Investigation of Polarimetric SAR Remote Sensing for Landslide Detection Using PALSAR-2 Quad-pol Data

Abstract

Recent SAR systems provide fully polarimetric SAR data, which is known to be useful in a variety of applications such as disaster monitoring, target recognition, and land cover classification. The objective of this study is to evaluate the performance of polarization SAR data for landslide detection. The detectability of different SAR parameters was investigated based on the supervised classification approach. The classifier used in this study is the Adaptive Boosting algorithms. A fully polarimetric L-band PALSAR-2 data was used to examine landslides caused by the 2016 Kumamoto earthquake in Kyushu, Japan. Experimental results show that fully polarimetric features from the target decomposition technique can provide improved detectability of landslide site with significant reduction of false alarms as compared with the single polarimetric observables.

1. Introduction

Many landslides occur simultaneously in a wide area due to triggers such as earthquakes and heavy rainfall (Keefer, 1984; Keefer et al., 1987). Space-borne remote sensing techniques can play an important role in detecting multiple landslides events. In particular, since the most landside can be triggered by heavy rain, Synthetic Aperture Radar (SAR) provides operational advantages over optical sensors for rapid survey of landslides. Although disaster monitoring including landslide can be one of the most important applications of radar remote sensing, an interpretation of SAR image is difficult and detection results are often ambiguous (Singhroy et al., 1998; Czuchlewski et al., 2003; Singhroy and Molch, 2004).

Recent SAR systems operate in fully polarimetric SAR (PolSAR) mode. The PolSAR system offers increased number of independent observation in comparison to traditional single-polarization radars. In addition, several advanced polarimetric information extraction techniques, such as Polarimetric Target Decomposition (PTD) method, provide supplementary information about microwave scattering properties. Although many useful polarimetric parameters have been proposed (Cloude and Pottier, 1996; Freeman and Durden, 1998; Yamaguchi et al., 2005), there have been a few studies about practical application of PolSAR data to landslide detection. Recently, Li et al. (2014) applied X-band air-borne PolSAR data to landslide areas caused by the Earthquake. Sibayama et al. (2015) used L-band space-borne PolSAR data for observing typhoon-induced landslides. Both studies used PTD parameters for recognizing landslides qualitatively. They proposed several polarimetric indices effective to distinguish landslide area from other land-cover types but there have been no quantitative evaluation on detection performance.

The objective of this study is to quantitatively compare the performance of polarization SAR features for detecting landslides. The detectability of different SAR parameters is evaluated based on the supervised classification approach. The classifier used in this study is the Adaptive Boosting (AdaBoost) algorithm (Freund and Schapire, 1996). To detect landslides, various polarimetric features obtained from the L-band PALSAR-2 data of ALOS-2 satellite acquired before and after the landslide events were investigated.

2. Study area and data

The 2016 Kumamoto earthquake of magnitude 7.3 occurred on April 16, 2016 in Kumamoto prefecture, Kyushu Island, Japan. The earthquake triggered numerous landslides around Mt. Aso. In this study, detection is conducted for a larges landslide close to Minami-Aso. Fig. 1 (a) shows the location of the study area and Fig. 1 (b) is an aerial photograph of the landslide selected in this study.

Fig. 1. (a) Location of Kumamoto study area (b) an aerial photograph of target landslide area.

Two temporal L-band PALSAR-2 images were used in this stud to detect landslide. Fig. 2 shows the ALOS-PALSAR images around the target landslide area. The images were acquired on April 21, 2016 (post-event) and December 3, 2015 (pre-event) both in the fully polarimetric mode. Details of the images used in this study are summarized in Table 1. Since all data used in the study have the same acquisition parameters, differences in the images can be assumed to be directly related to the changes of ground conditions.

Fig. 2. ALOS-PALSAR Pauli images. blue is HH+VV, green is HV, red is HH-VV. Red square is landslide occurrence area. (a) December 3, 2015 pre-event image and (b) April 21, 2016 post-event image.

Table 1. Details of PALSAR-2 data

3. Theoretical Background

In this Section, we recall several fundamental polarimetric parameters which were applied to the landslide detection. More details about the other fully polarimetric observables and advanced PolSAR processing techniques can be found in several books (Lee and Pottier, 2009; Cloude, 2010). The fully PolSAR system records the complex scattering matrix [S] in each resolution cell. The complex 2×2 scattering matrix can be expressed as follows using the most common horizontal (H) and vertical (V) polarization basis

\([S]=\left[\begin{array}{cc} S_{H H} & S_{H V} \\ S_{V H} & S_{V V} \end{array}\right]\)      (1)

where elements S_HV denotes the complex scattering coefficient of polarization transmitting horizontally and eceiving vertically. In the monostatic case, the target matrix is reciprocal with S_HV = S_VH.

In order to investigate scattering information of the natural targets, the 3×3 covariance matrix [C] and 3×3 coherency matrix [T] are more appropriate than the [S] matrix (Lee and Pottier, 2009; Cloude, 2010). The covariance matrix can be generated from the complex scattering vector \(\vec{\Omega}=\left[S_{H H} \sqrt{2} S_{H V} S_{V V}\right]^{T}\), such as

\(\begin{array}{c} {[C]=\left\langle\vec{\Omega} \cdot \vec{\Omega}^{* T}\right\rangle=} \\ {\left[\begin{array}{c} <S_{I I I} S_{I I I}^{*}> & \sqrt{2}<S_{I I I} S_{I I V}^{*}><S_{I I I} S_{V V}^{*}> \\ \sqrt{2}<S_{H V} S_{H H}^{*}> & 2<S_{H V} S_{H V}^{*}>\sqrt{2}\left\langle S_{H V} S_{V V}^{*}\right\rangle \\ <S_{V V} S_{H H}^{*}> & \sqrt{2}<S_{W} S_{H V}^{*}>\quad\left\langle S_{V V} S_{V V}^{*}\right\rangle \end{array}\right]} \end{array}\)       (2)

where < > is ensemble averaging. To obtain the coherence matrix, scattering matrix can be reduced to the complex scattering vector as follows

\(\vec{K}=\frac{1}{\sqrt{2}}\left[S_{H H}+S_{V V} \quad S_{H H}-S_{V V} \quad 2 S_{H V}\right]^{T}\)       (3)

The coherency matrix describing the distributed scatterers can be expressed as follows

\(([T]=\left\langle\vec{K} \cdot \vec{K}^{*}\right\rangle\\ =\frac{1}{2}\left[\begin{array}{c} \left\langle | S_{H H}+\left.S_{W}\right|^{2}\right\rangle & \left\langle\left(S_{H H}+S_{W V}\right)\left(S_{H H}-S_{W V}\right)^{*}\right\rangle &2<\left(S_{H H}+S_{V V}\right) S_{H V}^{*}> \\ \left\langle\left(S_{H H}-S_{V V}\right)\left(S_{H H}+S_{W}\right)^{*}\right\rangle & \left\langle\left|S_{H H}-\left.S_{W}\right|^{2}\right\rangle\right. & 2<\left(S_{H H}-S_{W V}\right) S_{H P}^{*}>\\ 2\left\langle S_{H H}\left(S_{H H}+S_{W}\right)^{*}\right\rangle & 2\left\langle S_{H I}\left(S_{H H}-S_{W}\right)^{*}\right\rangle & 4\left\langle\left| S_{HV}\right|^{2}\right\rangle \end{array}\right]\)       (4)

The PTD method aims to distinguish different scattering mechanisms in the observed covariance or coherence matrix. Since the PTD technique enables interpretation of radar images more easily, it has been one of the key research topics in PolSAR remote sensing community. In this study, PTD features from the eigenvalue-based decomposition technique (Cloude and Pottier, 1996; Cloude and Pottier, 1997) and from the model-based scattering power decomposition technique (Yamaguchi et al., 2005; Yamaguchi et al., 2011) are extracted and applied to machine learning algorithm to perform landslide detection.

1) Eigenvalue-based decomposition

The eigenvalue-based decomposition method proposed by Cloude and Pottier (1996) is a mathematical technique leading to an understanding of averaged scattering mechanisms in the scattering medium. Since the fully polarimetric coherence matrix is a Hermitian matrix, it can be decomposed into real eigenvalues and orthogonal unitary eigenvectors as follows

\([T]=\left[U_{3}\right]\left[\begin{array}{ccc} \lambda_{1} & 0 & 0 \\ 0 & \lambda_{2} & 0 \\ 0 & 0 & \lambda_{3} \end{array}\right]\left[U_{3}\right]^{T}\)       (5)

where λ_i is real eigenvalues (λ₁≥λ₂≥λ₃) and the unitary matrix [U₃] can be expressed as

\(\begin{array}{c} {\left[U_{3}\right]=} \\ {\left[\begin{array}{ccc} \cos \alpha_{1} & \cos \alpha_{2} & \cos \alpha_{3} \\ \sin \alpha_{1} \cos \beta_{1} e^{j \delta_{1}} & \sin \alpha_{2} \cos \beta_{2} e^{j \delta_{2}} & \sin \alpha_{3} \cos \beta_{3} e^{j \delta_{3}} \\ \sin \alpha_{1} \sin \beta_{1} e^{j y_{1}} & \sin \alpha_{2} \sin \beta_{2} e^{j_{2}} & \sin \alpha_{3} \sin \beta_{3} e^{\dot{\gamma}_{3}} \end{array}\right]} \end{array}\)       (6)

Based on the eigenvalue decomposition, the polarimetric entropy (H) can be defined as (Cloude and Pottier, 1997)

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

It is an index that measures randomness of a scattering process with 0 ≤ H ≤ 1. When H is low, it represents a non-depolarizing single dominant scattering mechanism. When H is high, a single equivalent point scatterer does not exist and a mixture of various scattering contribution must be considered. Another eigenvalue ratio parameter defined as the polarimetric anisotropy (A) is a complementary parameer to H, indicating the relative importance magnitudes of the second and third eigenvalues.

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

On the other hand, averaged scatterng mechanism can be extracted from the eigenvector. Among various eigenvector parameters in (6), angle α provides direct information on the scattering mechanism represented by each eigenvector. By using the pseudo-probability P_i in (7), mean α angle (\(\bar {a}\)) can be defined as

\(\bar{\alpha}=\sum_{i=1}^{3} P_{i} \alpha_{i}\)       (9)

It is related to the average scattering mechanisms from single-bounce scattering at the surface when  = 0°, dipole scattering when  = 45°, and double-bounce scattering when = 90°.

2) Model-based scattering decomposition

Common limitation of eigenvalue-based decomposition methods is that it is not physically based and cannot be easily interpreted by physical scattering mechanisms On the other hand, the model-based PTD technique is based on the physics of microwave scattering. It aims to separate the observed covariance or coherency matrix as a linear combination of elementary scattering matrices related to specific scattering moels. Among several approaches, the most representative method is Freeman and Durden’s three component decomposition (1998) that has been successfully applied to decompose PolSAR image under the assumption of reflection symmetry. However, there is a problem that the reflection symmetry condition is not satisfied in nonorthogonally-oriented man-made structures with respect to the line of sight and sloping mountainous forests.

Yamaguchi et al. (2005) proposed the four component decomposition method to relax the reflection symmetry constraint. This method divides the measured coherency matrix into four scattering types, such as surface scattering, double-bounce scattering, volume scattering, and helix scattering represented as

\(\begin{aligned} \text [T]=f_{S}\left[\begin{array}{lll} 1 & b^{*} & 0 \\ b & |b|^{2} & 0 \\ 0 & 0 & 0 \end{array}\right]+f_{D}\left[\begin{array}{ccc} |a|^{2} & a & 0 \\ a & 1 & 0 \\ 0 & 0 & 0 \end{array}\right]+\\ f_{V} \frac{1}{4}\left[\begin{array}{lll} 2 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{array}\right]+f_{C} \frac{1}{2}\left[\begin{array}{lll} 0 & 0 & 0 \\ 0 & 1 & \pm i \\ 0 & \mp i & 1 \end{array}\right] \end{aligned}\)       (10)

where f_S, f_D, f_V, and f_C represent the contribution of the surface scattering, double scattering, volume scattering, and helix scattering components, respectively. The parameter a and b are unknown double-bounce and surface scattering coefficients to be determined from the measured coherency matrix. Yamaguchi et al.(2011) extended the previous algorithm by considering the orientation angle compensation of the coherence matrix to improve the problem of overestimation of the volume scattering component.

3) AdaBoost Classifier

To compare landslide detectability of different polarimetric parameters quantitatively, an image classification was carried out in this study based on a boosting algorithm. Boosting is one of the ensemble learning methods widely used for the various classification problems. One of the widely known boosting algorithms is the AdaBoost which was firstly proposed by Freund and Schapire (1996) and has been practically used in many areas including the SAR Automatic Target Recognition (ATR) field. It runs iteratively and builds an efficient classifier during iteration. The weight for the wrongly classified samples increases in each iteration so that the classifier in the next iteration focuses more on the difficult samples to decrease the training error. Consequently, AdaBoost can build decrease of the global error of the classifier and hence builds a stronger and efficient classifier.

To adjust AdaBoost in different learning scenarios and to improve the classification performance, several modified versions have been proposed, such as Real Adaboost (Schapier and Singer, 1999), Logit Boost (Friedman et al., 2000), and Gentle Adaboost (Freund et al., 1999). Among the Adaboost variants, the Gentle AdaBoost algorithm is known to have less sensitivity to the presence of noisy patterns since it puts less emphasis on outliers during iteration process (Dietterich, 2000; Friedman et al., 2000). In this study, the Gentle AdaBoost algorithm was used in to evaluate the accuracy of each polarimetric feature in landslide detection problem. Learning process as conducted with manually selected regions of interest.

4. Experimental Results

In the experiment, we conducted two groups of experiments: one used the single-date post-event SAR image; the other used the multi-date pre-and post-event SAR images. The purpose of the first experiment is to evaluate whether the PolSAR parameters could discriminate landslide area from neighboring land- cover types. In the post-event image, there could be other areas showing similar backscattering properties with the landslide. Consequently, the second experiment was designed to use change augmented features in order to evaluate actual landslide detection capability of PolAR parameters.

1) Classification of single post-event image

As the first experiment, image classification was conducted for the post-event image acquired April 21, 2016. Among various polarimetric parameters discussed above, 8 different features (Fig. 3) were tested to assess landslide detection capability of polarimetric observables. The polarimetric features considered in this paper include two backscattering coefficients (first row of Fig. 3), three eigenvalue-based decomposition parameters (second row of Fig. 3), and three model-based decomposition parameters (third row of Fig. 3). Google Earth optical image taken after the landslide is also shown in Fig. 3(a) for reference.

Fig. 3. Selected polarimetric parameters of the study area obtained after the landslide.

As one can identify from the optical image, the bare surface exposed in the center of the image is the target landslide site. The land-cover types in the surrounding area include forests, grasslands, agricultural fields, and bare soils. However, the landslide-affected area and other bare soil areas probably have no significant difference in backscattering properties. Therefore, those two classes were assumed to be the same class in the classification of single post-event image.

The AdaBoost-based landslide classification results for each selected polarimetric feature are shown in Fig. 4. Here, the classification results are binarized into two classes, such as landslide or soil class (black area in Fig. 4) and other classes (white area in Fig. 4), in order to ascertain the detectability of polarimetric features. The landslide detection performance of each polarimetric parameter was evaluated with well-known metrics, such as Probability of Detection (POD) and False Alarm Rate (FAR). To have an overall idea about both POD and FAR, the Critical Success Index (CSI) (Schaefer, 1990) was also used in the performance evaluation. The CSI is defined as

Fig. 4. landslide or bare soil detection result of the study area (black: landslide; white: others).

\(C S I=\frac{1}{\frac{1}{(1-F A R)}+\left(\frac{1}{P O D}\right)-1}\)       (11)

Table 2 summarizes the performance evaluation metrics for selected polarimetric features. It indicates that HV-polarization backscattering coefficient and f_V features emphasize landslide. It can be explained by the dominant volume scattering mechanism in forest surrounding the landslide. The polarimetric entropy and the mean α angle of the eigenvalue-based decomposition can also separate landslide areas from surrounding forests in mountainous terrain. Nonetheless, the accuracy assessment results listed in Table 2 show relatively high false alarms over all polarimetric parameters. This is due to a lack of in-situ information as well as a limitation of L-band SAR in discriminating between landslide class and other land-cover classes. Landslide areas and bare soil or short grass areas are classified into the same class because they have no physical difference in single-date L-band SAR image. However accuracy evaluation was carried out exclusively for landslide because of a lack of reference information other than landslide area.

Table 2. Assessing accuracy of landslide detection from single-date post-event image

2) Classification of bi-temporal images

In the second experiment, bi-temporal change-based classification was carried out to resolve ambiguity in landslide detection from single-date SAR data. The change augmented polarimetric features were used as inputs of AdaBoost classifier. The differences of polarimetric parameters of two temporal acquisitions, i.e., post event – pre event, were used as a change augmented feature in this study. Based on the result obtained in the first experiment, six polarimetric parameters showing better performance in identifying landslide areas were further tested for the bi-temporal analysis. Fig. 5 shows temporal changes of the selected polarimetric parameters including the backscattering coefficients at HH and HV polarization channels, the polarimetric entropy, the mean α angle, the volume scattering component, and the helix scattering component.

Fig. 5. Difference image for selected polarimetric parameters.

Based on the change augmented polarimetric features in Fig. 5, the binary classification results were obtained as shown in Fig. 6. The accuracy assessment results of the bi-temporal approach are summarized in Table 3. The landslide detection accuracies are improved in all selected polarimetric parameters with significant reduction of FAR as compared with the single post- event classification. Among them, the polarimetric entropy in the eigenvalue-based decomposition revealed the best detection accuracy with the highest CSI value. The HV polarization backscattering coefficient shows the highest CSI among backscatter intensities.

Fig. 6. Landslide detection result of study area (black: landslide; white: others).

Table 3. Accuracy assessment of landslide detection for multi-temporal approach

Based on the experimental results, one can consider a synergy of polarimetric features showing good detection performance. Fig. 7 and Table 4 show the landslide detection result using both HV and H as an input to the AdaBoost classifier. The landslide detection accuracy can be improved with a combination of two polarimetric parameters. However, there are still significant amount of false alarms especially due to seasonal changes in grasslands in upper-left part of the image during the reference (pre-event) and post-event observations. In addition, there were some missed detection areas, for example the undetected part in the right side of the landlide (red circle in Fig. 7). Those areas were originally agricultural fields in the pre-event observation. The missed detection occurred because the learning process for landslide detection in this study only considered the land-cover change from forests to bare soil areas. Consequently, the landslide detection result obtained in this study can be further improved by considering landslides occurred in different land-use and land-cover conditions.

Fig. 7. Landslide detection result from ΔHV and ΔH features.

Table 4. Accuracy assessment of landslide detection for polarimetric parameters

5. Conclusion

In this study, detection of landslide affected area was performed with various polarimetric features observed by the fully polarimetric L-band PALSAR-2 system. The single post-event image classification experiment shows that polarimetric parameters indicating the volume scattering mechanism emphasized the difference of scattering properties between landslides and surrounding forests. In addition, the bi-temporal change-based image classification was conducted based on results obtained by the single post event image. As a result, the HV polarization backscattering coefficient and the polarimetric entropy features can detect landslides effectively from surrounding area. Furthermore, a combination of the two features can provide improved performance in landslide detection.

In the both experiment, the PolSAR parameters showed better performance than the single-polarization parameter. Nonetheless, there were significant amount of false alarms mainly due to the effect of seasonal changes in natural scatterers. The use of ancillary information, such as digital elevation model and GIS database can improve the detection result. However, more importantly, an increase in the number of fully polarimetric acquisitions in the SAR observation plan will be necessary to resolve such kind of problem in the disaster reaction.

Since the volume scattering mechanism or signal depolarization properties play an important role in the landslide detection, different radar frequency may provide significantly different results. Consequently, it is necessary to extend this study to other next generational polarimetric SAR systems operating at different frequencies, such as Kompsat-6 (X-band), RADARSAT Constellation (C-band), NiSAR (S-bad and L-band), and Biomass (P-band). Although it is not discussed in this study, the use of other machine learning algorithm can also affect the detection results. Therefore, the effect of classifier on the landslide detection will be investigated in the follow-up study.