Landslide mapping using optical and radar data: a case study from Aminteo, Western Macedonia Greece

ABSTRACT Landslide mapping is one of the most important steps which take place in the early stages of a landslide survey, contributing to the immediate assessment of slide’s effects. Nowadays in this direction, different remote sensing techniques have been developed providing effective results. This work deals with the exploitation of optical data as well as radar data aiming at the mapping of a landslide area located in Western Macedonia, Greece. The landslide occurred in June 2017 at the lignite open pit mine of Aminteo. In that context Sentinel-1, Sentinel-2 and Landsat-8 data were acquired in order to be processed appropriately for the imprinting of disturbed areas. In particular, Sentinel-1 images before and after the landslide were processed interferometrically, yielding to the estimation of vertical displacements. Concerning optical data, the images were submitted to diverse digital processing techniques the results were visualized in GIS environment. The derived results were compared and evaluated for their accuracy and identification. It is worth mentioning that the specific methodology may also be applied to more extensive landslides or landslides triggered by different factors. Additionally, it could be assumed that the methodology can be applied to mass movements as a result of earthquakes.


Introduction
Landslide assessment, prediction and monitoring are the key points in the long-standing landslide research considering the fact that landslides are the most widespread geo-risk on a global scale.Thus, several studies deal with the development of methodologies for landslide vulnerability assessment.In particular, a successful approach for landslide vulnerability assessment was proposed taking into account statistical indicators of the landslide density and leading to a zonation of the sliding areas (Sharma, Patel, Ghose, & Debnath, 2011).On the other hand, an attempt was made to define landslide's vulnerability zones using only current parameters such as entropy (Sharma, Patel, Ghose, & Debnath, 2012).Additionally, a new hybrid framework was developed for vulnerability assessment at high-risk seismic-prone areas which utilizes Analytic Network Process (ANP) and Artificial Neural Network (ANN) models (Alizadeh, Ngah, Hashim, Pradhan, & Pour, 2018).
However, one of the most important steps of a landslide research is the landslide mapping which takes place in the early stages of the survey.In recent years the interest of the scientific community has focused on the exploitation of earth observation data from different remote sensing sensors for the mapping, characterization and monitoring of landslides.
The early attempts of using earth observation data for the monitoring of a landslide were based on digital image analysis processing and detection of relief changes through the comparison of DEMs resulted from repeated campaigns (Chandler, 2001;Maas & Kersten, 1997;Mantovani, Soeters, & Van Westen, 1996;Weber & Herrmann, 2000), as well as on the measurement of the horizontal displacement via photogrammetric techniques (Baum, Messerich, & Fleming, 1998;Kääb, 2000;Powers, Chiarle, & Savage, 1996).The fundamental principle for the landslide risk assessment remains the systematic implementation of aerial photo interpretations and fieldwork, despite the advances in remote sensing technology and the existence of satellite data with significantly high spatial resolution (Wieczorek, 1984).
Nowadays in this direction, Earth observation has been developed providing sufficient and effective results, however through the integration of different remote sensing data and newer methodologies of information acquisition (Nikolakopoulos et al., 2015;Nikolakopoulos, Vaiopoulos, Skianis, Sarantinos, & Tsitsikas, 2005).Specifically, Landsat-8 data and PALSAR-2 data along with GIS techniques were used successfully in order to identify and map landslide zones triggered by intense rainfall in a high vegetation region at Peninsular Malaysia (Hashim, Pour, & Misbari, 2017).In more detail, the approaches contribute to the investigation and analysis of landslide phenomena (detect and map landslides, analyse the triggering factors, describe the failure mechanism, model the susceptibility, examine the risk assessment) using data from different remote sensing platforms.The effectiveness of these techniques has already been examined in numerous studies (Casagli et al., 2017;Zhao & Lu, 2018), while an example including a synergy of multispectral data, radar data and data from UAV campaigns has been elaborated demonstrating the efficiency of the remote sensing data for the monitoring of an active landslide (Nikolakopoulos et al., 2017).In addition, multispectral as well as radar data were utilized for the inventory of landslides and their monitoring for the successful completion of the EU-funded FP7-SPACE project SAFER (Services and Applications For Emergency Response) (Casagli et al., 2016).
Although the landslide prediction remains a complicated and difficult procedure even through the usage of terrestrial monitoring techniques, it has already been proven that radar interferometric techniques could be used as complementary data source providing especially useful information for wide areas.In that context, several studies have already considered the use of interferometric techniques in landslide analysis and monitoring.In particular, the outcomes of two different differential interferometric (DInSAR) approaches were compared with in situ monitoring data demonstrating the usefulness of the specific approach for providing information about the stability of areas suffering from slow movements (Calò, Calcaterra, Iodice, Parise, & Ramondini, 2012).The most widely used interferometric approach in landslide monitoring is Persistent Scatterer Interferometry (PSI) and thus several studies dealt with the specific technique with conventional techniques such as GPS campaigns or in situ monitoring instrumentation in order to achieve higher accuracy of the final results (Crosetto et al., 2013;Fan et al., 2014;Righini, Pancioli, & Casagli, 2012;Wasowski & Bovenga, 2014).
This work deals with the exploitation of optical data as well as radar data aiming at the mapping of a landslide area located in Western Macedonia, Greece.The landslide occurred in June 2017 at the lignite open pit mine of Aminteo, due to the interaction of the geological structure with tectonics and human activities and resulted in the movement of more than 80 million cubic meters of material as well as in the evacuation of an entire settlement.In particular, Landsat-8 and Sentinel-2 optical data and also Sentinel-1 radar data were used in order to map the sliding areas after the landslide that occurred at the lignite mine of Aminteo in Western Macedonia, Greece.These data were selected due to the fact that both missions provide continuous, freely available data with short revisit period that make them suitable for Earth observation and thereafter for landslide mapping and monitoring.
The main objectives of this work are: a) to assess the suitability of Sentinel-1 for landslide mapping in an open pit mine, b) to compare the effectiveness of the freely available multispectral data (Sentinel-2 and Landsat-8) on landslide mapping and c) to evaluate the results and to prove whether optical or radar data are more effective in the context of the specific work.The manuscript is structured as follows: In the next section, the study area and the landslide are described, while section 3 contains the materials that were used.In section 4, the processing methodologies are analysed, and results are presented.Finally, in section the results are interpreted while in section 6 conclusions are reported.

Study area and landslide description
The area selected for the study comprises the lignite mine of Aminteo which is located in the region of Western Macedonia at the North-Northwestern part of Greece (Figures 1 and2).Figure 3 shows the geological setting of the specific area.The tectonic trench of Western Macedonia contains the largest lignite reserves of the country that produce 70% of the electricity of Greece.The wider area is geologically structured by Pliocene lignite, which appears mainly in layers, consisting of alternating layers of low thickness lignite with clays and marls.The specific lignite is soft, brown to black and it has been deposited in lowland vegetation.
The landslide event occurred on 10 June 2017 due to the combination of different factors such as: a) the presence of low-strength materials layers within the geological formations surrounding the lignite deposit, b) the presence of faults, c) the presence of underground aquifers under the mine floor and the development of water pressures and d) the geometry and the individual slopes of the mining front.The landslide crown is estimated about 3 km and the total landslide material contains about 25 million tons of lignite (Figure 4).The white line in Figure 3 represents the boundaries of the lignite open pit mine before the landslide, while the red line encompasses the landslide material.The occurrence of the landslide caused the appearance of numerous cracks and several damages to the houses located near to the crown of the slide, so the authorities decided the evacuation of the settlement of Ag.Anargyroi.In addition, damages presented on power and water supply networks as well as on the equipment (excavators) of the mine and on the mine itself (reduction of lignite quantities).

Materials
As already said, this work focuses on the exploitation of multispectral as well as radar data for landslide mapping.Thus, five Landsat-8 data (Table 1) and four Sentinel-2 data (Table 2) covering the study area before and after the occurrence of the landslide were obtained and processed.Regarding the radar data, 15 Sentinel-1 data (Table 3) in ascending and descending pass covering the wider area of the mine were acquired and processed interferometrically.The processing methodologies for both types of data are described in the following paragraphs.

Optical data analysis
The approach of landslide mapping using multispectral data was based on the application of digital processing  techniques such as Principal Component Analysis (PCA) or Independent component analysis (ICA) and simultaneously on the implementation of automatic change detection procedure.In particular, Principal Component Analysis (PCA) is a mathematical transformation technique of the original image to a new set of uncorrelated components derived in decreasing order of importance by applying a Gaussian distribution.On the other hand, Independent component analysis (ICA) constitutes a technique of high order feature extraction performing a linear transformation of the spectral bands of the original image.In addition, the automated change detection procedure based on the exploitation of DeltaCue Site monitoring tool, which identifies changes between two dates or times of imagery using Tasseled Cap algorithm at a specific site rather than over broad areas.Figures 5 and 6 present Landsat-8 and

Radar data analysis
Concerning landslide mapping utilizing radar data, the approach was based on the estimation of the line of ight displacement (LOS displacement) as well as of the vertical displacement through measurements of the interferometric phase.InSAR is able to measure a onedimensional motion along radar's viewing geometry that is called line of sight (LOS).LOS measurements contain the vertical motion and its horizontal contribution and thus they can often be different from the real values of motion.Nevertheless, the combination of ascending and descending data contributes to the estimation of the real vertical value and east-west components.In a similar study, an attempt was made to perform for determination of three-dimensional ground displacements in mining areas through the integration of multiple interferometric synthetic aperture radar (InSAR) methods (Wang et al., 2018).In that context, Sentinel-1 data were acquired and processed interferometrically, while during the procedure the interferometric baseline as well as the coherence of the interferometric pairs were taken into consideration (Table 4).

Optical results
As previously mentioned, landslide mapping using multispectral data was implemented using digital processing techniques such as Principal Component Analysis (PCA) or Independent component analysis (ICA).In that purpose, diverse layer stack of images before and after the event were created.In particular, stacked images were generated, containing the first five bands of Landsat-8 and the 1, 2, 3, 4 and 8 bands of Sentinel-2 mission.Each stacked image consisted of ten bands in total, five from the image before the slide and five from the respective image after the event.The statistical analysis of PCA technique for the stacked images resulting from Landsat-8 and Sentinel-2 data is presented in Table 5.As it can be noted most of the  information is concentrated at pca1, pca2 and pca3 bands.The results of PCA ICA technique are presented in Figure 7-10.In more detail, Figure 7 displays the mapped areas that emerged from PCA procedure using Landsat-8 data obtained on different dates.The mapped areas are imprinted with purple to greenish shades by combining different band combinations, while in both cases they seem to fit with the presence of the landslide material which is enclosed by the red line.Additionally, the blue line of the images represents the boundaries of the lignite open pit mine.Same procedure was followed utilizing Sentinel-2 data and in Figure 8 the results are presented.In particular, the mapped areas are highlighted with purple to green shades in the left stacked image and on the other hand with purple to orange shades.It is worth mentioning that the mapped areas resulting from Sentinel-2 data represent in a better way the area where the material slipped which is also reflected in the presence of more distinctive colors in comparison with those that appeared in the more stable areas of the mine.Additionally, the same stacked images were submitted to digital processing via ICA technique.The mapped areas that resulted from the specific procedure using Landsat-8 data are shown in Figure 9, whilst Figure 10 displays the corresponding mapped areas utilizing Sentinel-2 data.Specifically, in the left image of Figure 9 the slipped material is stained with orange to blue shades, while on the right image is presented with pink to greenish shades.Furthermore, if we compare the mapped areas of PCA and ICA procedures using Landsat-8 data, it is obvious that the results of the second technique (ICA) are better suited to the actual boundaries of the landslide material.Additionally, the mapped areas from Sentinel-2 data using ICA technique are imprinted with blue-green   colour on the left image of Figure 10 and pink to blue-green shades on the right image of the figure.
Simultaneously, the stacked images of Landsat-8 and Sentinel-2 data were exploited for landslide mapping through the implementation of an automatic change detection procedure, using the DeltaCue Site monitoring tool.Figures 11 and 12 display the results of the specific procedure.In particular, the left image of Figure 11 presents a Landsat-8 color combination image of the area before the landslide while the right   image corresponds to the area where the differentiation occurs (cyan and red shades).The specific area coincides with the boundaries of slide.Furthermore, Figure 12 constitutes the result of a same processing, however utilizing Sentinel-2 data.The left part of Figure 12 shows a Sentinel-2 color combination image of the area before the slide and on the other hand on the right part the area which was subjected to deformation stained in bluish and reddish shades.

Radar results
Concerning landslide mapping from Sentinel-1 data, it has already been reported that the approach was based on the estimation of the line of sight displacement (LOS displacement) and the vertical displacement.Figure 13 presents a generated interferogram from Sentinel-1 data, where a color alteration appeared within the boundaries of the slide that has already been observed in similar studies (Barra et al., 2016;Kyriou & Nikolakopoulos, 2018).
Then, the measurements of the phase were used for the estimation of the LOS and the vertical displacement.Figure 14 displays the outcome of the aforementioned procedure.The greatest value of LOS displacement was calculated about 4.5 m and it is located within the boundaries of the slide (reddish areas of Figure 13), while the highest estimated vertical displacement was approximately 1.73 cm.In general, LOS measurement is more sensitive to the ground motion because of the side-looking geometry of the SAR sensors (Hu et al., 2014), although in this case the values of LOS displacement are similar to the actual displacement values.Concerning the lack of accuracy in the values of the vertical displacement is likely to be  related with the orientation of the landslide.The remarkable point in the specific work is that the is able to map the deformation related to the appearance of numerous cracks which are located between the settlements of Valtonera and Anargiri (Figure 14) and have already been recognized in a corresponding study (Tzampoglou & Loupasakis, 2017).

Discussion
The scope of the specific work was the landslide mapping utilizing freely available multispectral as well as radar data.In that context, Landsat-8 and Sentinel-2 data covering the study area were obtained, and their processing took place through the use of digital processing techniques (PCA, ICA) and also automated change detection procedure in Erdas Imagine software.On the other hand, the landslide mapping using Sentinel-1 imagery was based on the estimation of the LOS and vertical displacement.
In more detail, regarding the optical data analysis, the mapped areas of PCA and ICA technique using Landsat-8 and Sentinel-2 data seem to fit with the boundaries of the landslide.However, the mapped areas resulting from Sentinel-2 data represent in a better way the landslide area which may be a consequence of the fact that the mission provides data with short revisiting period and therefore acquisitions are closer to the date of the landslide.The automatic change detection procedure, using the DeltaCue Site monitoring tool is a quite quick process however the results are not as accurate as the respective results of PCA and ICA techniques.So, the tool can be used as an initial step for landslide mapping, but it needs a further and more detail approach i.d.digital processing methodologies.
Concerning radar data analysis, the outcomes seem to have a low accuracy regarding the actual displacement nevertheless they agree with field data from a respective study that indicates the presence of cracks in the wider area.Therefore, a further radar analysis could be carried out at a future stage containing more  Sentinel-1 data for the improvement of the accuracy of the final outcomes.
Finally, in order to compare the results all the approaches, the mapped areas were digitized and calculated in ArcGIS environment (Table 6).The calculated mapped areas are quite similar with the exception of the mapped area of Sentinel-1 data, which is much larger due to the fact that the approach could map a wider deformation resulting from the landslide event as well as the presence of cracking.

Conclusions
This work has focused on the exploitation of freely available optical data and radar data for the mapping of a landslide area located in Western Macedonia, Greece.Sentinel-1, Sentinel-2 and Landsat-8 were acquired and processed appropriately in order to imprint the disturbed areas.Optical data were submitted to digital processing techniques (PCA, ICA) and automated change detection procedure, whilst radar data were processed leading to the estimation of the LOS and vertical displacement.In general, it could be said that Sentinel-1 and Sentinel-2 data are quite efficient for landslide mapping, while Sentinel-1 imagery seems promising enough for landslide monitoring.In addition, PCA technique gave better results with Sentinel-2 data compared to Landsat-8 data.It is worth mentioning that optical data can be supportive to radar data providing better accuracy and reliability on the final results.

Disclosure statement
No potential conflict of interest was reported by the authors.

Figure 2 .
Figure 2. Geology of the study area.Source: Geological map produced by the Institute of Geology & Mineral Exploration, Ptolemais sheet.

Figure 3 .
Figure 3. Google Earth image of the area before the landslide.

Figure 4 .
Figure 4. Google Earth image of the area after the landslide.

Figure 5 .
Figure 5. Landsat-8 images covering the study area on different dates.

Figure 6 .
Figure 6.Sentinel-2 images covering the study area on different dates.

Figure 11 .
Figure 11.Mapped areas resulted from DeltaCue Site monitoring tool using Landsat-8 data.

Figure 12 .
Figure 12.Mapped areas resulted from DeltaCue Site monitoring tool using Sentinel-2 data.

Figure 14 .
Figure 14.Left image: the line of sight displacement (LOS displacement) resulted from the interferometric process using Sentinel-1 data.Right image: The vertical displacement resulted from the interferometric process using Sentinel-1 data.
corresponding Sentinel-2 images (natural color band covering the study area on different dates.

Table 5 .
Statistical analysis of PCA technique.