欢迎访问一起赢论文辅导网
机械论文
当前位置:首页 > 机械论文
Multi‐temporal synthetic aperture radar flood mapping using change detection
来源:一起赢论文网     日期:2019-03-09     浏览数:1878     【 字体:

Introduction

Satellite imagery can determine the extent of flooding over large geographical areas, providing an advantage over in situ data sources where the information can have limited spatial and temporal resolution whilst being costly to acquire. In recent years the quantity and quality of satellite products available to stakeholders during and after an event has greatly improved. The Sentinel series of satellites is a prime example of this, producing data with high spatial and temporal resolutions that is free to download. These advances in satellite data sets have led to the development of near realtime, automated flood mapping algorithms (Matgen et al., 2011; Martinis et al., 2015; Twele et al., 2016). As well as providing vital information during an emergency, information derived from satellite imagery can be used to calibrate and validate hydrodynamic models, improving the predictive accuracy, and subsequently increasing stakeholder's understanding of flood dynamics (Schumann et al., 2009; Grimaldi et al., 2016).

 

Two types of satellite imagery are available at spatial resolutions suitable for monitoring surface water dynamics: optical and synthetic aperture radar (SAR). However, the acquisition and image properties vary, providing challenges when used to monitor hydrology. Optical sensors, such as those onboard the Landsat8 and Sentinel2 satellites, collect data across a variety of spectral bands ranging from the visible spectrum through to shortwave infrared (SWIR). Different land covers display specific reflective characteristics for each spectral band, which can be used to identify areas of water (Xu, 2006; Feyisa et al., 2014). Optical sensors are passive, with the images capturing the solar reflectance of the earth's surface or atmosphere, resulting in the sensor being unable to penetrate cloud cover (Gan et al., 2012). This is the main disadvantage of optical satellites, with flood events potentially occurring without any images being captured, making these sensors a poor choice for monitoring.

 

SAR systems, such as those onboard the Sentinel1 satellites, are active sensors which emit a radar pulse and record the land surface return at the satellite. They provide an advantage over optical sensors by enabling collection of data through cloud cover and during the night (Alsdorf et al., 2007; Schlaffer et al., 2015). The strength of the radar return is dependent on a number of factors, notably surface roughness, dielectric properties, and local topography in relation to the radar look angle (Brivio et al., 2002; Gan et al., 2012). Water bodies are a specular reflector of the radar pulse, resulting in minimal signal returned to the satellite (Jung et al., 2010; Schlaffer et al., 2015). Various methods have been used within the literature to delineate water from SAR data, either as a singular process or in combination. These include histogram thresholding (Brivio et al., 2002; Henry et al., 2006; Brown et al., 2016), fuzzy classification (Martinis et al., 2015; Twele et al., 2016), region growing (Matgen et al., 2011; Mason et al., 2012; Martinis et al., 2015; Twele et al., 2016), and texture analysis (Pradhan et al., 2014). Unlike the above processes, which use a single SAR image, change detection highlights the temporal changes in land cover by comparing the flood scene to a previous dry image (Giustarini et al., 2013; Schlaffer et al., 2015). The difference between the images can be combined with other image segmentation techniques to identify areas producing an unusually low backscatter response, improving the reliability of the flood delineation when compared to the single image methodologies (Matgen et al., 2011).

 

Despite the operational advantages of SAR compared to optical systems, there are challenges in identifying flooding. Roughening of the water surface, created by heavy rainfall or wind, can cause backscattering of the radar signal, increasing the possibility of inundated areas not being highlighted (Alsdorf et al., 2007; Jung et al., 2010). SAR systems are side looking and, depending on the incidence angle, terrain features can produce radar shadow, overlaying, and foreshadowing (Rees, 2000; Giustarini et al., 2013). In particular, radar shadow can provide difficulties in hydrological studies, creating anomalous dark areas within the radar image which can be misclassified as water. Identification of flooding can also be problematic in areas where other structures protrude the water surface and interact with the radar signal. This causes the doublebounce effect, with the signal reflecting off the water surface before interacting with the vertical structure, creating a corner reflector, and a strong return signal at the satellite (Horritt et al., 2001; Jung et al., 2010; Giustarini et al., 2013). The high density of buildings in urban areas can cause both radar shadow and doublebounce, limiting the delineation of flood extents without the use of expensive ultrahigh resolution SAR data, such as TerraSARX, RADARSAT2 or COSMOSkyMed (Giustarini et al., 2013; Pulvirenti et al., 2016). However, the ability of SAR sensors to penetrate cloud cover, along with the improvements in the spatial and temporal resolution of freely available data sets, demonstrates the key role SAR has in rural flood monitoring and management.

 

The aim of this research is to move towards near realtime determination of flood extents using multitemporal satellite SAR data sets, with methods that are both quick and easy to apply over large areas as new data becomes available. The case study of the 20152016 UK winter storm season has been selected to test the reliability of the process. The derived flood extent will be validated against cloud free optical data and modelled flood maps. The results will provide an insight into the flood extent change over a 5week period, providing information on flood dynamics, as well as highlighting pluvial flooding away from the floodplains.

 

Methodology

Location

 

A 400 km2 study area was selected in Yorkshire, UK, for testing of the proposed methodology. The region, shown in Figure 1, stretches from the south of York down to Selby, and west beyond Tadcaster. The area is largely rural, with two major rivers flowing through it: the Wharfe and Ouse. The region suffered from spatially and temporally variable flooding during December 2015 and January 2016, when storms Desmond (56 December 2015), Eva (24 December 2015), and Frank (2930 December 2015) brought widespread rainfall across northern UK.

 

 image

Figure 1

Open in figure viewerPowerPoint

The 400 km2 study region, shown by the red box, for which flood extents have been determined for December 2015 and January 2016. Background and location maps throughout article © Crown Copyright/database right 2017. An Ordnance Survey/EDINA supplied service.

Data

 

The 13 Sentinel1 SAR images collected over the study region between 5 December 2015 and 10 January 2016 are listed in Table 1. Radiometrically calibrated and terrain corrected Sentinel1 images are stored within Google Earth Engine (GEE), which provides free cloud computing facilities for research, with the change detection processing completed within this infrastructure (GEE, 2015).

 

Table 1. List of Sentinel1 scenes used, the date, the percentage of study area covered, the satellite track ID and the number of images used to calculate the reference image

Sentinel1 image Date Footprint (%) Track ID No. of ref. images

S1A_IW_GRDH_1SDV_20151205T061404_20151205T061429_008903_00CBC9_2323 05/12/2015 100 81 18

S1A_IW_GRDH_1SDV_20151208T174942_20151208T175007_008954_00CD3E_349B 08/12/2015 100 132 4

S1A_IW_GRDH_1SDV_20151210T062205_20151210T062230_008976_00CDE1_1951 10/12/2015 100 154 10

S1A_IW_GRDH_1SDV_20151213T175808_20151213T175833_009027_00CF27_4F38 13/12/2015 68.1 30 7

S1A_IW_GRDH_1SDV_20151217T061404_20151217T061433_009078_00D09B_ECA6 17/12/2015 100 81 18

S1A_IW_GRDH_1SDV_20151220T174947_20151220T175012_009129_00D20A_C0F7 20/12/2015 100 132 4

S1A_IW_GRDH_1SDV_20151222T062204_20151222T062229_009151_00D2AF_17F0 22/12/2015 100 154 10

S1A_IW_GRDH_1SDV_20151225T175803_20151225T175828_009202_00D428_9464 25/12/2015 68.2 30 7

S1A_IW_GRDH_1SDV_20151229T061403_20151229T061428_009253_00D59B_CC2A 29/12/2015 100 81 18

S1A_IW_GRDH_1SDV_20160101T174941_20160101T175006_009304_00D70A_60DE 01/01/2016 100 132 4

S1A_IW_GRDH_1SDV_20160103T062204_20160103T062229_009326_00D7AC_C9F2 03/01/2016 100 154 10

S1A_IW_GRDH_1SDV_20160106T175807_20160106T175832_009377_00D920_8394 06/01/2016 68.4 30 7

S1A_IW_GRDH_1SDV_20160110T061404_20160110T061433_009428_00DA93_B5C5 10/01/2016 100 81 18

Sentinel1, part of the European Space Agency Copernicus programme, consists of two satellites launched on 3 April 2014 and 22 April 2016. The satellites are in opposite polar sunsynchronous orbits at an altitude of 693 km, with a repeat cycle of 12 days, containing 175 orbits. This results in a repeat frequency of 24 hours at high latitudes and 3 days at the equator. The SAR system operates within Cband (5.407 GHz) frequencies in one of four acquisition modes: Stripmap (SM), Interferometric Wide swath (IW), ExtraWide swath (EW), and Wave (WV). IW is the default mode over land, operating under the TOPSAR (Terrain Observation with Progressive Scans SAR) principle (Geudtner et al., 2014). During acquisition the radar scans in both the azimuth and range directions simultaneously to provide three subswaths with a 2 km overlap (De Zan and Guarnieri, 2006). Each subswath contains six bursts, which are processed individually as single look complex (SLC) scenes, before being resampled to 10 × 10 m pixel spacing, deburst and merged into one tile. Data is collected in 250 km swaths at incidence angles between 29.1° and 46.0°, providing a ground resolution of 5 × 20 m (range × azimuth). The user guide (https://sentinel.esa.int/web/sentinel/userguides/sentinel1sar) provides further information on the satellite's acquisition parameters.

 

The polarisation of SAR images refers to the geometric plane that the radar wavelength is transmitted and received along. In most systems these are either horizontal (H) or vertical (V) in relation to the satellite antenna, creating four common polarisations: HH, HV, VH, and VV. Although each polarisation can be used for flood delineation, the backscatter characteristics of the radar signal varies, impacting the accuracy of the inundation maps produced. Manjusree et al. (2012) compared the four polarisations, concluding that HH has the greatest potential for delineating flooding consistently and accurately, results mirrored in other research (Henry et al., 2006; Brisco et al., 2008). Sentinel1 collects images in VH and VV polarisation when in IW mode, both of which have the potential for classification errors. Crosspolarised data (VH and HV) produces a wider range of backscatter values from vegetated land surfaces compared to copolarised data (VV and HH), leading to potential overlap with the low backscatter values associated with water, causing misclassification of land as flooded (Manjusree et al., 2012; Twele et al., 2016). VV polarised wavelengths are more susceptible to roughening of the water surface, commonly caused by wind or rain, increasing the backscatter return to the satellite, resulting in inundation not being identified (Manjusree et al., 2012). The limitations of each polarisation as environmental conditions vary requires acknowledgement when using Sentinel1 for flood mapping. Previous research concluded that VV provides a slight advantage when identifying flooding when using Sentinel1 data (Twele et al., 2016). To allow for further comparison both polarisations have been processed using the same methods within this study.

 

Within the methodology a terrain filter is applied to remove areas where the topographical location suggests that flooding is unlikely, but where SAR image acquisition may result in misclassification. For this the Ordnance Survey 5 m Digital Terrain Model (DTM) was used to create Height above Nearest Drainage (HAND) and slope data sets. The slope aspect of the filter is required to remove areas of radar shadow, found when large vertical structures limit the ability of the SAR system to record data from the lee of the feature. The minimal radar response in these areas is similar to that of flat water. The HAND data set represents the topographic difference between a pixel and its hydrologically determined nearest water course (Rennó et al., 2008; Nobre et al., 2011). The addition of HAND reduces the impact of the slope filter in the lowlands by including features such as river banks, which would be otherwise removed. For this project a HAND threshold of 20 m, along with 3° slope, were combined to create the terrain filter.

 

A cloud free satellite optical image was collected by Sentinel2 on 29 December 2015, 12 hours after the Sentinel1 pass. The use of optical imagery to validate SAR water extractions has become common practice, despite the potential errors in classifying water using optical data. However, the lack of in situ data to act as a reference means the Sentinel2 image has been used to validate the SAR flood extents. To extract the water bodies from the optical image the modified normalised difference water index (MNDWI) was applied, defined by Xu (2006) as:

 

urn:x-wiley:1753318X:media:jfr312303:jfr312303-math-0001(1)

with band 3 and band 11 representing the green and SWIR wavelengths within the Sentinel2 instrumentation. The MNDWI highlights the strong absorption of SWIR radiation by water bodies, improving on other water extraction indices, notably the normalised difference water index (NDWI) (McFeeters, 1996), by providing better separation between water and urban areas. The MNDWI data set can theoretically be segmented at zero to identify areas of water, however, differences in sensor acquisition parameters and geographical image characteristics can create a differing range of values, necessitating the need for individual image thresholding. To achieve this Otsu's method was employed, maximising the variance between the water and land classes (Otsu, 1979). The SAR and MNDWI flood extents have been compared, with the producer's (flooding misclassified as land), user's (land misclassified as flood) and total accuracies calculated, along with Cohen's kappa coefficient (κ – agreement that is not caused by chance).

The Environment Agency Flood Maps for Planning (EA FMP) contain modelled indicative results of areas likely to be inundated during a 100year river or 200year sea flood event (known as Zone 3, referred as 100year event henceforth), as well as a 1000year event from either source (Zone 2) (Porter and Demeritt, 2012). The extents have been modelled using a DTM with the flood defences removed, allowing for a subsequent data set highlighting the areas protected by the current flood defences during a 100year event. A further map showing the areas designated for storage areas is available, with these locations used to attenuate the flood peak in vulnerable areas. The SARderived flood extents have been compared to the EA FMP, providing a comparison between the remote sensing data and modelled results. Within the study region, a 100year flood would inundate 53.4 km2 of the area (13.4% of the study region), including 10.4 km2 classed as water storage areas (2.6%), with an additional 24.9 km2 being actively protected by defences (6.2%).

 

An aerial photograph, taken on 27 December 2015 by the National Police Air Service (NPAS) Carr Gate helicopter, has been used to provide supplementary information about the hydrological conditions in the region prior to the satellite crossing. The image was taken as the helicopter was to the west of the study region, close to Tadcaster on the river Wharfe. The image looks eastwards, towards the confluence of the Ouse and the Wharfe.

 

Processing steps

 

A change detection and thresholding (CDAT) methodology, adapted from Long et al. (2014), was used to determine the flooding extent. Figure 2 provides a diagrammatic overview of the workflow. The first step requires a nonflood reference image for change detection. Selection of this image can influence the outcome, with seasonal differences in land use and variances in satellite acquisition parameters (e.g. orbit direction and incidence angle) requiring consideration (Hostache et al., 2012). The reference images in this study were calculated using a collection of 39 previous Sentinel1 images, dated between 3 July 2015 and 5 November 2015. Ideally the time period used to create the reference collection would be similar to that of the flooding. However, due to the relatively short time Sentinel1 has been operational, the majority of winter images suffer from either flooding or poor preprocessing within GEE, leading to the inclusion of summer images to ensure coverage for each satellite track. For each flood scene analysed, the images with the same satellite track are selected, with the median value from this subset taken for each pixel to create the final reference image. Four different satellite tracks provide coverage over the study area, with the number of images in each reference image collection ranging from 4 to 18, as summarised in Table 1.

 

 image

Figure 2

Open in figure viewerPowerPoint

Workflow used to extract the flooding extents for this study. μ and σ represent the mean and standard deviation.

SAR suffers from speckle due to the variation in the radar return within a pixel caused by multiple scattering sources, such as vegetation (Esch et al., 2011; Giustarini et al., 2015). Filters can be applied to remove speckle, leaving a smoother image that can be used more accurately in further processing. Both the flood and reference images had a median 5 × 5 pixel filter applied for this purpose. The difference between the flood and reference image is calculated under the premise that change detection highlights variations in the radar return to the satellite, and by proxy changes in land cover or conditions. It is expected that flooding will cause a large negative difference due to the specular reflection of the radar signal by water, compared to the normal, stronger land backscatter response.

 

The difference image is subsequently filtered based on terrain, with composition and parameters of the filter described previously. The application of the filter removes just over 8 km2, or 2%, of the region. A threshold approach is applied to the difference image to extract the largest negative change in backscatter, thus highlighting areas most likely to be inundated. Long et al. (2014) determined the ideal threshold to be:

 

urn:x-wiley:1753318X:media:jfr312303:jfr312303-math-0002(2)

where P F are the pixels identified as flooded, μ and σ the mean and standard deviation of D, the difference image, and f c is a coefficient. Optimal f c was found to be 1.5 by Long et al.

A second filter was applied to the extracted inundation extent based on the flooded SAR image, an additional step compared to the original CDAT methodology. This is due to seasonal changes in land cover occasionally producing similar decreases in backscatter as flooding within the difference image. For this filter, a global threshold for the landwater boundary was defined based on the histograms of the SAR images used in the study, with only the areas identified as flooding by both the SAR threshold and the CDAT process used as the final flood extent.

 

The results are mapped, allowing direct comparisons between the two polarisations, as well as the Sentinel2 optical data set and the EA FMP. An estimate is also made of the number of days each pixel was inundated during the 37day study period. Each satellite image has been allocated a number of days, calculated as an even distribution of the time between the preceding and following satellite passes. For each pixel, the image scores for the dates when flooding has been identified are summed to provide an estimate for the number of days the pixel was inundated. The study area has been subdivided for this purpose, with different weightings required when the images do not cover the full region.

 

Results

Polarisation comparison

 

Both polarisations display a similar sequence for the amount of flooding throughout the study period (Figure 3). The image collected on 29 December provides the maximum flood extent for both polarisations, showing the aftermath of storms Eva and Frank. On this date, 6.7% and 6.1% of the study region were inundated for VH and VV, respectively. Preceding this date is a slight downwards trend in flood extent from the initial image on 5 December, with extents of 2.3% (VH) and 2.0% (VV), to 25 December, with 0.7% (VH) and 1.0% (VV) of the region inundated. A decrease in flood extent is observed following 29 December, before an increase on 10 January to 4.9% (VH) and 4.2% (VV), the second greatest extent observed.

 

 image

Figure 3

Open in figure viewerPowerPoint

Percentage of region identified as flooded for the VH and VV polarisations. Dates with full satellite coverage are joined to show the approximate sequence of flooding. Other data points, labelled as partial, have 68% of the region covered by Sentinel1 and are likely to underestimate the extent of flooding.

Both VH and VV polarisations are available for all images, allowing a comparison of their ability to delineate flooding. The observed timeseries between the two data sets are similar, as seen in Figure 3. The satellite crossing on 17 December provides a match in the extent of flooding between the two polarisations, although only 80.4% of the identified areas correspond. All other images provide differing flooding extents between VH and VV, with an even split for the greatest estimator. Figure 4 shows the relationship between the two data sets. A strong linear distribution is observed, with an R 2 value of 0.87. At lower extents of flooding VV identifies a greater area of inundation, with the polarisations matching at 1.6%. As the extent of flooding grows, VH identifies an increasingly greater proportion of the region as inundated compared to VV, with 6.0% in VH equating to 5.1% in VV.

 

 image

Figure 4

Open in figure viewerPowerPoint

Relationship between the percentages of the study region identified as flooded for the VH and VV polarisations. Grey line represents y = x as reference.

Two dates show a considerable difference between the two polarisations. On 20 December, the flood extent from VH (1.9%) is almost double that of VV (1.0%), with the VH identifying potential pluvial flooding that is missed by the VV. Similarly on 1 January, VV (2.9%) identifies just 61.7% of the flood extent as estimated from VH (4.7%). On this date, the difference is largely within the main body of flooding, with the VH identifying a uniform water surface compared to the smaller separate areas seen in the VV (Figure 5). It can be hypothesised that the lack of consistency in the VV backscatter response is caused by the wind roughening of the water surface.

 

 image

Figure 5

Open in figure viewerPowerPoint

Comparison of the flood extents from 1 January 2016. (Left) VH SAR image. (Middle) VV SAR image. (Right) Derived flood extents. Note VV shows an inconsistent body of water compared to VH due to increased backscatter on the water surface in the SAR image, likely caused by wind roughening of the water.

Validation

 

The MNDWI water extent of the Sentinel2 optical image of 29 December has been used to validate the flood extents from the two polarisations of the SAR image collected on the same day. The accuracies of both polarisations are shown in Table 2. Producer's accuracy for identifying flooded pixels is slightly better with VH, showing a greater inclusion of Sentinel2 identified water pixels in the flood extent. However, VV produces 94.3% user's accuracy compared to 87.0% for VH, showing less misclassification of land as water using this polarisation. Overall the total accuracies are similar with 0.4% difference. The kappa coefficient (κ) varies from 0.778 for VH to 0.799 for VV, showing a good relationship between the optical result and the two polarisations, with minimal correlation caused by chance.

 

Table 2. Error matrices showing the accuracy of the methodology for both polarisations. The MNDWI computed from a Sentinel2 scene acted as a reference data set. Italic values represent km2 and bold values are percentages. Total accuracy for VH was 97%, with a Cohen's κ of 0.778 (0 = relationship is chance, 1 = perfect relationship). Total accuracy for VV was 97.4%, with a κ of 0.799

VH Reference Sentinel2

SAR Sentinel1 No flood Flood SAR total User's %

No flood 364.4 8.7 373.1 97.7

Flood 3.5 23.4 26.9 87.0

Ref. total 367.9 32.1 400.0

Producer's % 99.0 73.0

VV Reference Sentinel2

SAR Sentinel1 No flood Flood SAR total User's %

No flood 366.5 9.2 375.7 97.6

Flood 1.4 22.9 24.3 94.3

Ref. total 367.9 32.1 400.0

Producer's % 99.6 71.5

Figure 6 provides a mapped comparison between flood extents from the SAR and optical data sets. There is good correlation between the three data sets for the large area of water which represents the inundated floodplain next to the Wharfe and Ouse rivers. The differences between the data sets can be characterised in four ways: permanent water bodies identified in the optical image but not in the SAR, misclassification of shadow areas as water within the optical image, extraction of the edges of flat manmade features within the SAR, and potential misclassification of land within the VH SAR image.

 

 image

Figure 6

Open in figure viewerPowerPoint

Validation of the results for a subset of the region against the Sentinel2 image. (Top left) True colour composite Sentinel2 satellite image for 29 December. (Top middle) Sentinel1 SAR image for 29 December, VH polarisation. The location of the described airport is shown by the red box. (Top right) VV polarised SAR image. (Bottom left) MNDWI, calculated from the Sentinel2 image, with blue representing water. (Bottom middle) Comparison of extracted flood extents, with blue representing those from VH SAR, red from optical and black represents areas identified in both. (Bottom right) VV flood extents compared to reference data set.

Both VH and VV polarisations have identified the edges of some urban features as flooded. This is most notable with the flat tarmac associated with an airport, highlighted in Figure 6, which provides a similar specular reflectance as water. There are matching flood extents between the Sentinel2 validation data set and the VV polarisation in these areas on 29 December. The VH polarisation identifies additional sections of the runway edges as inundated compared to the other imagery. The suggested flooding may be correct, however, caution is required due to the potential for misclassification around these features.

 

Areas of pluvial flooding are highlighted in locations away from the floodplains. These areas, likely to be agricultural fields that have become partially inundated, are more readily identified with the VH polarisation. Both polarisations identify smaller fields not classed as water in the MNDWI, with VH also extracting some larger areas. The VV flood extents provide a better match to the MNDWI data set, suggesting the additional flood areas identified with the VH are unlikely to be accurate. However, the misclassification of shadows as water within the reference data set can complicate the accuracy assessment of the SARderived flood maps when using optical data as a reference.

 

An aerial image from 27 December, captured by the NPAS Carr Gate police helicopter, has been used to provide secondary validation of the results (Figure 7). The image has been geolocated based on road and railway locations, visible above the flood water in both data sets. Although the image is 2 days before the satellite crossing, the similarities are good, with the main flooded regions showing a match. Despite the lack of statistical metrics, there is still benefit in comparing the satellite data with other sources of imagery to confirm the results.

 

 image

Figure 7

Open in figure viewerPowerPoint

Comparison between an aerial photograph from 27 December (from @NPAS_CarrGate) (top), and the identified flood areas from VV polarised SAR on 29 December (bottom). Red square and arrow show approximate location and viewing direction of the helicopter. Locations A (railway embankment) and B (B1223) provide georeferencing examples. Point C shows The Foss joining the river Wharfe, and point D shows the confluence of the Wharfe and the Ouse. Field level flooding is visible in both data sets, with an example given at E. Differences in flood extent are potentially caused by the 2day time gap between the images.

Flood dynamics

 

The multiple satellite images of the region over the study period enabled tracking of the advance and retreat of the flood waters. A good example is the recession of the peak event on 29 December, through two satellite passes on 1 January and 3 January (Figure 8). At first inspection, the polarisations show a similar pattern of recession, particularly where the main body of flooding is concerned. There are three main areas where waters recede during the 5day period: to the east of the image on the Wharfe, towards the north of the image below York, and downstream of the confluence of the two rivers. However, as mentioned previously, VV polarisation produces an erroneously reduced flood extent for 1 January, with areas of flooding likely missing from this analysis.

 

 image

Figure 8

Open in figure viewerPowerPoint

The retreat of flood waters during the aftermath of storm Frank. VH (left) and VV (right) polarisations for satellite orbits on 29 December, 1 January and 3 January are shown.

The other main bodies of flooding around the Wharfe and Ouse display minimal change in surface area. However, it can be observed that on the 3 January locations within the wider flood boundary are being classed as land rather than flood, particularly along the river reach. This suggests a reduction in the depth of water, allowing features such as river banks to protrude the water surface.

 

To provide an overview of the most flood prone areas in the region an estimate has been made for the number of days each pixel was flooded during the 37day study period (Figure 9). Both polarisations display a similar pattern along the course of the rivers, with greatest inundation located within the floodplains of the Wharfe and Ouse before the confluence. Some of these areas are shown to have flooded for the whole study period. Downstream of the confluence flooding only occurred after the extreme rainfall experienced during storm Eva, when the peak flood extents are found. There are major differences with the mapping of inundated fields, with some areas identified as flooded for most of the study period within the VH map, whilst only being minimally highlighted by VV.

 

 image

Figure 9

Open in figure viewerPowerPoint

The estimated number of days each pixel is inundated for VH (left) and VV (right) polarisations. Total number of days in the study period is 37. Each image is assigned a number of days representing an even distribution of the time to the preceding and following satellite passes. For each pixel, the images identified as flooded have been summed to provide an estimate total days flooded.

Comparison with EA FMP

 

It has been reported that some rivers in the UK exceeded their 100year return period during the winter 20152016 floods (NHMP, 2016). Accordingly, the SARderived inundation extents have been compared to the EA FMP 100 year flood zone, as well as the areas designated for storage of flood water and protected by flood defences (Figure 10).

 

 image

Figure 10

Open in figure viewerPowerPoint

The EA FMP for a subset of the study region (left), showing 1 in 100 year flood, areas protected by flood defences, and flood storage areas outlines. Extracted flood extents for VH, VV and where they overlap for 29 December (right) for comparison.

An area of 53.4 km2 of the 100year flood area is undefended, and at risk of flooding during such an event. At peak flood 49.2% (VH) and 48.0% (VV) of these regions were inundated, suggesting the flooding was not a 100year event in this area at the time of the satellite pass (Figure 11). Within the 100year flood boundary, 10.4 km2 has been designated as water storage areas. At maximum, 79.3% and 79.1% of these areas were inundated for VH and VV, respectively, with Figure 10 suggesting the area downstream of the confluence was close to capacity. Arguably, the most important information within the EA FMP is the areas protected from the 100year event by the flood defences. At peak flood a total of 2.3 km2 became inundated using the VH data, with 2.2 km2 for VV, under 10% of the protected areas in the region (Figure 12). This is largely to the west of Figure 10. Overall, the timeseries for flooding for each designated area closely follows the overall flood sequence observed in Figure 3.

 

 image

Figure 11

Open in figure viewerPowerPoint

Area of flooding identified within the EA FMP 100 year flood boundary, and the proportion located in a designated flood storage area. Partial data points represent those without full satellite coverage. Total area for the 100year flood region is 53.4 km2 (13.4% of the study area), including 10.4 km2 designated as flood storage areas.

 image

Figure 12

Open in figure viewerPowerPoint

Amount of flooding within the protected areas identified in the EA FMP. Partial data points represent those without full satellite coverage. Protected areas cover 24.9 km2, or 6.2%, of the study region.

The EA FMP largely encompass the main bodies of flooding identified in the satellite data. The percentage of the flooding that occurred outside any of the designated FMP areas has been calculated for each date (Figure 13), with mean percentages of 18.0% (VH) and 9.5% (VV). There are some dates where identified flood outside of the FMP prediction is high. The 5 December (VH and VV), 20 December (VH), and 22 December (VH) have less than 75% of identified flooding within the FMP areas, with the SAR data all showing a large amount of pluvial flooding on these dates.

[返回]
上一篇:地球物理模式函数的SAR风速反演性能评估
下一篇:合成孔径雷达运动目标检测的研究进展