Geomatics, Natural Hazards and Risk

Vol. 1, No. 1, March 2010, 19–35

Flood monitoring and mapping using passive microwave remote sensing

in Namibia

TOM DE GROEVE*
Joint Research Centre of the European Commission, Ispra, Via Fermi 2147, 21020 Ispra,
Italy

(Received 10 January 2010; in final form 25 January 2010)

Space-based river monitoring can provide a systematic, global, timely and

impartial way to monitor disastrous floods. This paper describes a methodology
to use daily passive microwave observations to detect, map and size floods, both
for the purposes of global humanitarian organizations and national hydrological
services. In the best case, floods can be detected as early as 2 h after they occur.
Early warning is possible by monitoring upstream areas, with warning lead times
up to 30 days. Flood maps are of a low resolution but match maps derived from
high-resolution imagery. The interest lies in their daily availability, allowing us to
understand the dynamic aspects of floods. Finally, objective flood sizing is
achieved by integrating information over time and space. This paper details the
results of the technique for the 2009 floods in Southern Africa and experiences for
the 2010 flood season in Namibia.

1. Introduction
In 2008 and 2009, Namibia was affected by major floods, which caused more than
100 people to drown, affected more than 30% of the country’s population and
caused over US$200 million of losses (Van Langenhove 2009). In 2008 and 2009,
very high rainfalls in the catchments of the Cuvelai, Kavango, Kwando and Zambezi
rivers resulted in extreme floods (the worst in more than 40 years or longer). Because
the frequency and magnitude of floods have been markedly reduced in the previous
25 years, they came as a total surprise. Conventional hydrological monitoring
systems during floods, for both immediate and long-term purposes, are very
vulnerable and have limited use for emergency response: the former because of
ground inaccessibility and the latter because of lack of aerial information.
Although floods are the most frequent and recurring natural disasters (49%)
consuming up to one-third of humanitarian aid (FTS OCHA, 2009), causing $20
billion annually in damages and affecting up to 100 million people annually
(Rodriguez et al. 2009), there is no systematic, global and timely monitoring system
available yet.
Floods are difficult to monitor on a global scale, because they are determined by
local factors such as precipitation, slope of the terrain, drainage of the river,
protection devices in place, etc. Each river must be monitored at different places

*Email: tom.de-groeve@jrc.ec.europa.eu
Geomatics, Natural Hazards and Risk
ISSN 1947-5705 print/ISSN 1947-5713 online ª 2010 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/19475701003648085
20 T. De Groeve
along its course. Although some flood disasters occur annually, most happen
unexpectedly. About 2641 flood disasters were recorded in the Dartmouth Flood
Observatory catalogue (DFO 2009), one of the most complete archives of flood
events, between 2000 and 2007, affecting 2051 different rivers. Some rivers flooded
more than 10 times (at different places along their course), but 1133 rivers flooded
only once in 7 years. All rivers must therefore be monitored, and along their full
course. The number of rivers in the world is hard to determine, but even the Digital
Chart of the World (Danko 1992), a database at a scale of 1:1 million which shows
only major rivers, has close to 1 million records, with a total length of 10 million km.
Unlike for earthquakes where few measuring stations suffice to monitor the globe
(the United States Geological survey global Seismographic network has less than 150
stations outside America), an in situ global flood monitoring system would need a
dense network of gauging stations along all rivers. However, such stations are
expensive (the United States Stream gauging Network costs US$89 million per year;
USGS 1998), which makes this hardly feasible on a global scale.
On a global scale, humanitarian emergency responders do not have many
information sources at their disposal to learn about flood disasters. Currently, they
have to rely on media reports or their own global network of colleagues. One system
providing a global list of flood events of potential humanitarian concern is the
Global Disaster Alert and Coordination System (De Groeve 2007), but its data
source is the media-based catalogue compiled manually at the Dartmouth Flood
Observatory. Information is scattered and not standardized, making a global
compilation and relative comparison of flood events difficult. Organizations that
need to prioritize humanitarian aid between ongoing disasters need to do so on an ad
hoc basis. As a consequence, aid is not always proportional to the size of the disaster
(De Groeve and Riva 2009).
On a national scale, water authorities also face challenges to monitor their river
network. First, the cost of maintaining gauging stations can be a limiting factor, in
particular for large countries. While stations are out of service, there are gaps in the
hydrograph time series. Second, rivers are often shared between countries, but
information on floods in upstream countries is not always communicated to
downstream countries. For adequate flood warning, countries need to be able to
collect information over bordering countries independently. Third, gauging stations
measure the water height, but not the extent of the flood. Obtaining information on
the extent of the flood is a challenge for emergency managers, requiring aerial
reconnaissance or satellite imagery.
In situ measurements can be replaced by remote sensing measurements, from
airplanes or satellites. The change in surface water extent can be extracted from
aerial or satellite imagery. While the use of sensors in the visible or infrared portion
of the spectrum is limited due to cloud cover, the microwave portion of the spectrum
can penetrate clouds (Smith 1997). However, for most remote sensing solutions, the
revisit frequency (i.e. the time between two measurements in the same place) is too
low for monitoring purposes, or the spatial coverage is limited (Bjerklie et al. 2003).
For satellite based imagery, the revisit time depends on the orbit and the image size,
and at least a few sensors have a daily revisit time and global coverage, and provide
microwave data in near-real time free of charge. These are the Advanced Microwave
Scanning Radiometer – Earth Observing System (AMSR-E) instrument on board of
the NASA EOS Aqua satellite (launched in 2002) and the Tropical Rainfall
Monitoring Mission (TRMM).
 Flood monitoring and mapping using passive microwave remote sensing 21
Using AMSR-E data, De Groeve et al. (2006) developed a method for detecting
major floods on a global basis in a systematic, timely and impartial way
appropriate for humanitarian response. Brakenridge et al. (2005) demonstrated
that AMSR-E can measure river discharge changes in various climatic conditions.
The methodology uses the brightness temperature at 36.5 GHz H-polarization
during the descending (night) orbit of AMSR-E with a footprint size of
approximately 8 6 12 km2 and an average revisit time at 1 day. Brightness
temperature is related to the physical temperature T and the emissivity e of an
object: Tb ¼ eT. Due to the different thermal inertia and emission properties of
land and water, the observed microwave radiation in general accounts for a
lower brightness temperature values for water (Tb,water) and higher for land
(Tb,land 4 Tb,water). Since each observation of the satellite (or pixel) covers a
relatively large area of 8 6 12 km2, the observed brightness temperature is mostly
composed of both water and land values, in proportion to the relative area of
water (w) and land (1 – w) in the pixel.

Tb ¼ ð1  wÞTb;land þ wTb;water ; ð1Þ

where w is the water portion of the pixel.

If the physical temperature remains constant, changes in brightness temperature
will be related to changes in water surface extent in the pixel. However, in spite of
the great radiation dissimilarity of water and land cover, the raw brightness
temperature observations cannot be used to reliably detect changes in surface
water area. This is because brightness temperature (Tb) measures are influenced by
other factors such as physical temperature, differences in emissivity and
atmospheric moisture. While the relative contribution of these factors cannot be
measured, they are assumed to be constant over a larger area. As shown in
equation (2), the ratio between two nearby pixel values is a function of w alone.
Therefore, by comparing a ‘wet pixel’ received over a river channel of a potential
inundation location (w 4 0) with a ‘dry pixel’ without water cover (w ¼ 0), the
mentioned noise factors can be reduced. The brightness temperature values of the
measurement/wet signal are divided by the calibration/dry observations, referred
to as M/C ratio or signal s.

Tb;measurement ¼ ð1  wÞTb;land þ wTb;water

¼ Tmeasurement ðð1  wÞeland þ wewater Þ
Tb;calibration ¼ Tb;land ¼ Tcalibration eland : ð2Þ

If, for nearby pixels, we assume

eland;measurement  eland;calibration  eland ; Tmeasurement  Tcalibration ð3Þ

then
M Tb;measurement Tmeasurement ðð1  wÞeland þ wewater Þ
s¼ ¼ ¼
C Tb;calibration Tcalibration eland
ewater
1wþw ¼ fðwÞ: ð4Þ
eland
22 T. De Groeve
Typically, emissivity values for water are around 0.5, and for most minerals
between 0.75 and 0.95 (Rees 1990). When calibration pixels are chosen carefully
(without any open water), s varies between 1 (w ¼ 0, no water) and 0.3 to 0.5 (w ¼ 1,
fully flooded) depending on the mineral content of the soil. In a time series, s is
expected to vary with changes in water surface. Anomalies of the signal s can be
correlated well with flood events. Moreover, as demonstrated by Brakenridge et al.
(2007), if gauging data are available for the site, the signal can be calibrated to
staging height with accurate results (figure 1).
De Groeve and Riva (2009) showed that a local maximum in a moving window
(more specifically the local 95 percentile value in a window of 7 6 7 pixels) can be
used for the calibration value, avoiding the need for manual selection of the
calibration site. This finding extended the applicability of the method to any
arbitrary area on Earth, rather than in carefully chosen sites, opening the way for a
fully automatic global flood-detection system.
In order to distinguish between areas with permanent water (e.g. lakes or wide rivers)
and areas with flood waters, one has to consider changes in flood signal over time.
Based on a time series of 7 years (going back to June 2002 when the satellite was
launched), anomalies are automatically detected using a method described by De
Groeve et al. (2006). Since lower M/C signals generally account for increased water
coverage, extreme events, or major floods, should represent negative anomalies in the
time series of a given site. In order to detect anomalies, De Groeve et al. (2006) first
determined the reference value for normal flow, which varies for each site based on the
local emissivity properties and river geometry. This reference value was calculated as
the average M/C value for the site since the launch of the satellite. They then set flood-
level thresholds based on the statistics of the time series. Flood magnitude was defined
as the number of standard deviations (SD) from the mean (avg):

s  avgðsÞ
m¼ : ð5Þ
SDðsÞ

Floods appear typically for anomalies of 2 (small and regular flood) or 4 (large
and unusual flood). This is equivalent to probabilities of 2.1% and 0.003%.

Figure 1. Example of time series of Tb,measurement (M), Tb,calibration (C) and the flood signal s
(RatioMC) for site 2010 100 km north of Hyderabad on the Indus River in Pakistan. The
three peaks correspond to true flood disasters in 2003, 2005 and 2006.
 Flood monitoring and mapping using passive microwave remote sensing 23
The purpose of this paper is to demonstrate the usefulness and limitations of
passive microwave satellite-based flood monitoring for humanitarian response and
for national monitoring. While both communities have different requirements, they
can both benefit from satellite-based flood monitoring technology, either as their sole
source or as an additional, independent source to their toolset.
In particular, three aspects will be examined:

. Use for flood detection: can floods be reliably detected, and if so, is there a
significant time gain compared with alternative methods?
. Use for real-time flood mapping: can water signal or magnitude images be used
as flood maps, in order to visualise affected areas or the extent of the floods?
. Use for flood-size estimation: can floods of different years be compared in a
quantitative way?

2. Methodology
In order to demonstrate the usefulness of passive microwave sensing for flood
monitoring, an operational processing system was set up at the Joint Research
Centre of the European Commission (JRC). In the development of this system,
particular interest was focused on making the processing time as short as possible to
approach real-time monitoring. Also, issues of reliability and redundancy were
considered by using alternative sources of satellite data. This system has been in
place since early 2009, but all data back to 2002 were reprocessed.
First, we describe the processing chain in detail. Subsequently, we describe the
methodologies for addressing the three aspects: flood detection, flood mapping and
flood sizing.

2.1 Data processing

The methodology has been tested with AMSR-E data and TRMM data. However,
TRMM data are only available up to latitudes of 308. In order to have global
coverage, AMSR-E data are used.
The data used are the Level 1 brightness temperature at 36.5 GHz H-polarization
during the descending (night) orbit of AMSR-E with a footprint size of
approximately 8 6 12 km2 and an average revisit time of 2 days south of 308 and
1 day north of 308 latitude. The swath data are available about 2 h after acquisition
by the satellite from the Japanese Space Agency (Jaxa). About 24 h later, they are
also available from the National Snow and Ice Data Centre (NSIDC) in the USA,
which is used as a backup data provider.
The data are downloaded at JRC, resampled (using a nearest neighbour
approach) and mosaicked in daily grids with global coverage; see De Groeve and
Riva (2009) for a detailed description of the data-processing steps. The result is
global grid (4000 6 2000 pixels) of brightness temperature data, available by
date. The data of the current day are recalculated whenever new swath data are
available.
In order to compensate for missing data (in particular around the equator) and to
eliminate some noise, a 4-day average is calculated for each pixel, starting with the
day of interest and going 4 days back. The 4-day window ensures at least two data
points for each pixel.
24 T. De Groeve
Flood signals are calculated for each pixel as soon as new data are available (using
a local maximum as the calibration temperature), resulting in a ‘water signal image’
showing standing surface water. Using the statistics of a 7-year time series, the
magnitude for each pixel is calculated, too.
Figure 2 describes the systems involved in the flood disaster alert system. There are
five independent systems involved. Three systems for flood detection: Jaxa/NASA’s
AMSR-E system (itself composed of many systems), the Jaxa and/or NSIDC
distribution system and the JRC Global Flood Detection System (GFDS). Two
systems for flood impact analysis were used: a global spatial data infrastructure
(GSDI) with data on population distribution, populated places and infrastructure
globally; and the Global Disaster Alert and Coordination System (GDACS), which
is a multihazard alert system to which the Global Flood Detection System provides
flood alerts for further processing with impact models (De Groeve et al. 2010).
GDACS combines information on the hazard with information on the exposed
population to assess the hazard impact, and has the ability to compose and send alert
messages.
On average, the time between the flood event and the flood alert is between 2 and
26 h. Since the satellite passes most locations on earth once per day, the delay
between flood (t0) and observation (t1) is between 0 and 24 h, with an average of
12 h. From the time of observation to publication on the Jaxa ftp site (t2) on average
2 h elapse. A download at JRC of one AMSR-E track file takes on average 5 min.
Files are available at JRC for processing (t5) on average 2 h after observation.
Processing at JRC is a matter of minutes. Data for over 10 000 monitoring sites are
extracted from the data files in less than 10 s and inserted into a database. The flood
signal and magnitude are recalculated with the new data, and XML feeds and web
pages are published dynamically.

Figure 2. Overview of systems involved in providing global flood disaster alerts.

 Flood monitoring and mapping using passive microwave remote sensing 25

2.2 Flood detection

For detecting floods, we define a set of monitoring areas for which a virtual
hydrograph is calculated. A monitoring area can be a pixel of the grid, or a group of
pixels. For each pixel, two measurements are available daily: the water signal s (or
M/C ratio) and the magnitude m ¼ (s – avg(s))/sd(s). If a monitoring area consists of
more than one pixel, we consider the following quantities: average signal, minimum
signal, average magnitude, maximum magnitude. In addition, for a monitoring area
of more than 1 pixel, we store the number of pixels exceeding magnitude 2
(probability of 2.1%) and the number of pixels exceeding magnitude 4 (probability
of 0.003%).
All these quantities are indicative of the size of the flood in the observation area.
The average of minimum flood signal is an objective measure of the water surface in
the pixel. The average or maximum magnitude is an objective measure of the relative
size of the flooding compared for all events since 2002. The pixel count is a measure
of the extent of the floods. Since each pixel represents an area of about 100 km2, the
‘number of pixels exceeding magnitude 4’ is an approximate measurement of the
flooded area (number of pixels 6 100 km2).
We have defined several types of monitoring sites. First, we used 2634
locations defined by the Dartmouth Flood Observatory in locations that pre-
viously suffered major floods or that have a high flood risk (Brakenridge et al.
2007). The sites are defined by geographical latitude and longitude. We created
monitoring areas using the pixel covering that geographical location. Second, we
defined 4475 areas as linear segments of 50 km in length on major rivers. Third, we
created a web-based application where hydrological services or water authorities
can define their own ‘virtual gauging sites’. With intimate knowledge of the local
river network, they are able to understand where to best locate sites for early flood
alerting. In November 2009, 27 sites were set up by the Namibian hydrological
service.
In order to detect floods automatically, a set of threshold needs to be defined on
one or more quantities, distinguishing between various flow states. We defined three
flow states: green (normal or low flow), orange (flood) and red (major flood). We set
thresholds at magnitude 2 and 4, respectively, for orange and red floods.

2.3 Flood mapping

The term ‘flood map’ is used for many cartographic products, sometimes leading to
confusion with emergency managers. Typically, a flood map should show the
maximum extent of a flood event. Since floods often propagate downstream, they are
a dynamic phenomenon, and flood maps of different times are necessary to give a
complete picture. Satellite-based flood maps are often mapping the flood extent at a
given day, not necessarily at the peak of the floods, but rather when a satellite image
was acquired. Flood maps should also show the difference with reference water,
highlighting the abnormal surface water.
With a resolution of 10 km2, one can hardly speak of flood extent mapping for the
microwave derived products described in this article. At best, the maps indicate areas
where floods occur, not the precise extent. It is possible to indicate if the area around
a town is flooded, but not if the town is flooded. However, this method provides
maps with a high temporal resolution (1 day), which is suitable for mapping the
26 T. De Groeve
dynamic aspects of floods, such as the downstream propagation or the growth and
recession of flood waters.
To produce flood maps, colour palettes are defined to render the water signal and
magnitude values. The water signal is shown in blue shades ranging from dark blue
(s < 0:5 , w  100%) to white (s < 0:9 , w  0%). Pixels with values larger than
0.9 are made transparent. The magnitude is shown in shades ranging from yellow
(magnitude 2) to red (magnitude 5 or larger). Magnitudes below 2 are transparent.
Global flood map images are generated automatically as soon as new data arrive.
However, flood maps are most meaningful in a temporal and spatial context: around
an observation site, around a city of interest, showing a river or showing a regional
overview map. In addition, maps should not be created for periods when no flooding
occurs. Therefore, in case a flood event occurs, flood maps are generated
automatically for each of the defined monitoring sites defined in the previous
section. If the magnitude exceeds a threshold, map images are generated for an area
expanded 200 km in each direction around the monitoring site from 10 days before
to the end of the flood event, and published on the website. In addition, an
animation is created for those images, effectively creating a movie showing the daily
evolution of the flood. This ensures timely flood map (and animation) creation for
monitoring sites.
However, to ensure timely flood maps for other areas, a web-based application
was created where emergency managers or hydrological services can interactively
create flood maps or animations of an arbitrary area of interest for a given
period.

2.4 Flood size comparison

A challenge for humanitarian organizations, but also for national governments and
reinsurance companies, is to estimate the size of the floods, be it in terms of affected
population, damage or cost of reconstruction. There is no single measurement that
characterises the strength of the hazard, such as Richter magnitude for earthquakes
or Saffir–Simpson categories (based on maximum sustained wind speed) for tropical
cyclones. Floods are characterized by their return period, but while this measure is
related to the risk of occurrence, it is not related to the affected area.
We will define a measure for flood size based only on the spatial and temporal
extent of a flood event, as recorded by passive microwave remote sensing. For the
spatial dimension, we use the average flood magnitude for a monitoring site, as
defined earlier in the section on flood detection. Since the water signal is
proportional to the water content of a pixel (equation (4)), the sum of all pixel
values in an area is proportional to the water content of the whole area. The same
does not hold for the magnitude, since it is scaled by the standard deviation of each
pixel. However, the average magnitude in the area remains an objective measure of
the anomaly. To consider the time dimension, the data must be integrated over time.
This is achieved by averaging magnitude over time in a time window.
For real-time processing, we consider for each year a window of 30 days back
from the current day, and we average the magnitude of the monitoring site. This
allows the current flood status to be compared with a similar period in earlier years.
For interactive data processing (through a web application or a desktop application),
an arbitrary time window can be set. The result is a measure taking into account time
and space, and making floods directly comparable across years.
 Flood monitoring and mapping using passive microwave remote sensing 27

2.5 Validation
The methodologies for detection, mapping and sizing of floods were tested in various
flood events, but in particular in Namibia for the floods in early 2009. For flood
detection, the flood signal and magnitude were compared with hydrograph data
from Namibian hydrograph stations, media reports on flood arrivals and situation
reports of emergency responders. For flood mapping, the maps were compared with
high-resolution satellite flood maps. For flood sizing, the data were compared with
information from situation reports and humanitarian funding decisions.

3. Results
3.1 Flood detection
The Global Flood Detection System has been producing flood alerts since 2007, with
progressively improved methods. Previous validation studies (Kugler and De Groeve
2007, De Groeve and Riva 2009) have shown that, on a global scale, the system can
detect floods in some cases up to 10 days before the international media (which is the
usual source for information on floods). However, Kugler and De Groeve (2007)
showed that omission and commission errors are significant if single pixels are used
as observation sites.
To find the sources for omission and commission errors, we take a closer look at
the results in Namibia. Satellite-based flood data match closely in-site hydrograph
data. For the station of Katima Mulilo, on the border of Angola and Namibia,
where the Zambezi River enters Namibia, we established an observation area.
Unfortunately, the area yields rather noisy data (figure 3). Nevertheless, major flood
events are clearly identifiably in both systems. For flow rates below 3000 m3/s, there
is no correlation with flood magnitude, because the river stays in the bank. However,
for flow rates over 3000 m3/s, clear peaks are visible in the flood magnitude time
series.

Figure 3. Time series at Katima Mulilo (Namibia). Hydrograph data courtesy of the
Namibian Hydrographic Server. For flow rates over 3000 m3/s, the satellite signal is
equivalent.
28 T. De Groeve
Other monitoring areas yield less noisy signals, and these can be used for flood
detection. For instance, a site on Kalangola (150 km upstream of Katima Mulilo, in
Zambia) delivers a very clear signal, which correlates well with the Katima Mulilo
in situ gauging data (figure 4). To compare the timing of the peaks, we took the date
at which the magnitude exceeded 2 (for satellite based gauging) and the date at which
the flow rate exceeded 3000 m3/s (for in situ gauging). In 2007, Kalangola reached
magnitude 2 on 2 February, while Katima Mulilo reached 3000 m3/s on 10 February,
giving an 8-day lead in warning.
Because data are available anywhere, one can also consider sites even further
upstream to increase the flood warning lead time. For instance, when compared with
a monitoring site 500 km upstream on the Upper Zambezi River (near
Lungwevungu), a lead time of up to 30 days can be obtained. The magnitude in
Lungwevungu exceeded 2 on 8 January 2007, which is 33 days before the flow rate
exceeds 3000 m3/s in Katima Mulilo (figure 5).
Overall, care must be taken with defining monitoring sites. From the experience of
hydrologists in Namibia, some sites perform badly for a variety of reasons. When the
river is confined and where flow variations result mainly in water-level changes
without much expansion of the water area, the satellite-based signal is only noise and
the magnitudes (anomalies in the signal) are meaningless. From experience with
Namibian water authorities, river segments may do better than points, because they
eliminate some of the random variations. Local heavy rains also affect the signal, as
does surface water in local pools, which do not add to the river flow by adding to the
signal and the apparent flow.

3.2 Flood mapping

A unique feature of the AMSR-E based flood system is the regional extent. Daily
global coverage allows the creation of regional overview maps, which cannot be
created with high-resolution satellite images which have small footprints. For
instance, figure 6 shows the flood situation in spring 2009 in southern Africa. Using

Figure 4. Flood magnitude at Kalangola (dashed, 12-day moving average) and in situ data at
Katima Mulilo (solid).
 Flood monitoring and mapping using passive microwave remote sensing 29
the water signal (and the blue–white palette described before), the flood areas are
clearly visible. In addition, these images are available daily, and the dynamic aspects
of floods can be illustrated with animations.
With regards to the accuracy of the maps, we compared the maps in the Caprivi
region with high-resolution satellite maps created by UNOSAT (based on various

Figure 5. Flood magnitude at Lungwevungu (dashed, 12-day moving average) and in situ
data at Katima Mulilo (solid).

Figure 6. Test areas: floods in Southern Africa, spring 2009 (animation available at http://
www.gdacs.org/floods).
30 T. De Groeve
high-resolution satellite sensors, including Radarsat, ASAR and DMC data).
Figure 7 shows a map produced by UNOSAT of the Caprivi area, showing the flood
waters on 20 and 29 March in orange and red respectively. Figures 8 and 9 show the
corresponding flood maps created by rendering the water signal of those dates using

Figure 7. Flood map, courtesy of UNOSAT, showing the floods on 20 and 29 March 2009.

Figure 8. Flood signal for 20 March 2009, using colour map described in the text, for the
same area as figure 7.
 Flood monitoring and mapping using passive microwave remote sensing 31

Figure 9. Flood signal for 29 March 2009, using colour map described in the text, for the
same area as figure 7.

the colour palette described before. A visual comparison shows a good

correspondence between our flood maps and the reference map, with the later
floods on the Chobe (to the south-west in the maps) visible only on 29 March.
Because the resolution of the maps is very low, positional accuracy is not a good
measure of quality of the maps. However, map quality could be measured quantitatively
in terms of the flood extent. Although the flood extent cannot be mapped with such a low
resolution, it can be measured
P because the water signal w is proportional to the surface
water extent. The sum A A s over an area A is a measure for the water surface extent
in the area. For 20 and 29 March, we calculated the average water signal s in an area
of 110 pixels (covering the floods shown in figures 8 and 9), which was 0.919 and
0.887, respectively (compared with 0.990 for normal flow). If we assume ewater ¼ 0.5
and eland ¼ 1, then equation (4) can be written as w ¼ 2ð1  sÞ w ¼ 2(1 7 s). Applying
this formula to the observed values, we estimate 16.2% and 22.6% of land covered
by water, compared with a reference value of 2%. Multiplying by the area of
observation (110 pixels of approximately 100 km2), we obtain estimates the flooded
area of 1782 km2 on 20 March 2009 and 2486 km2 on 29 March 2009, with a
reference value of 220 km2. Since the authors did not have access to the flood
polygons of UNOSAT, they redigitized the flood extent and measured a flooded area
of 2700 km2 for 29 March, which is only 8% more than our estimates.

3.3 Flood sizing

Many case studies including those described in this paper show that the system is
able to provide precise data on starting and ending dates of floods and the size of the
floods. While the absolute values of the flood size have not been validated sufficiently
against alternative data sources, the relative values are consistent with anecdotal
evidence. For instance, for the three areas in southern Africa shown in figure 6, we
32 T. De Groeve
Table 1. Average flooded area (km2) from February to May (4 months) of each year.
Border Namibia Angola Caprivi Region Upper Zambezi

2009 2000 1330 1330

2008 1360 700 260
2007 300 60 190
2006 280 160 90
2005 80 140 30
2004 60 180 610
2003 120 15 450

calculated the average flooded area over 4 months (from February to May, table 1).
The 2008 and 2009 floods stand out clearly, as expected. However, it is also clear that
the 2009 floods are worse than the 2008 floods, in particular on the Upper Zambezi.
This information is helpful to compare floods objectively with previous floods in
order to substantiate humanitarian appeals and funding decisions.

4. Discussion
Overall, considering the wide range of climate conditions, hydrographic settings,
river sizes and flow rates, the technique based on passive microwave remote sensing
performs remarkably well. Correlation with hydrograph data has been shown to be
excellent. However, sources of noise remain. Under certain conditions, single pixels
(or punctual observations sites) will not yield usable data, because the signal-to-noise
ratio is close to unity (Kugler and De Groeve 2007). The main reason is related to
local conditions on the ground, for instance, if a pixel covers a confined part of a
river where flow variations result mainly in water level changes without much
expansion of the water area (as was the case for several sites in Namibia).
Combining pixels in larger monitoring areas partially solves this issue. Experience
from Namibia shows that larger sites perform better in general (cancelling out
sources of noise), but not always. For instance, segments defined in the wrong
direction (perpendicular to the river) did not yield an improved signal.

4.1 Flood detection

The challenge of automatically detecting floods has not been solved yet. In most
cases, magnitude thresholds of 2 (minor flood) or 4 (major flood) yield good results,
but not always. Magnitudes are effectively anomalies in a 7-year time series (2002 to
now), which is very short compared with return periods of 50 or 100 years of
disastrous floods. For many pixels, no major floods occurred in the past 7 years,
resulting in low standard deviations. These, in turn, create large magnitudes and
therefore many false flood alerts. On the other hand, certain categories of floods
cannot be detected, including flash floods (unless the water remain for a few days)
and small, local floods. Recent floods in the UK could not be detected conclusively
using passive microwave sensing. Although the impact was great, the flooded area
was limited to one town district, and much smaller than a pixel of our system.
A solution for false alerts could be to use a cutoff minimum standard deviation in
the calculation of the magnitude. This would address cases where no major floods
occurred before but would create more flood omissions for small floods like the UK
 Flood monitoring and mapping using passive microwave remote sensing 33
floods. Alternatively, one could consider the change in magnitude (rather than the
value of the magnitude) as a trigger for floods. Typically, magnitudes rise steeply at
the onset of floods, resulting in high values for first derivatives. This will be examined
in future work.
In practice, the system can be used to detect upstream floods in neighbouring
countries. As shown in figure 5, this can result in lead-time warnings of up to 30
days. The use of satellite data can be a cheap alternative for flood early warning
without the need for collaboration or approval of the neighbouring governments.

4.2 Flood mapping

With regard to mapping, passive microwave sensing can contribute uniquely in two
ways: it can provide large overview maps, and it can produce maps on a daily basis.
With global coverage and daily temporal resolution, the system provides
information that is useful during the response phase. Emergency managers get an
overview of the most affected areas, while disaster-mapping organizations can use
the overview maps to target high-resolution satellite acquisitions. The high temporal
resolution also allows field campaigns to be put into context. Assessment teams visit
an affected area during the flood but not necessarily at the peak. Using daily maps,
and a satellite hydrograph, it is possible for the assessment team to understand how
much worse the situation was before or after the field visit.
The spatial accuracy of the maps could be improved using a digital elevation model.
The water content w of a pixel can be derived from the water signal s. If one assumes
that the water will be located in the lowest part of the pixel area, a high-resolution
digital elevation model can be used to find the lowest portion w of the pixel, which
should then be the extent of the flood. This will be addressed in future work.

4.3 Flood sizing

The method to estimate the size of a flood event compared with previous years is
particularly useful. Because it integrates data over time and space, noise-related or
inherent variability of pixel values tends to be cancelled out. The result is a stable and
reproducible measure for the size of a flood. The measure is objective and related to
the flood extent. However, just as an earthquake’s magnitude is not the only factor in
an earthquake disaster, the flood magnitude does not consider what has been
affected by the water; nor does it take into account water depth. Other data remain
necessary to assess the flood impact.

5. Conclusions
This paper shows that global and national flood monitoring can benefit from passive
microwave remote sensing. The availability of daily global observations, in
combination with a fast data-delivery system and an efficient processing system,
allows the fast detection of floods. Quantitative and qualitative measures can be
derived to assess the size and extent of the floods. Start and end dates can be
objectively determined for any arbitrary place on a river. Extent mapping is only
available at a low resolution but at a much higher temporal resolution than other
high-resolution satellite maps. It is also the only current solution for flood mapping
over large areas (whole countries or continents). In addition, an objective measure of
34 T. De Groeve
flood extents can be derived. Applied to the recent floods in southern Africa, the
technique allowed a quantitative assessment of the size of the floods relative to
previous years to be produced, guiding humanitarian funding decisions and other
response decisions. For the 2010 flood season, the system is being fine-tuned in
collaboration with the Namibian hydrological authorities and hopefully will provide
improved monitoring in the region.

Acknowledgements
The author wishes to acknowledge the collaboration with the Ministry of
Agriculture, Water and Forestry, Namibia (in particular Mr Guido Van
Langenhove), who provided hydrological data and feedback on the use of
microwave satellite flood monitoring in Namibia.

References
BJERKLIE, D.M., DINGMAN, S.L., VOROSMARTY, C.J., BOLSTER, C.H. and CONGALTON, R.G.,
2003, Evaluating the potential for measuring river discharge from space. Journal of
Hydrology, 278, pp. 17–38.
BRAKENRIDGE, G.R., NGHIEM, S.V., ANDERSON, E. and CHIEN, S., 2005, Space-based
measurement of river runoff. Eos, 86, pp. 185–188.
BRAKENRIDGE, G.R., NGHIEM, S.V., ANDERSON, E. and MIC, R., 2007, Orbital microwave
measurement of river discharge and ice status. Water Resources Research. 43.
DANKO, D.M., 1992, The digital chart of the world project. Photogrammetric Engineering and
Remote Sensing, 58, pp. 1125–1128.
DE GROEVE, T., 2007, Global disaster alert and coordination system: more effective and
efficient humanitarian response. In The 14th TIEMS Annual Conference 2007 Book of
Proceedings. Trogir, Croatia (Hydrographic Institute of the Republic of Croatia), pp.
324–334.
DE GROEVE, T., ANNUNZIATO, A., KUGLER, Z. and VERNACCINI, L., 2010, Near real-time global
disaster impact analysis. In Information Systems for Emergency Management, B. Van
De Walle, M. Turoff and S.R. Hiltz (Eds), pp. 302–326 (London: M.E. Sharpe).
DE GROEVE, T., KUGLER, Z. and BRAKENRIDGE, G.R., 2006, Near real time flood alerting for
the global disaster alert and coordination system. In Proceedings ISCRAM2007,
B. Van De Walle, P. Burghardt and C. Nieuwenhuis (Eds), pp. 33–39 (Newark, NJ:
ISCRAM).
DE GROEVE, T. and RIVA, P., 2009, Early flood detection and mapping for humanitarian
response. In Proceedings of the 6th International ISCRAM Conference, J. Landgren,
U. Nulden and Bartel Van De Walle (Eds), pp. 1–13 (Gothenburg, Sweden: ISCRAM).
DE GROEVE, T. and RIVA, P., 2009, Global real-time detection of major floods using passive
microwave remote sensing. In Proceedings of the 33rd ISRSE Conference. Stresa, Italy,
pp. 1–4.
DFO, 2009, Global Active Archive of Large Flood Events. Available online at: http://
www.dartmouth.edu/*floods (accessed 25 February 2009).
FTS OCHA, 2009, Financial Tracking Service. Available online at: http://ocha.unog.ch/fts
(accessed 25 February 2009).
KUGLER, Z. and DE GROEVE, T., 2007, The Global Flood Detection System (Luxembourg:
Office for Official Publications of the European Communities).
REES, W.G., 1990, Physical Principles of Remote Sensing (Cambridge: Cambridge University
Press).
RODRIGUEZ, J., VOS, F., BELOW, R. and GUHA-SAPIR, D., 2009, Annual Disaster Statistical
Review 2008 (Brussels: Centre for Research on the Epidemiology of Disasters).
 Flood monitoring and mapping using passive microwave remote sensing 35
SMITH, L.C., 1997, Satellite remote sensing of river inundation area, stage and discharge: A
review. Hvdrological Processes, 11, pp. 1427–1439.
USGS, 1998, A New Evaluation of the USGS Streamgaging Network. A report to Congress.
Available online at http://water.usgs.gov/streamgaging/report.pdf (accessed 25 Feb
2009).
VAN LANGENHOVE, G., 2009, Crisis Mapping. Available online at http://crisismapping.ning.
com/xn/detail/3244260:Comment:4477 (accessed 12 June 2009).
Copyright of Geomatics, Natural Hazards & Risk is the property of Taylor & Francis Ltd and its content may
not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written
permission. However, users may print, download, or email articles for individual use.