A Remote Sensing – based Intensity-Duration Curve , Faifa Mountains , Saudi Arabia

Construction of intensity-duration (ID) curves and early warning systems for landslides (EWSL) are hampered by the paucity of temporal and spatial archival data. We developed methodologies that could be used for the construction of an 10 ID curve that could be used for the construction of an EWSL over the Faifa Mountains in the Red Sea Hills. The developed methodologies relies on temporal, readily available, archival Google Earth and Sentinel-1 imagery, precipitation measurements, and limited field data. These methodologies accurately distinguished landslide-producing storms from non– landslide producing ones and identified the locations of these landslides with an accuracy of 60%.


Introduction
Mountainous areas worldwide with steep slopes, high precipitation, and limited vegetative cover often experience landslides.
Two main types of landslides are often reported from Faifa Mountains in Saudi Arabia.The first are debris flows that occur when water-saturated soils (largely from weathered bedrock and fragmented rock) move down mountainsides, get channelled into streams, pick up objects along their paths, and deposit their thick load down the valley slopes (Iverson, 1997).The second results from failure along preexisting fracture planes (Lowell, 1990) that occurs when the following conditions are met (Norrish and Wyllie, 1996): (1) the strike of the planer discontinuity is similar (within 20°) to the strike of the slope face, (2) the dip of the planer discontinuity is less than that of the slope face and oriented in the same general direction, (3) the dip of the planer discontinuity is greater than the angle of the friction of the surface, and (4) the friction angle of the rock material is partially controlled by the size and shape of the grains exposed on the fracture surface and by the mass of the block above the planer discontinuity (Alharbi et al., 2014).These two types of landslides pose a substantial threat to human life and property in mountainous areas, especially in populated regions that are witnessing unplanned urbanization.
Unfortunately, monitoring programs for landslide occurrences (date and time) and conditions (intensity and duration of precipitation) are very limited in Faifa.If such programs existed, they would have generated the archival landslide data needed for the construction of intensity-duration (ID) relationships and for the development of an early warning system for landslides (EWSL).The paucity of such data in Faifa is largely caused by its rugged nature and its limited road network coverage; both factors rendered many areas inaccessible and hindered the development of monitoring systems.Despite the absence of organized monitoring programs for landslides in the study area, a few were recorded by the Saudi Geological Survey (SGS) in the past few years.Our field observations and those collected by the SGS revealed that debris flows are by far the most prominent (>85%) landslide type in the study area.
To compensate for the deficiencies in field data in Faifa, we complemented the available field data with observations extracted from readily available temporal remote sensing data.These include high-resolution Google Earth images (spatial resolution: 15 m to 15 cm), Sentinel-1A Ground Range Detected (GRD Level-1 radar data; spatial resolution: 10 m), satellite-based precipitation data including Tropical Rainfall Measuring Mission (TRMM, 3-hourly_3B42 v7, spatial resolution: 0.25°x 0.25°) and Global Precipitation Measurement (GPM) IMERG Final Precipitation L3 Half Hourly (spatial resolution: 0.1° x 0.1°0 V05), and a high-resolution digital elevation model (TanDEM-X DEM; spatial resolution: 12 m).Field observations were collected (period: 2014 to 2016) from the areas that witnessed landslides following precipitation events.In this manuscript, we develop procedures that take advantage of readily available remotely acquired data, and limited field data, to develop ID curves for the study area, procedures that could potentially be applied to similar areas with limited field data.The proposed procedures for the development of ID curves should not be considered as alternatives to the well-developed field-based ID relationships and to recently introduced advances in such applications, but could be used in absence of such field-based datasets.These advances include the development of ID relationships that account for the role of antecedent rainfall preceding the landslide-causing precipitation (e.g., Kim et al., 2014, Hong et al., 2017).

Geologic, topographic, and climatic setting of the Faifa Mountains
The study area (119 km 2 ) lies within the Red Sea Hills, and covers an area (17.20°N to 17.29° N and from 43.05° E to 43.16° E) proximal to the Saudi-Yemeni border (Fig. 1).The elevation is high (range: 259 to 1817 meters above mean sea level [m.a.m.s.l]) compared to the surrounding lowlands, the topography is steep (slope as high as 67º; Figs.3a and 3b), and vegetation is extensive over the mountains but sparse in the surrounding lowlands, as shown in the normalized difference vegetation index (NDVI) map (Fig. 3c).The Faifa region is located within the north to northeast trending Tayyah tectonic belt that consists of a complex of metamorphosed volcanic and pyroclastic rocks of basaltic, andesitic, and clastic metasedimentary rocks (Greenwood et al., 1983) that were generated and accreted in an island arc setting some 800 to 900 Ma (Stoeser and Camp, 1985).The Faifa Mountains are predominantly composed of intergranular syenite, biotite schist, and sericite schist, and are dissected by north-south, northwest-southeast, and east-west trending fault systems (Fairer, 1985;Alharbi et al., 2014;Fig. 3d).mm/year; 1998-2009: 227 mm/year) with the wettest year on record in 2016 (total annual rainfall: 450 mm).Two systems of wind regimes are responsible for the rainfall over Faifa: (1) northerly and northwesterly winter cyclonic regimes from the Mediterranean, and (2) summer monsoons from the Arabian Gulf and the Indian Ocean (Alsharhan et al., 2001).

Methods
The methodology we developed entailed two main steps.The ID curve for the Faifa Mountains was first generated to identify 5 storms that caused landslides (temporal analysis; Sect.2.1), and then pixel-based ID curves were constructed to identify the locations where movement is likely to occur (spatial analysis, Sect.2.2-2.4).The latter step involved: (1) selection, calibration, and pre-processing of radar images; (2) generation of backscatter coefficient difference images as a measure of surface roughness change due to precipitation-induced landslides; and (3) development, refinement, and validation of the model to identify pixels susceptible to movement under user-defined precipitation conditions.10 Data used for the study include: (1) Temporal Google Earth imagery, (2) Sentinel-1A radar imagery; (3) TRMM and GPM; (4) field observation of landslide (debris flow and failure along fracture plane) locations; and (5) the TanDEM-X DEM.A detailed flow chart is presented in Fig. 4.

Generation of the ID curve for the Faifa Mountains
Unfortunately, the distribution of rain gauges is inadequate in the study area (Fig. 1).There is only one station within Jazan province, and three more stations in its surroundings (Fig. 1).We utilized the GPM half-hourly (spatial resolution: 0.1° x 0.1°) and TRMM 3-hourly (spatial resolution: 0.25°x 0.25°) data to extract the intensity and duration of rainfall that caused landslides throughout the period 2007 to 2016 (Table 1 and Fig. 5).Rainfall events of less than 1 mm/h were omitted given that no landslides were reported from the area at these low rainfall rates and TRMM could mistakenly identify fog for low rainfall events (<1 mm/h; Milewski et al., 2009).Because of these reasons, the threshold of 1 mm/h was established for the precipitation data to develop the ID curve.Altogether, 131 precipitation events were extracted from TRMM and GPM data throughout this period, of which 19 events were identified as landslide-producing storms (Table 1).These storms were identified using spectral and morphologic variations associated with landslide development, variations detected in the field and/or extracted visually from pairs of Google Earth images bracketing large precipitation events.Google Earth images were favored over other readily available visible near-infrared (VNIR) satellite data sets given their high spatial resolution (15 m to 15 cm) and long temporal coverage for the study area (2007 to present).An area that witnessed landslides will be covered by spectrally dark vegetation on the Google Earth image preceding the landslide and by spectrally bright rocks and sediments on the image acquired after the landslide development.In many cases the latter image, not the former image, shows a major scar During a number of these storms, landslides were reported from the mountainous areas proximal to, but outside of, the study area (e.g., Youssef et al., 2014).Using detected storm-induced spectral and morphologic variations in pairs of archival temporal Google Earth imagery, two additional landslide-producing events (November 18 and December 1, 2014) were extracted; these were apparently not reported to, or verified by, the SGS researchers during their field campaign due to their location in inaccessible areas.Using the same techniques (storm-induced spectral and morphologic variations) 16 storms were detected in the period (2007 through 2014) preceding the SGS field campaign (2014)(2015)(2016).Given the paucity of Google Earth images (18 images in 10 years), a 5 number of precipitation events are likely to have occurred between consecutive Google Earth images.If landslides were detected within the period covered by the consecutive Google Earth image acquisitions, it was assumed that the largest of these storms caused the observed landslides.In doing so, a few landslide-producing storms and those that did not cause landslides were not identified.The latter type of storms were identified during the field campaign period (Table 1 The data presented in Table 1 were plotted to extract the ID curve for the Faifa Mountains.Landslide-producing storms (seven storms) were represented in Fig. 5 by solid circles (red and black), and the non-landslide producing storms by open circles.
The solid black circles are for field-verified landslide-producing storms, and the red circles are for landslide-producing events extracted from Google Earth images.The figure shows the extracted ID curve (red line) that provides the best visual separation 5 between the landslide-producing (solid circles above red line) and non-producing (open circles below red line) precipitation events.For comparison purposes, additional ID curves are reported from other areas; each is characterized by its unique hydrologic, geologic, climatic, and topographic settings and plotted in Fig. 5.These include Mediterranean cool climates, the shallow landslide and debris flows, wildfire-related debris flow, shallow landslides, soil slips, debris flows, and warm mountain climates, (Caine, 1980;Innes, 1983;Clarizia et al., 1996;Crosta and Frattini, 2001;Cannon and Gartner, 2005;Guzzetti et al., 10 2008).The closest ID curve to the Faifa Mountains curve is that of the shallow landslide.This is to be expected given that the shallow landslide curve was generated for areas with similar climatic (precipitation: Faifa 252 mm/year; shallow landslide: >200 mm/year) and topographic (slope: Faifa: 20° to 65°; shallow landslide: 25° to 40°) settings.

Selection, calibration, and preprocessing of radar images
The ID curve for Faifa separated the storms that caused landslides from those did not cause any landslides.The radar backscatter differences were used to determine the location of the landslide for the storm that caused the landslides.Sentinel-1A radar scenes were downloaded for ascending and descending acquisition modes from the Sentinel Data Hub (https://scihub.copernicus.eu/dhus/#/home),a download platform for the European Space Agency (ESA), for the period between October 2014 and October 2016.The scene acquired immediately after (1 day or less) the rainfall were not used in the generation of backscatter coefficient difference images to avoid differences in backscatter due to precipitation-related change in moisture content.The pre-processing steps that were applied to the downloaded scenes included radiometric calibration and calculation of beta naught (b 0 , the radar brightness coefficient), speckle filtering, terrain flattening and correction, and image co-registration.Ascending and descending scenes for the same area provide different degrees of visibility, and depending on the orientation and complexity of the topography, one acquisition mode may provide better visibility than the other.Ground range detected (GRD) images, acquired from the Single Look Complex (SLC) were downloaded and radiometrically calibrated using ESA's Sentinel Toolbox.The GRD images were used to calculate the b 0 (Small, 2011), a measure of radar backscatter energy in decibels (dB; Raney et al., 1994), for both ascending and descending modes.The speckle in the extracted radar backscatter coefficient images was minimized using the Lee Filter (window size: 3 × 3; Lee, 1983;Lee et al., 2009) and high-resolution DEM (TanDEM-X DEM; resolution: 12.5 m).The Terrain Flattening and the Range Doppler Terrain Correction (Small, 2011) was applied to the speckle filtered scenes to correct for radiometric biases introduced by the rugged topography of the study area.Each of the processed scenes was co-registered (sub-pixel co-registration) to the previously acquired one in Sentinel Toolbox (Press et al., 1992).

Generation of backscatter difference images
The generation of the backscatter difference images involves a number of steps: (1) calculation of radar visibility, (2) removal of corner reflectors, and (3) generation of backscatter difference images.
Ascending and descending scenes for the same area provide different degrees of visibility depending on the orientation and complexity of the topography given the variations in slope, aspect, radar incidence angle, layover, and shadow (Notti et al., 2014).The radar visibility index (R; Notti et al., 2014) was used to identify and mask out areas of low visibility in both the ascending and descending scenes.The index values range from 0 (low visibility) to 1 (high visibility).Pixels with R values below a threshold of 0.6 were found to be affected by overlays and shadowing and were thus identified as low visibility areas and masked out.
The distribution of pixels with backscatter coefficients exceeding 0.5 dB were found to correlate with that of buildings, construction areas, vehicles, and parking spaces.Such features can act as corner reflectors and produce high radar returns by reflecting waves towards the source.Pixels with backscatter coefficients exceeding 0.5 dB (corner reflectors) were masked out.
The filtered backscatter images were used to generate backscatter difference images between pairs of consecutive backscatter scenes, and only those pairs bracketing storm events were considered for further analysis.

Refinement and validation of the model
The refinement and validation of the model involved: (1) spatial refinement and standard deviation (SD) image generation, (2) field verification, and (3) batch processing of scenes.The population density in and around the road networks is high, and so are the risks for human and property losses if landslides occur in their vicinity.The construction of roads can trigger debris flows, especially in cases when roads intersect steep slopes (Fig. 6a) or terraces constructed on these steep slopes (Fig. 6b), ephemeral valleys, and fracture planes dipping towards the road (Fig. 6c; Alharbi et al., 2014).slopes, along first order streams, above and proximal to the main roads as shown in Fig. 5. Using these three criteria, areas susceptible to debris flows were identified by: (1) extracting stream networks using a stream delineation algorithm (Tarboton et al., 1991) in ArcGIS 10.5 over the steep slopes (>30º) and capturing first order streams, using a small flow accumulation value (10 pixels), ( 2) assigning a buffer zone (width: 20 m) around the extracted streams to delineate the areas that are likely to be triggered by runoff during and following rainfall events, and (3) assigning a buffer zone (width: 100 m) around the roads.
The use of the latter criterion allows the identification of areas susceptible to failure along preexisting fractures as well since our field observations showed that the majority of such failures were triggered by road construction.The selected width of the buffer zones was determined by examining the proximity of the historical landslides to roads and extracted streams.
The selection of the buffered zones for further investigation served two purposes: (1) targeting areas of high risk, and (2) capturing the backscatter variations that are related to landslides, variations that could have been confused with those caused by factors other than landslides (e.g., change in vegetation intensity or vegetative cover) if the entire area was considered.
Figure 7 shows several landslides within areas identified as being susceptible to landslides using the three above-mentioned criteria.The spatial refinement was followed by filtration to detect spectral anomalies in the resultant image.A standard deviationclassified image was generated for a backscatter difference image that bracketed the December event (images acquired on December 24, 2015 and February 12, 2016).The differences in backscatter were classified into four groups: area of ≤1SD (transparent), >1SD to ≤ 2SD (green), >2SD to ≤3SD (yellow); and >3SD (red).Within the buffered areas on any of the backscatter difference images, the largest variations are expected to correspond to areas that witnessed landslide-related changes in roughness.Field observations following the December 25, 2015 precipitation event (15 mm) were conducted (February 26 to March 7, 2016) to test this assumption.The investigation proved the examination of the variations in spatially refined and spectrally filtered backscatter difference images, and the effectiveness of the applied filtering techniques in omitting the false positives.

Pixel-based adaptation of Faifa ID curve as a predictive tool
The ID curve/line for any pixel should separate landslide-producing events (backscatter difference > 1SD) from non-landslide producing (backscatter difference ≤ 1SD) precipitation events.On these graphs, landslide-producing events plot above the line/curve, and the non-producing events plot below it.5 A pixel-based debris detection system was developed by adopting the slope of the extracted Faifa ID curve.The assignment of the ID curves to the individual pixels will depend on the relative stability of the individual pixel.The less stable pixels, such as those on steep slopes, are expected to experience movement in response to weak, moderate, and extreme storm events, whereas the more stable pixels will move during the extreme events only.Figure 9 demonstrates the ID curve for two locations: 10 pixel was assigned an ID curve whose slope is similar to that of the Faifa curve.In other words, the pixel-based ID curve uses the historical landslide response of a pixel to estimate the intensity and duration of the precipitation that would cause landslide in the future.Any event that plots above the pixel ID curve would produce landslides at that location, whereas the one that plots below would not produce a landslide.The pixel-based ID curve helps to predict the location that will witness landslides under any future storm event.

Validation of ID curve
Three precipitation events larger than the threshold (1 mm/h for 1 h) were recorded during the period from November 2016 to April 2017.These occurred on February 14 (intensity: 2.28 mm/h; duration: 3.5 h), February 17 (intensity: 1.07 mm/h; duration: 9.5 h), and April 28 (intensity: 1.75 mm/h; duration: 4.5 h) of 2017.The event on February 14 plotted above the Faifa ID curve, whereas those on February 17 and April 28 plotted below the curve.Landslides were reported following the February 14 event, but not for the two other storm events, an observation that supports the validity of the extracted ID curve for Faifa.
Using the precipitation intensity and duration for the February 14 storm, and the extracted pixel ID curves, we generated a map showing the areas (three or more pixels) that are likely to witness movement under the specified precipitation conditions.
We visually inspected these areas on the Google Earth images that were acquired before (October 2, 2016) and after (March 29, 2017) the February 14 storm.Spectral and morphological variations indicative of landslides were detected on the March 29 Google Earth image.Specifically, 13 landslides were predicted, out of which 6 were verified by inspecting the March 29 image, an accuracy of 60%.This accuracy is reasonable given: (1) the uncertainties associated with extracting the Faifa and pixel-based ID curves (refer to section 4.7), and (2) the fact that landslides in a particular area tend, in some cases, to stabilize the location and reduce the chances of landslide recurrence in the same area.

Conclusions, limitations, and future work
We developed a predictive system that shows whether a storm with a particular intensity and duration can cause landslides in the Faifa Mountains.For the identified landslide-producing storms, the developed methodologies will also select areas that are likely to witness landslide development.The extracted ID curve for the Faifa is used for the former and the extracted pixelbased ID curves for the latter.
Our ability to predict the landslide-producing storms and the locations of these landslides will depend largely on the accuracy of the extracted/adopted ID curves.The production of the Faifa ID curve was based on precipitation data for 30 storms, approximately 63% of which produced landslides.Precipitation was extracted from the earlier coarse TRMM (3 h; 0.25° × 0.25°) and later from the finer resolution GPM (1 h; 0.1° × 0.1°) data; field observations and temporal Google Earth images were used to identify which of these storms produced landslides.The temporal coarseness of the precipitation data, especially 3-hourly TRMM data, provides the uncertainty in the precise placement of the ID curve.In upcoming years, additional data points (landslide producing and non-producing storms), especially those extracted from GPM or rain gauge data with higher spatial and temporal resolution will be used to refine the initial Faifa ID curve.
Refining the Faifa ID curve will enhance the accuracy of the pixel-based ID curves as well, given that they are assumed to be parallel to the Faifa ID curve.In the construction of these curves, we were constrained by the following limitations in data availability and processing technique: (1) the limited number of storms (18) that occurred throughout the time period (October

Figure 1 .
Figure 1.Location of the Faifa Mountains (outlined by red box) within the Jazan province (outlined by purple polygon).The blue dots represent meteorological stations (Mashat and Basset, 2011) in Saudi Arabia.

Figure 3 .
Figure 3. Maps showing the study area.(a) Elevation map showing locations where field observations were collected for the period extending from February 26th to March 7th, 2016.(b) Slope map generated from TanDEM-X DEM.(c) NDVI map generated from Sentinel-2 data (date of acquisition: 2017-06-05).(d) Geologic map for the Faifa Mountains (after Fairer, 1985 and Alharbi et al., 2014).5Analysis of TRMM and GPM measurements for the period 1998 to 2016 revealed sparse precipitation over the Saudi Arabian landscape (MAP: 83 mm/year), with the exception of the Red Sea Hills in southern Saudi Arabia, which receive an MAP of 108 mm/year.A progressive increase in overall rainfall over Faifa was noted over the past six years(MAP: 2010(MAP:  -2016: 315   : 315 Nat. Hazards Earth Syst.Sci.Discuss., https://doi.org/10.5194/nhess-2018-282Manuscript under review for journal Nat.Hazards Earth Syst.Sci. Discussion started: 26 November 2018 c Author(s) 2018.CC BY 4.0 License.

Figure 4 .
Figure 4. Flow chart summarizing the developed methodology that could serve as important steps towards the construction of an EWSL.The developed procedures involved the analysis of temporal Google Earth images, Sentinel-1A radar scenes, and TRMM and GPM rainfall data.Analysis involved two main steps: generation of an ID curve for the Faifa Mountains to identify storms that are likely to produce landslides (landslide-producing storms), and generation of pixel-based ID curves to identify the locations where movement is likely to occur during landslide-producing storms.Step II involved: (1) selection, calibration, and pre-processing of radar images, (2) generation of backscatter coefficient difference images as a measure of surface roughness change due to precipitation-induced landslides and, (3) development, refinement, and validation of the model to identify pixels susceptible to movement under user-defined precipitation conditions.The downloaded scenes were processed using the ESA's Sentinel Toolbox software, ENVI, ArcMap, and Python.
in the source area (onset of landslide) that gives way to more linear scars in vegetation along the landslide path.It is worth noting that different parts of the Faifa Mountains have differing numbers of Google Earth image acquisitions, amount of coverages, and resolutions.Starting in October 2014, the SGS initiated a program to field-verify reports of landslide occurrences.Field observations were conducted by our research team following the December 25th, 2015 landslide-producing storm and by the SGS researchers throughout the period from October 2014 to October 2016.Our collective field investigations revealed extensive landslides following the events on December25, 2015, April 13, 2016, April 29, 2016, August 1, 2016, and August 25, 2016 events.

Figure 5 .
Figure 5. ID scatter plot generated from landslide-producing storm events (solid circles) and non-producing precipitation events (open circles) during the period 2007 to 2016.Faifa ID curve (equation: I = 4.89*D −0.65 ; duration threshold [x]: 0.5 to 12 h) separates the landslide-producing events from non-producing events.Field-verified landslide-producing storms are plotted as black solid circles and those extracted from Google Earth images as solid red circles.Also shown for comparison are published ID curves for areas with unique climatic, topographic, and geologic settings.
Nat. Hazards Earth Syst.Sci.Discuss., https://doi.org/10.5194/nhess-2018-282Manuscript under review for journal Nat.Hazards Earth Syst.Sci. Discussion started: 26 November 2018 c Author(s) 2018.CC BY 4.0 License.The distribution of historical debris flows in the study area was investigated to identify areas susceptible to debris flow.As described earlier, areas that witnessed recent debris flows are characterized by spectrally bright rocks and sediments, a major scar in the vegetation within the source area (onset of landslide) that gives way to more linear scars in the vegetation along the landslide path.As years go by, spectral and morphologic features indicative of debris flows can get obscured by encroaching vegetation making it more difficult to identify the older debris flows.Many of the historical debris flows were found on steep

Figure 6 .
Figure 6.Landslides proximal to, and possibly triggered by, road construction.(a) Debris flow caused by failure on steep slopes intersected by roads.(b) Debris flow caused by failure of terraces constructed on steep slopes.(c) Landslide flow caused by failure on fracture planes dipping towards the road.

Figure 7 .
Figure 7. Google Earth Imagery showing the spatial refinement procedures.(a) Areas showing debris flows within areas characterized by steep slopes (>30º), and proximity to roads (<100 m) and to the first order streams (<20 m).(b) The areas that satisfy these three conditions are outlined by the shaded polygon.

Figure 8 .
Figure 8.(a) Google Earth imagery showing the distribution of debris flows.(b) Backscatter difference image for two descending scenes bracketing (acquisition dates: December 24, 2015 and February 10, 2016) a precipitation event on December 25, 2015 showing correspondence of areas of low to negligible variations (≤1SD) with vegetation and areas exceeding 1SD (clusters of red, yellow, and green pixels) to debris flow locations that showed evidence for recent redistribution of boulders and cobbles in the field.
Nat. Hazards Earth Syst.Sci.Discuss., https://doi.org/10.5194/nhess-2018-282Manuscript under review for journal Nat.Hazards Earth Syst.Sci. Discussion started: 26 November 2018 c Author(s) 2018.CC BY 4.0 License.a more stable pixel and a less stable one.The dotted line represents the ID curve for more stable locations as it showed evidence for landslide-related movement (>1SD on the radar backscatter difference image) in response to the largest storm event on August 1, 2016, and no movement (<1SD on the backscatter image) during the remaining events.The dashed line represents the ID curve for the less stable locations that showed evidence for landslide-related movement during five events (December 1, 2014, December 25, 2015, April 13, 2016, August 1, 2016, and August 25, 2016) but none following the December 25, 2016 and November 18, 2014 storms.Thus, knowing the historical response of each individual pixel to these storms, each

Figure 9 .
Figure 9. Demonstration of pixel-based ID curves.Curve "A" (dotted line) is an ID curve for a more stable pixel that witnessed landslides in response to the storm on August 1, 2016, and curve "B" (dashed line) is a curve for a less stable pixel that witnessed landslides in response to four storms (December 1, 2014, December 25, 2015, April 13, 2016, August 1, 2016, and August 25, 2016).These ID curves for stable and less stable locations are parallel to the Faifa ID curve (red line) that separates landslide-producing storm events (solid circles) from landslide non-producing storm events (open circles) from Fig. 5.The four events (solid circle within open circle) plotting between the dotted and dashed lines are landslide-producing storms (solid circle) for the less stable pixel (curve B) and non-producing storms (open circles) for the more stable pixel (curve A).

Table 1 . Intensity and duration of the precipitation events used for the construction of the Faifa ID curve. Landslide-producing storms were verified through field observations and by examining spectral and morphologic variations in pairs of Google Earth archival images bracketing significant storm events.
). Nat.Hazards Earth Syst.Sci.Discuss., https://doi.org/10.5194/nhess-2018-282Manuscript under review for journal Nat.Hazards Earth Syst.Sci. Discussion started: 26 November 2018 c Author(s) 2018.CC BY 4.0 License.