A heuristic approach to global landslide susceptibility mapping
Landslides can have significant and pervasive impacts to life and property around the world. Several attempts have been made to predict the geographic distribution of landslide activity at continental and global scales. These efforts shared common traits such as resolution, modeling approach, and explanatory variables. The lessons learned from prior research have been applied to build a new global susceptibility map from existing and previously unavailable data. Data on slope, faults, geology, forest loss, and road networks were combined using a heuristic fuzzy approach. The map was evaluated with a Global Landslide Catalog developed at the National Aeronautics and Space Administration, as well as several local landslide inventories. Comparisons to similar susceptibility maps suggest that the subjective methods commonly used at this scale are, for the most part, reproducible. However, comparisons of landslide susceptibility across spatial scales must take into account the susceptibility of the local subset relative to the larger study area. The new global landslide susceptibility map is intended for use in disaster planning, situational awareness, and for incorporation into global decision support systems.
KeywordsLandslide Landslide susceptibility Remote sensing GIS Fuzzy logic
Landslides cause thousands of fatalities annually (Petley 2012; Kirschbaum et al. 2015b; Haque et al. 2016), as well as substantial property damage. The true risk may be higher than that observed in recent landslide catalogs due to the fact that most casualties are caused by rare catastrophic events (Petley et al. 2005). The first step in characterizing the potential impact of landslides (defined in this paper as any mass movements, including shallow debris flows, rock falls, and deep-seated rotational slides) is to identify where these events have occurred in the past. An ideal landslide inventory would provide both spatial and temporal information on all previous landslides over a certain domain. However, most inventories are limited to a short time period that may not fully reflect the probability of catastrophic landslides. In addition, many landslides go unreported. Therefore, it is helpful to consider not only the historical record of landslide occurrences, but also account for general principles of slope stability when predicting the spatial patterns of future landslide events.
Small-scale (defined in this paper as less than 1:1,000,000 scale) landslide susceptibility maps suffer from four main problems: (1) the lack of comprehensive and unbiased landslide inventories; (2) the coarse resolution or absence of data inputs; (3) regional differences in the importance of causative factors; and (4) the dearth of expertise on landscape processes across a vast region. This work addresses several of these limitations through a heuristic approach to represent relative landslide susceptibility at the global scale.
Summary of selected landslide susceptibility maps
Radbruch-Hall et al. (1982)
Brabb et al. (1999)
Günther et al. (2014)
BMTPC and CDMM (2003)
Liu et al. (2013)
Kirschbaum et al. (2015a)
Nadim et al. (2006)
Hong et al. (2007)
Analytical hierarchy process & Frequency ratio
Weighted linear combination
Weighted linear combination
Weighted linear combination
Weighted linear combination
Soil type or texture
Distance to stream/Drainage density
Köppen climate classification
Global landslide susceptibility map was created by combining information from four principal sources of information for five explanatory variables: slope, distance to fault, geological classification, presence of roads, and forest loss
Source and details
Viewfinder panoramas digital elevation data
3 arc seconds (~90 m)
de Ferranti (2014a) derived from 3 arc seconds SRTM DEM and several other sources.
Faults and geologic regions
Geological map of the world, 3rd edition
Distance to fault zones and geological classification
Presence of roads
OpenStreetMap contributors (2015) Data represents OSM on June 4th, 2015
Global forest change 2000–2013
Hansen et al. (2013)
There are relatively few sources of topographic information with global coverage. One of the best is the Shuttle Radar Topography Mission (SRTM). This dataset was initially released at a 3 arc seconds (approximately 30 m) resolution (Rabus et al. 2003), but has been released recently at a resolution of 1 arc seconds (approximately 30 m) and is available from 60° North to 56° South. Unfortunately, the Middle East was not available at this resolution at the time of writing. SRTM data contain substantial voids. Several attempts to address this problem have been made, including the SRTM 90 m Digital Elevation Database v4.1 (Jarvis et al. 2008), Global Land Survey Digital Elevation Model (USGS 2008) and HydroSHEDS (Lehner et al. 2008). While many of the SRTM void-filling techniques produce reasonably accurate elevations in flat areas, slope and other elevation derivatives can be severely affected—especially in mountainous terrain. Each product was evaluated by calculating slope over test areas in the Himalayas and the Sahara (where SRTM voids are common). The best global DEM for the purpose of calculating slope in complex topography was found to be Viewfinder Panoramas (de Ferranti 2014a). This is attributed to the use of several sources of topographic information in addition to SRTM, which are described below.
In order to better represent the size and shape of complex topographic features, de Ferranti (2014a) reviewed multiple series of topographic maps, as well as data from SRTM, the Advanced Spaceborne Thermal Emission and Reflection Radiometer (Hirano et al. 2003), the Ice, Cloud, and land Elevation Satellite (Schutz et al. 2005), and the RADARSAT Antarctic Mapping Project (Jezek 2002). Landsat imagery was also consulted. Then these data sources were combined in a manner designed to draw on the advantages of each (de Ferranti 2014b). Typically, SRTM DEM 1-degree tiles with 3 arc seconds resolution formed the basis for the map. Voids in each tile were filled by the most accurate alternative source. The first step in filling voids was to calculate topographic contours from the SRTM DEM. Next, the contours were connected across the no-data regions by referencing topographic maps, including spot heights. Then the map was searched for artifacts, which were corrected by hand. Finally, the contours were converted back to a raster DEM. In some cases, voids were filled directly with data from the ASTER Global DEM (GDEM) and then checked for artifacts. In Europe, most elevations are based on topographic maps or more precise sources, rather than SRTM data. Unfortunately, the tile-based contouring process seems to have introduced errors along some tile edges. The specific reason for this behavior is unclear, and the global effect is relatively minor, but it should be noted. Nevertheless, this process produces a global DEM with far better representation of SRTM no-data regions than other free elevation datasets.
Rating (Nadim et al. 2006)
Greenland ice cap
Extrusive volcanic rocks, Archean–Paleozoic
Endogenous rocks, Archean–Paleozoic
Old sedimentary rocks, Archean–Paleozoic
Extrusive volcanic rocks, Paleozoic–Mesozoic
Endogenous rocks, Paleozoic–Mesozoic
Sedimentary rocks, Paleozoic–Mesozoic
Extrusive volcanic rocks, Mesozoic
Endogenous rocks, Mesozoic–Cenozoic
Sedimentary rocks, Paleozoic–Mesozoic
Extrusive volcanic rocks, Mesozoic–Cenozoic
Extrusive volcanic rocks, Cenozoic
Seismicity increases landslide hazard by destabilizing the soil and debris on slopes, introducing additional fracturing that can allow water to penetrate and more rapidly influence the subsurface, and creating steeper or more marginal slopes as a result of seismic shaking and co-seismically triggered landslides (Keefer 1994; Okamoto et al. 2013). In addition, tectonically active areas may be prone to increased erosion, due to jointing, graben formation, volcanism, stresses (Scheidegger and Ai 1986), and uplift (Larsen and Montgomery 2012). To describe these effects, vector representations of major faults were obtained from the GMW. The distance to these faults was calculated to create a proxy for tectonic activity.
2.4 Forest loss
Land use is commonly used to explain patterns in landslide susceptibility (Korup and Stolle 2014). However, the association between specific land cover classes and the probability of landslides is challenging to characterize globally. While ontological difficulties may be avoided by use of a single global dataset, some error is likely introduced by grouping disparate biota into a relatively small number of classes. More importantly, there has not been clear consensus from the research community as to how to weight these classes. Most studies assign high susceptibility to urban areas and low susceptibility to forested areas, which might reflect the impact of anthropogenic disturbances on slope stability but could also reflect a bias toward urban areas in landslide inventories. The relationship between landslide initiation and land cover classes is more ambiguous. Empirically fitted weights would seem to obviate a research review, but biases in landslide inventories can generate incorrect associations between specific land cover classes (Steger et al. 2016b), as well as support a false sense of confidence in the resulting model (Steger et al. 2016a). Finally, it should be noted that land cover is a constantly changing variable (van Westen et al. 2008). The changes caused by fires, urbanization, etc. are likely to have more predictive power than the static land cover class itself. For these reasons, land use/land cover was eschewed in favor of forest loss.
Vegetation contributes to slope stability by binding soil particles together and enhancing evaporation (Sidle et al. 1985, 2006; Haigh et al. 1995). In a few cases, vegetation may increase hazard, but most slopes are strengthened by the presence of vegetation and weakened by its loss. To represent this variable, a Landsat-based global map of forest loss from 2000 to 2013 was evaluated (Hansen et al. 2013). The 30-m forest loss pixels were aggregated to a resolution of 30-arc seconds by treating the binary output pixel as “forest loss” if it contained any 30-m forest loss pixel. The resulting map represents forest cover change due to many causes, including timber harvesting, fire, and storms.
Roads may increase the frequency of mass wasting events (Haigh et al. 1989; Larsen and Parks 1997). Particularly in developing countries, roads built into and along steep mountain terrain often serve to destabilize the slope (similar to a river cut at a slope’s toe), which can increase the frequency of landslides. After visual comparisons with VMAP Level 0 [NIMA (National Imagery and Mapping Agency) 1993] and gROADS (CIESIN and ITOS 2013), the vector dataset OpenStreetMap (OSM) (OpenStreetMap contributors 2015) was selected to represent this factor, due its more comprehensive and accurate coverage. This roadway network was converted to a raster layer at a resolution of 30 arc seconds. Larsen and Parks (1997) observed that landslide scars were far more common within 85 m of roads. While rates remained slightly elevated at greater distances, the effects beyond 100 m from the road were less pronounced. Researchers working at the local scale typically classify distance to road by tens or hundreds of meters when mapping landslide susceptibility (Ayalew and Yamagishi 2005; Weirich and Blesius 2007; Dahal et al. 2008; Regmi et al. 2013; Bhatt et al. 2013; Rubel and Ahmed 2013). With a pixel size of approximately 1 square kilometer, the current susceptibility map cannot model the effect of road construction with the same specificity as local studies. Thus, the raster representation of road-related hazards was simplified to the presence or absence of a highway in any given pixel.
2.6 Landslide inventories
Landslide inventories used for validation of the landslide susceptibility map
Number of points/polygons
Landslides triggered by Hurricane Mitch
11,555 landslide initiation points
Bucknam et al. (2001)
Historical landslides in Nicaragua
Landslide inventory of El Salvador
Gerencia de Geología (2012)
Landslide maps of Utah
Elliott and Harty (2010)
Statewide landslide information database for Oregon, release 3.0 (SLIDO-3.0)
Guzzetti et al. (1994)
Badakhshan province inventory
Zhang et al. (2015)
Koshi river basin
Kirschbaum et al. (2015b)
A heuristic fuzzy approach has been taken at the continental (Kirschbaum et al. 2015a), regional (Ahmed et al. 2014), and local scales (Champati ray et al. 2007), but it has not been previously applied at the global scale. Fuzzy landslide models offer some advantages, which include the ability to combine similar datasets in a nested sequence prior to the final combination, the ability to use both continuous and discrete inputs, and widespread integration into GIS software. A disadvantage is that the output is a “possibility,” which is not strictly comparable to the probabilities generated by classical statistics. The heuristic fuzzy approach also enforces transparency, because all of the transformation functions are defined in advance. Unlike some machine-learning models, the hypothesis represented by the fuzzy overlay model must comport with prior knowledge, not just fit the data. This advantage is particularly important for landslide inventories that are known to have significant spatial biases.
Receiver operating characteristic (ROC) curves are commonly used to evaluate the performance of binary classifiers, i.e., tests that divide inputs into two outcomes (Zweig and Campbell 1993). Since landslide inventories are rarely complete, some locations are likely to contain unreported landslides. This is especially true for the current study area, where landslides have been recorded in less than 1% of the map’s pixels. Thus, ROC analysis will give only a rough guide to map performance, and other aspects of a landslide susceptibility map should also be considered. ROC curves were created for each of the landslide inventories described in Table 4 by calculating the number of historical landslides predicted by each possible susceptibility threshold (true positive rate) and the number of pixels above each susceptibility threshold (false positive rate).
A preliminary ROC analysis indicated that low gamma values generated a susceptibility map with a better fit to the GLC. However, inspection of the resulting maps showed that the low-gamma maps were dominated by the linear inputs, faults and roads. In contrast, the high-gamma maps identified broad regions of hazardous terrain. This discrepancy between quantitative and qualitative results can be explained by the fact that many events reported in the GLC are associated with road closures, leading to a false level of confidence in low-gamma maps that emphasize this feature. Because no single factor (other than slope, which was overlaid separately) is necessary for a landslide to occur, gamma was assigned a value of 0.9, which is consistent with the high values published in several previous studies (Tangestani 2004; Champati ray et al. 2007; Srivastava et al. 2010; Pradhan 2011; dos Santos Alvalá et al. 2013; Ahmed et al. 2014).
Performance of the global susceptibility map was analyzed with eight local landslide inventories
Koshi Basin, Nepal–India–China
5 Comparison with previous small-scale maps
The new global landslide susceptibility map resembles previous publications, both in methods and results. Landslide hotspots were identified by Nadim et al. (2006) in many of the same locations that the current study finds highly or very highly susceptible. However, the new map identifies a much larger portion of the world’s surface as highly susceptible than was shown as highly hazardous in the map of landslide and avalanche hotspots. The difference is probably due to the use of a classification system that relies upon “approximate annual frequency” in the prior work. The current study identifies some additional large areas as hotspots, including the Appalachian Mountains in the eastern United States, eastern Brazil, and Madagascar, which were previously classified as “negligible to very low”. This difference is important because many landslides, including fatal ones, have occurred in places like West Virginia, Minas Gerais, and Orissa. The new map also has much in common with the previous global landslide susceptibility map by Hong et al. (2007), including large hotspots in the Andes, Himalaya, and eastern Brazil. The most notable differences are the relatively low susceptibility ratings assigned to Indonesia, the Philippines, and New Zealand by the earlier map. The distribution of categories differs between the maps, with more pixels rated moderately susceptible in the map by Hong et al., and more pixels rated very low in the newer map. The significance of this is that very few areas can be excluded from future analysis on the basis of the older map, whereas the new global map can be used to exclude a majority of the Earth’s land surface from more detailed study. The spatial distribution of fatal landslides (Petley 2012) mostly confirms the patterns seen in all three global maps. Highly rated areas with few fatal landslides, such as the Southern Andes and the Canadian Rockies, tend to be sparsely populated, resulting in fewer reported fatalities.
This comparison suggests that maps produced with different methods, data, and scope may show largely similar results. However, maps focused on specific landslide hotspots are not directly comparable to broader overviews unless a single, rigorous classification method was applied to both maps.
While comparison with previous small-scale maps revealed strong similarities, this global landslide susceptibility map improves upon prior maps in four important ways. First, several new or updated datasets have been released in the last decade. In the current context, the most important of these is a DEM made with high-quality SRTM void-filling techniques. Second, the use of a conservative method for aggregating 90-m slope values means that all major topographic features were considered by this analysis. Third, the use of fuzzy overlay preserves the full information content of continuous variables like slope gradient. Fourth, the simple classification scheme will be familiar to users of other susceptibility maps, but the uneven pixel distribution should draw the user’s attention to the most critical sites.
Nevertheless, several features of the new map may limit its use. First, the resolution of the map is approximately 1 km, and terrain varies significantly within many pixels. The choice to aggregate slope by computing the maximum value means that some pixels may contain a very small area of steep terrain, while the remainder is not susceptible to landslides. Second, the use of biased and incomplete landslide inventories to evaluate the susceptibility map makes the results more difficult to interpret. Although this susceptibility model (Fig. 2) was not fitted empirically, landslide inventories informed the prior research on which it was based. Third, the Geological Map of the World is only appropriate for use over very large areas. At local and national scales, more detailed information is often available, but varies in quality, format, and cost. Fourth, this map models all mass movements with the same treatment. The real world is more complex, and factors which drive rock toppling in Canada are not the same as those which can cause debris flows in New Guinea. Fifth, this map does not provide an explicit hazard level in the form of an annual probability of slope failure. Therefore, it is very likely that landslides will occur at some date in all of the very highly susceptible locations, but the size, frequency, and timing of those events are not known. These limitations suggest that the global susceptibility map is best used for a few purposes: situational awareness of global landslide hotspots and potential occurrence, the development of global decision support systems, and prioritization of future landslide research. It is not appropriate for decisions about infrastructure design, building code legislation, or local land-use planning.
Excerpts from this landslide susceptibility map have already been used during the period leading up to potential disasters. In one such instance, the approach of Hurricane Madeline toward the Hawaiian Islands triggered a request for information on the potential for landslides. Although the global map is not tailored specifically to this location, it was still the most relevant and detailed dataset available to decision makers. This map has also been applied as one component of a global landslide nowcast system (Kirschbaum and Stanley 2016). The nowcasts are issued at two levels, high-hazard and moderate-hazard, which correspond to different classes of the global susceptibility map. After considering susceptibility, a 7-day antecedent rainfall index is compared to historical precipitation levels to identify hazardous locations in nearly real time. While this system focuses upon rain, other landslide triggers, such as melting snow or recent seismicity, could be considered in similar models.
This research assessed landslide susceptibility at a resolution of approximately 1 km with nearly global coverage. The map was evaluated with one Global Landslide Catalog and several local to regional landslide inventories. The geographic distribution of landslide susceptibility is very similar to that in previous small-scale maps, with the most dangerous terrain located around the Pacific Rim and along the Himalayan Mountains. Other hot spots can be found in Europe, Africa, and the Americas. While this map benefited from several excellent and free datasets, further improvements to thematic data, particularly in soil mapping of mountain regions and landslide cataloging, would improve the results of any future work. The global susceptibility map might be improved by incorporation of any future in situ and satellite-based datasets with improved resolution, accuracy, or completeness. The map may be useful for long-term risk assessment and disaster response planning, as well as in the development of real-time hazard models.
Thank you to all of the contributors to the Global Landslide Catalog since its creation in 2007. Thank you also to all of those who provided landslide inventories for analysis, including Deo Raj Gurung and Jianqiang Zhang (ICIMOD), Mauro Rossi (CNR IRPI), Graziella Devoli, Manuel Diaz (MARN), the Oregon DOGAMI, the USGS, and the Utah Geological Survey. This work was supported by NASA’s Precipitation Measurement Missions.
- BMTPC (Building Materials and Technology Promotion Council Ministry of Urban Development and Poverty Alleviation Government of India), CDMM (Centre for Disaster Mitigation and Management, Anna University) (2003) Landslide Hazard Zonation Atlas of India. New DelhiGoogle Scholar
- Bonham-Carter G (1994) Geographic information systems for geoscientists: modelling with GIS. Elsevier, AmsterdamGoogle Scholar
- Bouysse P (2009) Geological map of the world at 1:50 000 000. Commission for the Geological Map of the WorldGoogle Scholar
- Brabb EE, Colgan JP, Best TC (1999) Map showing inventory and regional susceptibility for Holocene debris flows and related fast moving landslides in the conterminous United States. Accessed http://pubs.usgs.gov/mf/1999/2329/
- Bucknam RC, Coe JA, Chavarría MM et al (2001) Landslides triggered by Hurricane Mitch in Guatemala—inventory and discussion. US Geological Survey Open File Report 01-443:38Google Scholar
- Center for International Earth Science Information Network, Information Technology Outreach Services (2013) Global roads open access data set, version 1. http://sedac.ciesin.columbia.edu/data/set/groads-global-roads-open-access-v1. Accessed 1 Jan 2015
- Cepeda J, Smebye, H, Vangelsten, B et al (2010) Landslide risk in Indonesia. Global assessment report on disaster risk reduction. United NationsGoogle Scholar
- de Ferranti J (2014a) Digital Elevation Data—with SRTM voids filled using accurate topographic mapping. http://www.viewfinderpanoramas.org/dem3.html. Accessed 17 Nov 2015
- de Ferranti J (2014b) Digital Elevation Data: SRTM void fill. http://viewfinderpanoramas.org/voidfill.html. Accessed 19 May 2016
- DOGAMI (Oregon Department of Geology and Mineral Industries) (2015) SLIDO: statewide landslide information layer for Oregon. http://www.oregongeology.org/sub/slido/data.htm. Accessed 11 Oct 2015
- dos Santos Alvalá RC, Camarinha PIM, Canavesi V (2013) Landslide susceptibility mapping in the coastal region in the State of São Paulo, Brazil. In: American Geophysical Union, Spring MeetingGoogle Scholar
- Elliott AH, Harty KM (2010) Landslide maps of Utah. Utah Geological Survey Map 246DM:14. 46 plates. 1:100,000 scale. DVDGoogle Scholar
- ESRI (2013) ArcGIS Desktop, version 10.2. Environmental Systems Research Institute, Redlands, CaliforniaGoogle Scholar
- Frolova JV, Gvozdeva IP, Kuznetsov NP (2015) Effects of Hydrothermal Alterations on Physical and Mechanical Properties of Rocks in the Geysers Valley (Kamchatka Peninsula) in Connection with Landslide Development. In: Proceedings World Geothermal Congress 2015, pp 1–6Google Scholar
- Gerencia de Geología (2012) Landslide inventory of El Salvador. Ministerio de Medio Ambiente y Recursos Naturales, El SalvadorGoogle Scholar
- Günther A, Van Den Eeckhaut M, Malet J-P et al (2014) Climate-physiographically differentiated Pan-European landslide susceptibility assessment using spatial multi-criteria evaluation and transnational landslide information. Geomorphology 224:69–85. doi:10.1016/j.geomorph.2014.07.011 CrossRefGoogle Scholar
- Hijmans RJ (2015) Raster: geographic data analysis and modeling. R package version 2.4-15. https://CRAN.R-project.org/package=raster
- Hosmer DW, Lemeshow S (2005) Assessing the fit of the model. In: Applied logistic regression, 2nd edn. Wiley, Inc., Hoboken, NJ, USA, pp 143–202Google Scholar
- ICIMOD (International Centre for Integrated Mountain Development) (1992) Landslides in Koshi River Basin of 1990. http://rds.icimod.org/Home/DataDetail?metadataId=23175&searchlist=True. Accessed 7 Jan 2015
- ICIMOD (International Centre for Integrated Mountain Development) (2010) Landslides in Koshi River Basin of 2010. http://rds.icimod.org/Home/DataDetail?metadataId=23176&searchlist=True. Accessed 7 Jan 2015
- Jarvis A, Reuter H, Nelson A, Guevara E (2008) Hole-filled SRTM for the globe version 4. Available from the CGIAR-CSI SRTM 90 m DatabaseGoogle Scholar
- Kirschbaum D, Stanley T (2016) A satellite-based global landslide hazard assessment model for situational awareness. In: Geological society of america abstracts with programs, vol 48. doi:10.1130/abs/2016AM-279271
- Larsen MC, Parks JE (1997) How wide is a road? The association of roads and mass-wasting in a forested montane environment. Earth Surf Process Landf 22:835–848. doi:10.1002/(SICI)1096-9837(199709)22:9<835:AID-ESP782>3.0.CO;2-C CrossRefGoogle Scholar
- NIMA (National Imagery and Mapping Agency) (1993) Vector map (VMap) level 0. http://earth-info.nga.mil/publications/vmap0.html. Accessed 1 Jan 2014
- OpenStreetMap contributors (2015) OpenStreetMap. http://osm-x-tractor.org/Data.aspx. Accessed 7 Jun 2015
- Petley DN, Dunning SA, Rosser NJ (2005) The analysis of global landslide risk through the creation of a database of worldwide landslide fatalities. In: Hungr O, Fell R, Couture R, Eberhardt E (eds) Landslide risk management. CRC Press, Boca Raton, p 776Google Scholar
- Radbruch-Hall DH, Colton RB, Davies WE et al (1982) Landslide overview map of the conterminous United States. U.S Government Printing Office, WashingtonGoogle Scholar
- Rubel Y, Ahmed B (2013) Understanding the issues involved in urban landslide vulnerability in Chittagong metropolitan area. Association of American Geographers (AAG), BangladeshGoogle Scholar
- Steger S, Brenning A, Bell R, Glade T (2016b) The impact of systematically incomplete and positionally inaccurate landslide inventories on statistical landslide susceptibility models. In: EGU general assembly conference abstracts 18:6666Google Scholar
- USGS (United States Geological Survey) (2008) Global land survey digital elevation model. Global Land Cover Facility, University of Maryland, College Park, Maryland. http://glcf.umd.edu/data/glsdem/
- Verdin KL, Godt JW, Funk C et al (2007) Development of a global slope dataset for estimation of landslide occurrence resulting from earthquakes: U.S. Geological Survey, Colorado. Open-File Report 2007–1188:25Google Scholar
- Zweig MH, Campbell G (1993) Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clin Chem 39:561–577Google Scholar