AUTOMATIC NEAR-REAL TIME FLOOD EXTENT AND DURATION MAPPING BASED

ON MULTI-SENSOR EARTH OBSERVATION DATA

Sandro Martinis, Marc Wieland, Michaela Rättich, Christian Böhnke, Torsten Riedlinger

German Aerospace Center, German Remote Sensing Data Center, Oberpfaffenhofen, 82234 Wessling,
IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium | 978-1-7281-6374-1/20/$31.00 ©2020 IEEE | DOI: 10.1109/IGARSS39084.2020.9324295

Germany

ABSTRACT multi-sensor satellite data is presented. It consists of four

automatic processing chains for the extraction of the
In order to support disaster management activities related inundation extent from Sentinel-1 (S-1) and TerraSAR-X
to flood situations, an automatic system for near-real time radar as well as from optical Sentinel-2 (S-2) and Landsat
mapping of flood extent and duration using multi-sensor EO data. In order to indicate the temporal stability of
satellite data is developed. The system is based on four flooding over time, the flood system automatically computes
processing chains for the automatic derivation of the two flood duration products, i.e. the Backward and Total
inundation extent from Sentinel-1 and TerraSAR-X radar as Flood Duration, together with corresponding quality layers
well as from optical Sentinel-2 and Landsat data. While the to indicate the temporal stability of an inundation over time.
systematic acquisition plan of the Sentinel-1/2 and Landsat The multi-sensor EO-based flood monitoring system is
satellites allows a continuous monitoring of inundated areas demonstrated on a flood situation in Mozambique caused by
at an interval of a few days, the TerraSAR-X processing the landfall of cyclone Idai in March 2019.
chain has to be triggered on-demand over the disaster-
affected areas. Beyond flood extent masks, flood duration 2. EO-BASED MULTI-SENSOR FLOOD
products are generated to indicate the temporal stability and SYSTEM
evolution of flood events. The flood monitoring system is
demonstrated on a severe flood situation in Mozambique An overview of the multi-sensor flood monitoring system’s
related to cyclone Idai in 2019. workflow is visualized in Fig. 1. Due to the systematic data
acquisition capability of S-1, the Sentinel-1 Flood Service
Index Terms— Flood Mapping, Flood Monitoring, [5, 2] has a key role within the proposed systematic NRT
Total Flood Duration, Backward Flood Duration, Flood flood monitoring system. The two processing chains based
Duration Quality, Sentinel-1/2, TerraSAR-X, Landsat on Landsat-8 and S-2 [6] complete the systematic
monitoring capability of the system. A TerraSAR-X Flood
1. INTRODUCTION Service [1] can be triggered on demand over a disaster area
to increase the revisit period and/or to derive the inundation
Flooding is a widespread and frequently a dramatic natural extent in higher detail than using the systematically
disaster that affects lives, infrastructures, economics and acquiring sensors of up to a spatial resolution of 1m. All
ecosystems worldwide. Floods can cause loss of life and flood services contain the following generic steps: automatic
substantial property damage. Further, the economic EO data ingestion (S-1: Copernicus Open Access Hub and
ramifications of flood damage disproportionately impact the direct downlink at the ground station of DLR in Neustrelitz,
most vulnerable members of society. Emergency responders S-2: Copernicus Open Acces Hub, TerraSAR-X: TerraSAR-
often request Earth Observation (EO) based crisis X delivery server, Landsat-8: USGS) and pre-processing,
information for flood monitoring to target the often limited computation and adaption of global ancillary data (digital
resources and to prioritize response actions during a disaster elevation models, topographic slope information and
situation. topographic indices, as well as reference water masks),
Single EO missions are in general not always suitable to automatic flood extent mapping, and dissemination of the
fulfill the requirements of emergency responders with crisis information, e.g. via a web-client. The radar-based
respect to revisit time and spatial coverage. Furthermore, processing chains make use of an automatic hierarchical tile-
disaster management especially benefits from the automation based thresholding approach in combination with fuzzy-
of algorithms to reduce time-consuming manual interactions logic-based post-processing for the unsupervised extraction
of imaging experts in extracting crisis information from EO of the flood extent [1, 5]. The thematic S-2 and Landsat
data. Therefore, an automatic multi-sensor satellite system services perform water segmentation with a convolutional
for near-real time (NRT) flood extent mapping based on neural network (CNN) based on the U-Net architecture that

978-1-7281-6374-1/20/$31.00 ©2020 IEEE 3243 IGARSS 2020

Authorized licensed use limited to: Institut Teknologi Sepuluh Nopember. Downloaded on March 21,2024 at 11:58:24 UTC from IEEE Xplore. Restrictions apply.
has been trained, tested and validated on a global reference checks stepwise backwards in time until it is not covered by
dataset representative for a variety of climatic, atmospheric water any more. The BFD is updated as soon as more up-to-
and land-cover conditions. Clouds and cloud shadows are date EO data gets available.
specifically handled by the network to remove potential The TFD is computed for each pixel for a user-defined
biases from downstream analysis [6]. time period. In contrast to the BFD it might cover more than
one flood event. The duration of n flood events is summed
in order to calculate the duration of flood coverage over
time. The quality of the flood duration products depends on
the observation frequency and temporal coverage of the used
virtual satellite constellation.
The FDQ is automatically estimated to indicate the
uncertainties related to these products. The computation
takes the acquisition frequency and distribution of
acquisitions during a flood event into account. By
computing the BFD the uncertainty is expressed by the sum
of the Pre-event (PreU) and Co-event Uncertainty (CoU). A
third term, the Post-event Uncertainty (PostU), might be
added by to PreU and CoU of each flood event in the
defined time period by calculating the TFD. CoU is
computed based on the number of gaps resulting from days
without satellite observation between two days with flooding
within a single inundation event. The influence of each gap
on the uncertainty is weighted according to the duration of
Figure 1: Workflow of the proposed EO-based system for the gap. Thereby, long observation gaps get penalized with
flood extent mapping and flood duration estimation. higher uncertainties.
The principle of the calculation of BFD, TFD, and FDQ
In order to provide emergency managers with spatio- is visualized for fictive flood events during a time period of
temporal information about the stability and the evolution of 20 days Dx in Fig. 2.
an inundation event or to assess the damage related to flood
events, the estimation of flood duration is a very crucial
parameter. In order to estimate the flood duration on a
regular basis in NRT in case of disaster situations a fully
automated approach is developed [3] and integrated into the
flood monitoring system.
Binary flood masks are automatically preprocessed
including re-projection to WGS84, clipping to a defined
AOI, merging (if there are several observations on the same
day), and resampling to a common spatial resolution. The
pre-processed flood masks serve as input for the subsequent
processing where a pixel-based examination of the time
series is carried out in order to compute two kinds of flood
duration products: the Total Flood Duration (TFD) and the
Backward Flood Duration (BFD). In addition, a Flood
Figure 2: Schematic illustration for computing a) BFD, b)
Duration Quality (FDQ) layer is generated in order to
TFD, c) FDQ for a fictive flood scenario over a 20-day
indicate the uncertainty related to both flood duration
period.
products. All available data sets are used for computing the
flood duration products. In the case that a pixel is covered
3. RESULTS
by cloud it is considered as no observation is available for
this raster cell during the acquisition time.
3.1. Study area and data set
BFD is computed for each image element backwards in
In this chapter, results of the EO-based flood system are
time from the latest incoming satellite scene until the start of
presented for the flooding related to the tropical cyclone Idai
the inundation event. It is an indicator for the duration of an
in 2019, which formed over the Northern Mozambique
ongoing flood situation in days and should be delivered to
Channel on March 9 and made landfall close to the city of
the user in NRT in order to provide an indication of the
Beira on the night of March 14 to 15. The study area is
persistency of the flooding. For each pixel the algorithm

3244

Authorized licensed use limited to: Institut Teknologi Sepuluh Nopember. Downloaded on March 21,2024 at 11:58:24 UTC from IEEE Xplore. Restrictions apply.
mainly situated in the province Sofala and has an extent of the landfall of Idai in the night of March 14 a significant part
~15,000km² (Fig. 3). of the AOI is flooded (~682 km²). On 19/03/2019 (Fig. 4a)
the flood peak is reached where more than 1,800km² are
inundated. These flooding occurred mainly in the catchment
area of River Búzi in the west of the AOI. Between 20/03/-
25/03/2019 the flood extent drastically receded by more than
1,400km² to 400km². The recession of water occurred
mainly in the central and southern parts of the AOI, whereas
inundations persisted in the North. In the following time
period until 07/04/2019 the flooding decreased slowly by
~300km² to an extent of ~105km². Due to the spatial
resolution of the used EO data and the side-looking imaging
geometry of SAR systems flooded areas might be
underestimated in urban areas.

Figure 3: Study area in Mozambique. Map data © Google

Earth and OpenStreetMap contributors.

In total 75 EO data are used, which cover the area of

interest (AOI) within the time 02/03/-07/04/2019:
 S-1: 33 Sentinel-1 GRD Interferometric Wideswath (IW)
data acquired on 11 days (spatial resolution ~20m)
 S-2: 12 L1C Sentinel-2 scenes acquired on 12 days (36
tiles, spatial res. 10m)
 TerraSAR-X: 4 ScanSAR data sets (spatial res. 18.5m)
 Landsat-8: 2 scenes of the Operational Land Imager (OLI)
onboard Landsat-8 (spatial res. 30m)
3.2. Flood extent masks
Due to the predicted landfall of Idai the International
Charter “Space and Major Disasters”
(https://disasterscharter.org/) was activated on March 14, Figure 4: Results of the automatic multi-sensor flood system
2019. Flood extent masks have been automatically derived in Mozambique: a) S-1 of 19/03/2019 (© ESA); b)
from S-1, TerraSAR-X, S-2, and Landsat-8 data using the TerraSAR-X of 23/03/2019 (© DLR); c) S-2 (RGB: Red,
flood monitoring system. The 37-day period between Green, Blue) of 25/03/2019 (© ESA) and d-f) derived
02/03/2019-07/04/2019 is covered in total on 25 days by single-temporal flood extents and reference water masks
these EO missions. Due to the imaging capability from Shuttle Radar Topography Mission Water Body Data
irrespective of atmospheric conditions the radar missions S- (SWBD) [4] and data of HOT. A mosaic of S-2 tiles
1 and TerraSAR-X allow a systematic monitoring of the acquired on 02/12/2012 serves as background in d-f.
flood extent and of the dynamic of the inundation. S-2 Modified from Rättich et al. (2020).
acquisitions are highly cloud covered on the days
during/after the landfall. However S-2 data, together with 3.3. Flood duration products
Landsat-8, provide an additional value to monitor the The proposed flood monitoring system is able to
recession of the water extent after the movement of the automatically generate the TFD and BFD products as well as
cyclone from 17/03/2019 onwards. Fig. 4 depicts examples a corresponding FDQ in NRT based on flood masks derived
of flood extent products based on S-1, TerraSAR-X, and S-2 from the respective flood services. In Fig. 5a the BFD is
data. visualized computed at the peak of the flooding on
10 radar acquisitions (9 S-1 and 1 TerraSAR-X data) 19/03/2019. At this stage especially the central and southern
cover the whole AOI and are therefore used to calculate the part of the AOI is already flooded since ~7 days. This shows
dynamic of the flooding over the whole period. S-1-based that regions have already been flooded some days before the
results on 14/03/2019 (03:09 UTC) show that also before landfall of Idai. This product can be used as indicator for the

3245

Authorized licensed use limited to: Institut Teknologi Sepuluh Nopember. Downloaded on March 21,2024 at 11:58:24 UTC from IEEE Xplore. Restrictions apply.
duration of an ongoing flood event and should be assessable weeks and ~130 km² for 2-3 weeks. The TFD can be used as
for emergency responders in NRT in order to provide damage indicator for insurance companies providing climate
information about the persistency of an ongoing flood risk insurance products or be useful for long-term
situation. The TFD is depicted for the observation period comparisons of annually recurring hydrological phenomena.
02/03/2019 – 07/04/2019 in Fig. 5b. It considers all flooded The FDQ for the TFD product is shown in Fig. 5c. It shows
pixels during this time period. Therefore, this product shows relative low uncertainty values. Due to the high temporal
a larger inundation extent than the BFD. Fig. 5b shows coverage of EO acquisitions during this disaster, relative
extensive inundations of 14-20 days in the southern and high certainty values between 1 and 19 are achieved for
northern part (along Pungwe river) of the AOI. In total an most regions. Higher uncertainty values are mainly related to
area of ~2,500 km² was inundated during 02/03/2019– the occurrence of clouds on one or more acquisition dates of
07/04/2019. More than 1,000 km² were flooded on only 1-2 the optical data, which cause a lower availability of valid
days and > 1,700 km² less than one week (1-7 days). Other pixels and the prolongation of observation gaps.
regions were inundated longer - more than 600 km² for 1-2

Figure 5: Flood duration products over the AOI in Mozambique derived from Sentinel-1, TerraSAR-X, Sentinel-2, and
Landsat-8 data: a) BFD (02/03/2019-19/03/2019), TFD (02/03/2019-07/04/2019), and FDQ for the TFD (02/03/2019-
07/04/2019). The reference water extent is based on SWBD [4] and data of HOT. Modified from Rättich et al. (2020).

5. CONCLUSION Photogrammetry and Remote Sensing, vol. 104, pp. 203-212,

2015.
In this study an automatic system for NRT flood extent and
duration mapping based on multi-sensor EO imagery is [2] Martinis, S., S. Plank, and K. Cwik, “The use of Sentinel-1
time-Series data to improve flood Monitoring in arid areas”,
presented. Flood masks are automatically derived using Remote Sensing, vol. 10 (582), pp. 1-13, 2018.
flood services based on S-1, S-2, TerraSAR-X, and Landsat-
8 data. Two different flood duration products are computed [3] Rättich, M., S. Martinis, M. Wieland, “Automatic flood
in order to provide 1) NRT information about the duration duration estimation based on multi-sensor satellite data. Remote
of an ongoing inundation and 2) the total duration of flood Sensing, vol. 12, no. 4, pp. 1-19, 2020.
coverage for a pre-defined time period. Furthermore, a
Flood Duration Quality layer is computed, which gives [4] SWBD, “Shuttle radar topography mission water body data
information about the reliability of the flood duration set”, http://www.dds.cr.usgs.gov/srtm/version2_1/SWBD/
(accessed 09.01.2020).
products by taking into account pre-event, co-event and,
post-event uncertainties. The proposed system is [5] Twele, A., W. Cao, S. Plank, and S. Martinis, “Sentinel-1
demonstrated on a flood situation related to cyclone Idai in based flood mapping: a fully automated processing chain,”
March 2019 in Mozambique. International Journal of Remote Sensing, 2016, vol. 37, no. 13, pp.
2990-3004, 2016.
6. REFERENCES
[6] Wieland, M., S. Martinis, “A modular processing chain for
[1] Martinis, S., J. Kersten, and A. Twele, “A fully automated automated flood monitoring from multi-spectral satellite data,”
TerraSAR-X based flood service”, ISPRS Journal of Remote Sensing, vol. 11, no. 9, pp. 1-23, 2019.

3246

Authorized licensed use limited to: Institut Teknologi Sepuluh Nopember. Downloaded on March 21,2024 at 11:58:24 UTC from IEEE Xplore. Restrictions apply.