Progress in operational flood mapping using satellite synthetic aperture radar (SAR) and airborne light detection and ranging (LiDAR) data

Abstract

During flooding, operational tools for mapping flood extent and depth of water in flood-prone areas are required by those planning emergency response, including UK statutory agencies such as the Environment Agency. Satellite data have become a source of information to map and monitor floods, but many of the methods developed to process the data are unsuitable for accurate, near real-time production of flood information products. This paper describes a new semi-automated methodology developed to provide operational mapping of flood extent and flood depth using satellite synthetic aperture radar (SAR) data combined with light detection and ranging (LiDAR) elevation data. In this method, an analyst uses the flood boundary derived from 8 m spatial resolution satellite SAR data to estimate the flood surface elevation at points around a flooded area using a digital terrain model derived from LiDAR data. This method is compared to a simple satellite ‘SAR-only’ method for generating flood extent and alternative, automated methods of generating flood extent and depth that also used SAR and LiDAR. TerraSAR-X and SPOT 5 data were used from an area including the UK Somerset Levels which suffered a major flood event in February 2014. The new semi-automated method produced similar overall accuracy to the SAR-only method (P_o = 95.8% and P_o = 95.3%, respectively), but was more accurate at mapping flood extent where large vegetation or other objects appeared in the satellite SAR data. The automated methods were relatively inaccurate (overall accuracy ranged from P_o = 83.4% to P_o = 88.8%), but were used to identify where further work on improving the semi-automated-elevation method could be carried out. In addition to the flood extent information provided by the semi-automated-elevation method, flood surface elevation data were produced that could be used to estimated flood depths and flood volumes. The accuracy of the flood elevation surface was tested using LiDAR data acquired of the water surface during the flooding (root mean square error = 0.152 m). The paper discusses progress towards operational flood monitoring using SAR and LiDAR remote sensing products.

Keywords

Operational flood mapping synthetic aperture radar LiDAR satellite data

I Introduction

In the UK, the Environment Agency (EA) has responsibility for taking a strategic overview of the management of all sources of flooding and coastal erosion in England. The EA also has operational responsibility for managing the risk of flooding from main rivers, reservoirs, estuaries and the sea, as well as being a coastal erosion risk management authority. During a flood event the EA requires operational tools to help assess the extent, risk and potential damage from flood water.

Developing such operational methodologies requires a different approach than research driven studies. For operational remote sensing, the output is the critical driver. The process, though important to the output, is much less important than developing a methodology that meets operational requirements. User requirements in the UK are that flood outputs are available as soon as possible after data acquisition, are updated regularly and are accurate in terms of the location of flooding and the flood height. The time requirement from data acquisition to product delivery for emergency response end-users varies. Some end-users require near real-time delivery (e.g. to route emergency access avoiding flooding), while for others up to 24 hours may be an acceptable time frame (e.g. for some strategic planning). Methodologies may have to be developed over short timescales, as end-user requirements develop during a flood event. Ideally, output products such as flood extent or flood depth would be available minutes after acquisition, but this is currently not possible, as data supply and processing both take hours. While the development of methods that are complex or require specialist software may meet operational requirements, ideally techniques should be:

accurate;

simple to apply for a user with basic Geographical Information Systems (GIS) experience. Image processing experience is limited, so a method that uses basic GIS skills increases the pool of available people to process the data;

easily applicable to a wide variety of satellite data including SAR and optical data;

fast to apply. Automating a methodology can increase speed, but if this involves complex processing, a simple to apply, manual methodology may be quicker.

1 Remote sensing solution

A variety of remotely sensed systems have the potential for producing flood mapping in the UK, but key factors such as areal coverage, reliability of suitable timely data acquisition and spatial resolution will have an impact on the operational suitability of these systems. In many cases cloud cover prevents optical systems from reliably providing suitable coverage for flood mapping.

One of the best opportunities for operational flood monitoring comes from satellite synthetic aperture radar (SAR) systems. These transmit pulses of radio waves that ‘illuminate’ a scene. These radio waves are reflected at the surface and the amount of energy received at the sensor is measured and can be used to infer information about the surface type and structure. Water, and therefore flooding, has a low backscatter in wavelengths measured by typical space-borne SAR sensors compared with most other surfaces. This allows flooding to be mapped using SAR data (Pierdicca et al., 2013). Due to their ability to penetrate cloud and acquire useable imagery at night, SAR satellite systems provide a more consistent method of mapping flooding (Brown, 2014; García-Pintado et al., 2013; Mason et al., 2009; Richman 2014a, 2014b). A limitation of SAR satellite systems has been that until the commissioning of TerraSAR-X in 2007, regularly updated SAR satellite data with a spatial resolution finer than 10 m were not available (Martinis et al., 2013).

There are still issues with SAR data, as isolating flooded areas can be more difficult than using optical data (Gan et al., 2012). There are also problems during certain weather conditions. For example, strong winds create wavelets on flood water that increase backscatter, making flooding more difficult to isolate (Mason et al., 2009; Pierdicca et al., 2013). Despite this, SAR systems offer a more consistent method of acquiring data during times when optical systems cannot ‘see’ flooding due to cloud or cloud shadow.

In recent years SAR satellite systems such as TerraSAR-X/TanDEM-X, RADARSAT 2 and COSMO-SkyMed have enabled more timely, fine resolution mapping during a flood incident, as it is now possible to acquire data every 1 or 2 days over large areas at spatial resolutions finer than 10 m.

2 Mapping flooding using satellite SAR data

Single threshold method

The simplest method of identifying flooding in SAR data is to use a single threshold value within the SAR scene, below which a surface is classified as water (Henry et al., 2006). There are automated ways of identifying the threshold (Martinis et al., 2009), but the threshold will depend on a range of environmental factors, such as atmospheric conditions and wind, and system factors such as viewing angle (Pierdicca et al., 2013). This variability means that automated methods of identifying the threshold will be unlikely to produce consistently accurate results. There are problems associated with artefacts such as speckling in the data, which is caused by interference between multiple signals returning from the ground (Giustarini et al., 2015) and the impact of terrain or slope angle on backscatter (Gan et al., 2012). Some of these issues can be mitigated by using additional terrain information and calculating surface incident angles of the RADAR beam (O’Grady et al., 2013).

Baseline method

Another commonly used method is to compare a baseline (pre-flooded) SAR dataset with a dataset acquired during flooding and calculate the change in backscatter. By identifying areas where the backscatter has reduced and where the backscatter is low in the second image it is possible to identify areas that have become flooded. This method has the advantage that areas of permanent water such as lakes and ponds are excluded, isolating areas of flooding from those that consistently contain water. This is conceptually simple and easy to apply, but the method has several limitations as an operational technique. Temporal differences in SAR images, such as changes in crops or different system parameters such as polarisation, orbit, viewing angle and viewing direction can result in a lowered accuracy of output flood mapping (Hostache et al., 2012). An example of a lack of suitable archive data with similar parameters was the July 2007 TerraSAR-X data used to map the River Severn near Tewkesbury in several studies (Matgen et al., 2011; Schumann et al., 2011). This was the first TerraSAR-X dataset to be acquired of the area, as the TerraSAR-X system was not fully commissioned at this time. Indeed, Matgen et al. (2011) had to use a post-flood TerraSAR-X image for their backscatter change methodology.

Effects of vegetation

Both the single SAR image thresholding and backscatter change methods have limitations in terms of the mapping flood extent, particularly with flooding in areas with objects that have a surface expression and rise above the level of the terrain. For example, where tall vegetation occurs, the flood surface is often obscured by the vegetation, even though the terrain surface under the vegetation is flooded (Mason et al., 2009; Pierdicca et al., 2013). This has the potential to be a particular issue where there are large amounts of woodland in flood plains, a major factor in some UK catchments. While the effect of large vegetation is not necessarily a major issue for small areas, such as single trees or hedges, there is the requirement that users of flood extents can interpret this effect for some critical decisions, which would mean that extra training for end-users would be required. As potential end-users may be in multiple organisations, with varying amounts of training in geographic information and data, this is unlikely to occur and would greatly increase the complexity of disseminating flood data. This means that ideally the impacts of larger objects, such as misclassification of flooded areas due to the presence of trees, are removed or minimised in output flood extents.

There are fundamental issues that restrict satellite data from being used to observe flooding under objects such as trees. Although complex image processing techniques can be applied (e.g. Gan et al., 2012) these methodologies will always be restricted in terms of mapping flooding, as they cannot be used directly to observe all of the flood water. To map flooding where objects restrict observation using remote sensing, the most suitable methodology will require additional data and/or modelling.

There are various methodologies that have been developed that use a fine spatial resolution digital terrain model (DTM) to increase the accuracy of flood mapping with satellite SAR data. However, the majority involve complex techniques and/or require substantial processing power. For example, Mason et al. (2012) used fine-resolution light detection and ranging (LiDAR) data to predict radar shadow and layover and use this information to increase the accuracy of the flood prediction where shadowing and layover are present using a rule based approach. Schumann et al. (2011) used an active contour model region growing method, which is constrained by LiDAR elevation. This approach expands areas that are identified as flooded in satellite SAR data, reducing the impact of objects such as large vegetation. These types of method are less likely to be applied operationally, as they have complex modelling and computing requirements.

3 Use of DTMs to improve flood retrieval from SAR

The use of DTMs to increase the accuracy of flood extents is a non-trivial task and requires accurate modelling of the flood surface elevation, though there are simple techniques that can be applied to incorporate this in flood mapping. Identifying the flood elevation is most easily done at the flood boundary, where the elevation at the flood boundary derived from the DTM is equal to the flood surface elevation (Schumann et al., 2011). Identifying the terrain heights at the flood boundary at many points will allow a flood surface to be interpolated. The interpolated flood surface elevation can be subtracted from the DTM to generate flood depths and flood extents. This methodology has the potential to minimise the effects of tall vegetation in the output flood extent data. In addition, the flood depth information has the potential to be used by emergency flood management to assess accessibility, for example, due to flooding on roads, and to calculate flood volumes, which is essential if flood water is being removed through pumping. However, geo-positioning errors in the flood extent derived from the remotely sensed data and errors in the DTM will result in errors in the flood surface elevation and therefore flood extent (Schumann et al., 2011).

4 Aims

This paper examines methodologies partially developed, used and tested during flooding in Southern England between December 2013 and March 2014. During this time there was extensive flooding in Southern England, with flooding of major river systems such as the Thames and Severn, as well as low-lying areas such as the Somerset Levels (BBC, 2014). During this period there was a need to provide those involved in managing emergency response with timely flood extent and flood depth data, less than 12 hours after data acquisition.

The aim of this paper is to compare the suitability of methodologies using satellite SAR data with and without a fine-resolution DTM for operational flood extent mapping based on the criteria described in the Introduction: accurate, fast and simple to apply, and applicable to a wide variety of satellite data.

II Methodology

1 Study area and data

Flooding affected the target study area in the Somerset Levels, UK with large amounts of surface water present between early January and March 2014.

TanDEM-X (SAR) and SPOT 5 (optical and infra-red) satellite data were acquired on 22 February 2014 (Figure 1) of the Somerset Levels in England. Another TanDEM-X data set of the Somerset Levels was acquired on 27 February 2014 (Figure 2). The satellite data were used to generate flood extent maps and flood surface elevation maps and test the accuracy of the flood extent mapping.

Figure 1.

Study area (white line) in Somerset Levels, England for flood extent analysis. Image shows TanDEM-X data acquired on 22 February 2014. TanDEM-X (© 2014 German Aerospace Center (DLR), 2014 Astrium Services / Infoterra GmbH). TanDEM-X were provided under the International Charter Space and Major Disasters.

Figure 2.

Study area in Somerset Levels, England for flood surface elevation analysis. Image shows TanDEM-X data acquired on 27 February 2014 at 0635 UT, with LiDAR data acquired on 26 February 2014 between 1730 and 1900 UT overlaid. LIDAR data were acquired to determine flood surface elevation of key areas and were not a complete coverage due to operational constraints. TanDEM-X data © 2014 DLR e.V., Distribution Airbus DS / Infoterra GmbH.

The TanDEM-X data were acquired at 0626 UT on 22 February 2014 in ScanSAR Mode with HH polarisation. The data were supplied as an enhanced ellipsoid corrected and radiometrically enhanced product with a spatial resolution of 8 m. Speckle suppression was applied using the ERDAS Imagine 2011 default, using a Lee-Sigma filter with a three by three window size, a coefficient of variation of 0.2 and a coefficient of variation multiplier of 2. The TanDEM-X data were then re-projected from UTM WGS 84 and re-sampled to an Ordnance Survey Great Britain (OSGB) 8 m grid.

SPOT 5 data were acquired at 1042 UT on the 22 February 2014. The SPOT data were provided in UTM WGS 84 with Ortho Basic processing applied by Centre National d’Etudes Spatiales and were re-sampled to an Ordnance Survey Great Britain (OSGB) 10 m grid.

The satellite data acquired on the 22 February 2014 used in the study were acquired under licence for the International Charter Space and Major Disasters (http://www.disasterscharter.org/home). All data were supplied freely to the UK Environment Agency to assist in estimating flood extent. The TanDEM-X and SPOT 5 data used in this study were processed by the authors to provide mapping for the emergency response during the flooding. Data were processed and flood extents supplied on the same day as satellite data were acquired.

For the methods that used terrain data to generate flood extent, a LiDAR DTM was used. The airborne LiDAR data were acquired by the Environment Agency between 6 January 2009 and 5 March 2009, when flooding was not present in the area, using an Optech Gemini ALTM (ALTM = Airborne Laser Terrain Mapper), which has a wavelength of 1064 nm. Processing was applied to identify objects such as large vegetation and buildings and these objects were removed from the data. The resulting gaps were interpolated across using triangulated irregular network (TIN) to form a DTM with a reported vertical root mean square error (RMSE) of 0.1 m (Schumann et al., 2011) and a spatial resolution of 2 m.

Suitable data to test the accuracy of the flood surface elevation derived from satellite SAR data were not available for 22 February 2014. However, Optech Gemini ALTM LiDAR data were acquired during the flooding on 26 February 2014 between 1730 and 1900 UT (Figure 2). These LIDAR data were acquired to test the accuracy of the flood surface elevation derived from TanDEM-X data programmed for the morning of 27 February 2014. Due to operational constraints, the LiDAR data were not acquired at the same time as the TanDEM-X data.

On 27 February 2014 TanDEM-X data were acquired at 0635 UT (Figure 2) in ScanSAR Mode with HH polarisation. The same speckle suppression, re-projection and resampling, as for the TanDEM-X on 22 February, were carried out.

2 Methods for acquiring flood extents

Geo-referencing error

To test the accuracy of the automated geo-positioning of the satellite data the geo-referencing error was estimated at 20 points in the vicinity of the study area, with Ordnance Survey 1:25,000 topographic data used as the reference. The RMSE of the SPOT data was 14.1 m. Using a different set of 20 points, the TanDEM-X geo-referencing error was also estimated. The RMSE of the TanDEM-X data was 11.9 m. Different points were used for the geo-referencing test, as differences in how surface features appear in optical and radar data made it difficult to identify the same points in the TanDEM-X and SPOT data.

Extracting flood extent from SPOT data

A normalised difference vegetation index (NDVI; Tucker, 1979) image was derived from the SPOT data. A manually derived threshold was applied to the image to identify flooded areas. This flood dataset was manually modified where the analyst believed there were errors. A standard methodology developed by the authors to filter remotely sensed flood data, contract expand flood (CEF), was used. The flood areas were shrunk using a three by three minimum value neighbourhood filter (no flood = 0, flood = 1). The flood areas were then expanded using a three by three maximum value neighbourhood filter. The CEF approach removes isolated pixels identified as flooding, but has a minimal impact on the shape and edge of flood extents. The output SPOT flood layer was used as the comparison reference dataset. It is acknowledged that this would not be 100% accurate and may have similar errors to some other satellite based approaches to mapping floods. Some of the general problems with satellite data identified above, such as areas of large vegetation being excluded from the flood layer, will apply to a flood dataset derived from SPOT data. However, ground data were not available and as the majority of the methods developed in this study used LiDAR elevation data to produce the final flood map, the SPOT data were felt to be broadly independent.

Deriving flood extents

Four different methods of calculating flood extents were applied to the TanDEM-X data: SAR-only, automated-elevation and SAR, slope-dependent-automated-elevation and SAR and semi-automated-elevation and SAR. These are detailed in the sections below.

SAR-only method

The SAR-only method involved applying a threshold to the TanDEM-X data, with values below the threshold considered as flooded. The threshold was determined manually by an interpreter. This was achieved by altering the threshold value iteratively and visualising the outputs. The threshold value that the interpreter thought maximised the accuracy of the output flood data map was applied. CEF filtering was applied to the flood data layer, as above. Manual editing of data was carried out to remove spurious areas of flooding, mainly due to terrain effects and object effects such as on the down-sensor side of woodland and buildings where shadowing occurs.

Automated-elevation method

The automated-elevation method used the SAR-only flood extent to identify the elevation at the edges of the flooding. Points were generated at 100 m intervals along the SAR-only flood extent. The elevation at each of these points was automatically extracted from the LiDAR DTM.

The flood boundary elevation points were then used to generate a flood surface using TIN interpolation. This subsampling of the flood extent results in an incomplete representation of the flood surface, as TIN interpolation does not extend beyond the outermost point. For this reason six additional elevation points were added to the TIN at corners of the study site. These points were allocated the value of the nearest neighbour. The TIN flood surface was subtracted from the LiDAR DTM to generate a flood depth dataset and from this a flood extent dataset. The flood extent was filtered using the CEF method described above.

Slope-dependent-automated-elevation method

In areas where the slope is steeper, positional errors in satellite positioning and boundary interpretation are more likely to result in a larger error in flood elevation estimation (Mason et al., 2009). For this reason, the automated-elevation methodology was repeated, but elevation points were excluded based on their slope value. Slope was derived from the LiDAR DTM and any points where the slope was greater than a specified threshold were excluded from the flood surface generation. The following slope threshold values in degrees were applied and used to generate the slope-dependent-automated-elevation method: 0.25, 0.5, 1, 2, 3, 4 and 5.

Semi-automated-elevation method

In the semi-automated-elevation method, the flood surface elevation was estimated by an analyst using the TanDEM-X backscatter image with LiDAR DTM contours overlaid. Elevation points were manually added to a TIN and the TIN was used to generate a flood elevation surface. The methodology after this point was the same as the automated-elevation method, but the flood extent was manually edited to remove areas erroneously identified as being flooded. The advantages of manually interpreting the flood boundary and linking this to LiDAR DTM elevation using the DTM, are that the initial stage in generating a flood layer is not required and positional errors in the satellite data are compensated for, as the shape of a elevation contour, as well as position could be used to estimate the boundary flood surface elevation. Slope thresholds were not used in this method.

3 Testing the flood surface elevation accuracy

When using a DTM to generate flood extent an additional product is produced, the flood surface elevation. This can be used to generate flood depth and flood volumes, parameters that are of considerable use in emergency response for determining accessibility and volumes of water to be pumped. The semi-automated-elevation method was used to generate flood extents and a flood surface elevation model derived from TanDEM-X data acquired on 27 February 2014. This flood surface elevation was compared to the flood surface elevation derived from LiDAR data flown on 26 February at the time of the flooding. The difference between the flood surface elevations measured using LIDAR data and modelled using semi-automated-elevation method were generated for the areas of overlapping data (Figure 2) that had been identified as having been flooded using the semi-automated-elevation method. From this elevation difference, an overall mean elevation difference and RMSE were generated.

III Results

1 Flood extent mapping results

Each of the four methods of generating flood extents from the SAR data was compared to the SPOT flood data, with a binary class assignment, “flood” and “no flood”, for each. A confusion matrix was generated for each of the methods and associated statistics, overall accuracy (P_o ), user’s accuracy and producer’s accuracy were calculated (Campbell, 2002; Storey and Congalton, 1986). The producer’s accuracy (Storey and Congalton, 1986) is the proportion of reference pixels that are correctly classified and provides an estimate of the proportion of pixels incorrectly omitted from a class and classified as another class during the classification. The user’s accuracy (Storey and Congalton, 1986) is an indication of how many pixels allocated to a particular class by the classification actually belong to that class and provides an estimate of the proportion of pixels from other classes that were incorrectly included or commissioned into a class during the classification.

The flood extents derived by the different methodologies are shown in Figures 3 to 6. The most accurate methods of generating the flood extents were the semi-automated-elevation (P_o = 95.8%; see Table 5) and SAR-only (P_o = 95.3%; Table 1). The automated-elevation method had a P_o of 83.4% (Table 2), while the slope-dependent-elevation method had P_o values between 85.5% (slope less than 5°) and 88.8% (slope less than 3°) (Table 3 and Table 4).

Figure 3.

Flood extents for 22 February 2014 derived using SAR-only method overlaid on TanDEM-X (© 2014 German Aerospace Center (DLR), 2014 Astrium Services / Infoterra GmbH). TanDEM-X were provided under the International Charter Space and Major Disasters.

Figure 4.

Flood extents for 22 February 2014 derived using automated-elevation method overlaid on TanDEM-X (© 2014 German Aerospace Center (DLR), 2014 Astrium Services / Infoterra GmbH). TanDEM-X were provided under the International Charter Space and Major Disasters.

Figure 5.

Flood extents for 22 February 2014 derived using slope-dependent-elevation method, with slopes less than or equal to 3° overlaid on TanDEM-X (© 2014 German Aerospace Center (DLR), 2014 Astrium Services / Infoterra GmbH). TanDEM-X were provided under the International Charter Space and Major Disasters.

Figure 6.

Flood extents for 22 February 2014 derived using semi-automated-elevation method overlaid on TanDEM-X (© 2014 German Aerospace Center (DLR), 2014 Astrium Services / Infoterra GmbH). TanDEM-X were provided under the International Charter Space and Major Disasters.

Table 1.

Confusion matrix for SAR-only method. P_o = 95.3%.

SAR-only	Reference – SPOT
SAR-only	No flood	Flood	User’s accuracy (%)
No flood	48772876	2034489	96.0
Flood	1461912	21745111	93.7
Producer’s accuracy (%)	97.1	91.4

Table 2.

Confusion matrix for automated-elevation method. P_o = 83.4%.

Automated-elevation	Reference – SPOT
Automated-elevation	No flood	Flood	User’s accuracy (%)
No flood	44592151	6632695	87.1
Flood	5642637	17146905	75.2
Producer’s accuracy (%)	88.8	72.1

Table 3.

Summary confusion matrix statistics for slope-dependent-elevation method.

Slope less than or equal to (°)	Producer’s accuracy		User’s accuracy		P_o
Slope less than or equal to (°)	No flood	Flood	No flood	Flood	P_o
0.25	97.2	66.6	86.0	91.7	87.3
0.5	96.9	70.0	87.2	91.3	88.2
1	95.5	72.7	88.1	88.5	88.2
2	94.6	76.1	89.3	87.0	88.7
3	94.0	78.0	90.0	85.9	88.8
4	89.4	77.8	89.5	77.6	85.7
5	89.2	77.6	89.4	77.2	85.4

Table 4.

Confusion matrix for most accurate slope-dependent-elevation method, with slope less than or equal to 3°. P_o = 88.8%.

Slope-dependent-elevation	Reference – SPOT
Slope-dependent-elevation	No flood	Flood	User’s accuracy (%)
No flood	47197401	5231747	90.0
Flood	3037387	18547853	85.9
Producer’s accuracy (%)	94.0	78.0

Table 5.

Confusion matrix for semi-automated-elevation method. P_o = 95.8%.

Semi-automated-elevation	Reference – SPOT
Semi-automated-elevation	No flood	Flood	User’s accuracy (%)
No flood	47402166	274767	99.4
Flood	2832622	23504833	89.2
Producer’s accuracy (%)	94.4	98.8

Though the SAR-only method had a similar overall accuracy value to the semi-automated-elevation method, the errors appear to be different. The SAR-only method appears to have a higher user’s accuracy for the flood class (93.7% compared to 89.2%; Tables 1 and 5). However, this may be an artefact of the SPOT comparison data, as hedgerows and trees appear not flooded in the SPOT and SAR-only data, but not in the semi-automated-elevation results (Figure 7). The areas with trees and hedgerows are likely to have similar terrain heights as the surrounding areas and would be likely to be flooded. This would indicate that errors in the SPOT derived map are matched by similar errors in SAR-only method and that the semi-automated-elevation method may be more accurate than the overall accuracy calculated in this study indicates.

Figure 7.

TanDEM-X data and flood extents on 22 February 2014 for subsection of the study area. (a) TanDEM-X data. (b) SAR-only derived flood extent. (c) Flood extent derived from SPOT data. (d) semi-automated-elevation derived flood extent. © 2014 German Aerospace Center (DLR), 2014 Astrium Services / Infoterra GmbH. TanDEM-X were provided under the International Charter Space and Major Disasters.

The automated-elevation method was the least accurate of those tested (P_o = 83.4%). As the semi-automated-elevation method was the most accurate, the use of LiDAR combined with satellite data has been shown to have potential to provide accurate flood mapping. The reason that the automated-elevation method failed to map the flooding as accurately as other methods is likely to be due to inaccurate positioning of the flood boundary in the input SAR-only flood extent. This will be partially due to incorrect classification of the flood extent at the boundary and may be due to geo-correction errors in the TanDEM-X data used to identify the flood extent. In areas with a larger variety in terrain elevation, positional errors in the predicted flood boundary position are likely to result in greater errors in the predicted flood surface elevation. This may be seen in the results from the slope-dependent-elevation method, as the flood class user’s accuracy when the slope was less than 0.25° was the highest (91.7%, Table 3) of all methods apart from the SAR-only (93.7%, Table 1). The automated-elevation method predicted flooding in several areas where there was no evidence from the satellite data that flooding had occurred and was not indicated in any of the other flood extent datasets (Figure 4).

Of the methodologies tested, the SAR-only was the fastest in terms of processing time, taking 125 minutes for the area in Figure 1 (Table 6). The automated-elevation and slope-dependent-elevation methods both rely on the outputs from the SAR-only method and are very similar methodologies, both taking 140 min in total. The slowest methodology was the semi-automated-elevation, taking 200 min.

Table 6.

Time taken for each of the methods of flood extent estimation for the area indicated in Figure 1. Time to the nearest 5 min.

Method	Time taken (min)
SAR-only	125
Automated-elevation	140
Slope-dependent-elevation	140
Semi-automated-elevation	200

2 Flood surface elevation accuracy

The mean difference in flood surface elevation between the LiDAR acquired during the flooding and the semi-automated-elevation method was very low at 0.033 m and an RMSE of 0.152 m. However, there was spatial variation in the offsets between the flood surface elevation values, as may be seen from Figure 8, with some areas having offsets between the LiDAR and semi-automated-elevation flood surface elevation values consistently greater than 0.25 m. Accurate data on the extent of hydrologically distinct areas was not available, but such information may have allowed areas to be processed separately, reducing the possibility of using flood surface elevations from one hydrologically distinct area to define the flood surface elevation of another area.

Figure 8.

Elevation differences between water surfaces directly measured from LIDAR data acquired on 26 February 2014 and that predicted using the semi-automated-elevation method using TanDEM-X data acquired on 27 February 2014. Difference data overlaid on TanDEM-X data (© 2014 DLR e.V., Distribution Airbus DS / Infoterra GmbH).

IV Discussion

1 Results

Of the methodologies, the semi-automated-elevation method was the most accurate (P_o = 95.8%), with a small increase in P_o (0.5%) compared with the SAR-only (Tables 1 and 5). Though this method took the longest processing time of those tested (Table 6), it provided an accurate flood surface elevation (RMSE = 0.152 m; Figure 8) making it suitable for flood water volume estimations, a critical factor when flood management involves pumping water from flooded areas. The automated-elevation and the slope-dependent-elevation methods tested in this paper produced much lower accuracy flood extents, with large contiguous areas incorrectly included in the predicted flooded areas (Figures 4 and 5, Tables 2, 3 and 4).

The time advantages to these methods (Table 6) were outweighed by the relative inaccuracy, making them unsuitable for operational use. However, automation could be an important factor in reducing time taken for processing, a critical issue in increasing the use of satellite based flood data in flood emergency response management. Determining why the points used to generate the flood elevation automatically did not represent the flood elevation accurately would be a first step to identifying which points could be used. This would potentially allow development of a methodology that generates some, but not all, of the flood elevation surface points automatically. It is likely that misregistration uncertainty between the elevation data and the satellite data should be taken into account (Schumann et al., 2011) and that the terrain within a potential misregistration footprint should be considered when determining if a particular elevation point is suitable to describe the flood surface elevation.

2 Methods

One advantage of the semi-automated-elevation method was that the analyst could take into account the shape of the flood boundary, which in some cases was as important as its position. Where the geo-referencing was less accurate or where the terrain was more complex, the shape of the flood boundary could be matched to a contour of similar shape. In several cases the position of the flood boundary in the satellite data did not match the position of the elevation contour used, but provided a way of minimising the impact of geo-referencing error in the output flood extent data. In these areas automating the extraction of flood surface elevation values would be more difficult.

It is possible that the semi-automated-elevation method was more accurate than estimated in this paper (P_o = 95.8%; Table 5). The SPOT flood extent, used as the baseline dataset, would have contained similar flood omission errors to the SAR-only method, as areas of flooding under large vegetation such as trees would have been likely to have been excluded from the predicted flood extents. However, the semi-automated-elevation method was more likely to predict that these areas were flooded. Data were not available to test the hypothesis that areas underneath objects such as large vegetation were accurately mapped using the semi-automated-elevation method. However, it is likely that areas of flooding not included in the SAR-only method due to the presence of large vegetation were correctly classified as being flooded using the semi-automated-elevation method.

All the techniques used were simple to apply for someone with GIS experience and many of the processes are easy to automate. The method with the most user intervention, the semi-automated-elevation method, was still simple to apply, though some experience using SAR data for flood mapping was useful.

Though the semi-automated-elevation method was applied using TanDEM-X, other platforms or sensors could be used to acquire the flooding data. Optical or SAR data could be used from satellite or airborne platforms. The methodology effectively changes the resolution of the flood extent from that of the data used to identify the flooding to that of the DTM. However, it is likely that there will be an inverse relationship between resolution of the satellite data and the accuracy of both the flood extent and the flood surface elevation (Schumann et al., 2011). Further work would be required to quantify these relationships, and the coarsest resolution that may be used operationally for flood mapping. This relationship is likely to be dependent on the complexity of the terrain and land cover. Coarser resolution data with a larger coverage would allow flood information to be derived over larger areas, probably at the cost of reducing the accuracy of the output flood data. However, if the increase in error is minimal, this may be a compromise worth accepting.

3 Recommendations

Of the methods tested, the semi-automated-elevation method was the slowest, but was the most accurate and it was used operationally to produce flood outlines, depths and volumes during the February 2014 flooding. However, for some emergency management and response there is a requirement to speed up all stages of the acquisition and processing chain. A shorter time between mission planning and satellite acquisition would allow targeting of areas that were considered most at risk or subject to the most significant flooding. The relative risks of flooding can change over short timescales when rainfall amounts or distributions are not as predicted. Faster download of satellite data and distribution of satellite data are required, with data providers currently supplying data over hours rather than minutes after acquisition. In addition, further work is required to speed up the process of converting image data to flood extent and depth.

One limitation of the semi-automated-elevation method is that it requires accurate elevation data of the majority of the flood plain, which is likely to be an issue in many countries. However, where precision, fine spatial resolution data are available, the semi-automated-elevation method can be used operationally. It also has the potential for further development, reducing processing time and increasing the accuracy of flood mapping. The semi-automated-elevation method was used operationally in the Somerset Levels in 2014 to reduce the errors in flood extent due to large vegetation, but initial work in other catchments shows that the method is less successful in heavily wooded catchments where the flood boundary cannot be observed in the satellite data. During the 2014 flooding, the urgency with which flood extent data were required meant that the SAR-only method was used to provide initial flood extent information and the semi-automated-elevation method was used to provide more accurate flood information and flood depths at a later stage. The semi-automated-elevation method also has limitations in urban areas, where the elevation at the flood edge is difficult to determine, as it may be on a vertical surface, which cannot be observed from satellite. There are also problems where temporary flood defences are erected, as the semi-automated-elevation method would predict flooding behind these defences. Further work is required to develop this method in areas with topography that is more characteristic of UK rivers, as the Somerset Levels contain much larger flood plains than are typical in the UK.

The flood outline data derived using the SAR-only and the semi-automated-elevation methods were used by the Environment Agency and other UK organisations at local and national level to provide an understanding of the impact of flooding and to prioritise resources. Flood surface elevation data from the semi-automated-elevation method were used to derive flood depth data, which in turn was used to derive flood volumes to help in the management of the flood water pumping operation (Wood M, personal communication, 8 October 2014).

V Conclusion

Of the methods tested in this paper, the semi-automated-elevation method was the most accurate at determining flood extent and could be used to accurately produce flood surface elevation data. The semi-automated-elevation method took 60% longer (200 min) to apply than a SAR only method (125 min) and was slightly more accurate (difference P_o = 0.5%). Though this increase in accuracy is not significant, it produced a flood surface elevation dataset, which the SAR-only method did not. This flood surface elevation data set was used to derive additional flood products, such as flood depth and volume. In addition, it is likely that some of the error reported in the semi-automated-elevation method was due to errors under large vegetation in the reference flood dataset derived from SPOT. This would result in an underestimate of the error in the SAR only method and an overestimate in the error in the semi-automated-elevation method.

The intervention by an analyst during processing greatly reduced output errors in the semi-automated-elevation method when compared to the automated methods. While this method took longer to apply than the more automated methods tested, the semi-automated-elevation method meets the criteria defined in the introduction. It is relatively accurate, simple to apply, has the potential to be applicable to a range of input satellite data and fast enough to be used during an emergency flood response. The use of the semi-automated-elevation method to produce outputs during February 2014 flood events highlights its suitability as an operational technique.

Footnotes

Acknowledgements

The satellite data used in the study were acquired under licence for the International Charter Space and Major Disasters (). TanDEM-X acquired on 22 February 2014 were provided by German Aerospace Centre (DLR), Astrium Services / Infoterra GmbH. SPOT 5 data acquired on 22 February 2014 were supplied by CNES, distribution SPOT Image S.A. TanDEM-X acquired on 22 February 2014 were supplied DLR e.V., distribution Airbus DS / Infoterra GmbH.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors

References

BBC (2014) 10 key moments of the UK winter storms. Available at: www.bbc.co.uk/news/uk-26170904 (accessed 16 December 2014).

Brown

(2014) International Charter Space and Major Disasters: Project Managers Report: Charter Activation 404 Flooding Southern England, UK. International Charter Space and Major Disasters.

Campbell

(2002) Introduction to Remote Sensing. 3rd ed. New York: Guilford Publications.

Gan

Zunic

Kuo

. (2012) Flood mapping of Danube River at Romania using single and multi-date ERS2-SAR images. International Journal of Applied Earth Observation and Geoinformation 18: 69–81.

García-Pintado

Neal

Mason

. (2013) Scheduling satellite-based SAR acquisition for sequential assimilation of water level observations into flood modelling. Journal of Hydrology 495: 252–266.

Giustarini

Vernieuwe

Verwaeren

. (2015) Accounting for image uncertainty in SAR-based flood mapping. International Journal of Applied Earth Observation and Geoinformation 34: 70–77.

Henry

Chastanet

Fellah

. (2006) ENVISAT multi-polarised ASAR data for flood mapping. International Journal of Remote Sensing 27: 1921–1929.

Hostache

Matgen

Wagner

(2012) Change detection approaches for flood extent mapping: How to select the most adequate reference image from online archives? International Journal of Applied Earth Observation and Geoinformation 19: 205–213.

Martinis

Twele

Strobl

. (2013) A multi-scale flood monitoring system based on fully automatic MODIS and TerraSAR-X processing chains. Remote Sensing 5: 5598–5619.

10.

Martinis

Twele

Voigt

(2009) Towards operational near real-time flood detection using a split-based automatic thresholding procedure on high resolution TerraSAR-X data. Natural Hazards and Earth System Sciences 9: 303–314.

11.

Mason

Bates

Dall’Amico

(2009) Calibration of uncertain flood inundation models using remotely sensed water levels. Journal of Hydrology 368: 224–236.

12.

Mason

Davenport

Neal

. (2012). Near real-time flood detection in urban and rural areas using high resolution synthetic aperture radar images. IEEE Transactions on Geoscience and Remote Sensing 50: 3041–3052.

13.

Matgen

Hostache

Schumann

. (2011) Towards an automated SAR-based flood monitoring system: lessons learned from two case studies. Physics and Chemistry of the Earth 36: 241–252.

14.

O’Grady

Leblanc

Gillieson

(2013) Relationship of local incidence angle with satellite radar backscatter for different surface conditions. International Journal of Applied Earth Observation and Geoinformation 24: 42–53.

15.

Pierdicca

Pulvirenti

Chini

. (2013) Observing floods from space: experience gained from COSMO-SkyMed observations. Acta Astronautica 84: 122–133.

16.

Richman

(2014a) International Charter Space and Major Disasters: Project Managers Report: Charter Activation 408 Flooding England, UK. International Charter Space and Major Disasters.

17.

Richman

(2014b) International Charter Space and Major Disasters: Project Managers Report: Charter Activation 401 Flooding Southern England, UK. International Charter Space and Major Disasters.

18.

Schumann

Neal

Mason

. (2011) The accuracy of sequential aerial photography and SAR data for observing urban flood dynamics, a case study of the UK summer 2007 floods. Remote Sensing of Environment 115: 2536–2546.

19.

Storey

Congalton

(1986) Accuracy assessment: a user’s perspective. P hotogrammetric Engineering and Remote Sensing 52: 397–399.

20.

Tucker

(1979) Red and photographic infrared linear combinations for monitoring vegetation. Remote Sensing of the Environment 8: 127–150.