Landslide recognition and mapping in rural areas represents one of the main challenges faced by the research community. The spatial and temporal distribution of landslides is well known mainly in urban areas, where they often cause severe consequences to anthropic structures and population. On the contrary, landslides in rural and remote areas often remain unknown, limiting environmental evaluations like hazard and risk assessments (Guzzetti et al., 2012). Understanding where landslides have occurred may provide useful indications to forecast future events. In particular, the knowledge of the spatial distribution of landslides in a given region is essential to implement, calibrate, and validate statistically and physically based methods (Rossi and Reichenbach, 2016; Mergili et al., 2014a, b) aimed at predicting the possible location of future mass movements or identifying areas where the probability of failure is negligible (Marchesini et al., 2014). As stated by Reichenbach et al. (2018), the quality and completeness of the landslide inventories may affect the reliability of the landslide susceptibility assessment (Steger et al., 2016). To produce inventory maps with limited errors and uncertainties (Santangelo et al., 2015), the mapping techniques should be selected by taking into account a series of factors: the purpose of the inventory, the extent of the study area (Bornaetxea et al., 2018), the scale of the base maps, the resolution and characteristics of the available data, the skills and experience of the investigators, and the available resources (Guzzetti et al., 2000; van Westen et al., 2006; Casagli et al., 2017).

Besides conventional techniques (field mapping, visual interpretation of aerial photographs), remote sensing technologies based on satellite optical imagery, airborne and terrestrial laser scanning, and digital photogrammetry represent innovative solutions for landslide detection and mapping. In addition, multispectral and Synthetic Aperture Radar (SAR) satellite images have been also used with great success. To recognize landslides in multispectral and very high resolution (VHR) optical images, the most commonly applied methods consist of visual interpretation (Fiorucci et al., 2011; Ma et al., 2016) or semiautomatic classification (segmentation) that exploits different radiometric signatures of stable and failed areas (Martha et al., 2011; Mondini et al., 2011; Alvioli et al., 2018). However, optical images have some disadvantages and cannot be used when the analyzed areas are covered by persistent clouds or affected by shadow effects. To overcome these issues, SAR data can represent an effective alternative, since they are not influenced much by weather conditions.

Several techniques allow for extracting information from SAR data to identify and map slope failures. The Differential Interferometric Synthetic Aperture Radar (DInSAR) (Gabriel et al., 1989) has been widely used to detect surface displacements over large areas with sub-centimeter accuracy. DInSAR is aimed at calculating phase differences between two or more multi-temporal images and has been successfully applied to analysis of landslides (Calò et al., 2012; Zhao et al., 2012; Cigna et al., 2013; Calvello et al., 2017; Tessari et al., 2017), earthquakes, subsidence, soil consolidation, volcanoes and tectonic deformations (Plank, 2014, and references therein). Other techniques exploit the amplitude information contained in the pixels of the SAR images. Amplitude of the backscattered signal is influenced by the type of target and varies according to several factors, such as the type of land use (e.g., water bodies, ice cover, forest type, bare soil), the surface roughness and the terrain slope. According to Colesanti and Wasowski (2006), amplitude SAR imagery potentially represents a very useful source of information, which can complement high-resolution optical imagery and aerial photography in feature detection. Generally, amplitude-based methods analyze the correlation of the speckle pattern of two images (e.g., pre- and post-event, where the terms “event” can refer to a major natural and/or human-induced hazard affecting a given area, such as an earthquake, a hurricane, a forest fire) to map the land cover changes (Raspini et al., 2017). To date, the landslide mapping community has shown a poor attitude toward using this type of product, and thus only a few studies have demonstrated the valuable contribution of SAR amplitude changes to landslide detection and mapping. According to Mondini (2017), this is due to a series of problems and drawbacks represented by (1) the complex preprocessing procedures; (2) geometric distortions, such as layover and shadowing due to the side-looking acquisition geometry of SAR sensors, which can affect the quality of the images over mountainous areas where landslides are likely to occur; and (3) the difficulty in using the SAR signal in traditional statistical classification approaches mainly due to speckling. A successful example of the use of amplitude variations of the radar signal to analyze landslides is described by Zhao et al. (2013), which inferred the occurrence of the Jiweishan rock slide in China using changes in SAR backscattering intensity in ALOS/PALSAR images. Tessari et al. (2017) verified that when the phase information cannot be exploited, the amplitude of the reflected signal is very useful to detect and map rapid-moving landslides that cause significant variations in the ground morphology and land cover. Mondini (2017) proved that both landslides and flooded areas can be detected by verifying changes in the spatial autocorrelation in a multi-temporal series of SAR images. Konishi and Suga (2018) also identified a series of landslides in Japan by analyzing intensity correlation between pre- and post-event SAR images.

Besides the described techniques, recent advances in SAR technology are promoting the use of polarimetric SAR data (PolSAR) characterized by full polarimetric information (i.e., acquired in single-polarization, dual-polarization and fully polarimetric modes) for a target in the form of the scattering matrix (Skriver, 2012). According to Plank et al. (2016), these data provide more information on the ground, which enables a better land cover classification and landslide mapping. Successful applications were described by Yamaguchi (2012), Shimada et al. (2014), Li et al. (2014) and Plank et al. (2016).

The use of SAR data to analyze landslides and/or potentially unstable slopes should hence increase in relation to a series of other valuable technical innovations as well. The improved revisiting times and spatial resolution of the images, for example, represents a key factor during disaster response operations, when a preliminary localization of areas potentially affected by major landslides is crucial. Revisiting times have in fact been reduced from 35 d for European Remote Sensing (ERS) and Envisat satellites, to 12 h (at 40∘ latitude, in case of emergency response) for the COSMO-SkyMed constellation (Casagli et al., 2017). The enhanced spatial resolution (azimuth or along-track resolution × range or across-track resolution) of images spans on the order of few meters (i.e., 1–10 m), resulting in more detail with respect to the coarser resolution of the first-generation satellites characterized by pixel sizes up to 100 m (Plank, 2014).

Among the most advanced SAR spaceborne systems (Casagli et al., 2017), there are those of the mission Sentinel-1 operated by the European Space Agency (ESA) in the frame of the European Union's Copernicus Programme. Satellites Sentinel-1A and 1B acquire images characterized by pixels with sizes ranging from 5 m (range) × 20 m (azimuth) in the default acquisition mode for land observations (Interferometric Wide Swath mode – IW), up to 5 m×5 m in the Strip Map mode. The temporal resolution ranges from 6 to 12 d according to the surveyed geographic area. The Sentinel-related products have a global coverage and are freely available to all users registered on the ESA data hub (https://scihub.copernicus.eu/, last access: 24 January 2020). This is a considerable benefit that is leading many research institutions and public administrations to use Sentinel data to investigate landslides and other natural processes (Salvi et al., 2012; Dai et al., 2016; Twele et al., 2016; Intrieri et al., 2018). According to Raspini et al. (2017), the future increased number of available satellites characterized by shorter revisiting times and high spatial resolution will offer relevant information for decision support and early warning systems. Currently, significant limitations concern the real-time and/or quasi-real-time detection of rapid flow-like mass movements, rock failures and flash floods characterized by evolution times ranging from minutes to hours. This poses a challenge for the geohydrological risk management based on satellite technologies.

In this article, we present an automatic procedure aimed at supporting the detection of rapidly moving landslides by performing a periodic survey of unstable slopes using spaceborne radar imagery. We focus on rapidly moving landslides since they determine evident land cover changes with respect to slow-moving failures. The main purpose of the implemented procedure is to emphasize areas where land cover changes (potentially related to slope failures) have occurred, facilitating the following possible phases of mapping and/or field surveys. In other words, the procedure allows for producing a map that highlights the land cover changes observed by comparing two consecutive spaceborne SAR images. Decision makers and emergency managers can use this map to organize possible verifications and field investigations.

The procedure is implemented in a processing chain based on free data and software and exploits radar backscattered signals recorded within the Sentinel-1 SAR images. The values of some parameters related to the used algorithms must be provided by the user. As an alternative, they can be set based on the values derived from other similar areas. The processing chain was applied, in back analysis, to an area in Papua New Guinea that was struck by a severe seismic sequence in February–March 2018; the most powerful seismic events triggered numerous landslides, generating widespread land cover modifications.

We selected an area located in central Papua New Guinea (Fig. 2) that was affected by a severe seismic sequence at the beginning of 2018 as a test site. On 25 February, the area was hit by a main seismic event (M7.5) followed by several aftershocks, including a M6.7 earthquake on 6 March. The strong mainshock, which was rather superficial with a hypocentral depth at 25.2 km (USGS, 2018), caused buildings to collapse, road damage and widespread landslides, mostly along the Tagari river valley and the slopes of Mount Sisa (McCue et al., 2018). According to the International Federation of Red Cross and Red Crescent Societies (IFRC, 2018), more than 100 people died, most of them due to landslides.

To test the implemented procedure, we analyzed an area of about 3000 km2 in the mountainous region close to the epicenter of the mainshock (AoI in Fig. 2), where preliminary information on landslides were available (Petley, 2018a, b).

To simulate a periodic survey covering pre- and post-earthquake periods, we downloaded 36 Sentinel-1 images from the Sentinel Data Hub (https://scihub.copernicus.eu/, last access: 24 January 2020) acquired along the satellite track no. 82 with a temporal frequency of 12 d, from 12 November 2017 to 6 June 2018. Considering that the majority of the slopes in the study area are exposed towards the west, to limit geometrical distortions in the single images and in the change detection estimation, we preferred to use IW-SLC products acquired in ascending mode, with a VV-VH polarization. Each IW product is collected with a swath characterized by a width of 250 km, subdivided in turn to three sub-swaths containing one image per polarization, consisting of a series of bursts that are processed as independent SLC images.

The downloaded images were used to perform a total of 17 change detection analyses, which resulted in likewise LR layers, with a pixel size of about 14 m. The values of the segmentation parameters were defined with an interactive manual analysis (see Sect. 2.3) by segmenting the “pre-post M7.5 earthquake” LR layer, selecting the spatial kernel size (hs) of 10 pixels (see Sect. 2.3), and setting the maximum number of iterations to 200. This size of the spatial kernel was set to 10 pixels to detect significant differences of LR values during the smoothing stage of the segmentation process, taking into account the approximate expected size of the land cover changes. In the interactive (manual) analysis, we selected bandwidth sizes (hr; see Sect. 2.3) ranging from 0.0005 to 0.016 and thresholds (th) from 0.001 to 0.016 (Fig. 3), obtaining 20 different parameter combinations. For each pair of parameters, the number of generated segments and their average size was plotted in Fig. 3. Points highlight the major impact of the hr parameter with respect to the role played by the threshold (th) parameter in defining the number of total generated segments. Below an hr value of 0.004 over-segmentation occurs, whereas for hr values equal or larger than 0.004, the number of generated segments tends to become small and constant. With the aim of avoiding over-segmentation while maintaining a reasonable average size of the segments (to be able to also delineate small patches of the terrain where changes occurred), and considering a visual inspection of the segmentation results obtained with the different combinations, we decided to run i.segment in the automatic processing chain using the following set of parameters values: hs = 10; hr = 0.004; th = 0.008; minsize = 2; iterations = 200 (see Sect. 2.3).

After the segmentation of the 17 LR layers, areas affected by layover and shadowing effects were masked out in order to avoid errors in the statistical analysis described below and in the localization of potential landslides. The mask was developed in SNAP by means of the SAR Simulation Terrain Correction tool, exploiting the SRTM 1Sec DEM.

The segments with a minimum size of 5 pixels were extracted in the area of interest (an example is shown in Fig. 4d), and statistics were calculated according to the confidence intervals described in the methodology section. We decided to select only the segments with a minimum size of 5 pixels, corresponding to a minimum area of about 980 m2 (i.e., a single-pixel area roughly equal to 196 m2×5 pixels), after a general evaluation of the preliminary landslide-related images published on news websites and social networks and also considering that the detection of smaller segments in the test area was not significant at the scale of our analysis. It is worth noting that the accuracy of such a minimum area is not accurate due to the use of a geographic (and thus not projected) reference system (WGS84).

In Fig. 5, statistics of the selected segments are displayed for each change detection. The analysis of the histograms revealed two main peaks corresponding to change detections 9 and 10. Change detection 9 considers images acquired before and after the M7.5 earthquake, whereas change detection 10 considers the images acquired on 28 February and 12 March 2018. The first peak highlights widespread changes related to landslides that were extensively documented after the M7.5 event (Petley, 2018a). The second peak was instead unexpected and was probably due to the occurrence of further landslides triggered by the M6.7 event on 6 March 2018. In Fig. 6, segments related to these two peaks are displayed (red pixels relate to change detection 9; blue pixels relate to change detection 10). To check whether these segments were effectively located in areas characterized by a high concentration of seismically induced landslides, we analyzed optical images available on the Planet explorer application (Planet Team, 2017). By means of a visual interpretation, we identified the zones (the yellow polygons shown in Fig. 6) where clusters of landslides occurred, verifying a general accordance with the spatial distribution of both red and blue segments.

Small widespread segments outside the areas affected by landslides, mostly related to stream changes and noise, also resulted in the AoI. Similar random segments occurred in the pre- and post-event change detections, as also displayed by statistics in Fig. 5.

In this article, we describe a processing chain aimed at identifying SAR amplitude changes that can be partially explained by the occurrence of mass movements. We have selected SAR data since they have the advantage of not being affected by the cloud cover disturbance. In fact, as described by Mondini et al. (2019), the use of SAR amplitude data can mitigate the cloud coverage issue and can allow for detecting landslides that otherwise might remain unknown or unnoticed for a long time. In this way, the procedure can be exploited for a “continuous” (in terms of time) slope monitoring activity, even if failures occur during long-lasting periods of precipitation and persistent cloud cover that do not allow us to use optical data for a rapid and detailed landslide recognition. In the selected study area, widespread cloud cover persisted for several weeks during and after the seismic sequence. The first cloudless optical image of the area damaged by the seismic shaking was published by the daily monitoring service delivered by © 2019 Planet Labs Inc. (https://www.planet.com/, last access: 24 January 2020) on 25 March, almost 1 month after the M7.5 mainshock that triggered numerous landslides. The high cloud persistence is quite common in Papua New Guinea, and in fact this is one of the cloudiest regions of the world, with an annual cloud frequency (proportion of days with a positive cloud flag) higher than 80 % percent (Wilson and Jets, 2016; Mondini et al., 2019). As a consequence, the use of optical data in this area, and in other mountainous regions exposed to prolonged rainfall related to monsoons, cyclones, or other persistent meteorological systems, is tricky.

The obtained results depend on the definition of the image preprocessing and segmentation parameters that should be calibrated a priori (see Sect. 2.3 and 2.4). While the geometric and radiometric corrections to the images are quite standard and well-accepted procedures, the SAR multiplicative noise filtering remains a largely discussed point in the scientific literature, and there is not a consensus on the selection of strategies. We chose the Frost filter because it has already been proven to work properly in mountainous environments (Schellenberger et al., 2012) and it was used successfully in previous studies dealing with landslides (Mondini, 2017). We acknowledge that different filters might have produced different results or requested a different tuning of the segmentation procedure. The impact of different filters on our procedure might be an interesting follow up to this work. Another improvement may consist of the use of images acquired in both ascending and descending geometries. The use of ascending images was only related to focus this first step of the work on the implementation of the entire processing chain, which we tried to simplify as much as possible. In fact, there is no doubt that combining images acquired in ascending and descending geometries can improve the quality of results, representing a nontrivial advancement of the procedure that was out of the aim of this first implementation. The a priori choice of using ascending products was based on the findings that most of the slopes in the study area are exposed towards the west, with the aim of limiting the inclusion of geometrical distortions in the change detection products.

The tuning of the segmentation parameters is the key element for identifying areas affected by significant land cover changes that are also induced by rapid slope movements. This process is event dependent, requiring a well-known landslide event that occurred in the past in the analyzed area or in zones with similar topographic and land use characteristics. In the case study described here, definite values of the segmentation parameters were obtained by segmenting the pre-post M7.5 earthquake LR layer and by testing different value combinations (Fig. 3). This may represent a limit of the proposed procedure if it is applied to different geomorphic settings without past landslide events or for identifying different types of slope failures. On the other hand, if a proper event-based tuning operation is performed, a continuous monitoring of slopes can be efficiently carried out without temporal limitations, exploiting both available pre- and post-event images, as has been done for the current case history. The described application in fact highlighted that by keeping the same parameters values, landslides and other land cover changes triggered by the M6.7 aftershock were also detected. Overlapping between the calculated segments (i.e., change detections 9 and 10) to ground truth data revealed that largest SAR amplitude changes often corresponded with landslides (Fig. 6). Further evidence was provided by the statistical distributions shown in Fig. 9, with results similar to those estimated by other landslide-related studies (Stark and Hovius, 2001; Malamud et al., 2004; Rossi et al., 2012; Schlögel et al., 2015). The segments located mostly outside the areas affected by landslides were caused instead by other land cover changes that were out of the aims of this study or by random noise effects. Segments related to these changes can be easily identified because they are composed of an average number of pixels close to 10, as detected in all the change detections, whereas segments related to landslides (i.e., change detections 9 and 10) are characterized by a higher number of pixels (Fig. 5c).

The occurrence and location of secondary failures (blue pixels in Fig. 6) were not known before our analysis because they were not reported by news and local government websites and were also missing in the maps of the Copernicus Emergency Management Service (https://emergency.copernicus.eu/mapping/list-of-components/EMSR270, last access: 24 January 2020) activated for the disaster response. The general lack of information related to these failures was likely due to a series of issues affecting both the field and the satellite surveys in the aftermath of the M=6.7 earthquake. In fact, an effective assessment in the field was impeded by the road damage also caused by the mass movements triggered by the previous major M7.5 event, whereas the use of optical satellite images was hampered by a widespread cloud cover that, as stated before, persisted during several weeks after the two main seismic shakings. The first information about the occurrence of these landslides was provided online by Petley (2018b) about 1 month later, without a clear indication of its relationship with the M6.7 earthquake. The detection of this second set of failures in areas poorly affected by previous slope movements triggered by the mainshock demonstrates the relevant usefulness of the proposed processing chain.

A suitable segmentation can hence allow us to get statistical evidence of the occurrence of event landslides. Statistical distributions of the three parameters shown in Fig. 5 provided distinctive signatures of widespread land cover changes triggered by the M7.5 mainshock and by the M6.7 aftershock. It is worth noting, however, that the 95th percentile highlighted in the plots is also exceeded by other peaks (e.g., change detection 13 in the segment count) that cannot be considered as diagnostic of landslide occurrence since they are ephemeral and are not as steady in all three plots as those of the change detections 9 and 10. In the case of small-scale landslides occurring in localized portions of a wide area, the related statistical signals may be imperceptible if these are of the same magnitude as other previous and successive signals that are not related to landslides. In cases like this, distinctive evidence of slope failures can be achieved by starting the processing chain with a smaller subset of the LR layer (i.e., monitored area).

The outcomes of this study represent a concrete example of how to exploit the relevant advantages of open-source software with a command line interface (i.e., SNAP and GRASS GIS) to implement automatic processing chains. Moreover, the proposed methodology can be properly adopted to monitor areas on the order of thousands of square kilometers if powerful hardware resources are available. In fact, the preprocessing and segmentation steps require significant amounts of calculation power and memory. It is well known that the Mean Shift is a time-consuming algorithm for large datasets (Wu and Yang, 2007) and that convergence for large areas can be reached in dozens of hours. Segmentation times are proportional to the dimensions of the monitored area and to the selected spatial kernel size (hs).

A final remark concerns the occurrence of landslides in the study area. Generally, landslides in the mountainous sectors of Papua New Guinea are very common processes. Earthquakes with a magnitude greater than 5 are among the dominant factors triggering widespread landslides. According to Robbins and Petterson (2015), such earthquakes occur regularly in the country, but records of the triggered landslides are surprisingly lacking. The lack of systematic reporting and the remoteness of communities affected by such events, also impeded an adequate characterization of landslide hazard and risk (Blong, 1986). Robbins et al. (2013) stated that landslides occur annually, and failures tend to range from few cubic meters of material to mass movements with estimated volumes of 1.8×109 m3, varying from debris slides, avalanches and flows to translational and rotational slides. In this framework, the landslide detection procedure described in the article may result in a relevant tool for local authorities of countries characterized by extensive remote areas repeatedly affected by slope failures, and for the humanitarian organizations operating in response to geohydrological disasters.

This study presents a procedure aimed at supporting the detection of landslides inducing sharp land cover changes on vast mountainous areas. It is based on SAR data acquired systematically by the Sentinel-1 satellites. The computation of the LR index and segmentation of the consequent raster layers allow for detecting areas affected by multi-temporal variations of the radar backscattered signal. Among them, areas potentially related to rapidly moving landslides can be identified with a robust statistical analysis. The performance of the implemented procedure was tested in back analysis in an area of about 3000 km2 in Papua New Guinea. Here, in 2018, two consecutive earthquakes (M7.5 and M6.7) triggered widespread slope failures causing more than 100 fatalities and severe damage to roads and buildings. The simulation of a multi-temporal survey of about 7 months, before and after the earthquakes, revealed the ability of the implemented procedure to detect statistical evidence of significant land cover changes in correspondence with the two events. Moreover, results demonstrated that the zones characterized by significant backscattering changes resulted in a reasonable agreement with those affected by landslides, as compared to the ground truth data.

The study highlights advantages of free SAR products that may guide the scientific community and the local authorities to develop archives of freely accessible data suitable for implementing streamlines of information aimed at monitoring natural and urbanized areas. As demonstrated in the case study, the proposed procedure has the potential to be a valid support in landslide emergency management, providing relevant information in near real time for civil protection authorities and scientists involved in the emergency response. Future improvements may limit the user decisions in the model parameterization, optimizing the processing times and refining the filtering of landslide-related changes by also considering geological and geomorphological factors.