Landslide detection based on height and amplitude differences using pre- and post-event airborne X-band SAR data
The recognition of landslides and making their inventory map are considered to be urgent tasks not only for damage estimation but also for planning rescue and restoration activities. Owing to the capability of synthetic aperture radar (SAR) for day-and-night and all-weather imaging, various studies utilizing SAR data for landslide detection have been reported to date. Among the detection methods utilizing SAR data, those based on height differences accompanying landslides are attractive and should be further improved, since they can directly contribute to damage estimation through a volumetric estimation of landslides. In this context, we propose in this paper a landslide detection method utilizing height differences derived from pre- and post-event SAR digital elevation models (DEMs) combined with amplitude differences. The proposed method was applied to the landslides triggered by the 2016 Kumamoto earthquake. The application results demonstrate that SAR DEMs with a high altitudinal resolution can improve the detection ability and that the incorporation of the amplitude differences is effective for decreasing the number of false detections. Although the reliability of the proposed method is deemed moderate when evaluated on the basis of the kappa coefficients derived through an accuracy assessment, most of the outliers are correctly filtered out and large- and medium-scale landslides are detected. Therefore, the inventory maps derived from the proposed method are thought to be effective at the initial stage of planning rescue and restoration activities.
KeywordsLandslide Height difference map Amplitude difference map InSAR DEM
According to Cruden (1991), a landslide is defined as the mass movement of rock, earth or debris down a slope. Landslides can be induced by natural events and human activities such as heavy rainfall, earthquakes, deforestation and the excavation of slopes (Dai et al. 2002). Landslides wreak havoc on human lives and infrastructures in mountainous areas (Dai et al. 2002; Metternicht et al. 2005; Scaioni et al. 2014); thus, the recognition of landslides and making their inventory map are urgent tasks not only to support damage estimation but also to plan rescue and restoration activities (Plank 2014; Plank et al. 2016). Owing to its capability of wide-area observation, remote sensing based on spaceborne and airborne sensors is a powerful tool for performing this task. Landslide recognition has been performed using various remotely sensed data such as optical passive sensor data, thermal infrared data, laser scanning data and synthetic aperture radar (SAR) data (Scaioni et al. 2014).
SAR data hold a unique position among these remotely sensed data owing to the capability of SAR for day-and-night and all-weather imaging. The processed SAR data are expressed as a complex image that consists of amplitude and phase at each resolution pixel (Oliver and Shaun 2004). The amplitude data alone can be directory utilized for visual interpretation. On the other hand, the phase data are often combined with two or more SAR data acquired from similar orbits to derive a digital elevation model (DEM) through the interferometric SAR (InSAR) technique, as well as to derive minute changes in the slant range direction through the differential InSAR (DInSAR) technique (Lu et al. 2010). Moreover, polarimetric SAR data have been studied for terrain and land use classification. Various studies utilizing SAR data for landslide detection and characterization have been reported, even for these ten years (e.g., Cao et al. 2008; Liao and Shen 2009; Christophe et al. 2010; Dong et al. 2011; Kawamura et al. 2011; Furuta and Sawada 2013; Zhao et al. 2013; Liu and Yamazaki 2015; Shibayama et al. 2015; Tang et al. 2015; Plank et al. 2016).
These methods can be roughly classified into two groups: those dependent on the coherency of coevent SAR data and those that are not. One of the advantages of the former is that they can treat surface changes or movements smaller than the wavelength of a SAR system. For example, Zhao et al. (2013) analyzed the L-band Advanced Observing Satellite (ALOS) Phased Array-type L-band SAR (PALSAR) using the short baseline subset (SBAS) InSAR. They detected pre-rockslide movement of less than 50 cm within 184 days for the 2009 Jiweshan rockslide in China. However, the requirement of coherency between coevent data imposes a severe constraint on their spatial and temporal baselines. On the other hand, for the latter methods, the constraint on the baselines is less rigorous, although the scale of the phenomena that they can treat depends on the spatial resolution of the SAR system, which generally becomes coarser than those of the coherency-based methods. For example, to detect landslides in combination with a slope map derived from an external DEM, Liu and Yamazaki (2015) analyzed and visually assessed the difference in polarimetric backscattering coefficients for the coevent data obtained from different airborne L-band SAR sensors having a time lag of more than 10 years. Their research indicated that the coherency-independent methods have more chances of being applied to landslides detection than the coherency-dependent ones. Moreover, Plank et al. (2016) proposed and assessed quantitatively a landslide detection method that utilized high-resolution satellite optical imagery as the pre-event data and very high-resolution satellite SAR imagery as the post-event data.
Now, we further classify the coherency-independent methods into two groups on the basis of their viewpoints: One viewpoint is the land cover change, and the other is the height change in association with landslides. The former methods assume that landslides replace vegetated areas with bare soil or rock. However, such a land cover change can be triggered not only by landslides but also by human activities such as the cultivation of flatlands. Thus, some previous studies involved attempts to filter out the detected land cover changes independent from landslides using a slope map calculated from an external DEM or a post-event SAR DEM (e.g., Cao et al. 2008; Liu and Yamazaki 2015; Plank et al. 2016). On the other hand, the latter methods treat the height changes of landslides. Arturi et al. (2003) generated a post-event SAR DEM from the European Remote Sensing (ERS)-1 and ERS-2 tandem pairs and subtracted it from an external pre-event DEM. They mentioned that the DEM differences showed a good fit to landslide features qualitatively, although they also showed strong differences where the landslide did not produce evaluable effects. The methods utilizing the SAR DEM seem to be attractive since they may successively contribute to damage estimation through a volumetric estimation of landslides and thus to the planning of subsequent rescue and restoration activities. To improve the detection accuracy of landslides from the subtraction of pre- and post-event DEMs, the following two ideas can be pointed out. The first is to incorporate other factors given that the landslide detection methods based on land cover changes are often combined with a slope map. This would contribute to a decrease in the number of false detections. The second is to utilize more accurate DEM datasets. For example, to generate a height change map for the 2009 Jiweshan rockslide, Zhao et al. (2013) generated a SAR DEM by stacking several post-event ALOS InSAR data and then subtracted Shuttle Radar Topography Mission (SRTM) DEM from the obtained SAR DEM. They reported that the accuracy of their height change map is approximately 14 m. Tang et al. (2015) also generated a height change map and performed a volumetric estimation for the 2008 Wenjiagou landslide using SRTM and SAR DEM calculated from TerraSAR-X and TanDEM-X.
Considering the aforementioned two ideas, in this paper, we propose a landslide detection method utilizing height differences derived from pre- and post-event SAR DEM combined with amplitude differences. The rest of this paper is organized as follows. In Sect. 2, we introduce a study event and datasets. The proposed method is described in Sect. 3. In Sect. 4, we demonstrate the proposed method and discuss the application results. Finally, conclusions are given in Sect. 5.
2 Study area and datasets
A large earthquake, called the 2016 Kumamoto earthquake, occurred on April 16, 2016, at 01:25 JST (Japanese standard time, JST = UTC + 9 h) in Kyusyu, Japan. The epicenter and magnitude of this earthquake were 32.8°N, 130.8°E and M7.3, respectively, as determined by the Japan Meteorological Agency (Yagi et al. 2016). Yagi et al. (2016) constructed a rupture process model of this earthquake. According to their model, the mainshock rupture mainly propagated northeastward from the epicenter and terminated near the southwest side of Aso volcano. A large number of landslides occurred because of this earthquake. In addition to field surveys, various observations such as aerial and satellite photographs and SAR data acquisition and aerial laser scanning were performed after this earthquake occurred.
The National Institute of Information and Communications Technology (NICT) performed SAR observation on April 17, 2016, the day after the earthquake occurred, using the airborne X-band SAR called Pi-SAR2, which has been developed by NICT since 2006 (Nadai et al. 2009). Pi-SAR2 is a left-side-looking SAR having the capabilities of full polarimetric and cross- and/or along-track interferometric observations (XTI and/or ATI) simultaneously (Kojima et al. 2014). Among the operation modes of Pi-SAR2, the finest spatial resolution of 0.3 m is achieved when operated in modes 0 and 1. In these operation modes, the center frequency and bandwidth are 9.55 GHz and 500 MHz, respectively. Pi-SAR2 collects XTI data using the main and subantennas suspended separately at the base of the left and right wings of a Gulfstream II aircraft. The baseline length for XTI is approximately 2.6 m, and the standard deviation of height measurement of XTI was estimated as 2.4 m using corner reflectors on a runway in a preliminary evaluation (Kobayashi et al. 2012).
Observation parameters for pre- and post-event data
3.1 Landslide features on amplitude and height difference maps
The expected features of landslides on the height and amplitude difference maps are as follows: At the source area, the ground is gouged, and thereby, the height decreases. At the same time, land cover objects such as trees are scoured away. This results in the increase in the amplitude of vertical polarization since the land cover change from vegetation to bare soil or rock makes the surface scattering dominant. In addition, steep slopes might form, resulting in the formation of radar shadow areas where tree canopies were observed in the pre-event data. Therefore, the amplitude can increase or decrease. On the other hand, at the deposit area, flowed-in soil or rock is heaped on the former land cover objects. Then, the height increases and the amplitude also increases. On the basis of the consideration mentioned above, the landslide features are summarized into the following three patterns: The height decreases and the amplitude increases, the height decreases and the amplitude also decreases, and the height increases and the amplitude also increases. The proposed method is designed to extract these three patterns. Note that meaningful information cannot be extracted from low-coherence areas in both pre- and post-event data. Thus, we define the coherence threshold of 0.8 to exclude these areas from landslide detection.
3.2 Data processing for each pre- and post-event data
3.3 Coregistration between pre- and post-event data
3.4 Difference processing
3.5 Fusion processing
To fuse the height and amplitude differences, we calculate the ZNCC coefficient between them in the proposed method. Thresholding based on the ZNCC coefficient is expected to be effective for noise reduction, since there is no need for the unchanged areas to correlate with each other. To do so, the height and amplitude difference maps are binarized using the thresholds mentioned above; then, the ZNCC coefficients are calculated for the three patterns described in Sect. 3.1. The pixels having a ZNCC coefficient of 0.2 and more are regarded as the candidate areas of landslides. The candidates are then segmented into polygons based on the region-growing method after performing opening and closing processes. Finally, as performed by Liu and Yamazaki (2015) and Plank et al. (2016), small blocks less than 400 m2 are excluded.
4 Results and discussion
In this section, we describe and discuss the results of the application of the proposed method to the scenes listed in Table 1.
4.1 Amplitude differences
4.2 Height differences
Except for small-scale landslides, the areas of the landslides as well as the slope failures are colored in Fig. 6. Most areas are colored blue. The red-colored area that stands out is at the middle part in Fig. 6b at the most. From a comparison between Figs. 6b and 1b, this red-colored area can be recognized as a piled up area of flowed-in soil. However, we can find no such area around the gullet. This might be due to the existence of radar shadows. The cultivation areas, which are prominent in Fig. 5a, b, are scarcely colored in Fig. 6a, b. In addition, although the residential areas are colored in spots, the coherence between the amplitude and the height difference maps appears to be not so high.
4.3 Fused differences
4.4 Accuracy assessment
Confusion matrices with the overall, producer’s and user’s accuracies and the kappa coefficient for Scenes B and C
The overall accuracies and kappa coefficients for Scenes B and C are 87%, 0.60 and 95%, 0.46, respectively. On the basis of these kappa coefficients, the reliability of the proposed method is evaluated as moderate. It is worth noting that the producer’s accuracy in Scene C is noticeably low. This is because of two reasons: One is that the small-scale landslides are not detected, and the other is the imperfect detection of the medium-scale landslides. The detected areas for medium-scale landslides are half of them at best, as seen in Fig. 9c. As shown in Fig. 6c, it can be seen that these landslides are not properly identified in the height difference map. This means that the height difference of a part of medium-scale landslides is measured as not greater than the threshold of 5 m in addition to the small-scale ones. The following two possibilities can be considered. One is that the height difference is actually not greater than the threshold. The height can decrease when the land cover objects such as trees are scoured away, while it can increase when flowed-in soil or rock accumulates. Their synergistic effects result in the decrease in the magnitude of height changes. The other reason is possibly the limitation of height measurement accuracy. The baseline errors are not estimated, the offset phase of ϕ0,pre in Eq. (5a) is enforced to be 0, and we only estimate the offsets independent from the pixel positions in the proposed method, as mentioned in Sect. 3.4. Thus, the estimated height differences would contain systematic errors depending on the slant range position and/or the elevation height. These systematic errors might cause underestimation to some extent. Through the comparison between the measured height differences and truth data, these two possibilities should be discriminated in future studies.
The recognition of landslides and making their inventory map are urgent tasks not only to support damage estimation but also for planning rescue and restoration activities (Plank 2014; Plank et al. 2016). To date, various studies utilizing SAR data for landslide detection have been reported. Among the methods utilizing SAR data, those based on height differences accompanying landslides seem attractive since they may successively contribute to damage estimation through a volumetric estimation of landslides. In this context, we propose a landslide detection method utilizing height differences derived from pre- and post-event Pi-SAR2 DEMs combined with amplitude differences. The proposed method is applied to the landslides accompanying the 2016 Kumamoto earthquake, and the accuracy of the detection result is assessed using truth data. Through the application and accuracy assessment, it is demonstrated that the detection ability improves with the utilization of SAR DEMs with a higher altitudinal resolution. In addition, it is demonstrated that the incorporation of amplitude differences is effective for reducing the number of outliers that appear in the height difference map. The overall accuracy and kappa coefficient for the two scenes including landslides are 87%, 0.60 and 95%, 0.46, respectively. Although the reliability of the proposed method is moderate, as evaluated on the basis of these kappa coefficients, most of the cultivation and residential areas are correctly filtered out, and large- and medium-scale landslides are detected so that the inventory maps derived from the proposed method should be effective at the initial stage of planning rescue and restoration activities. The reasons why the reliability is not good but only moderate are that small-scale landslides cannot be detected and medium-scale ones are imperfectly detected. There are two possible reasons for this. One is that the height differences in these areas are actually not greater than the 5 m threshold. The other is that they are underestimated owing to problems of the estimation method itself. To improve the landslide detection accuracy of the proposed method, these two possibilities should be discriminated in future studies though the comparison between the estimated height changes with truth data.
The aerial photographs and digital elevation model of Fundamental Geospatial Data utilized for orthorectification were provided by the Geospatial Information Authority of Japan (GSI.) The polygon data of landslides utilized as the truth data were provided by the National Research Institute for Earth Science and Disaster Resilience (NIED.) This research was carried out under the collaborative research agreement between the National Institute of Information and Communications Technology (NICT) and Nagasaki University. The authors would like to thank Professor Y. Yamaguchi of the Department of Information Engineering, Niigata University, Japan, for encouragement and valuable comments to promote this research.
Compliance with ethical standards
Conflict of interest
The authors declare that they have no conflict of interest.
- Arturi A, Del Frate F, Lategano E et al (2003) The 1998 Sarno (Italy) landslide from SAR interferometry. In: Proceedings of FRINGE 2003 workshop. Frascati, Italy, 1–5 December 2003Google Scholar
- Cao YG, Yan LJ, Zheng ZZ (2008) Extraction of information on geology hazard from multi-polarization sar images. Int Arch Photogramm Remote Sens Spat Inf Sci 37:1529–1532Google Scholar
- Christophe E, Chia AS, Yin T, Kwoh LK (2010) 2009 Earthquakes in sumatra: The use of L-band interferometry in a SAR-hostile environment. In: Proceedings of the IEEE IGARSS. Honolulu, HI, USA, 25-30 July 2010, pp 1202–1205Google Scholar
- Furuta R, Sawada K (2013) Case study of landslides recognition using dual/quad polarization data of ALOS/PALSAR. In: Proceedings of APSAR. Tsukuba, Japan, 23-27 September 2013, pp 481–484Google Scholar
- Ghiglia DC, Pritt MD (1998) Two-dimensional phase unwrapping: theory, algorithms, and software. Wiley, New YorkGoogle Scholar
- Kawamura M, Tsujino K, Tsujiko Y, Tanjung J (2011) Detection method of slope failures due to the 2009 sumatra earthquake by using TerraSAR-X images. In: Proceedings of the IEEE IGARSS. Vancouver, BC, Canada, 24–29 July 2011, pp 4292–4295Google Scholar
- Kobayashi T, Umehara T, Uemoto J et al (2012) Performance evaluation on cross-track interferometric SAR function of the airborne SAR system (PI-SAR2) of NICT. In: International Geoscience and Remote Sensing Symposium (IGARSS)Google Scholar
- Kobayashi T, Kojima S, Umehara T et al (2016) Pleliminary results of repeat pass interferometry of the X-band airborne SAR system (pi-sar2) of NICT. In: Proceedings of the IEEE IGARSS. Beijing, China, 10–15 July 2016Google Scholar
- Kojima S, Umehara T, Kobayashi T et al (2014) Preliminary analysis of cross-track interferometry by synthetic beam of two antennas of along-track interferometric SAR. In: Proceedings of the IEEE IGARSS. Quebec City, QC, Canada, 1–18 July 2014, pp 366–369Google Scholar
- Liu W, Yamazaki F (2015) Detection of landslides due to the 2013 Thypoon Wipha from high-resolution airborne SAR images. In: Proceedings of the IEEE IGARSS. Milan, Italy. 26-31 July 2015, pp 4244–4247Google Scholar
- Nadai A, Uratsuka S, Umehara T et al (2009) Development of X-band airborne polarimetric and interferometric SAR with sub-meter spatial resolution. In: Proceedings of the IEEE IGARSS. Cape Town, South Africa, 12–17 July 2009, pp II-913–II-916Google Scholar
- Oliver C, Shaun Q (2004) Understanding Synthetic Aperture Radar Images. SciTech Publishing, RaleighGoogle Scholar
- Tobita M, Fujiwara S, Murakami M et al (1999) Accurate offset estimation between two SLC images for SAR interferometry. J Geod Soc Jpn 45:297–314Google Scholar
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.