Gravitational mass movements such as landslides constitute a major natural hazard in all hilly or mountainous regions throughout the world. Although these movements are mostly a very local phenomenon, they cause damage to all types of man-made structures and affect infrastructures from local to regional scales and even on a national scale. The floods and landslides in China from May to August 2010 ranked second highest in terms of economic damage caused by natural disasters, with US$ 18 billion worth of damage [1]. In the course of this natural disaster, mass movements killed 1,765 persons and ranked in the top 10 of the most important disasters by number of persons killed [1]. Landslide triggering conditions, such as heavy rain falls and typhoons or earthquakes, can affect very large areas and sometimes cause several thousand landslides per event; for example Tsai et al. [2] reported over 9,000 detected landslides after typhoon Morakot in Taiwan in August 2009.

Within the framework of the EC-GMES-FP7 project SAFER (Services and Applications For Emergency Response), where an integrated Landslide Monitoring (LM) service has been established, a semi-automated object-based approach for landslide detection has been developed for Mont de la Saxe area, Aosta Valley, Italy. Landslide Monitoring represents a thematic reference service carried out to retrieve past ground movements of single large landslides affecting built-up areas with a high level of risk. For the site of Mont de la Saxe, LM was based on the integration of object-based analysis of optical satellite images with InSAR (Interferometric Synthetic Aperture Radar) and PSI (Persistent Scatterer Interferometry) measures of ground displacements obtained from interferometric processing of radar satellite images. These input data were subsequently integrated with further ancillary data (e.g., detailed geological and geomorphological information) and in situ measurements following a consolidated methodology, in order to obtain detailed information about the spatial and temporal distribution of landslide movements within the study area. This approach is particularly useful for investigating highly hazardous landslides, which threaten exposed, built-up areas. In such situations, the assessment of the future evolution of slope instabilities by means of integrated monitoring systems plays a major role in the adequate set up of early warning systems (EWS) and emergency plans, fundamental tools for landslide hazard and risk reduction.

The SAFER services, which are part of the GMES (Global Monitoring for Environment and Security) initiative, are targeted at the integration of users (e.g., public authorities such as civil protection agencies, humanitarian aid organizations, NGOs), who are supposed to make use of the provided services and products in emergency situations. By expressing their explicit needs as well as proposing improvements through feedback questionnaires, the users support the deployment of an operational emergency response service. The user engaged in the LM monitoring service in the Mont de la Saxe area is the Italian Department of Civil Protection (DPC), the national institution in charge of risk management and mitigation over the whole Italian territory. The investigated area represents one of the most hazardous areas of the Valle d’Aosta Region due to widespread slope instability affecting very steep slopes overlooking narrow, but densely urbanized valleys and exposing local population and infrastructure at risk.

Today, the wide range of available Earth observation (EO) data implies the need for accurate and fast methods for detecting, analyzing and monitoring landslides and to facilitate the generation of landslide inventory maps and databases to assist hazard and risk analysis. Landslide inventories are traditionally derived by visual interpretation of aerial photographs and field surveys. This kind of inventory is, however, time-consuming, fraught with the subjectivity of the visual interpreters and very costly in terms of data and workload. Satellite imagery offers a fast and economical opportunity to monitor slopes and map landslides over large and inaccessible areas, especially since the spatial resolution is ever increasing [3–6]. Applying automated methods can contribute to more efficient monitoring and timesaving as well as cost-efficient updating of existing landslide inventories.

Supervised and unsupervised classification methods of multi-spectral SPOT-5 imagery for mapping landslides were successfully used by Borghuis et al. [7] and then compared to a manual delineation. Nicol and Wong [8] used a Maximum Likelihood classifier with SPOT XS images and were able to detect approximately 70% of the existing landslides. A multiple change detection (MCD) technique to semi-automatically recognize and map landslides triggered by typhoons has been developed by Mondini et al. [9]. However, in many cases object-based methods seem to capture the complexity of natural phenomena and geomorphologic processes such as landslides in a more appropriate way than traditional pixel-based ones [10]. Thus, salt-and-pepper effects are avoided, which is especially important when dealing with complex features in terms of shape and size as landslides, where pixels show quite different spectral values. Semi-automated methods for landslide detection and analysis, especially object-based image analysis (OBIA) techniques [11,12], although still not very common, are able to deliver fast and accurate results as demonstrated by several studies. An object-based approach using a combination of high spatial resolution satellite imagery and DEM derivatives has been proposed by Barlow et al. [3]. According to the classification scheme of Cruden and Varnes [13] they distinguished between debris slides, debris flows and rock slides and achieved accuracy rates between 60% for debris flows and 90% for debris slides. Barlow et al. [3] pointed out that per pixel spectral response patterns are ineffective for delineating mass movements and instead applied an image segmentation and classification approach combining high spatial resolution satellite imagery and digital elevation derivatives. Similar approaches were demonstrated by Lahousse et al. [14], who applied a multi-scale object-based landslide detection technique based on optical imagery supported by digital elevation information to map shallow landslides in Taiwan, or Martha et al. [10], who characterized landslides based on their spectral, spatial and morphometric properties, while Martha et al. [15] elaborated the objective selection of suitable segmentation parameters to outline landslides as individual segments for subsequent landslide classification. A supervised workflow based on a Random Forest machine learning algorithm was developed by Stumpf and Kerle [5] and successfully tested on various optical data sets reaching accuracies between 73% and 87%. Further approaches for object-based landslide identification were shown by Martin and Franklin [16] for classifying soil and bedrock-dominated landslides in British Columbia, Moine et al. [17], who semi-automatically detected landslides in the French Alps using aerial and satellite images, Rau et al. [18] by using imagery acquired by a fixed-wing unmanned aerial vehicle (UAV), Hölbling and Füreder [19], who carried out preliminary work in the Aosta Valley, Northern Italy, distinguishing different mass movement types, or by Aksoy and Ercanoglu [20], who identified landslides on Landsat ETM+ (Enhanced Thematic Mapper Plus)imagery applying fuzzy classification. In the context of rapid landslide mapping a semi-automatic object-based change detection analysis was suggested by Lu et al. [4]. An automated classification system of morphological landform elements based on OBIA has been established by Drǎgut and Blaschke [21]. Their approach focuses on the use of Digital Terrain Models (DTMs) and its derivatives and they were able to transfer the approach to different landscapes and datasets. A related object-based method was recently proposed by Drǎgut and Eisank [22], who automatically classified topography from SRTM (Shuttle Radar Topography Mission) data to decompose land-surface complexity into homogeneous domains.

OBIA is making considerable progress towards a spatially explicit information extraction workflow [11] as it offers a methodological framework for addressing complex classes, defined by spectral, spatial and structural as well as hierarchical properties [23]. Thus, it provides suitable methods for analyzing landslides by using remote sensing data. Main assets of OBIA are the general potential to tackle the complexity and multi-scale characteristics of very high spatial resolution (VHSR) imagery and to allow the integration of various data sources [11,12]. Object-based methods have a high potential to monitor the evolution of landslide-prone areas over time, as spectral, spatial, contextual as well as morphological parameters can be considered [19]. In the present study the potential of an object-based landslide detection and classification approach is evaluated in an operational context aiming for the development of integrated solutions for mapping and monitoring landslides.

The study area ‘Mont de la Saxe’ is located in the Aosta Valley, North-Western Italy, near the Mont Blanc massif and covers approximately 70 km² (see Figure 1). Major villages within the study area are Cormayeur and Entréves. The Dora Baltea river originates here as the confluence of Dora di Ferret and Dora di Vèny streams. The study area is characterized by steep terrain with altitudes ranging from approximately 1,100 m to over 4,000 m above sea level.

The Western Italian Alps constitute a collisional belt developed from the Cretaceous onwards by subduction of a Mesozoic ocean and continental crust of the Adriatic (Austroalpine-Southalpine) and European (Penninic-Helvetic) continental margins [24].

In the study area, from the external to the internal sectors, three main tectono-stratigraphic units crop out (Figure 2):

Monte Bianco massif and Mont Chètif wedge complex, belonging to the Helvetic Domain, consist of paragneisses, migmatites, orthogneisses, granites and porphyries; in the SE part of the massif (Val Veny-Val Ferret) the contact with surrounding units is tectonic.

Penninic units, which comprise the lower distal clastics and pelagic Cretaceous deposits of the Courmayeur Zone, the more internal ocean-continent transition zone (Valais zone) and the middle Penninic nappes consist of Zone Huillère Permo-Carboniferous deposits (black schist with coal measures, quartzites and conglomerates) and a Pre-Permian crystalline basement (paragneiss and micaschists with amphibolites and metabasites intercalations). The described units dip towards SE forming an imbricated structure, and are arranged as large belts oriented NE-SW.

Quaternary alluvial, glacial and debris deposits directly lie unconformably and discontinuously on the bedrock. The recent and actual alluvial deposits rest on the valleys’ bottom and consist of gravels, pebbles and boulders in coarse-grained sandy matrix along the river channels and finer-grained deposits (sandy silts and silts) in floodplain areas. Old glacial deposits crop out discontinuously along the slopes at elevations up to 2,000 m above sea level, while the recent ones become particularly widespread along secondary tributary valleys at higher altitudes. They consist of massive boulders, pebbles and gravels in an abundant sandy-silty matrix (diamicton). The debris deposits include talus, alluvial fan and flow deposits and crop out at the base of rocky cliffs (talus), along the stream incision (flow deposits) and at the outlet of the minor valleys (alluvial fans).

The structural setting of the study area is strictly influenced by the Penninic frontal thrust and the Ultrahelvetic front, two regional compressive structures oriented NE-SW which separate the main tectono-stratigraphic units previously described. These lineaments are associated with secondary shear planes, milonitic and cataclastic bands, further complicating the structural setting of the area. In particular, the Penninic frontal thrust is a tectonic alignment of regional importance and its presence is detectable along most of the Western Alps [25–27].

Due to the general high relief energy and the steep slope gradient, slope instability is quite diffused in the study area. There is a clear relationship between the material involved in the instability and development of a particular type of movement. Rock falls and topples, resulting in formation of wide talus deposits are, in general, widespread along the steep rocky slopes and cliffs characterized by the presence of fractured hard rock mass (granites, gneiss, porphyries, quartzites). In some cases complex large falls (i.e., rock avalanches) have been reported [28]. Deep-seated large landslides affect high steep slopes shaped in metamorphic rocks with marked anisotropy such as gneiss and schists of the Ultrahelvetic and Penninic units. Within the displaced mass the development of sudden and rapid secondary landslides (topples, rotational and planar slides, and flow-like movements) is a common feature, giving rise to complex deformational patterns. Shallow landslides and channelized flows affect in general colluviums, weathered bedrock, soils and the more erodible lithologies of the Ultrahelvetic and Penninic units, such as argillaceous schists and black schists with coal measures (Zone Houillère). In some cases, movements could be initialized by sliding on the bedding or schistosity planes and then evolve as flow-like phenomena, thus generating complex landslides.

The Mont de la Saxe is affected by the Bois de Plan Cereux landslide (Figure 3), which constitutes an active portion of a large deep-seated gravitational slope deformation (DSGSD) extending along the whole Mont de la Saxe ridge. The instability, which can be classified as a complex landslide [13] with an estimated volume of a few Mm³, involves a significant portion of the slope and poses high risk to the Entrèves and La Palud villages as well as to the infrastructure (i.e., A5 motorway, SS26 national road, Monte Bianco cableway). The SE slope of the Mont de la Saxe ridge (Val Sapin) is characterized by the presence of complex and planar/rotational slides.

The object-based detection of landslides was based on optical, very high spatial resolution satellite imagery (SPOT-5, acquisition date: 5 May 2005) and digital elevation data (DEM with 20 m GSD) including its derived products slope, aspect, curvature and plan curvature. The multispectral SPOT-5 image with two bands in the visible spectrum (red, green) and a near infrared (NIR) band was available in pan-sharpened mode with a spatial resolution of 2.5 m. Although the satellite image was supplied in geo-referenced format, it had to be co-registered to remove distortions between the SPOT-5 image and available orthophotos, which were used as reference for the visual interpretation of the PSI results. Due to the steep terrain features within the study area, some mountain sides, including the northern hillside of Mont de la Saxe, are shaded on the SPOT-5 image and could not be taken into account for the classification.

The DEM showed some gaps and data errors, which had to be corrected before integrating it in the further analysis. Single DEM tiles were mosaicked and false values adjusted taking into account neighborhood statistics calculated for each affected input cell using map algebra in ArcGIS 10 software. The following morphological parameters were derived from the DEM to support the object-based classification of landslides: slope, aspect, curvature and plan curvature. A low pass 3-by-3 filter was applied to the curvature derivative to generate a smoother layer with less noise.

Data used for validating results of the object-based image analysis included the Italian national landslides inventory: IFFI (Inventario dei Fenomeni Franosi in Italia), and the Hydrogeological Asset Plan: PAI (Piano Stralcio per l’Assetto Idrogeologico) of the Po River Basin Authority. In addition, datasets of persistent scatterers (PS) were obtained by the processing of satellite SAR (Synthetic Aperture Radar) data acquired by three different C-band satellites (wavelength of 5.6 cm) along both ascending and descending orbits, and using monthly temporal frequency. ERS1/2 data acquired in ascending mode from 11/06/1993 to 25/12/2000, and in descending mode from 17/04/1992 to 25/12/2000 with a nominal repeat cycle of 35 d and ENVISAT ASAR (Advanced SAR) and RADARSAT-1 scenes with 35 and 24 d repeat cycle were used. ENVISAT ascending and descending data cover the intervals from 07/06/2004 to 19/10/2009 and 20/09/2004 to 23/11/2009, while RADARSAT-1 ascending and descending images span the intervals from 05/05/2006 to 28/10/2009 and 25/08/2005 to 15/10/2009. These data were processed by Altamira Information with the Stable Point Network (SPN) technique [29,30].

The distinction of the two classes flow-like landslides and other landslide types (falls, topples, slides and spreads) follows the kinematic-based classification of Cruden and Varnes [13]. Within the study area the class flow-like landslides mainly covers channelized debris bodies and their depositional areas. The source areas of the flows are included in the class other landslide types since the movements very often initiate at the head of the stream incisions by sliding or by concentrate erosion. The detection of flow-like landslides and other landslide types was conducted by applying object-based image analysis in eCognition 8 software, which offers a modular programming environment for (image-)object handling in a vertical and horizontal hierarchy [31]. Rulesets for automated analysis are coded in CNL (Cognition Network Language) within the mentioned software. OBIA constitutes the methodological framework for machine-based interpretation of target features defined by spectral, spatial, structural, contextual and also hierarchical properties [23], while reducing the influence of per pixel spectral response patterns. Knowledge is represented through the use of rule-based classifiers, making explicit the required spectral and geometrical properties just as spatial relationships for advanced class modeling [32]. Class modeling, a cyclic process of segmentation and classification, allows addressing objects individually in a region-specific manner at any stage in the ruleset [32].

In order to objectively select a suitable scale parameter for the appropriate image segmentation a statistical tool called ESP (estimation of scale parameter), developed by Drǎguţ et al. [33], was used. The ESP tool iteratively generates image-objects at multiple scale levels in a bottom-up approach and calculates the local variance of object heterogeneity within a scene for each scale [33]. By plotting the local variance against the respective scale the variation in heterogeneity can be evaluated. The scale levels at which the image can be segmented in the most appropriate manner, relative to the data properties at the scene level, are indicated by the thresholds in rates of change of local variance (ROC-LV) [33]. Based on this evaluation a scale parameter of 36, indicated by a peak in the ROC-LV graph (see Figure 4), was chosen for the initial multi-resolution segmentation.

The three multispectral bands of the SPOT-5 image were used for segmentation. A higher weight was given to the color criterion than to the shape criterion. A coarser segmentation level with scale parameter 82 and the same shape (0.3) and compactness factors (0.5) as for the finer segmentation level, also determined by making use of the ESP tool, resulting in larger image-objects, was created to facilitate a first, rough delineation of target areas affected by landslides.

By using OBIA, the influence of variant spectral reflectance of single pixels can be minimized. This seems to be applicable especially for natural, complex phenomena such as landslides, which tend to exhibit varying spectral response patterns. The creation of rulesets relied on the transformation of expert knowledge into machine-based rules, whereby the SPOT-5 image constituted the a priori data used for landslide detection and classification. During class modeling absolute spectral threshold values were kept to a minimum, whereas relational features and spatial characteristics were used instead. Vegetation indices as the NDVI (Normalized Difference Vegetation Index) and the MSAVI (Modified Soil-Adjusted Vegetation Index, see Equation (1)) proved to be very useful for differentiating target areas from areas of no interest. (1) MSAVI = 2 * N I R + 1 − ( 2 * NIR + 1 ) 2 − 8 ( NIR − RED ) 2where NIR is the near infrared band and RED the red band.

As a first step, potential landslides were identified as target objects on the finer segmentation level by applying the mentioned vegetation indices in combination with other feature characteristics, for example slope, standard deviation of the near infrared band, the maximum spectral difference within image-objects, neighborhood information, shape properties (e.g., length/width) or relief, the maximum change in elevation within an object. Objects with slope gradients above 60 degrees and below 10 degrees were excluded from the target objects. The range of these thresholds was defined relatively wide, but should ensure not to miss any areas of interest. Additionally, aspect, curvature and plan curvature appeared to be appropriate for improving the detection of target areas. By making use of these characteristics it was possible to exclude falsely classified target areas, for example rocky ridges in high-lying areas, as well as add some missed objects. The pre-classified objects were further divided into two classes: flow-like landslides and other landslide types. As the segmentation levels follow a strict image-object hierarchy the initial classification of target areas on the finer level could be transferred to the level above. The differentiation of classes was initially done on the coarse segmentation level, where shape parameters in particular, but also morphological characteristics appeared to be more significant for a first, rough separation than on the fine segmentation level (cf. Figure 5). Mainly the mean slope was used to distinguish flow-like landslides from other landslide types. Image-objects with a mean slope higher than 25° were allocated to the class other landslide types, those with a lower value were classified as flow-like landslides. Besides slope, also aspect, curvature and plan curvature were used for this first differentiation. Latter mentioned morphological characteristics turned out to be useful for supporting the class separation on the coarse segmentation level (e.g., low plan curvature values were considered for the detection of flow-like channelized landslides), but as the spatial resolution of the available DEM (20 m GSD) was not fully sufficient for determining unique thresholds, they were mainly used in combination with shape characteristics such as length/width or the brightness of objects (flow-like landslides tend to appear brighter on the SPOT-5 image) to facilitate the stepwise refinement of the classification.

Subsequently, the intermediate classification results were transferred to the finer segmentation level for further improvement of the classification. The refinement of the classification comprised the use of selected feature characteristics: besides the NDVI, spectral and spatial properties (length/width or compactness), morphological parameters as aspect and slope and especially context information (e.g., relative common border to neighboring objects) and hierarchical information (relations to super objects on the higher segmentation level) were taken into account. As natural phenomena like landslides can hardly be described using one absolute threshold, several image-object feature characteristics combined with context information were used for the further improvement of the classification. For instance, if objects were primarily classified as flow-like landslides, but their classification was fraught with uncertainty and the objects only slightly missed the parameter thresholds attached to the class other landslide types, they were assigned to a temporary class in the first step. Secondly, if these temporary objects, which were most often relatively small ones, showed a high relative border to the class other landslide types or were almost completely enclosed by this class, they were finally reclassified as other landslides type. Using neighborhood characteristics, in this case the relative border to neighboring objects, contributed essentially to the improvement of the classification.

In a few areas (e.g., partially shadowed regions on the SPOT-5 image) some misclassifications still occurred. In such cases, we made use of the capability of eCognition software to address specific regions individually. Thereby it is possible to adapt certain rules only to pre-defined regions within a larger study area. This allows for more detail and quickly eliminates classification errors and also reduces analysis time as it is not necessary to find valid rules for the whole area under investigation.

Finally, the delineation of ‘meaningful’ image-objects was further enhanced through applying split and merge functions as well as smoothing of image-object boundaries by using pixel-based resizing algorithms. In order to keep the preliminary existing result, the refinement of the boundaries was done on an independent segmentation and classification layer. Working with such sub-projects in one eCognition project allows running processes independently, for example performing independent segmentation and classification while not changing the initial image-object borders.

For smoothing the boundaries of classified objects, specific pixel-based shrinking algorithms were applied. Edge pixels between flow-like landslides and other landslide types with an NDVI value below 0.1 were allocated to a temporary class applying a surface tension of larger than 0.4 within a 5 × 5 pixel window. Surface tension is calculated based on the relative area of a given class within a moving window around the current candidate pixel. The ratio of the relative area of seed pixels to all pixels inside the moving window is computed [34]. The temporary image-objects were subsequently merged to those neighboring objects with the highest relative common border. Consequently, smoother borders between flow-like landslides and other landslide types were achieved (Figure 6). Further shrinking was done for edge pixels between classified landslide objects with NDVI values higher than 0.2 and unclassified objects, keeping the same size of the moving window, but increasing the surface tension value to 0.5.

As one of the last refinement steps, classified image-objects smaller than 200 m², having no common border to other classified objects, were removed, and unclassified objects smaller than 300 m², which were enclosed by flow-like landslides or other landslide types, were assigned to the respective enclosing class. While a few objects, which probably constituted objects of interest, still remained unclassified and some others seemed to be assigned to the wrong class, a minimal manual refinement was done at this stage. As this affected only a minor number of small objects it was decided to perform a manual improvement instead of applying additional rules.

Finally, the classified objects were partly merged to obtain more ‘meaningful’ objects. Therefore, several rules, considering aspect, size or maximum change of slope, were used to ensure that not all objects of one class were merged, but rather the most similar ones, which ideally represent one landslide.

The result of the object-based landslide detection for the Mont de la Saxe study area, which covers approximately 70 km², is shown in Figure 7. In total, an area of 3.77 km² was classified as landslide-affected; comprising an area of 1.68 km² flow-like landslides (mainly channelized debris bodies and their depositional areas) and 2.09 km² classified as other landslide types (falls, topples, slides and spreads). The smallest other landslide type detected is 556 m², the largest 94,500 m²; the smallest and largest flow-like landslides are 1,800 m² and 262,000 m², respectively.

Due to the illumination conditions at the acquisition time and the orientation of the mountainous ridges and valleys (NE-SW), all slopes facing towards NW resulted in shadowed areas (see Figure 7). For these areas, it was hardly possible to detect all landslides on the basis of the optical satellite image. Other than these obstacles, minor errors of omission (i.e., a few small landslides or sub-areas of landslide complexes could not be detected) and commission (i.e., wrongly classified parts) occurred (see Figure 8). To correct these errors minor manual editing was conducted.