Application of modified Segment Tracing Algorithm (mSTA) Method to Identify Landslide Susceptibility Zones Around Mt. Sinabung, Indonesia

Lineament could present the fractured zones and increase the landslides occurrence probabilities. This study aims to assess landslides susceptibility zones based on recorded landslide events corresponded to the detected lineaments around Mt. Sinabung, Indonesia. In this study, we use Synthetic Aperture Radar (SAR) data to increase the extraction accuracy of fracture-related lineament at Mt. Sinabung. The lineament densities were calculated to identify the landslide susceptibility zones including frequency, length, and intersection using modified Segment Tracing Algorithm (mSTA) method for area about 10.9 × 9 km. Based on visual observations and technical report, a total of 61 landslides were occurred around Mt. Sinabung from 2016, 2018, and 2021. An Ordinary Kriging density map of the lineament frequency using 800 × 800 m cells showed agreement between the lineament density and landslide distribution. Comparing landslide location and lineament density, landslide occurrences were located mainly at high and moderate lineament densities. We have classified the landslide density for three years into low, moderate, and high based on histogram quartile analyses. This classification could be used to mitigate the landslide hazard in a large area.


Introduction
Indonesia has a humid tropical climate with an average annual rainfall of 2500 mm and average temperatures of 22 -28ºC.The high intensity tropical rainfall coupled with hilly areas and a deeply weathered rock profile comprise near-ideal conditions for landslide development [1,2].Landslide is one of the most common geomorphological hazards caused by the acceleration of slopes dynamics in the river basins leading to long-term irreversible changes affecting the natural environment, human life properties, and infrastructure facilities [3].The severity of the landslide impact is related to the number of exposed elements and their linked vulnerabilities, the intensity of the landslide event [4,5].Among Indonesia's physiographic divisions, the volcano region, which consist of mountain belts is particularly vulnerable to landslide due to its high seismicity, young and fragile geology, geomorphology, and heavy precipitation.In this regard, landslide susceptibility zonation to categorizing the land surface into Landslides are reported in many cases to be controlled mostly by structural trends such as discontinuities [6,7].These discontinuities can be identified and measured by satellite images; therefore, lineament extraction and analysis are valuable tool to identified discontinuities in landslides studies.A discontinuity plane can be defined as separation or fracture in geological formation which divides the rock into two or more pieces.This structural weakness originated mostly around physical discontinuities such as faults, joint sets, or dykes are important elements to controlled landslide event [8,9].However, lineament extraction from optical sensor images as hindered by low contrast due to the sun illumination effect and high humidity, cloud cover, and demonstrate poor performance in the detection of outcrop or geological lineaments in areas of vegetation [10].To avoid such non fractures related lineaments, this study uses images of the radar Sentinel 1. SAR imagery can be acquired under cloud cover with all weather conditions and can identify small topographic relief with relatively high spatial resolution by using oblique microwave radiation [11].Mt.Sinabung is an andesitic stratovolcano elevated about 2460 m above sea level [12].This volcano is located in North Sumatra, 40 km northwest of lake Toba formed by catastrophic caldera eruptions [13].Volcanic and tectonic activities around Mt. Sinabung cause potential secondary hazards such as landslides and mass movements around the edifice.Volcanic landslide is a catastrophic feature and represents an enormous danger to the neighbouring population [14].Landslides following the eruption affect the stability of the slopes around Mt. Sinabung and produce hazards in certain areas [14].The prime purpose of this study is to develop a method that automatically extracts lineaments from SAR images and identifies landslide susceptibilities zone based on lineament distribution at Mt. Sinabung, Indonesia.

Data Collection
We have used Sentinel-1 data, and landslide inventory from the Center for Volcanology and Geological Hazard Mitigation (CVGHM) and visual images from Google Earth.The Sentinel-1 is equipped with twin polar-orbiting satellites operating up to 25-min per orbit with a repeat cycle of 12 days.The satellites are to operate day and night and perform a synthetic aperture with radar imaging.Sentinel-1 bands allow us to get imagery in all weather conditions.Sentinel-1 was composed of three data products, and we use level-1 SLC product focused on the slant range by azimuth imaging plane, single looks, which contain phase and amplitude information.Sentinel-1 has azimuth and range.Azimuth is the satellite track or so-called orbit.Both orbits have opposite ranges.It can provide images of the opposite line of sight (LOS) utilizing two orbit modes: the ascending mode in which the satellite moves from south to north direction, and the descending is when the satellite moves from north to south direction.We combined a pair of Sentinel images of the ascending and descending orbit modes to reduces the effect of geometrical distortion.Then, the combination can equalize the contrast of both flanks, because the back slope at one mode changes to the fore slope at the other mode [15].We used six scenes level-1 dual-polarization (VV and VH) C-band Sentinel-1 Single Look Complex (SLC) for each ascending and descending orbit.The detail of SAR data used in this study was listed in Table 1. for this research, only the 2016, 2018 and 2021 landslides were used for the analysis and these landslides were also digitized in a point format (see Fig 1).The 61 landslide locations have been identified well visually using Google Earth engine and soil movements inspection reports from CVGHM.The identified landslide locations are not primary hazards from volcanic products such as debris avalanche or pyroclastic flows but originated to the secondary hazards following the eruption.A volcano-triggered landslide is a landslide occurrence following the eruption period passing a steep slope.

Lineament extraction
Linear features represent the geological structure with a different hue contrast and relief in the images.analysis using satellites for lineament detection is currently an effective option, but cannot be performed for large areas [16].Lineament detection is automatically referred to extraction which is carried out based on the examination of digital numbers from satellite images with certain criteria.An effective method of automatic lineament detection and extraction is mSTA (modified Segment Tracing Algorithm).The mSTA works based on the STA principle [17].The STA principle detects lines from pixels that are read as vector elements by examining local variations in the gray level of digital numbers [18].The program used for the linear features extraction is developed using MATLAB syntax.
In this study, the automatic lineaments extraction was carried out on the basis of two processing: first, preparation images of Sentinel-1 were performed through three main steps, multilooking, speckle filtering, and geometric correction.Multilooking was increasing focus without losing resolution, speckle filtering was minimizing the noise from surrounding objects, and the last step was the correction of the geometrical distortions attributed to the topographic variation in the scene as well as the inclination of the radar.Consequently, geometric corrections task is especially necessary so that the image will be represented geometrically as close as possible to the reality, and subsequently be ready for use.These pre-processing is indispensable to mSTA application.Next, a line element is identified by examining the local variation along 16 directions at 11.25 o intervals within a small window size of 11 × 11 pixels, following the principle of STA.In the preceding STA applications [17], the direction that minimizes the variation is assumed the valley direction and expressed by a symbol kmin.kmin is also defined to minimize εj that represents the difference between the values at the central pixel in the window.Since the image may produce more than one kmin due to remaining speckle noises [18] average the three neighbouring εj values in the window as: Where   is the average value of three εj to obtain kmin value.The direction J that minimizes   was assigned to kmin.Next, we transform the segments into a lineament.A search ellipsoid window is set along a target segment whose start and middle points are inside the window with angle differences from the segment less than 11.25 o are selected.Segments that only have endpoints inside the window are excluded.The window size was determined by considering the agreement of the lineaments and major faults length in the study area.Then, one line is defined as lineament using the least squares method, which approximates the coordinates of the start and end points of the selected segments.For the final step, the lineaments from the ascending and descending mode images were merged and the overlaps were grouped into one line using the least square method.
Therefore, to quantify the spatial concentration of the extracted lineaments, the study area was covered by a grid cell size 800×800-m and the three indices related to the density: the lineaments intersection (Li) that's the number of the intersected lineaments which may extend the estimated damage zone by the multiple effect of fracturing, lineament frequency (Lf) that's the number of extracted lineaments, and lineament length (Ll) that's the total length of lineaments in the cell per unit area were calculated.

Landslide distribution and identification
Landslide occurrences were collected visually from the Google Earth engine explained as follows.The locations of evidence were identified based on image color, tone, surface roughness, and texture.These image parameters depend on the light reflected by the surface which useful to distinguish rock type, soil, and vegetation.Landslides are easily recognizable immediately after a landslide event.Their boundaries are usually recognizable easily in the image that the new occurrence of the landslides is less vegetation covers or opened area.As time passes, vegetation starts to cover the landslides and humans might work the land thus destroying the boundary and the landslides might be reactivated and move as time progresses.In addition, we also used the soil movement inspection reports from CVGHM for 2016 -2021 to ensure that landslide locations are correctly digitized.The areas of these landslides were also digitized in a point format.Therefore, the landslide density map was calculated based on the number of landslides in every 800 m × 800 m grid and using the fishnet tools to create a grid set that was later overlaid with a lineament map.Until now, there is no standard landslide density classification to indicate an area's high or low landslide density.For that reason, this study uses three density classes are low (for one landslide/640000 m 2 ), moderate (2 -3 landslides/640000 m 2 ), and high (≥ four landslides/640000 m 2 ).
Next, the landslide susceptibility map is produced by combining both lineament and landslide density maps.The overlapping areas indicated as high lineament density and high landslide density could be considered highly susceptible areas for the landslide.This susceptibility map is validated by calculating the numbers and percentage of landslides in the high susceptibility zone.The level of susceptibility indicated by the Australian Geomechanics Society [19] is used to indicate the level of susceptibility based on the proportion of landslides in the high susceptibility class with the total landslides in the study area.The hazard level was also calculated for the study area by calculating the number of landslides in the high susceptibility zone for the assessment year.The levels of hazard and susceptibility based on the guidelines are given in Table 1.Susceptibility is the proportion of the total landslide population in the study, and area hazard is the number of landslides/year/800m 2 .Modified Segment Tracing Algorithm (mSTA) was used to extract the linear feature, which was interpreted as fault or fractured based on Sentinel-1A data with Ascending and Descending modes.The improvement of mSTA was achieved by excluding short lineaments of mSTA, which originated from speckle noises in the black slope of LOS with low backscattering intensities [18].The mSTA data is gained by combining two orbit mode lineaments, grouping them, and eliminating short lineaments.In the rossete diagram, ratios of the counts of individual linear feature and the total amount of all linear features for ascending and descending images along each direction was calculated.There were 1152 on ascending and 1186 on descending detected lineaments in total using mSTA method.In the ascending and descending image, since 2016 and 2018, more variated in all directions but the most dominant is NW-SE.However, the linear features direction at 2021 NW-SE direction is the most dominant direction than the others corresponding to geologi structure map of Sinabung [20] (See Fig 3).The number of extracted linear features in descending images was more significant than in descending images, where counts of individual linear features to the total amount of all linear features for ascending and descending images along each direction were also calculated.Therefore, to map the three indices of lineament density; Li, Lf, and Ll, their semivariograms were calculated omnidirectionally and all the data were well-fitted to the spherical or exponential model.Furthermore, the density of intersections between lineaments has to 0 to 7 km/km 2 .While the lineament length density is spread from 0 to 8 counts/km 2 .The number of lineament with density has values from 0 to 8 counts/km 2 .The lineament results could be seen in Fig 4.   The three ordinary kriging maps show large and small index values by red-toned and blue-toned colours, respectively.The good match between the major faults and the lineament densities for the three indices verified the ability of mSTA to detect faults and the effectiveness of the ordinary kriging mapping.from the three indices of lineament density, Li, Lf, and Ll, we found Lf has a good indicator that caused a good match between the distribution of landslides with high-density values compared to the others.Comparison between landslide occurences and the lineament density map showed that most landslide occured in high and moderate lineament density area.Thus, from the result, we conclude that Lf is a good indicator of the degree of near-surface fracturing (See Fig 4).
A lineament map that was produced from the interpretation using mSTA is used to observe the association between landslide occurrences (landslide density) with the lineament density in the study area.The corrected lineament map that was transformed into a lineament density map is shown in Fig 3 .Distribution of landslides have been known to located in highly fractured zones with high permeability over the shallow to deep depth range.To quantify the spatial concentration of the extracted lineaments, the conversion on the lineament map from vector to density in a grid format is based on the total length of lineament in each of the 800 m 2 .Subsequently, the lineament density map was classified into three density classes, namely: low, moderate, and high.

High
The degree of susceptibility and hazard can be quantified and classified based on the method and proposed classes of susceptibility and hazard by AGS (Table 1).To proceed with the quantification process, the total landslide from the three assessment years were intersected with 8 zones of high susceptibility classes shown in Fig 4 .From the total 61 landslides acquired from the three assessment years, 36 landslides are within the 8 high susceptibility zones, which indicates that if the degree of susceptibility and hazard are calculated based on the guidelines by AGS [19] the area at Mt. Sinabung can be classified as highly susceptible (0.59) but moderate in terms of hazard (0.9).In terms of landslides that occurred in every 800 m 2 , which is also can be indicated as high based on the classification of landslide density used in this study (Fig 4).Table 3 shows the degree of susceptibility, hazard, and landslide density (per km 2 ) that are based on the 36 landslides in the 8 high susceptibility zones at the Mt.Sinabung.

Discussion
In landslide susceptibility mapping, the selection of landslide factors is one of the most important elements.However, there is no well-defined standard to select the most significant landslides factors.The factors that initiate the landslide incidence in the study area selected based on data availability, literature review and field evaluation.On this study, we use manual land use classification of the Google Earth imagery and found that be effective as it has a high spatial resolution to identify landslide location.Methodologies for producing landslide susceptibility maps are mainly based on GIS technology.Previous published works has been explained to solve the deficiencies and difficulties in landslide susceptibility assessment [19,[21][22][23].Practically, the procedure for preparing landslide susceptibility maps should be applicable and accurate to the field scale.To validate the high susceptibility zones generated for the study area, 15 landslides in 2022 that occurred close to the area Mt.Sinabung were used to validate the map (Fig 6).However, due to a lack of data from google earth and a report from the Centre for Volcanology and Geological Hazard Mitigation (CVGHM), the landslide interpreted from the 2022 dataset only covered the area and only 15 landslides were recorded based on the availability of the satellite photographs.Nevertheless, the landslides from 2022 are adequate to validate the landslide susceptibility map (see Fig 6).As a result, from the intersection between the high susceptibility zone and the 2022 landslides, 15 out of 21 landslides are located within the high susceptibility zones.This means that the susceptibility map accurately captures almost most of the landslides in 2022 at Mt. Sinabung.This also demonstrates that the combined high lineament and landslide densities maps can be used to indicate locations that are highly susceptible to landslides in the future.
Regarding the landslide occurrences in 2022 and landslide zone susceptibility using lineament density results, we obtained 15 landslides based on visual observations using the Google Earth engine match with the predicted landslide zone.Generally, the landslide occurred in several zones (see Fig 6) as shown in L14 located northeast of Mt.Sinabung.Based on the geological map of Mt.Sinabung, this type of landslide is in pyroclastic flow deposits [24].The landslides are reported to be mostly controlled by structural trends and made low shear strength effects of infill materials and decrease the slope strength thus causing landslides to occur.The rock types played a significant role in the rainfall triggered landslides, as they underwent intensive weathering to form thick soil profiles observed at landslide ridges and flanks/headwalls.The geological conditions, including physical and weathering characteristics can control the triggering mechanism, type, and pattern of landslides.Bartelliti et al. [25] reported that bedrock lithology seems to be an important factor that influences the localization of shallow landslides.The effect of geological conditions on landslides may be related to the mechanical properties of the soil present in the slope and originating from parent rock.The physical and mechanical properties of soil vary depending on geological conditions because they are the results of the physical and chemical weathering of bedrock.Soil properties are also affected by the composition of minerals present, which depend on the geological conditions of the site [26,27].The content and geochemical characteristics of the minerals constituting the soils are important factors in landslide susceptibility and size [27].Based on the geological map, the bedrock in Mt.Sinabung in age and consist of layers of lava andesite-dacite, pyroclastic flow deposits, and alluvium [24].After weathering, andesite also has characteristics of clay on the surface.The bedrocks of the landslides and their constituent mineral types are prone to rapid and deep chemical weathering by hydrolysis and produce weakened clay-rich fractured rocks and soils.Clay, a weathering product of rock mass, contributes to landslide occurrence because of their chemical and physical properties.The presence of clay minerals in the altered andesitic can be considered one of the factors that induced Landslides at Mt. Sinabung.Moreover, a high landslide also occurred west of Mt.Sinabung located around the volcanic crater and that occurrence is triggered by the eruption of Mt.Sinabung (see Fig 6B).Based on the geological map of Mt.Sinabung, this type of landslide developed in a thick layer of lava andesit-dasit where erosion is experienced.This condition makes landslides on volcanic cones common due to the weakening of the volcanic rock hydrothermal caused alterations as well as eruptions.The intercalation of lava unconsolidated pyroclastic rocks can also lead to the weakening of slope stability, which triggers landslides.We also found the landslide located near the river Lou Borus, Mt.Sinabung (See Fig 6C).Based on the geological maps of Mt.Sinabung, this type of landslide is in alluvium deposited [24].The area is predicted to be a zone with high landslide intensity and close to residential areas that were about 92-m away from the landslide location.This is of greater concern to the local community and government and makes this information useful for preventive or mitigating actions for the surrounding community.

Conclusion
In this study, a simple but useful technique using the landslide density method was used to carried out a temporal landslide assessment at the Mt.Sinabung.We used the mSTA method to extract linear features derived from ascending and descending images producing 1152 and 1186 lineaments, respectively.Lineaments related to geological structures were presented by continuous lineaments associated with morphology.The detected lineaments using the mSTA method are more effective to identify geological structures than visual method especially for large coverage area.The agreement between statistics of lineament direction with confirmed geological structure was used as verification approach.The major lineaments directions agreed to the fault directions of the NW-SE corresponding to geologi structure map of Sinabung.The densities of the extracted lineaments were mapped by ordinary kriging.The high density of lineaments frequency (Lf) is concordance to the 48 out from 61 landslide location.We have also obtained that the 36 landslides located at the high and moderate potential hazard at Mt. Sinabung.
The landslide occurrences in this study were associated mainly with high lineament density potential to the instability.Integrating the landslide and lineament densities, we calculated the accuracy of the produced susceptibility map is about 75%.This map could be used to mitigate landslides potential and hazard mitigation purposes.

Figure. 1 .
Figure.1.Study area located at Mt. Sinabung in the middle of North Sumatra (A), showing footprints of Sentinel-1 in ascending and ascending orbits (B), and shaded topographic map using DEMNAS 9 m data showing the eruption centre, the distribution of faults, and landslide locations (C).

4. Result and Discussion 4 . 1
Landslide and lineament densityThe interpretation of visual observations using the Google earth engine resulted in 11 landslides in 2016, 22 landslides in 2018, and 28 landslides in 2021.The total landslides recorded for these 3 years are 61 landslides that spanned along the 119 km Mt.Sinabung area.The distribution of landslides based on each assessment years are presented in Fig 4.The landslide was classified to examine their density in every 800 m 2 of the study area.

Figure 2 .
Figure 2. The occurrences map of landslide distribution at study area in 2016, 2018, and 2021 (A) and landslide density map calculated from landslide occurrences (B).The classification was divided into three classes representing low landslide occurrence, moderate occurrences, and high occurrences.Although the study area is dominantly characterized by low landslide occurrences all three assessment years, there are several of the area that were classified as high landslide density area.The indicated high density sections of the area shown with red pixel at Fig 2 in three different locations, on around the crater of Mt.Sinabung, at the west of Mt.Sinabung, and in the southwest aligned to the Lou Borus River close to community settlements.Although there were just few high landslide density zones in each of the assessment year, it can be observed that at least 1 landslide occurred in every 2 -3 km of the area Mt.Sinabung.To examine the landslide density as a total density for the three assessment years, landslides recorded in those three years were combined and analysed together to produce the final landslide density map.

Figure 3 .
Figure 3.The final lineament map produced by grouping the lineaments from the two observation modes in 2016 (A), 2018 (B), and 2021 (C).The rose diagrams (lower panels) show the frequency of lineament directions.

Figure 4 .
Figure 4. Ordinary kriging maps of the lineament density indices for length Ll (A), frequency Lf (B), and intersection Li (C) in 2021 indicate the spatial concentration of lineaments overlaid by landslide distribution (white dots) and faults (white dashed lines).

Figure 5 .
Figure 5. High landslide density maps around Mt. Sinabung (A), overlaid with high lineament density (B), and overlap of high landslide and lineament densities zones (C) with the black dashed lines are faults at Mt. Sinabung [20].

Figure 6 .
Figure 6.The landslide susceptibility map presents the distributions of landslide occurrences in 2022 overlaid by faults in the black dashed lines (A) and verification of the high susceptibility zone at Mt. Sinabung presented visually by Google Earth Engine (B, C, D).

Table 1 .
Details of Sentinel 1 SAR data used in this study.

Table 2 .
[19]descriptor for the susceptibility and hazard degrees used in this study[19] The Second International Seminar on Earth Sciences and Technology (ISEST-2023)

Table 3 .
Description on the degree of susceptibility, hazard, and landslide density of the study area