Polarimetric Decomposition Analysis of ALOS PALSAR Observation Data before and after a Landslide Event

Radar scattering mechanisms over landslide areas were studied using representative full polarimetric parameters: Freeman–Durden decomposition, and eigenvalue–eigenvector decomposition. Full polarimetric ALOS (Advanced Land Observation Satellite) PALSAR (Phased Array type L-band Synthetic Aperture Radar) datasets were used to examine landslides caused by the 2008 Iwate-Miyagi Nairiku Earthquake in northern Japan. The Freeman–Durden decomposition indicates that areas affected by large-scale landslides show dominance of the surface scattering component in both ascending and descending orbit data. The polarimetric parameters of eigenvalue–eigenvector decomposition, such as entropy, anisotropy, and alpha angle, were also computed over the landslide areas. Unsupervised classification based on the H-  plane explicitly distinguishes landslide areas from others such as forest, water, and snow-covered areas, but does not perform well for farmland. A landslide area is difficult to recognize from a single-polarization image, whereas it is clearly extracted on the full polarimetric data obtained after the earthquake. From these results, we conclude that 30-m resolution full polarimetric data are more useful than 10-m resolution single-polarization PALSAR data in classifying land coverage, and are better suited to detect landslide areas. Additional information, such as pre-landslide imagery, is needed to distinguish landslide areas from farmland or bare soil. OPEN ACCESS Remote Sens. 2012, 4 2315


Introduction
SAR (Synthetic Aperture Radar) is capable of observing the Earth's surface in all weather conditions, day and night.Full polarimetric SAR data provide a scattering matrix of a terrain's surface.The scattering matrix consists of magnitude and phase at four polarizations, HH, HV, VH, and VV, that are sent and receive horizontally (H)-and vertically (V)-polarized waves by radar antenna [1].Some methods have been suggested in order to obtain earth surface characteristics from full polarimetric data.Eigenvector decompositions of a coherency or covariance matrix were proposed by Cloude and Pottier [2].An approach to the three-component scattering mechanism model was presented by Freeman and Durden [3], and a four-component approach was presented by Yamaguchi et al. [4].These and other decomposition methods were reviewed [5].The usefulness of full polarimetric SAR data for land cover classification was examined by Alberga et al. [6], and some of the techniques were applied to disaster monitoring-and assessment applications.A landslide caused by the 1999 Chi-chi earthquake in Taiwan was recognized using L-band airborne SAR data acquired after the disaster, and the characteristics of radar scattering mechanism were examined.The extent of the landslide was identified using data such as scattering entropy and anisotropy [7,8].Other airborne L-band SAR data with full polarimetry were analyzed by Watanabe et al. [9] to detect a landslide area in Japan using three-and four-component decomposition techniques.It was shown that the landslide areas could be effectively identified on the basis of the dominance of the surface scattering component of the three-component scattering model.
The Iwate-Miyagi Nairiku earthquake (magnitude 7.2) struck an inland area of northern Honshu, the main island of Japan, on 14 June 2008, causing large landslides within mountainous forest areas.Several large landslides locally transformed forest to bare surface areas.Watanabe et al. [10] analyzed full polarimetric data obtained from ALOS (Advanced Land Observation Satellite) data after the earthquake, and showed that the landslide areas were well detected by combining the surface scattering and backscattering coefficients in VH polarization ( 0 VH ).ALOS carries an L-band (23.5-cm wavelength) SAR, termed PALSAR (Phased Array type L-band Synthetic Aperture Radar).PALSAR is the first space-borne SAR to have full polarizations mode.It also has three different polarimetric observation modes: single polarization (mainly HH), dual polarization (mainly HH and HV), and full polarization.Its range resolutions are 7-44, 14-88 and 24-89 m, respectively.In this paper, we used PALSAR full polarimetric data, acquired in both ascending and descending orbits before and after the earthquake-induced landslide.The pre-and post-landslide data used the same observation mode (off-nadir angle, flight direction).Since relatively few global-scale observations have been carried out by air-borne SAR, acquiring the data before the disaster with the same observation mode was hardly achievable for a landslide area observation.We adopted two representative full polarimetric decomposition models: a physically-based Freeman-Durden decomposition model, and a mathematically-strict eigenvalue-eigenvector decomposition model.These models are applied not only to the landslide areas, but also to whole scenes, including forest, water surface and snow-covered areas, and areas of farmland or bare soil.Furthermore, results from the full polarimetric analysis were compared with those for single polarization.The comprehensive analysis results for landslide and neighboring areas are presented to clarify the radar scattering characteristics over the landslide areas by comparing other categories from the "actual" data, and to assist the identification of landslide areas from the pre-and post-event data.

Study Area
The epicenter of the Iwate-Miyagi Nairiku earthquake is located in the southern, inland part of Iwate Prefecture, at latitude 39°1.7′N and longitude 140°52.8′E(Figure 1).
caused by the earthquake [11].The landslide areas were distributed in valleys adjacent to a mountainous area.At 1,627 m, Mt.Kurikoma is the highest peak in the disaster zone.The top of the mountain is covered by a snowcap from fall to spring.After the earthquake, many aerial photographs and much satellite data were obtained.Figure 2 shows the affected areas observed by ALOS AVNIR-2 (optical sensor, Advanced Visible and Near-infrared Radiometer type 2) on 2 July 2008.Figure 2 is a false-color composite image with landslides recognizable from the bare surfaces characterized by blue-green color.We use this image as a reference for land coverage.The areas of earthquake damage are plotted on a damage status map [12,13].The damage status map was generated from visual interpretation of aerial photographs obtained on 15, 16 and 18 June with 0.4 m ground resolution by the Geospatial Information Authority of Japan (GSI); and 16 and 18 June by Asia Air Survey Co., Ltd.GSI also conducted a field survey on 17 and 18 June.The affected regions were previously covered by forest and some farmland.Three major landslide event areas were selected and analyzed during this time.One of the largest landslide areas is located at the upstream end of the Aratozawa Dam reservoir.The affected area measured 900 × 1,300 m, and the displaced soil mass was estimated as 70 million m 3 by ground survey.A mudslide at Dozou-sawa, on the eastern slope of Mt.Kurikoma, was analyzed, and its soil mass by slope failure was estimated as 1 million m 3 by aerial photography [13].The third landslide event area is Ubasume-gawa, where the area is 400 × 300 m [14], and soil mass by slope failure is estimated as 5 million m 3 .Field observations were carried out on 23 August 2008 at Aratozawa Dam reservoir.Several photos were taken, and we observed that fallen trees were mixed in some parts of the landslide area.

PALSAR Data
PALSAR data of the disaster area were acquired via different observation modes.Many observations have been carried out in single-and dual-polarization modes.Full polarimetric datasets, including the epicenter before the earthquake, were acquired on 19 May 2006 in descending orbit.Those after the earthquake were observed on 24 August 2008, 9 October 2008, and 24 November 2008 in descending orbit; and on 4 June 2009 in ascending orbit.The off-nadir angle of 4 June 2009 data was 23.1°, and that of the other data was 21.5°.The observation areas are shown in Figure 1, and the datasets are summarized in Table 1.

Data Analysis
The polarimetric data were analyzed using PolSARpro software [15].A Lee filter [16] was applied with a window size of 3 × 3 pixels, to reduce speckle noise to Level 1.1 PALSAR data, and was converted to the coherency matrix.We attempted a three-component polarimetric decomposition model [3] for full polarimetric data analysis.This model describes the scattering contributions from double-bounce, volume, and surface scattering by phase and intensity information in the polarimetric SAR data, as indicated below.
Here, Sij is the scattering coefficient, transmitted in the j polarization plane, and received in the i polarization plane; and f v , f d , and f s are the contribution to the power in VV polarization from volume, dihedral, and surface scattering.Parameters  and  are related to double-bounce and surface scattering.Determining whether double-bounce or surface scattering is the dominant contribution by using the sign of Re(S HH S * VV ) [17] enables us to identify the contribution of each scattering mechanism solely from SAR polarization data, without any field data.
We also attempted eigenvalue analysis of the coherency matrix.An eigenvalue-eigenvector decomposition was performed and scattering entropy H, polarimetric anisotropy A, and scattering mechanism parameter   (alpha) were computed; unsupervised classification based on the H-  plane [2] was attempted.H-A-  parameters were independently defined.The elements of the scattering matrix are defined as The coherency matrix is defined as follows: Parameter  i is directly related to the angle of incidence and dielectric constant of the surface with i ranging from 1 to 3. The  i angles can be interpreted as orientation angles.γ i and δ i account for the phase relations.
The appearance probability of each  i contribution is given by The polarimetric scattering entropy is defined as follows: The randomness of the scattering process is measured by entropy.
The dominant scattering mechanism for each pixel is given by the   (alpha).
Angle   is associated with the type of scattering mechanism, and yields direct information about the scattering mechanism represented by each eigenvector.Consider a scattering matrix or rotated by .In case of a = b, which corresponds to surface scattering, the alpha parameter is 0. In case of a = −b, which corresponds to dihedral scattering, the alpha, parameter is /2.In case of a >> b, which corresponds to highly anisotropic or dipole scattering, the alpha parameter is /4.The other is in between them.Polarimetric anisotropy, A, is defined as follows: Anisotropy is used to characterize the scattering phenomenon.One of the important characteristics for the eigenvalue-eigenvector decomposition is that the parameters are rotation invariant, and the parameters are constant for rotation about the radar line of sight.Table 2 shows a nine-zone segmentation scheme and represents scattering mechanisms on the H-  plane.

Freeman-Durden Decomposition
Figure 3 shows ortho-rectified Freeman-Durden decomposition images with the double-bounce shown in red, volume scattering in green, and surface scattering in blue for the data obtained before the disaster on 19 May 2006 in descending orbit; after the disaster on 24 November 2008 in descending orbit; and on 4 June 2009 in ascending orbit.Volume scattering is dominant within the extensive areas of forest coverage.
Landslide areas upstream of Aratozawa Dam, Dozou-sawa, Ubusume-gawa, and Hiya-sawa on the post-earthquake image show a larger surface scattering component than that on the pre-earthquake image, and are consistent with the results derived by Watanabe et al. [10].Surface scattering at the landslide areas east of Mt.Kurikoma, Ubusume-gawa, and Dozou-sawa are recognized in the post-earthquake images (see Figure 3(b,c,e-h)).Strong surface scattering upstream of Aratozawa Dam also appeared on the post-earthquake images Figure 3(b,c), but this area was beyond the observation area on 19 May 2006 (Figure 3(a)).
Strong surface scattering is found in the upper-left of Figure 3(a), in the area corresponding to the upper slopes of Mt.Kurikoma.The AVNIR-2 image for 21 May 2006 shows that this area was covered by snow, which may induce strong surface scattering.Some farmland in the middle of the target area shows dominant surface scattering.Table 3 shows the contribution of the scattering components derived from Freeman-Durden decomposition and double-bounce, volume, and surface scattering within the landslide areas denoted "A" (Dozou-sawa) and "B" (Ubusume-gawa), shown by the yellow rectangles in Figure 3(a-c).The images obtained of both areas on 19 May 2006, prior to the earthquake, show that volume scattering accounts for more than 81% and surface scattering represents 14%.On the other hand, three datasets from the descending orbit obtained after the earthquake show 63-67% for the volume scattering and 29-33% for the surface scattering components at these two landslide areas.After the earthquake, there was a shift of between 15 and 19%, from the volume scattering to surface scattering component.There are three possible causes for stronger volume scattering ratio in landslide areas.
(A) Volume scattering from fallen trees within the landslide areas.The presence of fallen trees within the landslide areas was confirmed by aerial photographs and field observations.
(B) Very rough surfaces induce a  0 hv component, leading to an erroneous volume scattering component.(C) Small contribution from forest areas, even if the areas enclosed by the yellow rectangles in Figure 3 are occupied by the landslide areas.
Since the regions enclosed by the yellow rectangle are largely occupied by the landslide sites, and the values change between the pre-and post-disaster measurements, we regarded the values in Table 4 as representative of landslide areas.The data obtained on 4 June 2009 in the ascending orbit shows volume scattering at 69% and surface scattering at 25% in the Dozou-sawa area, and volume scattering at 79% and surface scattering at 17% in the Ubusume-gawa area.The surface scattering ratio in the Ubusume-gawa area in the ascending orbit image is smaller than that of the later image, obtained in descending orbit.The discrepancy between the ascending and descending observations is not outstanding for H-  parameters shown in Table 4 of the next section, and no significant difference was observed for the land surface of the landslide areas.This is explained by the rotation invariance of H-  parameters, which are robust to the difference in observation direction.The percentage of double-bounce scattering is 4-6% within both landslide areas on every image.

H / A / Alpha Decomposition
Figure 4 shows an H-  classification map with range-azimuth projection.Entropy is lower over the landslide area than it is over the surrounding forest areas.Entropy comes from the randomness of the scattering process.High entropy is caused by the depolarizing effects of volume scatter from the structural elements of trees, which produce greater disorder in scattering.Alpha is high for forest areas and is low on bare surfaces, snow, and water surfaces.No significant differences were found in the anisotropy image.These results are consistent with those obtained by Czuchlewski et al. [7].Table 4 shows the entropy and alpha values, averaged over 4 × 16 windows from forest, agricultural, landslide, water surface, and snowcapped areas, and the values are plotted in Figure 4(e).For comparison, the same area sizes are picked from each category.The window size of 4 × 16 ensures that pixels of the same category are included in a site.Since the width of the landslide areas is narrow, the window size is almost decided by the size of the landslide areas.Each set of windows is selected from every dataset, including the corresponding areas denoted on Figure 4 as "A" (forest), "B" (agricultural), "C" (landslide), "D" (landslide), "E" (landslide), "F" (snow), and "G" (snow).Every dataset shows a similar behavior, with relatively high entropy and alpha on the forest areas.The landslide areas show lower entropy and alpha than the forest areas.Polarimetric entropy is more than 0.7 for forest areas, compared with 0.52 to 0.63 for the landslide areas.The alpha angle for the forest areas is 47-52 and is 29-37 for the landslide areas.The agricultural areas show 0.59-0.70 for entropy and 30-40 for alpha, and these values are similar to those of the landslide areas.This finding implies that additional information, such as additional parameters [10] or pre-disaster land use classification, is required to identify the landslide areas.Snow surface on 19 May 2006 shows lower entropy (0.14-0.28) and alpha (8)(9)(10)(11)(12)(13)(14)(15)(16) than the other land cover types.The entropy value of the water surface at Aratozawa Dam is 0.43-0.26,and is lower than that of the forest and the landslide areas.The relatively low entropy values correspond with the areas where surface scattering is dominant.The effect of different observation directions is evident from agricultural and water surface areas, and both the entropy and alpha values of ascending data are larger than those of the descending data for these areas.Ridges on agricultural fields and waves on the water surface may cause direction-dependent backscattering.4.
The H-  classification space has nine zones according to scattering mechanisms (Table 2, Figure 4(d)).Large-scale landslide areas with bare surfaces are categorized as 6, 7, 8, and 9, whereas they were categorized as 5, 6, and 9 in a previous study [7].The extension to high alpha angle (zones 7 and 8) is observed in our data, which we take as evidence for the mixture of fallen trees in the landslide area, as observed in the field.Forest areas are categorized as zones 4 and 5, whereas they were categorized as 5 in a previous study [7].The discrepancy may be explained by differing forest types.Water surfaces show strong surface scattering areas.A large proportion of the water surface is classified as category 9, surrounded by category-6 areas.This is consistent with the airborne SAR this radar shadow, and the resultant images of eigenvalue-eigenvector decomposition do not include the effect of the radar shadow.Yonezawa et al. [19] suggested that PALSAR single polarization data are useful to detect changes in terrain on landslides upstream of Aratozawa Dam and the temporal sequence of drawdown of the dammed lake.Single polarization data detects changes to land features, but a pre-landslide image is essential for the comparison.Spatial resolution of a full polarimetric image is coarser than that of a single polarization image.However, full polarimetric data can assess land surface coverage from the scattering mechanism.For example, the landslide at Ubusume-gawa is not recognized from visual interpretation of single polarization data; however, it is clearly extracted on the full polarimetric data obtained after the earthquake.

Conclusions
We employed Freeman-Durden decomposition and eigenvalue-eigenvector decomposition for ALOS (Advanced Land Observation Satellite) PALSAR (Phased Array type L-band Synthetic Aperture Radar) full polarimetric data.Data observed from different directions, ascending and descending orbit with small incidence angles (21.5° and 23.5°) were examined in this study.Land surface change from forest to bare soil due to landslides causes the scattering mechanism to change.The surface scattering component is dominant in landslide areas.The volume scattering component is strong in the forest areas.Agricultural, snow, and water surface areas also show a large surface scattering component compared to forest areas.Unsupervised classification using the H-  plane makes it possible to distinguish landslide areas from water and snow surface areas.Snow coverage and water areas generally show lower H and   values compared to landslide areas.Agricultural areas show similar scattering characteristics to those of landslide areas.Areas of bare soil, that were present before the earthquake, should show the same scattering properties as the landslide area.The difference in the direction of geometric distortion associated with the differing observation direction has a slight effect on the scattering ratio, H, and   values; however, the overall scattering properties are consistent with each other.We conclude that landslide areas are well recognized by the post-earthquake full polarimetry image, except for areas of farmland or bare soil.Classification accuracy is improved if the land coverage before the disaster is known from optical sensors, aerial photographs, or other references.Single polarization PALSAR images show landslide areas as topographic features that differ before and after the earthquake.However, single polarization data is unsuitable for efficient detection of landslide areas because of its difficulty in land cover classification.

Figure 1 .
Figure 1.Northern part of Honshu Island, Japan, with earthquake epicenter of the Iwate-Miyagi Nairiku earthquake (cross) and Mt.Kurikoma (square).Blank squares show observation areas of ALOS (Advanced Land Observation Satellite) PALSAR (Phased Array type L-band Synthetic Aperture Radar) full polarimetric mode."Asc."orbit data was obtained on 4 June 2009.Observation dates of "Desc."orbits were 19 May 2006, 24 August 2008, 9 October 2008, and 24 November 2008.Shaded square corresponds with the area shown in Figure 2.

Figure 2 .
Figure 2. ALOS AVNIR-2 (optical sensor, Advanced Visible and Near-infrared Radiometer type 2) false-color composite (Bands 4, 3, 2 in red, green and blue, respectively) obtained on 2 July 2008 after the 14 June 2008 earthquake.The white rectangle corresponds to the area shown in Figure 3.

Figure 4 .
Figure 4. Unsupervised classification image using H-  segmentation with eight looks in azimuth for (a) 19 May 2006, (b) 24 November 2008, and (c) 4 June 2009; (d) Partitioning and color legend, with numbers referring to the region described in Table 4; (e) H-  plot

Figure 5 .
Figure 5. (a) Single-polarization intensity image for Aratozawa Dam reservoir area observed on 23 March 2008, divided by the image from 23 June 2008; (b) On-site photograph taken around Area "B".

Table 1 .
Summary of Observation Mode for PALSAR (Phased Array type L-band Synthetic Aperture Radar) Data.

Table 3 .
Percentage of dominant scattering mechanism in areas associated with landslides following the 14 June 2008 Earthquake.