Pi-SAR-L2 Observation of the Landslide Caused by Typhoon Wipha on Izu Oshima Island

Transcription

1 remote sensing Article Pi-SAR-L2 Observation of Landslide Caused by Typhoon Wipha on Izu Oshima Island Manabu Watanabe 1, *, Rajesh Bahadur Thapa 1 and Masanobu Shimada 1,2 1 Earth Observation Research Center, Japan Aerospace Exploration Agency, Sengen, Tsukuba, Ibaraki , Japan; rajesh.thapa@jaxa.jp (R.B.T.); shimada@g.dendai.ac.jp (M.S.) 2 Tokyo Denki University, Ishizaka, Hatoyama-machi, Hiki-gun, Saitama , Japan * Correspondence: watanabe.manabu@jaxa.jp; Tel.: ; Fax: Academic Editors: Zhenhong Li, Roberto Tomas, Ioannis Gitas and Prasad S. Thenkabail Received: 24 December 2015; Accepted: 22 March 2016; Published: 29 March 2016 Abstract: Pi-SAR-L2 full polarimetic data observed in four different observational directions over a landslide area on Izu Oshima Island, induced by Typhoon Wipha on 16 October 2013, were analyzed to clarify most appropriate L-band full polarimetric parameters and observational direction to detect a landslide area. Japanese airborne Pi-SAR-L2 and PiSAR-L data were used in this analysis. Several L-band full polarimetric parameters, including backscattering coefficient (σ ), coherence between two polarimetric states, four-component decomposition parameters (double-bounce/volume/surface/helix scattering), and eigenvalue decomposition parameters (entropy/α/anisotropy), were calculated to determine most appropriate parameters for detecting landslide areas. The change in land cover from forest before disaster to bare soil after disaster was detected well by α, and coherence between HH and VV. Observational data from bottom to top of landslide detected landslide well, whereas observations from opposite sides were not as useful, indicating that a smaller local incident angle is better to distinguish landslide and forested areas. Soil from landslide intruded into urban areas; however, none of full polarimetric parameters showed any significant differences between landslide-affected urban areas after disaster and unaffected areas before disaster. Keywords: polarimetry; disaster; L-band SAR 1. Introduction Full polarimetric SAR data are capable of identifying radar scattering mechanisms on ground, and y have been used to estimate land cover class by connecting radar backscattering mechanism to land cover condition both by day and by night in all wear conditions. Such characteristics make se data applicable to detection of a disaster area, especially for emergency observations made soon after a disaster happens. Watanabe et al. [1] used Japanese L-band satellite SAR (PALSAR; Phased Array type L-band Syntic Aperture Radar) full polarimetric data to detect landslide areas induced by Iwate-Miyagi Nairuku earthquake of 2008, using surface scattering component of a three-component decomposition model. Furrmore, σ HV has also been used to distinguish landslide areas with rough surfaces from or surface scattering areas, such as pastures and vacant pieces of land with smooth surfaces. Polarimetric decomposition analysis was conducted on data before and after a landslide event with ALOS PALSAR data [2]. For detection of landslides areas, 30-m resolution full polarimetric data using unsupervised classification based on Entropy-α plane are more useful than 10-m resolution single-polarization data. Czuchlewski et al. [3] use L-band airborne SAR polarimetry data, and identify extent of landslide, using scattering entropy, anisotropy, and pedestal height. They also pointed out that one post-event single polarized SAR image is insufficient for distinguishing and Remote Sens. 2016, 8, 282; doi: /rs

2 Remote Sens. 2016, 8, of 13 mapping landslides. Rodriguez et al. [4] use L-band airborne SAR polarimetry data, and show landslide scar areas are dominated by single-bounce scattering and surrounding forested regions are dominated by volume scattering. Radar vegetation index, pedestal height, and entropy are used to identify forest, to separate landslide area. Shimada et al. [5] used Japanese L-band airborne SAR (Pi-SAR-L2; Polarimetric and Interferometric Airborne Syntic Aperture Radar L2) data to show that change of land cover from forest before a disaster to bare soil after a disaster was detected well by polarimetric coherence between HH and VV (γ (HH)-(VV) ). Shibayama et al. [6] confirm usefulness of γ (HH)-(VV) for detecting a landslide. They also pointed out that in landslide areas, polarimetric indices of normalized surface scattering power (ps), normalized volume scattering power (pv), and γ (HH)-(VV) change drastically with local incidence angle, whereas in forested areas, se indices are stable, regardless of change in local incidence angle. Several full polarimetric parameters have been suggested to detect a landslide area since now. In this study, airborne full polarimetric L-band SAR data, obtained both before and after a landslide event, were used to determine most appropriate full polarimetric parameters and observation direction for identifying an area affected by a landslide induced by heavy rain. The data used in our analysis are unique for two reasons: (1) hey comprise full polarimetric data observed just after disaster (landslide). (2) They were observed from four different observational directions at same time after disaster. One of directions was also observed before disaster. To generalize our method, simple radar scattering models from rough surface were applied, as discussed in Section 4. The models were evaluated for two sites using three different local incident angles with two polarizations, simultaneously, which has rarely been undertaken for a landslide area. 2. Pi-SAR-L2 Data and Field Experiment On 16 October 2013, Typhoon Wipha struck Izu Oshima Island, which is located 100 km south of Tokyo (Figure 1), generating a rainfall rate that was recorded at mm/h. This heavy rain induced a large-scale landslide that affected an area of 1.14 million m 2 and led to 39 people being killed or missing. The Geospatial Information Authority of Japan (GSI) used aerial photographs taken after disaster to produce a landslide map [7], and main landslide areas are identified in Figure 2. The locations of many landslides can be observed in mountain area, and some material displaced by landslides intruded into residential areas. These data were used as validation data. The study area was observed before and after disaster using Japanese airborne SAR (Pi-SAR-L and Pi-SAR-L2) (Table 1). The Pi-SAR-L2 observations were acquired in four different observational directions (L L203204, Figure 1) six days after disaster. The time required for four flights was about one hour. Before disaster, one Pi-SAR-L observation (L03801) had been made on 30 August 2000, in same observational direction as L Three of four data (L203201, L203202, and L203203) were used to determine parameters and directions most appropriate for detecting landslide areas. L was not used, because its configuration (incident and azimuth angles to landslide area) is almost same as L data. These parameters for detecting landslide areas included backscattering coefficient (σ ), polarimetric coherence (γ), eigenvalue decomposition [8], and four-component decomposition parameters [9]. The γ is calculated from correlation between two polarimetric states (HH-HV-VV base, (HH+VV)-(HH-VV)-(HV) base). The eigenvalue decomposition parameters consist of entropy/α/anisotropy, and y were obtained using PolSARPro [10]. Entropy represents randomness of a scatterer, α represents scattering mechanism (0 for surface scattering, 45 for dipole scattering or single scattering by a cloud of anisotropic particles, and 90 for double-bounce scattering), and anisotropy represents relative importance of second and third eigenvalues. The four-component decomposition parameters (double-bounce/volume/surface/helix scattering) are related to surface, volume, double-bounce, and helix scattering components on earth s surface, and y were obtained using a program of our own making. The processing window size for calculating parameters was 7 ˆ 7 pixels.

3 Remote Sens. 2016, 8, of 13 Remote Sens.2016, 2016,8,8, Remote Sens. 3 of of Figure 1. After landslide: configuration of Pi-SAR-L2 observations performed using four different Figure After landslide: configuration observationsperformed performedusing using four different Figure After landslide: configuration of of Pi-SAR-L2 Pi-SAR-L2 observations four different observational directions (L L203204) on 22 October Before landslide: Pi-SAR-L observational directions (L L203204) on 22 October Before landslide: Pi-SAR-L observational directions (L L203204) on 22 October Before landslide: Pi-SAR-L observation (L03801) performed using same flight course as L on 30 August observation (L03801) as L L203201on on30 30August August observation (L03801)performed performedusing usingsame sameflight flight course course as Figure 2. Optical image of disaster area. Red polygon represents landslide map, produced by GSI [8]. (a) Before disaster (1 June 2010); (b) after disaster (17 October 2013). The green line Figure Opticalimage imageofof disaster disasterarea. area. Red Red polygon represents map, produced byby Figure Optical polygon represents landslide landslide map, indicates border between forest and or areas obtained from GSI, and light blueproduced polygon GSI [8]. (a) Before disaster (1 June 2010); (b) after disaster (17 October 2013). The green line GSI represents [8]. (a) Before experiment disaster (1 sites. June 2010); (b) after disaster (17 October 2013). The green line field indicates border between forest and or areas obtained from GSI, and light blue polygon indicates border between forest and or areas obtained from GSI, and light blue polygon represents field experiment sites. Table 1. Specification of Pi-SAR and Pi-SAR-L2. represents field experiment sites. Pi-SAR Pi-SAR-L2 Table 1.Items Specification of Pi-SAR and Pi-SAR-L2. Band width 50 MHz 85 MHz Table 1.Items Specification of Pi-SAR and Pi-SAR-L2. Pi-SAR Pi-SAR-L Sampling frequency 100 Band width 50 MHz 85 MHz MHz MHz Itemsfrequency Pi-SAR Pi-SAR-L2 Operation height 6 12 km 6 12 km Sampling 100 MHz MHz Spatial resolution (slant) 2.5 m 1.76 m Band width 50 MHz 85 MHz Operation(azimuth, height 6 12 km 6 12 km Spatial resolution 3.2 m 3.2 Sampling frequency 4look *) MHz 100mMHz Spatial resolution (slant) 2.5dB m 1.76dBm Noise equivalent sigma zero Operation height 6 12 km 6 12 km Spatial resolution (azimuth, 3.2 m m 3.2 mm Incidence angle deg deg. Spatial resolution (slant)4look *) 10~ Noise equivalent sigma zero 30 db 35 db Polarimetry full full Spatial resolution (azimuth, 4look *) 3.2 m 3.2 m Power 3.5 KW KW Incidence angle zero 10~60 deg deg. Noise equivalent sigma 30 db db Polarimetry fulldeg. full deg. * Number of multi-look. Incidence angle 10~ Power 3.5 KW 3.5 KW Polarimetry full full Power * Number of multi-look. 3.5 KW * Number of multi-look. 3.5 KW

4 Remote Sens. 2016, 8, of of A forest area map, obtained from national land numerical information download service managed A forest by area GSI (Figure map, obtained 2), was from used to national identify land forest numerical area information before disaster download [11]. service Field experiments managed bywere GSI (Figure performed 2), on was23 used and to 24 identify March forest The value area of before σ obtained disaster from [11]. bare soil Field is determined experimentsfrom were performed surface roughness, on 23 anddielectric 24 Marchconstant The of value soil of (equivalent σ obtained to fromvolumetric bare soil soil is determined moisture, from Mv), and surface local roughness, incident angle. dielectric The constant parameter of Mv soil was (equivalent not measured to in volumetric field experiments soil moisture, because Mv), and its local value incident was different angle. The from parameter that when Mv was Pi-SAR-L2 not measured observation in field was performed. experimentshowever, because itssurface value was roughness differentdoes fromnot thatchange when much, Pi-SAR-L2 if re observation is no disturbance. was performed. It was measured However, surface using a roughness needle profiler does not at two change sites much, (Sites if1 re and 2, is no Figure disturbance. 2) within It waslandslide measuredarea using to evaluate a needle profiler radar atbackscattering two sites (Sitesfrom 1 and bare 2, Figure soil (Figure 2) within 3). The red landslide points shown area to in evaluate photos radar were used backscattering to evaluate from bare surface soil roughness (Figure 3). The (s) and red correlation points shown length in (l) photos normalized were used by wave to evaluate number k (e.g., surface ks roughness and kl). Site (s) 1 and was correlation a slightly length rough (l) surface normalized covered by by wave soil number with a slope k (e.g., of ks 5, and for kl). which Siteks 1 was and akl slightly were evaluated rough surface as 0.38 covered and 6.17, by soil respectively. with a slope Site of 2 5, was for a rough which surface ks and kl covered were evaluated by soil and as volcanic 0.38 and 6.17, rocks respectively. with a slope Site of 20, was for a rough which surface values covered of ks byand soilkl and were volcanic determined rocks with as 1.85 a slope and 5.03, of 20, respectively. for which values of ks and kl were determined as 1.85 and 5.03, respectively. Figure 3. Site photos with a needle profiler. (a) Site 1; (b) Site 2. There are three well known and simple surface scattering models. The first of se is small There are three well known and simple surface scattering models. The first of se is small perturbation model (SPM) [12], which is valid for smooth surfaces (ks 0.3). The second, physical perturbation model (SPM) [12], which is valid for smooth surfaces (ks < 0.3). The second, physical optics model (POM) [13], is valid for slightly rough surfaces within parameter ranges Mv 0.25, optics model (POM) [13], is valid for slightly rough surfaces within parameter ranges Mv < 0.25, 2.76 sλ, and kl > 6. The third, geometric optics model (GOM) [13], is valid for rough surfaces l 2 > 2.76 sλ, and kl > 6. The third, geometric optics model (GOM) [13], is valid for rough surfaces and predicts that σ HH = σ VV at all incidence angles; this model is valid within parameter ranges and predicts that σ HH = σ VV at all incidence angles; this model is valid within parameter ranges kl 6, > 2.76 sλ, and ks cosθ > 1.5. The POM [13] is valid for Site 1, and is described by: kl > 6, l 2 > 2.76 sλ, and ks cosθ > 1.5. The POM [13] is valid for Site 1, and is described by: 2 [ ] ( 2ks cosθ ) p2ks cosθq 2n ( ) ( ) 2n σ pp( θ) = k cos θγp θ exp 2ks cosθ I (1) σ 0 pppθq k 2 cos 2 θγ p pθq exp p2ks cosθq 2ı ÿ 8 n= 1 n! n! n 1 2 nl Iexponential = I (2) (2) nl [ n + 2( klsinθ) ] n 3/ 2 2 ` 2 pklsinθq 2ı 3{2 In se equations, Γ(θ,ε) represents Fresnel reflection coefficient, p represents In se equations, Γ(θ,ε) represents Fresnel reflection coefficient, p represents polarization polarization (h or v), pp represents any combination of h and v, such as hh, hv, vh, vv, l is (h or v), pp represents any combination of h and v, such as hh, hv, vh, vv, l is correlation length, θ is correlation length, θ is local incident angle, and ε is relative dielectric constant, which is local incident angle, and ε is relative dielectric constant, which is related to soil moisture. related to soil moisture. The GOM [13] is almost valid for Site 2, and described by: The GOM [13] is almost valid for Site 2, and described by: Γ( 0) exp ( tan θ 2m ) σ pp ( θ ) = (3) 2 4 σ 0 pppθq Γ p0q ` tan exp2 2 θ{2m 2 2m cos θ 2m 2 cos 4 (3) θ ε Γ p0q ( 0 ) = 1?ε (4) 1 `?ε (4) 1+ where m is surface slope and represented by? 2s{l. where is surface slope and represented by 2 s l. ε I (1)

5 Remote Sens. 2016, 8, of 13 Remote Sens. 2016, 8, of 13 These models models and and parameters parameters obtained obtained in field experiment field experiment were used were toused evaluate to evaluate observed σ observed in Section σ in 4. Section Results 3.1. Landslide Area Detection with Full Polarimetric Parameters The four-component decomposition image obtained by Pi-SAR-L2 (ID: L203201) is presented in Figure 4. Maximum and minimum digital number values for each four-component decomposition parameter were assigned to to values of of 00 and in in RGB RGB scale. The The actual landslide is is represented by by blue blue color, color, indicating that that surface surface scattering scattering is is dominant. However, However, similar similar surface surface scattering scattering can also can also be observed be observed near near top of top Mt. of Mihara, Mt. Mihara, volcanic volcanic mountain mountain located located in in center center of island. of The island. same The situation same situation was alsowas observed also observed for or for parameters or parameters (entropy, α, (entropy, polarimetric α, polarimetric coherence between coherence (HH+VV) between (HH+VV) and (HH VV) and (HH VV) (γ (HH+VV)-(HH VV) (γ(hh+vv)-(hh VV)), ), γ HH-VV γhh-vv), and and two of of representative parameters, γ γ(hh+vv)-(hh VV), (HH+VV)-(HH VV) and, and α are αshown are shown in Figure in Figure 5a,c. 5a,c. The differences in se parameters before and after disaster ( γ (Δγ(HH+VV)-(HH VV), (HH+VV)-(HH VV) Δα), α) can can be be established by byvisual inspection (Figure 5b,d). It can be be seen seen that that γ (HH+VV)-(HH VV) Δγ(HH+VV)-(HH VV) shows differences for both landslide area and top of mountain, where no landslide occurred, whereas α Δα shows differences in landslide area only. This indicates that α is better than γ γ(hh+vv)-(hh VV) (HH+VV)-(HH VV) for detecting for landslide areas areas in inlarge scale images. Two or parameters ( γ (Δγ(HH)-(VV) (HH)-(VV) and Δentropy) entropy) also showed same characteristics as Δα; α; however, four-component decomposition parameters ( σ Double, (Δσ Double, Δσ Volume, σ Volume, Δσ Surface, σ Surface, Δσ Helix σ Helix in digital in digital numbers), Δσ σ (in (indigital number, and Δanisotropy, anisotropy, showed differences not only for landslide area, but also bare soil area, where no landslide occurred. The The visual visual interpretation revealed revealed that entropy, that entropy, α, and γ HH-VV α, and were γhh-vv prospective were parameters prospective for parameters detection for of landslide detection areas of landslide from large areas scale from images. large scale images. Figure Four-component Four-component decomposition decomposition image image obtained obtained by Pi-SAR-L2 by Pi-SAR-L2 (ID: L203201, (ID: L203201, R: Doublebounce scattering scattering G: Volume G: scattering Volume scattering B: Surface B: scattering). Surface scattering). The main landslide The mainarea landslide is surrounded area is R: Double-bounce surrounded by red rectangle. by red The rectangle. field experiment The fieldsites experiment are represented sites are represented by red circles. by red circles.

6 Remote Sens. 2016, 8, 282 Remote Sens. 2016, 8, of 13 6 of 13 Figure 5. Full polarimetric parameters after disaster and difference between before and after Figure 5. Full polarimetric parameters after disaster and difference between before and after disaster. (a) γ(hh+vv)-(hh VV); (b) Δγ(HH+VV)-(HH VV); (c) α; and (d) Δα. disaster. (a) γ(hh+vv)-(hh VV) ; (b) γ(hh+vv)-(hh VV) ; (c) α; and (d) α. The user s and producer s accuracies (Story and Congalton, [14]) for entropy, α, and γhh-vv were evaluated fromproducer s error matrices to estimate classification accuracy forentropy, detectingα, and landslide The user s and accuracies (Story and Congalton, [14]) for γhh-vv were areas shown in red rectangle in Figure 5. Threshold levels were estimated from a cross-over point evaluated from error matrices to estimate classification accuracy for detecting landslide obtained from histogram of landslide and non-landslide areas for each parameter. areas shown in redwas rectangle in Figure 5. Threshold wereclasses estimated fromand a cross-over The classification conducted based on threshold levels level. Two (landslide no point obtained set, histogram of accuracy landslide non-landslide areas for each landslide) from have been and classification for and case using parameters obtained after parameter. disaster, and was case using difference parameter before andclasses after (landslide disaster are and no The classification conducted based onin thresholdvalues level. Two evaluated. The results are summarized in Table 2. User accuracy is a measure indicating landslide) have been set, and classification accuracy for case using parameters obtained after probability that a pixel is grouped into Class A, given that classifier has labeled pixel as disaster, and case using difference in parameter values before and after disaster belonging to Class A. The producer accuracy is a measure indicating probability that classifier are evaluated. The results are summarized in Table 2. User accuracy is a measure indicating has labeled an image pixel as belonging to Class A, given that ground truth is Class A. probability that a pixel is groupedareinto given classifierwhen has labeled pixel as The accuracies for γ(hh+vv)-(hh VV) alsoclass shown A, in samethat table for reference. data before belonging Class A. Thewere producer accuracy a measure indicating probability that58.7% classifier andto after disaster used, values ofisδα and ΔγHH-VV show user s accuracies of about 60.9% an andimage producer s of aboutto33.8% 35.8%. The producer s accuraciestruth are better than A. The has labeled pixelaccuracies as belonging Class A, given that ground is Class accuracies for γ(hh+vv)-(hh VV) are also shown in same table for reference. When data before and after disaster were used, values of α and γhh-vv show user s accuracies of about 58.7% 60.9% and producer s accuracies of about 33.8% 35.8%. The producer s accuracies are better than those of 25.9% 26.8% for entropy and γ(hh+vv)-(hh VV). The accuracies are almost same as those achieved when using parameters obtained after disaster, i.e., values of α and γhh-vv show user s accuracies of about 59.5% 61.8% and producer s accuracies of about 34.0% 38.4%. The

7 Remote Sens. 2016, 8, of 13 Remote Sens. 2016, 8, of 13 producer s accuracies of about 11.7% 11.8% for entropy and γ (HH+VV)-(HH VV) obtained after those of 25.9% 26.8% for Δentropy and Δγ(HH+VV)-(HH VV). The accuracies are almost same as those disaster are 14.2% 15% lower than entropy and γ (HH+VV)-(HH VV). This may be due to poor achieved when using parameters obtained after disaster, i.e., values of α and γhh-vv show classification user s accuracies accuracyof between about 59.5% 61.8% a forest and landslide producer s areas inaccuracies se parameters. of about 34.0% 38.4%. The producer s landslideaccuracies area detection of about was 11.7% 11.8% performedfor for entropy forest and γ(hh+vv)-(hh VV) area before obtained landslide after using forest disaster map and are 14.2% 15% results are lower alsothan presented Δentropy in Table and Δγ(HH+VV)-(HH VV). 2. Producer sthis accuracies may be are due especially to poor improved, and classification accuracy accuracy changes between from 35.8% a forest toand 52.2% landslide for areas α in parameter se parameters. (improved by 16.4%), and from 33.8% to The 49.5% landslide for area γ detection (HH)-(VV) parameters was performed (improved for by forest 15.7%). area before landslide using forest map and results are also presented in Table 2. Producer s accuracies are especially improved, and accuracy changes from 35.8% to 52.2% for α parameter (improved by 16.4%), Table 2. User s and producer s accuracies for detecting landslide areas using full polarimetric parameters. and from 33.8% to 49.5% for γ(hh)-(vv) parameters (improved by 15.7%). Using Difference in Table 2. User s and producer s accuracies for detecting Using landslide Parameters areas Obtained using full polarimetric parameters. Parameter Values before and Full Polarimetric after Disaster Area Using Parameters Obtained after Using Difference after in Parameter Disaster Full Parameters Disaster Values before and after Disaster Area Polarimetric User s Producer s User s Producer s User s Accuracy Producer s User s Producer s Accuracy Parameters Accuracy (%) Accuracy (%) Accuracy (%) Accuracy (%) (%) Accuracy (%) Accuracy (%) (%) α α All areas near γ (HH)-(VV) All areas near γ(hh)-(vv) landslide Entropy landslide Entropy γ (HH+VV)-(HH VV) γ(hh+vv)-(hh VV) Forest Forest areas areas before α landslide identified. α γ 66.7 HH-VV before 49.5 landslide γhh-vv identified Landslide Area Detection in Three Different Observational Directions 3.2. The Landslide entropy, Area α, and Detection anisotropy in Three obtained Different Observational from three Directions different observational directions (L203201, L203202, The andentropy, L203203) α, with and anisotropy forest mask, obtained where from only three forest different area is observational picked up, directions are presented in Figure (L203201, 6. TheL203202, main landslide L203203) areawith is delineated forest mask, by where red rectangle. only forest Thearea forest is picked area isup, indicated are as yellow presented area in and Figure 6. blue The main area represents landslide area is bare delineated soil area, by which red rectangle. also indicates The forest landslide area is area. indicated as yellow area and blue area represents bare soil area, which also indicates A visual inspection reveals that landslide area is detected well in L203201, wherein observations landslide area. A visual inspection reveals that landslide area is detected well in L203201, wherein are made from bottom to top of landslide, whereas clarity is lower in L203203, with observations are made from bottom to top of landslide, whereas clarity is lower in observations L203203, with made observations from top made tofrom bottom top to of bottom landslide. of landslide. Figure 6. (a c) entropy, α, and anisotropy obtained from three different observational directions. The landslide area used for validation [7] is shown by red lines in (a). The field experiment sites are represented by light blue circles.

8 Remote Sens. 2016, 8, of 13 The α and γ (HH)-(VV) values derived from L and L are presented in Table 3 for Site 1, Remote Sens. 2016, 8, of 13 Site 2, and a forest near Site 2. The values obtained from L lie between those for L and L and Figure have 6. been (a c) entropy, omitted α, and from anisotropy table. obtained Small from three directional different observational dependency directions. canthe be observed for Site 1, becauselandslide slope area used is 5. for Small validation directional [7] is shown dependency by red lines in (a). is also The field observed experiment forsites are forest area near represented by light blue circles. Site 2, because random scattering in forest canopy induces less directional dependency. However, relatively larger The directional α and γ(hh)-(vv) dependency values derived can from be L observed and L for Site are 2, presented becausein Table slope 3 for is Site 20. 1, Site 2, and a forest near Site 2. The values obtained from L lie between those for L and L Tableand 3. αhave andbeen γ (HH)-(VV) omitted values from derived table. Small from directional L and dependency after can be observed disaster. for Site 1, because slope is 5. Small directional dependency is also observed for forest area near Site 2, because random scattering in α forest canopy induces less γ (HH)-(VV) directional dependency. γ (HH+VV)-(HH VV) However, relatively larger directional dependency can be observed for Site 2, because slope is 20. Mean St. dev. Diff. Mean St. dev. Diff. Mean St. dev. Diff. Site 1 L Table 3. α and 33.1γ(HH)-(VV) 2.6 values derived from L and after 0.06 disaster. L Site 2 L α 0.77 γ(hh)-(vv) 0.05 γ(hh+vv)-(hh VV) L Mean 4.1St. dev. Diff. Mean 0.45 St. dev Diff. Mean 0.33 St. dev. Diff L L Forest near Site Site L L Site 2 L L Histograms of α for Site L and49.7 forest 2.1 near Site are compared to examine detectability of Forest near Site landslide area in a forested L region; 48.1 results 2.5 are presented 0.35 in 0.08 Figure 7. The 0.23 histograms 0.08 for forested area shows same for L and L203203, indicating re is no local incident angle dependency. Histograms of α for Site 2 and forest near Site 2 are compared to examine detectability of On or landslide hand, area histograms in a forested for region; landslide results are area presented showsin difference, Figure 7. The indicating histograms for local incident angle dependency. forested area The shows results same arefor consistent L and with L203203, one indicating obtained re by Shibayama is no local incident et al. [6]. angle The α value from landslide dependency. area in On L or hand, is clearly histograms different for from landslide value area obtained shows difference, from indicating forest, whereas local incident angle dependency. The results are consistent with one obtained by α value from landslide area in L overlaps that from forest. Some of landslide area Shibayama et al. [6]. The α value from landslide area in L is clearly different from value is misclassified obtained asfrom forest, forest, which whereas reduces α value classification from landslide accuracy area in L in L overlaps that The from same pattern is observed forest. Some γ (HH)-(VV) of landslide case. This area indicates is misclassified thatas a forest, smaller which local reduces incident classification angle is better accuracy to distinguish landslide in and L forested The areas. same pattern is observed for γ(hh)-(vv) case. This indicates that a smaller local incident angle is better to distinguish landslide and forested areas. The user s and producer s accuracies for detecting landslide areas based on different The user s and producer s accuracies for detecting landslide areas based on different observational observational directions directions by αby areα are summarized in Table L L shows shows good accuracy good accuracy with user s with user s and producer s and producer s accuracies accuracies of 66.7% of 66.7% and and 52.1%, respectively. L shows shows worst accuracy worst with accuracy with user s anduser s producer s and producer s accuracies accuracies of of 59.1% and 16.4%, respectively, as was asexpected was expected from from visual visual inspection in Figure 6. inspection in Figure Figure Figure 7. Histogram 7. Histogram of of α for Site Site2 2and and a forest a forest near Site near 2. Site 2. Table 4. User s and producer s accuracies for detecting landslide areas based on different observational directions. Accuracy (%) L L L User s Producer s

9 Remote Sens. 2016, 8, of Discussion The accuracy of detection of landslide area was improved when forest area was picked up. The landslide area is picked up and shows difference of entropy/α/anisotropy only before and after disaster in Figure 8. Maximum and minimum values for each parameter are assigned to values of 0 and 255 in RGB scale. Yellow areas seen in upper right of Figure 8 indicate significant change in entropy and α before and after disaster. To a large extent, this area changed from forest to bare soil after landslide. However, significant parameter change is not observed in upper-left area of Figure 8. This is residential area, which changed from a normal residential area before disaster to a residential area with mud induced by landslide after disaster. No significant change could be detected by full polarimetric parameters in this area. There are or areas without significant parameter change, such as in lower right of Figure 8. Here, re are many narrow valleys and layover/shadowing might prevent any significant change in full polarimetric parameters in this region. An outline of detection accuracy by using γ HH-VV and α is presented in Figure 9. The misidentification of residential area and narrow valleys result in a low producer s accuracy of 33.8% 35.8%. On or hand, higher user s accuracy of 58.7% 60.9% is due to detection of change from forested area to landslide area. Remote Sens. 2016, 8, of 13 - Figure 8. The difference in entropy/α/anisotropy before and after disaster for landslide area. Red: entropy Δentropy Green: α Δα Blue: Δanisotropy anisotropy.

10 - Figure 8. The difference in entropy/α/anisotropy before and after disaster for landslide area. Remote Sens. 2016, 8, of 13 Red: Δentropy Green: Δα Blue: Δanisotropy Figure 9. An outline of detection accuracy using γ ΔγHH-VV andδα. α. The accuracy is almost same when using α and γ(hh)-(vv) γ after disaster, and using (HH+VV)-(HH VV) for landslide area is lower than γ (HH)-(VV) (Table 3). Very little double-bounce difference between α and γ(hh)-(vv) before and after disaster. This indicates that α and γ(hh)-(vv) scattering, indicated by HH VV component, is expected from landslide area, and this led to are good smaller parameters γ to distinguish forest and bare soil areas. This is also supported by accuracy (HH+VV)-(HH VV). Since a smaller γ (HH+VV)-(HH VV) indicates a smaller γ (HH+VV)-(HH VV) between landslide and forested area, detection accuracy for γ (HH+VV)-(HH VV) parameter is lower than that for γ (HH)-(VV) parameter. The γ (HH+VV)-(HH VV) parameter highlights changes at top of Mt. Mihara (Figure 5b). Small changes in vegetation and soil moisture are expected between August and October observations; γ (HH+VV)-(HH VV) may be sensitive to such changes, unlike γ (HH)-(VV) and α. The accuracy is almost same when using α and γ (HH)-(VV) after disaster, and using difference between α and γ (HH)-(VV) before and after disaster. This indicates that α and γ (HH)-(VV) are good parameters to distinguish forest and bare soil areas. This is also supported by accuracy improvement after forest area is identified. The α and γ (HH)-(VV) obtained before disaster are essential for distinguishing between areas of bare soil before disaster and landslides induced by disaster, and between bare soil and forest. The values of σ obtained from Pi-SAR-L2 and estimated from oretical models are plotted against local incident angle in Figure 10. The L203201, L203202, and L observations correspond to minimum, median, and maximum local incident angle plots in figure. Pi-SAR-L2 observations were performed six days after disaster, and value of Mv is assumed as 100% for Site 1. The values of σ obtained from Pi-SAR-L2 show same pattern against local incident angle, but with an offset of a few db from POM. If value of ks was not 0.38 (measured on site), it was assumed as 0.5, and model describes data obtained from three different local incident angles well. There was a 1.5-year interval between Pi-SAR-L2 observations and surface roughness measurements and, thus, changes of 30% in smoothness of surface during that period could account for difference of a few db in σ values. The GOM works well for Site 2 if Mv = 25% is assumed, except for L data. The dielectric constant of volcanic rock, which partly covers surface of Site 2, is generally 5 10, and this corresponds to an Mv of 8% 18.8%. The assumption of Mv = 25% could be explained by mixture of rock and soil. At Site 2, a few larger rocks could not be characterized by surface roughness measurements and shadowing of se rocks might

11 roughness measurements and, thus, changes of 30% in smoothness of surface during that period could account for difference of a few db in σ values. The GOM works well for Site 2 if Mv = 25% is assumed, except for L data. The dielectric constant of volcanic rock, which partly covers surface of Site 2, is generally 5 10, and this corresponds to an Mv of 8% 18.8%. The assumption of Mv = 25% could be explained by mixture of rock and soil. At Site 2, a few larger Remote Sens. 2016, 8, of 13 rocks could not be characterized by surface roughness measurements and shadowing of se rocks might unexpectedly have affected values of σ for L observations. However, unexpectedly values of σ observed have affected for values landslide of σ areas for are L represented observations. moderately However, by simple values surface of σ observed scattering for models. landslide areas are represented moderately by simple surface scattering models. Figure 10. σ σ obtained from Pi-SAR-L2 and estimated from oretical models. 5. Conclusions Conclusions Pi-SAR-L2 full full polarimetric polarimetric data dataobserved observedin infour fourdifferent differentobservational directions directionsover over a alandslide area area were were analyzed analyzed to clarify to clarify most most appropriate appropriate L-band L-band full full polarimetric polarimetric parameters parameters and and observational observational direction direction to detect to detect a landslide a landslide area. area. Data Data from fromone onepi-sar-l Pi-SAR-Lobservation performed before disaster occurred were were also also used used in in this this analysis. A summary of of preferable parameters to detect landslide area area was was added added in in Table Table data before and after disaster were used, Δα and ΔγHH-VV When data before and after disaster were used, α and γ HH-VV showed user s accuracies of about 58.7% 60.9% and and producer s accuracies of of 33.8% 35.8%, indicating better performance than or or parameters, parameters, such such as as four-component four-component decomposition decomposition parameters ( σ Double, (Δσ Double, Δσ Volume, σ Volume, Δσ Surface, σ Surface, and Δσ Helix), and σ Helix), Δσ, Δγ(HH+VV)-(HH VV), σ, γ Δentropy, and Δanisotropy. (HH+VV)-(HH VV), entropy, and anisotropy. Or two two knowledge knowledge obtained obtained from from our our analysis analysis are: are: The detection accuracy is almost same when using parameters after disaster, and The detection accuracy is almost same when using parameters after disaster, and using difference between parameters before and after disaster. using difference between parameters before and after disaster. Producer s accuracies are improved, and accuracy changes from 35.8% to 52.2% for Producer s accuracies are improved, and accuracy changes from 35.8% to 52.2% for α parameter (improved by 16.4%), and from 33.8% to 49.5% for γ(hh)-(vv) parameters (improved parameter (improved by 16.4%), and from 33.8% to 49.5% for γ (HH)-(VV) parameters (improved by 15.7%), when evaluated by α and γ (HH)-(VV) parameters, if forested area before disaster is identified. Table 5. Summary for preferable parameters to detect landslide area. Parameters Local Incident Angle ForestÑLandslide α, γhh-vv Ó Ó Low Good High Moderate Entropy, σ Surface, γ(hh+vv)-(hh VV) σ Double, σ Volume, σ Helix, σ, Anisotropy All All Moderate Not good Land Cover Change Residential areañresidential Area with Mud Induced by Landslide Not good.

12 Remote Sens. 2016, 8, of 13 However, land cover change from residential area before disaster to residential area with mud induced by landslide after disaster could not be detected by full polarimetric parameters. The landslide area was clearly identifiable using data observed from bottom of landslide to top. The clarity was degraded when using data observed from top of landslide to bottom, indicating that smaller local incident angle is better to distinguish landslide and forested area. The observed σ for landslide areas was moderately represented using two simple models: POM for slightly rough surfaces, and GOM for rough surfaces. The α and γ HH-VV obtained from full polarimetric L-band SAR data are capable of identifying landslides, which is especially useful for emergency observations taken just after a disaster occurs; however, parameters only detect change from forest cover to bare soil. None of representative full polarimetric parameters showed any significant differences between landslide-affected urban areas after disaster and unaffected areas before disaster. The α and γ (HH)-(VV) obtained before disaster are essential for distinguishing between areas of bare soil before disaster and landslides induced by disaster. Acknowledgments: Agency (JAXA). Pi-SAR-L2 and Pi-SAR data are copyrighted by Japan Aerospace Exploration Author Contributions: M. Watanabe conceived, designed, and performed experiments, analyzed data, wrote paper. R. B. Thapa performed experiments, and R. B. Thapa and M. Shimada contributed to discussion andinterpretation of results. Conflicts of Interest: The authors declare no conflict of interest. References 1. Watanabe, M.; Yonezawa, C.; Iisaka, J.; Sato, M. ALOS/PALSAR full polarimetric observations of Iwate-Miyagi Nairiku earthquake of Int. J. Remote Sens. 2012, 33, [CrossRef] 2. Yonezawa, C.; Watanabe, M.; Saito, G. Polarimetric decomposition analysis of ALOS-PALSAR observation data before and after a landslide event. Remote Sens. 2012, 4, [CrossRef] 3. Czuchlewski, K.R.; Weissel, J.K.; Kim, Y. Polarimetric syntic aperture radar study of Tsaoling landslide generated by 1999 Chi-Chi earthquake, Taiwan. J. Geophys. Res. 2003, 108, [CrossRef] 4. Rodriguez, K.M.; Weissel, J.K.; Kim, Y. Classification of landslide surfaces using fully polarimetric SAR: Examples from Taiwan. IEEE Proc. Int. Geosci. Remote Sens. Symp. 2002, 2002, Shimada, M.; Watanabe, M.; Kawano, N.; Ohki, M.; Motooka, T.; Wada, W. Detecting mountainous landslides by SAR polarimetry: A comparative study using Pi-SAR-L2 and X band SARs. Trans. Jpn. Soc. Aeronaut. Space Sci., Aerosp. Technol. Jpn. 2014, 12, [CrossRef] 6. Shibayama, T.; Yamaguchi, Y.; Yamada, H. Polarimetric scattering properties of landslides in forested areas and dependence on local incidence Angle. Remote Sens. 2015, 7, [CrossRef] 7. Geospatial Information Authority of Japan. Available online: index.html (accessed on 18 March 2015). 8. Cloude, S.R.; Pottier, E. A review of target decomposition orems in radar polarimetry. IEEE Trans. Geoscie. Remote Sens. 1996, 34, [CrossRef] 9. Yamaguchi, Y.; Yajima, Y.; Yamada, H. Four-component scattering model for polarimetric SAR image decomposition. IEEE Trans. Geosci. Remote Sens. 2005, 43, [CrossRef] 10. Pottier, E.; Ferro-Famil, L.; Allain, S.; Cloude, S.; Hajnsek, I.; Papathanassiou, K.; Moreira, A.; Williams, M.; Minchella, A.; Lavalle, M.; et al. Overview of PolSARpro V4.0 software. The open-source toolbox for polarimetric and interferometric polarimetric SAR data processing. IEEE Proc. Int. Geosci. Remote Sens. Symp. 2009, 2009, Geospatial Information Authority of Japan. Available online: (accessed on 18 March 2015). 12. Chen, M.F.; Fung, A.K. A numerical study of regions of validity of Kirchhoff and small-perturbation rough surface scattering model. Radio Sci. 1988, 23, [CrossRef]

13 Remote Sens. 2016, 8, of Ulaby, F.T.; Kouyate, F.; Fung, A.K.; Sieber, A.J. A backscatter model for a randomly perturbed periodic surface. IEEE Trans. Geosci. Remote Sens. 1982, GRS-20, [CrossRef] 14. Story, M.; Congalton, R.G. Accuracy assessment: A user's perspective. Photogramm. Eng. Remote Sens. 1986, 52, by authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under terms and conditions of Creative Commons by Attribution (CC-BY) license (