Evaluation of Polarimetric Parameters for Flood Detection Using PALSAR-2 Quad-pol Data

Abstract

This study aims to evaluate the usability of polarimetric SAR measurements for discriminating water-covered area from other land cover types and to propose polarimetric parameters showing the better response to the flood. Flood-related changes in the polarimetric parameters were studied using the L-band PALSAR-2 quad-pol mode data acquired before and after the severe flood events occurred in Joso city, Japan. The experimental results showed that, among various polarimetric parameters, the HH-polarization intensity, the Shannon entropy, and the surfaces scattering component of model-based decomposition were found to be useful to discriminate water-covered areas from other land cover types. Particularly, an unsupervised change detection with the Shannon entropy provides the best result for an automated mapping of flood extents.

1. Introduction

With an increasing number of extreme climate eventssuch as heavy rainfall and storms,flood risk has been increased over the last few decades. In order to mitigate the flood risk, it is highly significant to understand the full extent of the damage. Remote sensing with satellite observation has been playing an important role in flood detection and monitoring. Particularly, Synthetic Aperture Radar (SAR) sensors are considered as one of the most useful tools for observations of flood events owing to their all-weather observation capability, whereas optical image is difficult to be obtained due to cloud coverage on the time of floods.

Various SAR sensors have been practically used for flood detection problems. Mason et al.(2012) proposed a near real-time flood detection algorithm including image segmentation techniques with high resolution synthetic aperture radar images, such as TerraSAR-X. Also, Mason et al. (2014) detected flooded urban area using double scattering of high resolution SARimages. Giustarini et al. (2013) used TerraSAR-X data with combination of thresholding, region growing, and change detection approaches to detect floods extent. Long et al.(2014) usedENVISAT/ASARandRadarsat-2 images for a flood detection based on change detection and histogram thresholding approaches.

Recently, in addition to the conventional single polarization SAR data, polarimetric SAR observations have been tested for a flood detection. Yamazak and iu (2016) used SARdata acquired from Phased Array type L-band SAR-2 (PALSAR-2) sensor onboard Japanese ALOS-2 satellite to delineate flooded area in Joso city, Japan by backscatter thresholding. Plank et al. (2017) detected floods around the Evros river near Greek and Turkish using fully polarimetric PALSAR2 data and dual-polarimetric Sentinel-1 data based on several classification scenarios. Previous studies have focused on exploiting polarimetric intensitiesfor better classification of flooded areas. However, polarimetric SARobservation provides not only intensity at different polarization channels but also correlation information between co- and cross-polarization channels. The polarimetric target decomposition methods can offer additional information on the type of scatterers and scattering processes.

This study aims to evaluate the usability of polarimetric SAR measurements for discriminating water-covered area from other land cover types and to propose polarimetric parameters showing the better response to the flood. Flood-related changes in the polarimetric parameters were studied using the L-band PALSAR-2 quad-pol mode data acquired before and after the flood events. In addition, unsupervised flood detection methods were evaluated for automatic mapping of flood extents.

2. Study area and data

The study area is Joso city located in Ibaraki prefecture, the northeast of Tokyo, Japan. A severe flood occurred in this area on September 11, 205 when bans of Kinugawa river collapsed by heavy rainfall. The inundated areas were gradually decreased, but lasted almost one week. Fig. 1(a)showsthe location of Joso city study area. Fig. 1(b) and Fig. 1(c) show Google Earth optical images obtained before (March 10, 2015) and after (September 11, 2015) the flood event.

Fig 1. (a) Location of Joso city study area, and Google Earth optical images obtained (b) before (Mar. 10, 2015) and (c) after (Sept. 11, 2015) the flood event. Boxed area in the optical image indicates target study site analyzed by the PALSAR-2 data.

In order to respond to the flood disaster, emergency observations were obtained by the PALSAR-2 sensors in several different modes. Among the emergency data sets, a fully polarimetric quad-pol mode data was acquired on September 16, 2015, about 5 days after flooding. Since the flooded area of the southern part of the Joso city (marked by a black rectangle in Fig. 1) was prominent even 5 days passed after flooding, this area was selected as the target study site. A pre-event polarimetric SAR data on August 11, 2015 was also obtained in this study. It is worth noting that the post-event SAR observation was carried out with left-looking mode instead of normal right-looking observation.

Table 1. Acquisition details of PALSAR-2 data

3. Polarimetric parameters for detection of water-covered area

1) Polarimetric parameters

Several poarimetric parameters hat have been widely used are introduced in this section. The basic measurement of polarimetric SAR system,such asthe 2×2 complex scattering matrix [S] or equivalently the target vector \(\vec{k}_{L}\), provides polarimetric scattering characteristics of the target.

\([S]=\left[\begin{array}{l} S_{H H} S_{H V} \\ S_{V H} S_{V V} \end{array}\right] \text { or } \vec{k}_{L}=\left[S_{H H} S_{H V} S_{V H} S_{V V}\right]^{T}\)       (1)

Here, S_HV indicates the scattering component of polarization transmitting horizontally and receiving vertically. S_HH and S_VV are scattering responses at copolarized channels and S_HV and S_VH representscattering responses at cross-polarized channels.

In natural surfaces or distributed targets, scattering characteristics can be better described by the covariance matrix [C] which contains effective information in quantifying geophysical properties based on stochastic processes (Lee and Pottier, 2009). For mono-static backscattering case, the covariance matrix can be described by

\([C]=<\vec{k}_{L} \cdot \vec{k}_{L}^{T},>\)     (2)

\(=\left[\begin{array}{c} <\left|S_{H H}\right|^{2}>\sqrt{2}<S_{H H} S_{H V}^{*}>\left\langle S_{H H} S_{V V}^{*}>\right. \\ \sqrt{2}<S_{H V} S_{H H}^{*}>2\left\langle\mid S_{H} \eta^{2}\right\rangle \quad \sqrt{2}<S_{H V} S_{V V}^{*}> \\ \left.\left.<S_{V} S_{H H}^{*}\right\rangle \quad \sqrt{2}<S_{V} S_{H V}^{*}\right\rangle \quad\left\langle\left|S_{W}\right|^{2}>\right. \end{array}\right]\)

where <…> representsthe spatial ensemble averaging. It is possible to direct analyze not only co- and crosspolarization backscatter intensities through diagonal terms of covariance matrix but also their correlation properties through off-diagonal terms. One of the important polarimetric correlation parameters is the co-pol correlation btween HH and VV polarization channels ρ_HHVV, which can be represented as

\(\rho_{H H V V}=\frac{\left\langle S_{H H} S_{V V}^{*}\right\rangle}{\sqrt{\left.<\left|S_{H H}\right|^{2}><|S_{V V}\right|^{2}\ >}}\)       (3)

The magnitude of ρ_HHVV can be a good indicator of signal depolarization. It will be zero for a completely random signal and one for a pure polarized signal.

One of the advantages of polarimetric SAR observation isthat once a targetresponse is acquired in one orthogonal polarization basis, the response in any orthogonal basis can be obtained by a simple unitary transformation. In this context, the coherency matrix [T] defined in the Pauli-basis is also widely used for characterizing scattering process. It cotains the same information asthe covariance matrix with a reordering of the product terms, defined as

\([T]=<\vec{k}_{P} \cdot \vec{k}_{P}^{*} T_{>}\)       (4)

where the Pauli scattering vector \(\vec{k}_{P}\)is given by

\(\vec{k}_{P}=\frac{1}{\sqrt{2}}\left[S_{H H}+S_{V V} \quad S_{H H}-S_{V V} \quad 2 S_{H V}\right]^{T}\)       (5)

Cloude and Pottier (1996) proposed a polarimetric parameters on the basis of eigenvalue and eigenvector of the coherency matrix. From the Hermitian nature of the coherency matrix, [T] can be expanded in terms of its eigenvalues and eigenvectors, such as

\(<[T]>=\sum_{i=1}^{3} \lambda_{i} \overrightarrow{e_{i}} \cdot \overrightarrow{e_{i}}^{*}\)       (6)

where λ_i and \(\vec{e}_{i}\) stand for eigenvalues and eigenvectors, respectively. One of the eigenvalue parameters named polarimetric entropy (H) has been widely used for
characterizing randomness of the scattering process (Cloude and Pottier, 1997), which can be represented as

\(H=-\sum_{i=1}^{3} P_{i} \log _{3} P_{i} ; \quad P_{i}=\frac{\lambda_{i}}{\labda_{1}+\lambda_{2}+\lambda_{3}}\)       (7)

The polarimetric entropy is close to 0 for weakly depolarizing target and increases as an increase of scattering randomness. Another eigenvalue parameter named polarimetric anisotropy (A) is related to the influence of secondary scattering mechanisms, which can be represented as

\(A=\frac{\lambda_{2}-\lambda_{3}}{\lambda_{2}+\lambda_{3}}=\frac{P_{2}-P_{3}}{P_{2}+P_{3}}\)       (8)

Recently, Réfrégier and Morio (2006) proposed a new eigenvalue parameter named Shannon entropy (SE), defined as

\(S E=\log \left(\pi^{3} e^{3}|[T]|\right)=\log \left(\pi^{3} e^{3} \prod_{i} \lambda_{i}\right)\)       (9)

It can be interpreted as a another statistical measure ofrandomness occurring in the scattering media on the basis of information theory.

On the other hand, Cloude and Pottier (1996) proposed a parameterization of the eigenvectors in terms of five angles, such as

\(\overrightarrow{e_{i}}=\left(\cos \alpha_{i} \sin \alpha_{i} \cos \beta_{i} e^{i \delta_{i}} \sin \alpha_{i} \sin \beta_{i} e^{\dot{\gamma}_{i}}\right)\)       (10)

Among them, the alpha angle contains useful information about scattering mechansm occurring in each scattering contribution.Th statistical interpretation of the alpha angle leads a form of the averaged alpha (\(\bar{\alpha}\)), such as

\(\bar{\alpha}=\sum_{i=3}^{3} \alpha_{i} P_{i}\)       (11)

Although the polarimetric eigenvalue and eigenvector parameters provide supplementary information on the scattering process, the interpretation of those parameters are not straightforward especially for multiple or volume scattering problems. For an efficient interpretation of the polarimetric measurements, another processing tool so called polarimetric modelbased decomposition methods have become a popular choice in the analysis of polarimetric SAR data.

Among several target decomposition methods(e.g., Freeman and Durden, 1998; Yamaguchi et al., 2005; Van Zyl et al., 2011), one widely-used methods is recalled in this study, such as the four component decomposition (Yamaguchi et al., 2011). According to this method, the measured coherency matrix can be decomposed into four elementary scattering types,such asthe surface scattering, the double-bounce scattering, the volume scattering, and the helix scattering, represented as

\(\begin{array}{r} {[T]=f_{S}\left[\begin{array}{lll} 1 & b^{*} & 0 \\ b & \left.b\right|^{2} & 0 \\ 0 & 0 & 0 \end{array}\right]+f_{D}\left[\begin{array}{ccc} |a|^{2} & a & 0 \\ a & 1 & 0 \\ 0 & 0 & 0 \end{array}\right]+} \\ \frac{f_{V}}{4}\left[\begin{array}{lll} 2 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{array}\right]+\frac{f_{C}}{2}\left[\begin{array}{ccc} 0 & 0 & 0 \\ 0 & 1 & \pm i \\ 0 & \mp i & 1 \end{array}\right] \end{array}\)       (12)

where f_S, f_D, f_V, and f_C denote the contribution of the surface, double, volume, and helix scattering components, respectively. The parameter α and b are unknown duble-bound and surface scattering coefficients to be determined from the measured coherency matrix.

2) Detection of water-covered area

For the flood detection with polarimetric SAR data, it is essential to figure out which polarimetric parametersrepresentstrong response to water-covered areas. Accordingly, among the aforementioned polarimetric parameters, following parameters were selected for analyzing detectability of water-covered areas:

– Covariance matrix parameters: |S_HH|²,|S_HV|²,ρ_HHVV

– Eigenvalue parameters: H, SE, \(\bar{\alpha}\)

– Model-based decomposition parameters: f_S, f_D

In order to remove speckle noise effect, the refined Lee speckle filtering (Lee et al. 1999 was applied to both data and the results of each parameters are shown in Fig 2 and Fig 3. The reference water-coverage map for the before and after the flood events (Fig. 2(i) and Fig. 3(i)) were manually generated with a reference information obtained from the Geospatial Information Authority of Japan (GSI map, 2015).

Fig 2. Polarimetric parameters of study area obtained before the flood.

Fig. 3. Polarimetric parameters of study area obtained after the flood.

The selected polarimetric parameters obtained before and after flood events were used for the detection of water-coered areas by the histogram thresholding method.In orderto evaluate the detectability regardless ofthe threshold values, performance of each parameter were assessed usingReceiver OperatingCharacteristic (ROC) curve and the Area Under the Curve (AUC) as shown in Fig 4 and Fig 5, respectively. Results show that, among selected polarimetric parameters, the HHpolarization intensity, the Shannon entropy, and the surface scattering component of the model based decomposition showed high detectability of water surface in both non-flooded and flooded conditions.

Fig. 4. The ROC curves of each polarimetric parameter obtained (a) before and (b) after the flood.

Fig. 5. The AUC of each polarimetric parameters obtained (a) before and (b) after the flood.

4. Automatic flood detection

Based on th characteristicsresponse of polarimetric parametersin water-covered area, the HH-polarization intensity, the Shannon entropy, and the surface scattering component were further used for an automatic mapping of flood extents. To ensure rapid disaster response, two types of unsupervised change detection approaches both based on the histogram thresholding were applied in this study.

1) Image differencing approach

In order to map flooded areas, a image differencing between pre- and post-disaster conditions using different polarimetric features was conducted. This method is composed of two steps: (1) subtraction of the image afterthe flood from the image before the flood, and (2) unsupervised binary classification of flooded area.

Fig. 6(a)-(c)show the image differencing results and Fig. 6(d)-(f) show the unsupervised classification results for selected three polarimetric parameters. In order to generate a binary flood map, a method based on the adaptive Kittler-Illingworth (KI) thresholding with bimodality test (Park, 2016) was applied to the histogram of difference image. The detection result was compared with the manually generated reference flood map. The accuracy assessment resultsin Table 2 show that the Shannon entropy outperforms other polarimetric parameters with the detection rate of about 70% and the false-alarm rate of about 7%. Indeed, the Shannon entropy can be represented by the sum of the intensity fluctuation and the signal depolarization(Réfrégier and Morio^, 2006). Consequently, the experimental result demonstrates that both the backscattering power and the degree of polarization are important parameters in the flood detection.

Fig. 6. (a)-(c) image differencing and (d)-(f) binary classification results.

Fig. 7. Reference data for flood mapping

Table 2. Accuracy assessment of flood mapping from different polarimetric parameters based on image differencing approach

2) Post-classification approach

The second unsupervised change detection tested in this study is the post-cassification method. It is composed of two steps: (1) Binary classification of water-covered area for pre- and post-disaster images separately, and (2) Detection of flooded area by comparing temporal classification results.

As a first step, the adaptive KI thresholding with bimodality test algorithm was applied to each temporal polarimetric data asshown in Fig 8.Results correspond to the binary map of water-covered area at each temporal acquisition. Then, the flooded area can be identified by comparing two temporal classification results and retaining classes changed from non-water class to water class. Fig. 9 shows the final flood mapping results obtained from selected polarization parameters. The detection accuracy was calculated by comparing the classification result with the reference data as summarized in Table 3.The flood mapping results from post-classification approach show better performance than the results of image differencing approach with higher detection rates (above 78%) for all three polarization parameters.Considering the falsealarm rate as well as the detection rate, the Shannon entropy again provides the best flood mapping result among selected polarization parameters.

Fig. 8. Detection of water-covered area by thresholding method for different polarimetric parameters acquired (a)–(c) non-flooded and (d)–(f) flooded conditions.

Fig. 9. Flood mapping results by post-classification approach.

Table 3. Accuracyassessmentforthepost-classificationapproach

5. Conclusion

The polarimetric SAR observations and secondary polarimetric scatter type indicators have been proven to be very usefulfor various applications.In thisstudy, the performance and usability of polarimetric parameters in detecting water-covered areas were evaluated by using L-band fully polarimetric SARdata. The experimental results showed that, among various polarimetric parameters, the HH-polarization intensity, the Shannon entropy, and the surfaces scattering component of model-based decomposition were found to be useful to discriminate water-covered areas from other land cover types.

In order to generate a binary map on flood extents, two types of simple unsupervised change detection approaches, suh as an image differencing approach and a post-classification approach, were also tested in thisstudy. The accuracy assessmentresultsindicate that a post-classification method will be a better strategy than an algebraic method. It is probably due to the fact that an algebraic method is more affected by temporal fluctuations of ground condition as well as variations ofimaging geometry between the pre- and post-disaster acquisitions.

Among selected polarimetric parameters,theShannon entropy providesthe bestflood mapping result with the detection rate of about 78% and the false-alarm rate of about 2% in the case of current experimental data. It is worth noting tht it was obtained with a simple thresholding-based approach. Therefore, the results presented in this study can be further improved by applying other advanced change detection methods or ancillary information, such as the surface topography.