more than 470,000 people. For the decade ending in 2009, West Africa as a whole experienced 105 major floods that resulted in around 1150 deaths and affected nearly 3.5 million people. Before ingestion into the BRDF algorithm, the SEVIRI top-ofatmosphere reflectances were masked to remove areas affected by cloud cover by applying the EUMETCAST MPEF cloudmask that is distributed along with the raw SEVIRI data. The reflectances were then corrected for atmospheric effects by using a modified version of the Simplified Method for Atmospheric Correction (SMAC) (Rahman and Dedeiu, 1994; Proud et al., 2010). Data was collected between 0600 and 1800 UTC each day, a time span that covers

538


Fig. 1. Locations of study areas and test sites in West Africa overlaid on a Blue Marble image from NASA’s Earth Observatory. Region ‘A’ is the Inner Niger Delta, Region ‘B’ is in Burkina Faso, with precipitation stations highlighted by circles and Region ‘C’ is the Niger River upstream of Niamey. The test pixels 1–4 are, respectively, on the Sokoto River (1), a grassland scene dominated by low-lying vegetation (2), a water-filled area in Lake Volta (3) and a sparsely vegetated scrub land pixel (4).

almost the entire sunlit period for West Africa. These images, along with information about the position of the Sun for each image, were then fed into the ‘inverse mode’ of the BRDF algorithm to produce the BRDF parameters. In turn, the parameters were fed back into the model, this time operating in ‘forward mode’, to produce simulated reflectances for each time slot. The relative difference between the measured and modelled reflectances gives a good indication as to the success of the BRDF algorithm at modelling the angular dependence of the land surface reflectivity. As such, any time slot in which the relative difference exceeds 10% is ignored and the parameters re-calculated with the remaining data. The final result is a BRDF model (made up of the Iso, Geo and Vol parameters) on a three day timescale that accurately describes the angular sensitivity of the land surface reflectance to the three scattering modes described previously. The three-day BRDFs are produced at the University of Copenhagen on a semi-operational basis, beginning on January 1st of each year, meaning that 122 BRDFs are produced annually. Production is not linked to flooding events, so if flooding is visible within a 3-day period then it may have occurred on any of the three days that make up that BRDF. All raw data are archived, however, so to accurately determine the start of flooding events it is possible to reprocess the data using different acquisition dates. 2.3. Study areas To determine the parameter values that can signify the presence of flooded land, a pixel on the Sokoto River in Nigeria was chosen as a test site, and is labelled as point 1 in Fig. 1. At this location, the Sokoto River is situated on a flood plain approximately 10 km across, and the river is braided into many smaller waterways. The flood plain is therefore clearly visible to MSG, but the river itself is smaller than the 3 km MSG pixel size, and so will only partially influence the measured reflectance. The pixel is dry for much of the year but between late July and October the river fills with water and, in 2009, burst its banks – flooding a large area. The pattern of diurnal reflectance in the dry, wet and flooded seasons was examined and compared to other pixels nearby that represent known land cover types. Point 2 in Fig. 1 is a typical grassland pixel, whilst point 3 is a good example of a bare soil area and point 4 is a pixel covered by water.

Furthermore, three regions were defined for use in a comparison between the MSG flood mapping technique and a variety of other methods, described in the following paragraphs, that have been used to detect the presence of flooded land. All three regions experienced one or more severe floods in 2009 and are labelled as A, B and C in Fig. 1. Region A is located in Mali, and covers from the Inner Niger Delta in North-Central Mali to the Burkina Faso border in the East. The area (13–16◦ N) belongs to the central and northern parts of the Sahel and experiences a typical semi-arid climate. The rainy season extends from June to September (Roncoli et al., 2007), with July and August being the wettest months. Annual rainfall ranges from 300 mm/year in the northern part of the region to 500 mm/year in the South (Nicholson, 2005). However, the rainfall regime is characterised by great variability in both time and space, even within short distances rainfall may be very different (Rasmussen et al., 2001). Due to the presence of the Inner Niger Delta in this region flooding is a common occurrence – most frequently in August, September, December and January as a result of both local rainfall and the amount of water transported from other areas by the Niger River itself (Diarra et al., 2004). For this site a comparison was made to data gathered by the MODIS instrument at 500 m resolution averaged over an 8-day period, substantially longer than the MSG 3-day timescale. As most flooding in the Inner Niger Delta is caused by the river rather than precipitation, there was at least one clear-sky opportunity every 8 days that was usable to examine the land surface with MODIS. The technique developed at the Dartmouth Flood Observatory (Brakenridge and Anderson, 2006) was used to pinpoint the areas of land covered by water, with the areal extent of surface water being output. This method compares the NDVI from MODIS reflectance data during a specific 8-day period to that gathered at other times. By comparison to data gathered in previous years it is possible to map the normal extent of a river during a particular time period and compare it to the current extent. If the extent is significantly larger compared to previous years then it is likely that flooding is occurring. Area B constitutes the easternmost part of Burkina Faso, stretching from the Centre region in the North to the Centre-East region in the South – close to the border with Togo. The area (10–13◦ N) is situated in the southern part of the Sahel and experiences a


semi-arid climate with a short rainy season, largely limited to the months of July, August and September. The annual rainfall ranges from 500 mm/year in the North of the region to 800 mm/year in the South and the general trend seen elsewhere in the Sahel towards an increased annual precipitation in the last decade is also evident in this region (Nicholson, 2005). The region differs from Mali as it contains no significant river system, meaning that any flooding is primarily due to large amounts of precipitation within a short time. Because of this, there is frequent cloud cover during the flooding events and it is not possible to employ the MODIS instrument as was used for Area A. Instead, precipitation data from 3 sites spread across Area B was averaged and used as a proxy, as major rain events are typically accompanied by large flooding on a scale visible to MSG (Bracken et al., 2008; Messager et al., 2006). The locations of the three precipitation stations are marked by circles in Fig. 1. Area C is a 70 × 22 km corridor along the Niger river upstream of the city of Niamey in Niger. In this area river flow data on the Niger was compared to the MSG flood map. Data from a river station near Niamey was provided by the Niger Basin Authority, allowing the volume of water passing through the station (the river discharge) to be compared to the flooded land mapped by MSG. Sudden increases in the discharge can signal a flooding event, as the river may not be able to contain the increased volume of water within its banks. This means that a dependence of the flood area upon the differential of water discharge should be noticeable when examining data gathered in Area C (Usachev, 1983; Smith et al., 1996). 3. Detection of flooded land Within this study we examine the BRDF parameter values produced from diurnal trends in land surface reflectance produced by MSG. The physical basis for this approach to flood detection is straightforward, as water displays a very different diurnal reflectance curve to other land cover types. Because of this, the BRDF parameters will vary dependant upon whether or not water is present within a pixel. The isotropic parameter values specify the reflectance of the land surface and, as is expected, at times when flooding occurs the land will display a markedly different reflectance to that seen outside of flooding events. By introducing the volumetric parameter a measure of the variability in reflectance is gained. The variation in the geometric parameter is small in relation to the land surface type, and so is not used as an indicator of flood extent. The following analysis examines this approach, and shows that waterlogged areas of land display very different diurnal reflectance trends to dry areas. Because of this, they therefore show a different value for the isometric and volumetric BRDF parameters, meaning that a combination of the isotropic and volumetric values enables a clear signal of flooding events to be generated. Fig. 2(a) shows the spectral characteristics of the Sokoto river on the 4th April 2009: the dry season. At this time, the riverbed is almost completely dry and much of the vegetation has died back. This leads to diurnal reflectance trends that very closely match those for bare soil, as shown in Fig. 2(d), that contain low reflectances in the morning and evening with a reflectance peak near midday. Channel 3 displays much higher values than channels 1 and 2 due to the strong reflectance of bare soil in the near-IR (Jacquemoud et al., 1992). For a dry river the isotropic and volumetric parameters are shown in row 1 of Table 1. The isotropic parameter is substantially higher than the volumetric value, and typically somewhat higher than the maximum surface reflectance. During the wet season the river fills with water, changing the diurnal reflectance trend, as shown in Fig. 2(b) for the 15th August 2007. Channel 3 reflectances have decreased since the dry season, with a peak of around 0.15, and the distinctive shape visible in the dry season has been replaced by almost constant reflectance values throughout the day. Nevertheless, a small midday peak is visible

539

Table 1 BRDF parameters for the Sokoto pixel in three sets of conditions: dry season, wet season and flooded land. Case

Dry Wet Flood

Channel 1

Channel 2

Channel 3

Iso

Vol

Iso

Vol

Iso

Vol

0.284 0.195 0.201

0.019 0.055 0.018

0.436 0.453 0.421

−0.021 0.070 0.046

0.602 0.522 0.375

0.021 0.167 0.092

for all channels, and in the morning there is a reflectance increase. Channel 1 displays a slight ‘bowl’ shape in which the morning and evening reflectances are higher than those near midday. This is typical of grassland (Fig. 2(e)) and water (Fig. 2(f)). For channels 2 and 3 water and grass display opposing trends, with water showing a slight bowl shape and grass presenting a small midday increase. A combination of these two reflectance trends results in the almost flat reflectance trend visible during the wet season, and therefore signifies that the pixel contains both water and grassland. For wet season conditions the parameters are located in row 2 of Table 1. The channel 2 and 3 isotropic parameters are now higher than the corresponding reflectances, and the volumetric parameters in all three channels are positive and non-negligible. The Sokoto river broke its banks in late August 2009 and the new diurnal trend is shown in Fig. 2(c). Channel 3 is now very low, and displays a trend almost identical to the water pixel in Fig. 2(f). Channel 2 is broadly similar to the normal wet season conditions, but is somewhat flatter in the early morning, and channel 1 now resembles its diurnal trend present in Fig. 2(a). The large peak visible at 07.30 is due to cloudiness, not a surface feature. The change in channel 3 reflectance trend is a good indicator of a substantial amount of water being present. The reversion of the channel 1 trend to that for bare soil indicates large amounts of sediment in the water, as at this wavelength there is little difference in reflectance between wet and dry soil (Jacquemoud et al., 1992), unlike for channels 2 and 3. For this case the parameter values are in row 3 of Table 1. The return of the channel 1 volumetric parameter to a value lower than 0.05 is a useful flooding indicator. The isotropic parameters in channels 2 and 3 have now swapped over, with channel 2 being higher than channel 3. The BRDF parameters for the Sokoto Pixel are shown for 2009 in Fig. 3(a). It shows that the channel 1 volumetric parameter is frequently less than 0.05, so it alone cannot be used as a flooding indicator. However, it is also clear that at the time of the floods the channel 1 isotropic parameter was low, whilst for channel 2 it was high. Similarly, during the flooded period the channel 3 isotropic parameter became lower than that for channel 2. By producing two new indices, each known as a Water Index (WI), that are a combination of these parameters it is possible to pinpoint times when the Sokoto is flooded: WI32 =

Iso3 − Vol3 Iso2 − Vol2

(2)

WI21 =

Iso2 − Vol2 Iso1 − Vol1

(3)

where Ison and Voln are the isotropic and volumetric parameters for channel n. The isotropic parameter describes the majority of reflectance variation caused by flooding, but by including the volumetric terms in the indices the accuracy of flood detection was increased. Even though the volumetric parameter is typically small compared to the isotropic term it contains much information about the shape of the reflectance trend – and hence the presence of flooded land. This is demonstrated by the effects of the Vol2 parameter on the Sokoto reflectance trend. Fig. 3(b) shows the WI32 and WI21 values for the Sokoto pixel. A strong flood signal is seen at times when WI32 is less than 0.9, WI21 is greater than 2.45 and

540


0.5

Ch01 (600nm) Ch02 (800nm) Ch03 (1600nm)

Reflectance

Reflectance

0.4 0.3 0.2 0.1 0

6

7

8

9

10

11

12 13 Time (hr)

14

15

16

17

18

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0

Ch01 (600nm) Ch02 (800nm) Ch03 (1600nm)

6

7

8

a Dry season river 0.5

0.5

Ch01 (600nm) Ch02 (800nm) Ch03 (1600nm)

0.3 0.2 0.1

14

15

16

17

18

14

15

16

17

18

14

15

16

17

18

0.3 0.2 0.1

6

0.5

7

8

9

10

11

12 13 Time (hr)

14

15

16

17

0

18

6

7

8

9

10

11

12 13 Time (hr)

d Bare Soil 0.5

Ch01 (600nm) Ch02 (800nm) Ch03 (1600nm)

Ch01 (600nm) Ch02 (800nm) Ch03 (1600nm)

0.4 Reflectance

0.4 Reflectance

12 13 Time (hr)

Ch01 (600nm) Ch02 (800nm) Ch03 (1600nm)

c Flooded River

0.3 0.2 0.1 0

11

0.4 Reflectance

Reflectance

10

b Wet season river

0.4

0

9

0.3 0.2 0.1

6

7

8

9

10

11

12 13 Time (hr)

14

15

16

17

18

e Grassland

0

6

7

8

9

10

11

12 13 Time (hr)

f Water

Fig. 2. Diurnal reflectance trends for a variety of land cover types. (a) Dry season river, (b) wet season river, (c) flooded River, (d) bare Soil, (e) grassland, and (f) water.

Vol1 is less than 0.05, and these values can be used as thresholds to indicate flooding. By automated comparison of the two index values, WI32 and WI21 , to their thresholds across a whole image, pixels that are flooded should be highlighted, enabling the large scale detection of flooded land. To examine the validity of this automated flood detection, several test areas were defined in West Africa, allowing analysis of the BRDF parameters in 2009. For all three test areas there is a strong relationship between the MSG derived flood extent and that measured by other sources. 4. Results and validation 4.1. Comparisons to polar orbiting satellite data The Niger Delta study region, Area A, shows a correlation of 0.882 between the flood extent measured by the MSG parameter method and that from MODIS, meaning that there is a good fit between the two techniques. Fig. 4(a) shows the variation in the mapped flood area as a percentage of the entire 83,000 km2 area of the region. Both MSG and MODIS show little flooded land during the

dry season with the exception of day 163 when MSG shows a spike that is caused by cloud contamination. From day 200 the beginning of the wet season is seen. Flooding becomes visible in the MSG data on a small scale (approximately 1–2% of all observed pixels). The MODIS data lags behind MSG due to its longer compositing period, but a gradual increase in water-covered land is also visible. On day 241 there is a dramatic increase in the area of flooded land that corresponds both to the arrival of water from upstream on the Niger and a series of heavy rainstorms in the preceding days. A peak is then seen on day 247, with the flood waters gradually receding after this day. The MODIS data show a similar trend, but the peak flooding occurs on day 258, 11 days later than MSG. Again this is due to the long compositing time for MODIS, with the actual peak flooding occurring some time within the previous 8 days. This highlights one of the primary drawbacks of using MODIS as a flood detection tool. Using one day data means clouds are an issue, but in the 8day data the exact times at which flooding occurs becomes unclear. The shorter compositing time of MSG helps overcome this limitation. The actual areas detected as flooded by both methods closely match, although an exact comparison is hampered by the different temporal and spatial scales of MSG and MODIS. Over the entire


0.5 0.4 0.3 0.2 0.1 0 -0.1 -0.2 -0.3

Iso. Ch01 Vol. Ch01

0

30

Iso. Ch02 Vol. Ch02

60

90

Iso. Ch03 Vol. Ch03

120

150

180

210

240

270

8

MSG Flooded Area MODIS Flooded Area

7 6 5 4 3 2 1 0 -1 150

1 NDVI

0.9

2.4

0.8

2.1

0.7

1.8

0.6

1.5

0.5

1.2

0.4

0.9

0.3

0.6

0.2

0.3

0.1

0

Flooding

0

30

60

90

120

150

180

210

240

270

0 300

Day of Year (DOY) Fig. 3. Variation in MSG BRDF parameters and index values for the pixel on the Sokoto River in Nigeria between January and October 2009. The two vertical lines on each figure denote the extent of a period in which the Sokoto was known, from local sources, to have flooded. (a) MSG BRDF Parameter Values for channels 1, 2 and 3, and (b) MSG Water and Vegetation Index Values.

year 85% of flooded pixels are detected by both MSG and MODIS. During the wet season this is reduced to 82%, primarily due to the obscuring influence of clouds in the MODIS images. Overall the MSG time series fits well with the MODIS equivalent, although cloudiness can cause unexpected changes in the area extent of flooded land. Additionally, MODIS retains a better spatial resolution than MSG (one MSG pixel contains at least thirty six 500 m MODIS pixels) but is let down by the long compositing time due to the infrequent overpasses of the Aqua and Terra satellites that carry the MODIS instrument. By using the MSG BRDF parameters instead there is a gain in resolution within the temporal domain but a loss in the spatial domain. As flooding events usually occur very rapidly this increased temporal resolution can be most useful in flood detection and assisting those who have been affected by the flooding. The over-estimation in flooded land by MSG evidenced within parts of Fig. 4(a) also highlights another important point. The MSG BRDF data is capable of being used to detect flooding events that are of smaller spatial extent than the 9 km2 pixel size. Comparison with the MODIS data shows that the MSG flood flag is raised, even if only 25% of the pixel is classed as inundated by MODIS. This means that the flood area will be overestimated somewhat as the entire pixel will be classified as flooded. By performing a more detailed pixel examination it may be possible to extract a more accurate area for the flooded land on a subpixel scale. For instance, by comparison of the BRDF parameters at the time of flooding to those from a non-flood period it may be possible to produce an estimate of the fraction of the pixel affected by flooding. 4.2. Comparison of flooded area Here the differences between the MSG and MODIS methods in detecting flooded land are examined. Fig. 5 shows the mapped flood extent for the Inner Niger Delta (Area A in Fig. 1) on the 19th of September 2009, which is a typical scene from this time of year in terms of flood extent – but is atypical in that it is one of the

4500

270

300

27

MSG Flooded Area Precipitation

4000

Flooded Area (km2)

WI21

24

3500

21

Major Flooding Event in Ouagadougou

3000

18

2500

15

2000

12

1500

9

1000

6

500

3

0

150

175

200

225

250

275

0 300

Day of Year (DOY)

b Burkina Faso, from May 30th until October 27th. A large spike is visible on day 245, signifying the serious flooding experienced in Ouagadougou on that day. A previous, but less well reported, flooding event in other parts of Burkina Faso is visible on day 225. 900

1800

MSG Flooded Area Niger discharge

810 2 Flooded Area (km )

WI32

240

a The Inner Niger Delta in Mali. The flooded area is calculated as a percentage of the whole region, and the MODIS Area is relative to the water level on 1st January. As MODIS is on an 8-day timescale there can be a lag between the two data-sets, depending on when in the 8-days the flooding occurred.

NDVI Value

Water Index Value

3

210

Day of Year (DOY)

Day of Year (DOY)

2.7

180

300

Precipitation (mm)

Flooding

1620

720

1440

630

1260

540

1080

450

900

360

720

270

540

180

360

90

180

0 150

165

180

195

210

225

240

255

270

285

Discharge (m3/s)

Parameter Value

0.6

Percentage of Scene Flooded

0.7

541

0 300

Day of Year (DOY)

c

A 70x22km portion of the Niger upstream from the city of Niamey showing the correlation between flooded area and the discharge of the Niger River, as measured by a station near Niamey. Sudden jumps in the discharge are accompanied by a corresponding rise in the MSG flooded area. Fig. 4. Comparison of the total size of the MSG flooded area to a variety of other flood detection techniques for several test sites in West Africa during 2009. A percentage is used rather than areal extent in (a) due to the differing spatial resolutions of MODIS and MSG. (a) The Inner Niger Delta in Mali. The flooded area is calculated as a percentage of the whole region, and the MODIS Area is relative to the water level on 1st January. As MODIS is on an 8-day timescale there can be a lag between the two data-sets, depending on when in the 8-days the flooding occurred. (b) Burkina Faso, from May 30th until October 27th. A large spike is visible on day 245, signifying the serious flooding experienced in Ouagadougou on that day. A previous, but less well reported, flooding event in other parts of Burkina Faso is visible on day 225. (c) A 70 × 22 km portion of the Niger upstream from the city of Niamey showing the correlation between flooded area and the discharge of the Niger River, as measured by a station near Niamey. Sudden jumps in the discharge are accompanied by a corresponding rise in the MSG flooded area.

542


Fig. 5. A comparison of the graphical flood extent from both the MODIS and MSG detection methods on the 19th of September 2009 for Area A, the Inner Niger Delta. (a) MODIS Channel 1 reflectances, with the Inner Niger Delta in the center of the image. (b) The MSG and MODIS classification of flooded land for the Inner Niger Delta.

few days in which the majority of the area is cloud-free within one of the two daily MODIS overpasses. Fig. 5(a) shows the Channel 1 reflectance from the MODIS sensor aboard Aqua at 250 m resolution, and it is noticeable in the North of the image (around 16◦ N) that there is significant cloud contamination within the data, visible as very bright areas. This is also noticeable close to 14◦ N, 3◦ 30 W, as well as in other smaller areas. The MODIS cloudmask was added to the data in order to remove these areas from further analysis. Fig. 5(b) shows the flood classifications by both MSG and MODIS, in which the MODIS data has been downscaled to the 3 km MSG spatial resolution. In total 8822 pixels within the scene are identified as non-flooded by both the MSG and MODIS methods, whilst 161 are detected as flooded by both methods. The MSG method highlights 154 pixels as flooded that MODIS does not highlight, and 54 pixels are flagged by MODIS as flooded – but not by MSG. All but one of the 54 pixels flagged only by MODIS are on the edge of a region in which the MODIS data is cloud contaminated. The remaining 1 pixel was inspected manually and does indeed show a flooded area that is not detected by the MSG method. Of the 154 pixels that only MSG flags as cloudy, 129 are in areas for which the MODIS data was masked as cloudy, and therefore no flooding indicator was present. 16 out of the remaining 25 pixels seem to be false detections by MSG, and all occur in areas that are covered by dense vegetation that is close to the normal flow area of the river, indicating that the method may require some additional work in order to be successful in such conditions. A possible solution to this problem would be the implementation of a database of seasonal NDVI and BRDF parameter values for each pixel. These could then be subtracted from the measured data, which would leave only the residuals between normal and current pixel conditions – thus providing a clearer measure of any flood signal that may be present. The down-side of this addition would be that the method no longer relies solely on the current MSG data, but will also require historical information. The final 9 pixels flagged as flooded only by MSG were examined in the MODIS data and do appear to be flooded. These pixels are also densely vegetated, but in this case the vegetation has raised the NDVI above the threshold used within the MODIS method to indicate flooding – meaning that the MODIS flood map erroneously lists these pixels as non-flooded.

This indicates that overall the MSG method is good at detecting the spatial extent of flooded land when compared to the MODIS method, with 86.56% of the 186 clear-sky pixels flagged as flooded by MSG also being flagged by MODIS. An additional 4.84% of pixels being successfully detected as flooded by MSG, but not by MODIS, and 8.60% of pixels being false positives within the MSG data. These false positives are highly correlated with land cover type, and so modification of the parameter values, or the inclusion of NDVI or land cover maps into the method, may well help to rectify the false detections. 4.3. The relationship between MSG flooding and precipitation For Area B the MSG data is compared to that from precipitation stations, and the results are shown in Fig. 4(b). The correlation between these two flood indicators is 0.58, significantly less than the MODIS to MSG correlation within Area A. This is caused by a number of precipitation events that do not result in any flooded land in the MSG flood map. Such rainfall is evident between days 170 and 180. It is possible that these rainfall events do not produce a peak in flooded area as the land is very dry, enabling absorption of large amounts of water and thus produce no significant flooding. Conversely, there is a flooding peak on day 206 that does not correspond to a rainfall peak. This flooding peak is caused by the MSG map showing significant flooding at the southern edge of the examination area, a location not covered by any of the precipitation gauges used within this study. Examination of precipitation measurements from northern Ghana – slightly outside the Southerly extent of Area B – shows a peak in rainfall around this day, however. Additionally, analysis of monthly data from the Tropical Rainfall Monitoring Mission (TRMM) shows a large amount of rain fell in southern Burkina Faso at this time. The primary flooding events of 2009 are both noticeable in the MSG and precipitation data. The first occurs on day 226, whilst the second is the more widely known floods on day 244 that inundated much of the Burkinan capital, Ouagadougou. For both flooding events there are peaks in the MSG mapped area and in the precipitation, particularly for the mid-August flood. It is therefore likely that the MSG mapped flood locations do correspond to actual floods within the region,


although the addition of more precipitation measurements covering a wider extent would increase the confidence in MSG flood mapping techniques in the Burkina Faso area. This highlights one of the primary problems with using ground based measurements for flood detection: Reduced spatial scale. Ground measurements can be very useful, and also highly accurate, when examining small areas that are prone to flooding but they are less useful on a regionwide scale. Lack of data from some areas may mean that flooding events are not recorded, whereas the large view afforded by MSG allows flood monitoring that, whilst slightly less accurate, covers a much broader area. 4.4. MSG detection of flooding resulting from river flow For the final area, C, along the Niger river the discharge and MSG flooded area are shown in Fig. 4(c). Examination shows that two rapid increases in river discharge that could signal a flood event are noticeable. The first occurs between days 227 and 229 when the discharge increases from 550 to 900 m3 /s. At the same time the area of flooded land increases from 86 to 350 km2 and continues to increase for 6 days before peaking at 440 km2 , during which time the Niger discharge remained almost constant. The flooded area then shrinks until a minimum of 156 km2 on day 156. The apparent decrease in area is because of the land surface characteristics changing due to long term water coverage, altering the BRDF parameters and no longer producing the flood signal. Another rapid increase in river discharge occurs between days 255 and 263 and again there is a corresponding rise in the flooded area, although the peak flooded area is reached before the peak discharge. The BRDF parameter based flood mapping method is best at detecting exceptional areas of flooded land rather than areas that are normally water covered (such as lakes), primarily due to the large amounts of sediment present in flood water. As with the August floods the MSG flooded area drops despite high discharge levels being retained, again due to the transformation of surface reflectance characteristics. This area shows that it is possible to map flooding in a way that closely matches the known state of a river, and it is clear that the MSG technique is more suited to highlighting transient, short term, flooding events rather than longer term waterlogged landscapes. 5. Conclusion and perspectives This study has shown that it is possible to derive a map of flooded land based upon the BRDF parameters produced from data gathered by the MSG-SEVIRI instrument. While not yet ready for operational use, the results derived from the BRDF parameters show a good fit to flooding events that have been detected through other means. The MSG based BRDF method shows numerous advantages over other techniques, however. The high rate at which the satellite captures land surface images can provide a much faster mapping of flood events than is possible with other space instruments, such as MODIS. Additionally, the fact that the satellite provides a continental scale view simultaneously means that flood mapping is possible over a wide area. This is not achievable using ground based detection procedures such as precipitation measurements and river flow data, as is illustrated in the case of the Burkina Faso study area. Flood mapping using instruments such as AMSR-E is confined to a coarse spatial resolution (De Groeve, 2010). MSG’s 9 km2 resolution, whilst able to map smaller flooding events than AMSR-E, is still of insufficient quality to map small local scale flooding. However, as has been shown in other studies (Hervé et al., 2007), combining the MSG data with those from other sources, such as MODIS data at 500 m resolution (Brakenridge and Anderson, 2006) or data from the COSMO-SkyMed series of satellites (Caltagirone et al., 2002), could avoid this issue and enable the high temporal resolution of

543

MSG to be fused with the high spatial resolution of other sensors. Additionally, by a detailed investigation of known flooded areas, more accurate knowledge of the BRDF parameters associated with flooded land can be gained – allowing adjustment of the threshold values and increasing the accuracy of flood detection. By combining the BRDF parameter values with a land cover database it would be possible to derive optimal thresholds on a per-pixel basis. This may remove some of the false positives that are visible in the time series, particularly for the Inner Niger Delta area. This BRDF based flood detection scheme is also capable of detecting subpixel flooding, even if only 25% of a pixel is flooded. This enables more flooding events to be detected, but naturally leads to an over-estimation in the flooded area. Finally, although this technique has been demonstrated in West Africa it is equally applicable in all areas covered by high temporal resolution satellite instruments. The Americas are covered by the GOES series of satellites, whilst the Indian Subcontinent and surrounding area is viewed by the CCD instrument aboard INSAT-3A. By using these satellites in addition to MSG it is possible to gain worldwide flood mapping for the tropical regions. Acknowledgements The authors would like to thank the anonymous reviewers for their useful and detailed comments that they provided during the review process. The Niger Basin Authority (NBA) are thanked for providing the Niger river discharge data to the African Monsoon Multidisciplinary Analyses (AMMA) project, from where it was retrieved for use in this study. Furthermore, they would like to thank UNOSAT, the Dartmouth Flood Observatory and ReliefWeb for providing background information on the 2009 West Africa floods. H. Nieto is thanked for his assistance in producing the flood maps used within this paper. C. Schaaf, Q. Zhang and the BRDF group at Boston University are thanked for their assistance in modifying the MODIS BRDF algorithm to function with MSG data. References Alcántara-Ayala, I., 2002. Geomorphology, natural hazards, vulnerability and prevention of natural disasters in developing countries. Geomorphology 47 (2–4), 107–124, http://www.sciencedirect.com/science/article/B6V9345MDT0B-4/2/be7bac8fc415b50b5bf270121d1306c1. Aminou, D., 2002. MSG’s SEVIRI instrument. ESA Bulletin 111, 15–17. Beget, M., Di Bella, C., 2007. Flooding: the effect of water depth on the spectral response of grass canopies. Journal of Hydrology 335 (3–4), 285–294. Boni, G., Castelli, F., Ferraris, L., Pierdicca, N., Serpico, S., Siccardi, F., 2008. High resolution COSMO/SkyMed SAR data analysis for civil protection from flooding events. In: Geoscience and Remote Sensing Symposium, 2007. IGARSS 2007. IEEE International ,. IEEE, pp. 6–9. Bracken, J., Cox, N.J., Shannon, J., 2008. The relationship between rainfall inputs and flood generation in south-east Spain. Hydrological Processes 22, 683–696, http://dx.doi.org/10.1002/hyp.6641. Brakenridge, G., Nghiem, S., Anderson, E., Mic, R., 2007. Orbital microwave measurement of river discharge and ice status. Water Resources Research 43 (4), W04405. Brakenridge, R., Anderson, E., 2006. MODIS-based flood detection, mapping and measurement: the potential for operational hydrological applications. Transboundary Floods: Reducing Risks Through Flood Management 1, 1. Caltagirone, F., Spera, P., Vigliotti, R., Manoni, G., 2002. SkyMed/COSMO mission overview. In: Geoscience and Remote Sensing Symposium Proceedings, 1998. IGARSS’98. 1998 IEEE International, vol. 2 ,. IEEE, pp. 683–685. Chen, J., Chen, X., Ju, W., Geng, X., 2005. Distributed hydrological model for mapping evapotranspiration using remote sensing inputs. Journal of Hydrology 305 (1–4), 15–39. Coulson, K.L., 1966. Effects of reflection properties of natural surfaces in aerial reconnaissance. Applied Optics 5 (6), 905–917. De Groeve, T., 2010. Flood monitoring and mapping using passive microwave remote sensing in namibia. Geomatics, Natural Hazards and Risk 1, 19–35. De Groeve, T., Kugler, Z., Brakenridge, G., 2006. Near real time flood alerting for the global disaster alert and coordination system. In: Van de Walle, B., Burghardt, P., Nieuwenhuis, C. (Eds.), Proceedings ISCRAM2007. , pp. 33–39. Diarra, S., Kuper, M., Mahé, G., 2004. Mali: Flood Management – Niger River Inland Delta. WMO Report. http://www.apfm.info/pdf/case studies/cs mali.pdf. Diner, D., Braswell, B., Davies, R., Gobron, N., Hu, J., Jin, Y., Kahn, R., Knyazikhin, Y., Loeb, N., Muller, J., et al., 2005. The value of multiangle measurements for

544


retrieving structurally and radiatively consistent properties of clouds, aerosols, and surfaces. Remote Sensing of Environment 97 (4), 495–518. EM-DAT, 2011a. The OFDA/CRED International Disaster Database. Universite Catholique de Louvain, Brussels, Belgium, Online. http://www.emdat.be/resultdisaster-profiles. 2011b. The OFDA/CRED International Disaster Database. EM-DAT, Catholique de Louvain, Brussels, Belgium, Online. Universite http://www.emdat.be/disaster-list. Fensholt, R., Anyamba, A., Stisen, S., Sandholt, I., Pak, E., Small, J., 2007. Comparisons of compositing period length for vegetation index data from polar-orbiting and geostationary satellites for the cloud-prone region of West Africa. Photogrammetric Engineering and Remote Sensing 73 (3), 297–309. Gao, F., Schaaf, C., Strahler, A., Jin, Y., Li, X., 2003. Detecting vegetation structure using a kernel-based BRDF model. Remote Sensing of Environment 86 (2), 198–205. Hahmann, T., Martinis, S., Twele, A., Roth, A., Buchroithner, M., 2008. Extraction of water and flood areas from SAR data. EUSAR. Hervé, Y., Bernard, A., Rémi, A., Stéphanie, B., Claude, B., 2007. Synergy of High SAR and optical data for flood monitoring; the 2005–2006 Central European floods gained experience. In: Proc. ‘Envisat Symposium 2007’. Jacquemoud, S., Baret, F., Hanocq, J., 1992. Modelling spectral and bidirectional soil reflectance. Remote Sensing of Environment 41 (2), 123–132. Jain, S., Singh, R., Jain, M., Lohani, A., Aug. 2005. Delineation of flood-prone areas using remote sensing techniques. Water Resources Management 19 (4), 333–347, http://dx.doi.org/10.1007/s11269-005-3281-5. Jonkman, S.N., 2005. Global perspectives on loss of human life caused by floods. Natural Hazards 34 (February 2), 151–175, http://dx.doi.org/10.1007/s11069004-8891-3. Justice, C., Vermote, E., Townshend, J., Defries, R., Roy, D., Hall, D., Salomonson, V., Privette, J., Riggs, G., Strahler, A., et al., 1998. The Moderate Resolution Imaging Spectroradiometer (MODIS): land remote sensing for global change research. IEEE Transactions on Geoscience and Remote Sensing 36 (4), 1228–1249. Kleinen, T., Petschel-Held, G., 2007. Integrated assessment of changes in flooding probabilities due to climate change. Climatic Change 81 (3), 283–312. Kriebel, K.T., 1978. Measured spectral bidirectional reflection properties of four vegetated surfaces. Applied Optics 17 (2), 253–259. Lucht, W., Schaaf, C., Strahler, A., 2000. An algorithm for the retrieval of albedo from space using semiempirical BRDF models. IEEE Transactions on Geoscience and Remote Sensing 38 (2), 977–998. McFeeters, S., 1996. The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. International Journal of Remote Sensing (Print) 17 (7), 1425–1432. McGranahan, G., Balk, D., Anderson, B., 2007. The rising tide: assessing the risks of climate change and human settlements in low elevation coastal zones. Environment and Urbanization 19 (1), 17. Messager, C., GallÃe, H., Brasseur, O., Cappelaere, B., Peugeot, C., SÃguis, L., Vauclin, M., Ramel, R., Grasseau, G., LÃger, L., Girou, D., Aug. 2006. Influence of observed and RCM-simulated precipitation on the water discharge

over the Sirba basin, Burkina faso/niger. Climate Dynamics 27 (2), 199–214, http://dx.doi.org/10.1007/s00382-006-0131-y. Meyer, D., Verstraete, M., Pinty, B., 1995. The effect of surface anisotropy and viewing geometry on the estimation of NDVI from AVHRR. Remote Sensing Reviews 12 (1), 3–27. Nicholson, S., 2005. On the question of the “recovery” of the rains in the West African Sahel. Journal of Arid Environments 63 (3), 615–641. Ottlé, C., Vidal-Madjar, D., 1994. Assimilation of soil moisture inferred from infrared remote sensing in a hydrological model over the HAPEX-MOBILHY region. Journal of Hydrology (Amsterdam) 158 (3–4), 241–264. Pierdicca, N., Chini, M., Pulvirenti, L., Candela, L., Ferrazzoli, P., Guerriero, L., Boni, G., Siccardi, F., Castelli, F., 2010. Using COSMO-SkyMed data for flood mapping: Some case-studies. In: Geoscience and Remote Sensing Symposium, 2009 IEEE International, IGARSS 2009, vol. 2. IEEE. Proud, S.R., Rasmussen, M.O., Fensholt, R., Sandholt, I., Shisanya, C., Mutero, W., Mbow, C., Anyamba, A., 2010. Improving the smac atmospheric correction code by analysis of meteosat second generation NDVI and surface reflectance data. Remote Sensing of Environment 114 (8), 1687–1698. Rahman, H., Dedeiu, G., 1994. SMAC: a simplified method for the atmospheric correction of satellite measurements in the solar spectrum. International Journal of Remote Sensing 15, 123–143. Rasmussen, K., Fog, B., Madsen, J., 2001. Desertification in reverse? Observations from northern Burkina Faso. Global Environmental Change 11 (4), 271–282. Roncoli, C., Jost, C., Perez, C., Moore, K., Ballo, A., Cissé, S., Ouattara, K., 2007. Carbon sequestration from common property resources: lessons from communitybased sustainable pasture management in north-central Mali. Agricultural Systems 94 (1), 97–109. Sandholt, I., Andersen, J., Dybkjær, G., Nyborg, L., Lô, M., Rasmussen, K., Refsgaard, J., Jensen, K., Touré, A., 2003a. Integration of earth observation data in distributed hydrological models: the Senegal River basin: Applications of remote sensing in hydrology. Canadian Journal of Remote Sensing 29 (6), 701–710. Sandholt, I., Nyborg, L., Fog, B., Lô, M., Bocoum, O., Rasmussen, K., 2003b. Remote sensing techniques for flood monitoring in the Senegal River Valley. Geografisk Tidsskrift, Danish Journal of Geography 103 (1), 71. Sanyal, J., Lu, X.X., Oct. 2004. Application of remote sensing in flood management with special reference to monsoon asia: a review. Natural Hazards 33 (2), 283–301, http://dx.doi.org/10.1023/B:NHAZ.0000037035.65105.95. Schaaf, C., Gao, F., Strahler, A., Lucht, W., Li, X., Tsang, T., Strugnell, N., Zhang, X., Jin, Y., Muller, J., et al., 2002. First operational BRDF, Albedo nadir reflectance products from MODIS. Remote Sensing of Environment 83 (1), 135–148. Smith, L., Isacks, B., Bloom, A., Murray, A., 1996. Estimation of discharge from three braided rivers using synthetic aperture radar satellite imagery: Potential application to ungaged basins. Water Resources Research 32 (7), 2021– 2034. Usachev, V., 1983. Evaluation of flood plain inundations by remote sensing methods. Hydrological Applications of Remote Sensing and Remote Data Transmission, 475–482.