Introduction

A massive low-pressure cyclone, “Tamara”, hit the Western Balkan countries in May 2014, resulting in extensive flooding and landslide damages in Serbia and Bosnia and Herzegovina. The floods, flash floods and landslides resulted in 51 casualties, approximately 32,000 people being evacuated and more than 1.6 million people being directly or indirectly affected in Serbia alone. The Serbian government declared a state of emergency covering the whole territory of Serbia on May 15, 2014. The United Nations Disaster Assessment and Coordination Team (UNDAC) assisted immediately and estimated that roughly more than 2000 landslides were activated in the western and central parts of Serbia (UNDAC 2014). Some of the locations in western Serbia were affected by many flow-type landslides, which had never previously been reported in these areas. These landslides caused severe damage to the local municipalities (residential areas, roads, infrastructure facilities, cultivated lands, pastures and forests), as well as to the highly urbanized areas in the city of Belgrade, the capital of Serbia. In the framework of the post-disaster recovery led by the United Nations Development Programme (UNDP) Office in Serbia, several actions were taken with the aim of supporting landslide risk assessment and management in the affected municipalities. Following these activities, attention was focused on developing a fast and efficient methodology for a post-event landslide inventory, which had to be completed within 5 months for more than 14,500 km2, with limited financial resources and personnel. It entailed integrating and homogenizing all of the landslide data in the National Landslide Database, which is yet to be completed.

Many recent studies have highlighted the importance of using remote sensing technologies for fast landslide mapping in similar situations (Tralli et al. 2005; Lin et al. 2011; Joyce et al. 2014; Kwan and Ransberger 2010; Guzzetti et al. 2012; Lira et al. 2013; Bhambri et al. 2016; Martha et al. 2015; Ray et al. 2016). In particular, during the response and recovery phases of the disaster and crisis cycle, only very fast delivery of up-to-date, accurate and comprehensive image analysis products can significantly assist in the assessment of large-scale disaster damage (Voigt et al. 2007; Denis et al. 2016). Event-based inventories are important for documenting the full extent of a landslide disaster caused by rainfall (Bucknam et al. 2001; Guzzetti et al. 2004; Cardinali et al. 2006; Lin et al. 2011; Marjanović and Abolmasov 2015) or earthquakes (Sato et al. 2007; Sato and Harp 2009; Tang et al. 2015; Yang and Chen 2010; Xu et al., 2014; Martha et al. 2016; Shafique et al. 2016) and for damage assessment related to post-disaster management activities (Martha and Kumar 2013; Ciampalini et al. 2015). Landslide event-based inventory maps in the post-disaster phase are usually created through the interpretation of stereoscopic aerial photographs taken after the event (Bucknam et al. 2001; Guzzetti et al. 2004, Tang et al. 2015) using visual (Marjanović et al. 2016, Lin et al. 2011) or semi-automatic analysis (Martha et al. 2010, 2012; Mondini et al. 2011) of the multispectral and multitemporal satellite images, using high resolution digital elevation model analysis (Chen et al. 2006; Booth et al. 2009; Kwan and Ransberger 2010; Iwahashi et al. 2012), using extensive field surveys (Cardinali et al. 2006; Marjanović and Abolmasov 2015) or using a combination of these techniques (Mondini et al. 2011; Xu et al. 2013, Murillo-García et al. 2015; Alkevli and Ercanoglu 2011). In relation to all the abovementioned case studies, this case study differs in some respects. Firstly, the regional extent of similar studies is much smaller than the one considered here (Lin et al. 2011, Martha and Kumar 2013), except for the earthquake-induced landslide cases (Tang et al. 2015; Martha et al. 2016). This finding implies that very high resolution (VHR) image coverage of sufficient quality was difficult to obtain for the entire area of interest, as highlighted in Xu et al. (2014). Secondly, previous case studies combined the same resolution of pre-event and post-event (VHR or high resolution (HR)) images for visual (Martha et al. 2015; Lira et al. 2013) or VHR satellite data for semi-automatic recognition (Martha and Kumar 2013). Although highly recommended, this is a demanding and expensive approach for large areas, such as the one in this case study. Instead, it is herein proposed that only some parts of the whole area should have VHR coverage, and these should be used for control of the semi-automated landslide mapping, performed by using lower resolution images. Thereby, semi-automated mapping can be extrapolated outside the VHR coverage. The pilot area in the Krupanj municipality was herein used to explore this option.

The main objective of this research was an analysis of the available multiresolution satellite image data for a fast post-event landslide inventory of the 23 municipalities affected by the extreme rainfall event during May 2014. VHR and HR Pléiades, SPOT 6 and WorldView-2 (WV2) multispectral satellite images were used for visual landslide recognition and to build an inventory to support an extensive landslide mapping field campaign. The VHR and HR satellite images covered a 9167.11-km2 territory, and a total of 1298 landslides of various types were visually recognized in these images. Additionally, the free-to-access HR Landsat 8 multitemporal satellite images and open source QGIS software were used for the analysis of the Krupanj municipality (341 km2) as a test area for the semi-automatic recognition of landslide occurrences. The objective of this additional analysis was to show that the semi-automatic landslide recognition in HR satellite images could be a useful tool for other affected territories where VHR satellite images are unavailable. In addition, another objective was to show how a preliminary landslide inventory obtained by remote sensing analysis can be an essential support for a field campaign, as well as a reliable source for landslide inventorying in inaccessible terrains. An event-based landslide inventory for a particular rainfall trigger event was undertaken for the first time on a previously unheard of scale in Serbia. Additionally, for the first time, the entire database has been made publicly available via an interactive web portal (http://geoliss.mre.gov.rs/beware/).

Study area

The study area covered 11.840 km2, i.e. 23 of the 27 municipalities included in the UNDP post-disaster BEWARE Project activities in the western and central parts of the Republic of Serbia (Fig. 1). Four municipalities were excluded from the analysis because there were no landslides that occurred during the May 2014 rainfall event, only floods and flash floods. In the 23 municipalities that were included, the total population was approximately 1,000,000 people. It is also important to mention that the majority of these municipalities are characterized as underdeveloped and undeveloped (11 underdeveloped and 4 undeveloped of the 23 chosen municipalities), with fragile economies, poor industrial potentials and high unemployment rates (http://www.regionalnirazvoj.gov.rs/ data from 2014, accessed January 2017). All these factors make these areas even more vulnerable to disastrous landslide events, such as the ones that occurred in May 2014.

Fig. 1
figure 1

Geographical position of the Republic of Serbia and the position of the included municipalities in the research

Full size image

Geological settings

The geological and geomorphological settings and other environmental conditions are very complex. The central and western parts of Serbia are situated in the region of the so-called Dinaric fold-and-thrust belt, encompassing several continental and oceanic tectonic units that were left behind after the extensive processes of the Jurassic closure of the Neotethys Ocean (Schmid et al. 2008) and the Cretaceous–Oligocene nappe stacking (Schefer 2010; Ustaszewski et al. 2010; Mladenović et al. 2015). The oceanic tectonic units show a typical sequence of peridotite, gabbro and basalt covered by an ophiolitic mélange (Dimitrijević 1997). The continental units are mostly composed of slightly metamorphosed sandstones and shales covered by shallow limestone sequences (Dimitrijević 1997). All the tectonic units are covered by Cretaceous flysch deposits. During the Miocene, this area was affected by an extension, which provided the conditions leading to smaller basins in the so-called Dinaric Lake system (Krstić et al. 2003). These basins are mostly filled with limestones, sandstones, marls and shales (Dimitrijević 1997).

The geomorphological settings are mostly characterized by the fluvial and fluvial-denudation features that are dominant in western and central Serbia. Erosional features prevail on the slopes of the hills and mountains, while depositional features occupy basins, valleys and other depressions. These deposits are mainly within the clay fraction and are hence considered to be responsible for the development of the most common types of landslides in this region (Menković et al. 2003; Mihalić Arbanas et al. 2013; Abolmasov et al. 2015). However, the landslides in Serbia, with all their diverse mechanisms and types of mobilized material (Cruden and Varnes 1996; Cruden and VanDine 2013; Hungr et al. 2014), involve many other factors, such as the local geomorphological characteristics, engineering geological material properties, degree and depth of the weathering substratum, land cover and land use as well as the characteristics of the factors that trigger them (e.g. the unusually high precipitation rate in May 2014).

The municipality of Krupanj, one of the most affected areas in May 2014, was selected as a test site for more detailed research. This area is located in western Serbia and covers 341 km2, with a population of 17,295 inhabitants (Fig. 1). The majority of the territory consists of Devonian-Carboniferous weathered low-crystalline metamorphic rocks, including mud and clay shales, phyllites and argillaceous schist (approximately 80%). Permian meta-sediments (clay shales, sandstone and flysch), Triassic limestone and sandstone and Cretaceous limestone make up the remaining 20% of the territory. The weathering zone thickness in these rocks is between 1 to 5 m, depending on the dominant lithological units and structures. Quaternary deposits are generally represented by alluvial and diluvial sediments. The topography is a typical low mountainous to hilly relief landscape with predominantly gentle to moderate slopes (7–21°) and many micro-relief forms enclosed in relatively small catchments.

Rainfall event

In the third week of May 2014, the massive low-pressure cyclone Tamara swept through the Western Balkans, resulting in extreme precipitation over a short period, which caused floods, flash floods and massive landslides in the western and central parts of Serbia. The highest 48-h rainfall was registered at the Loznica Main Meteorological Station (MMS) in the western part of Serbia, exceeding 160 mm/48 h, which corresponds to an extreme rainfall that statistically occurs once every 1000 years over that area (1000-year return period). MMS in Valjevo and Belgrade recorded precipitation of a 400-year return period for the same duration (Prohaska et al. 2014). The highest values of 72-h rainfall were also recorded in Loznica (213 mm), Valjevo (190 mm) and Belgrade (174 mm). The following flood events and landslides were triggered instantly, but it is possible that the massive landslides were also initiated by the antecedent rainfall, which was unusually high from April 15 to May 13, 2014 (Fig. 2), when the monthly precipitation exceeded half of the previous annual average. Namely, the annual precipitation averages for the Loznica, Valjevo and Belgrade MMSs were 868, 788 and 691 mm/year, respectively (reference period, 1981–2010). Even though these values are seemingly low in comparison to other latitudes (e.g. tropical or Mediterranean scenarios), 48–72-h rainfall of approximately 100–200 mm (10–20% of annual total) at these three MMS represents an extreme level of soil saturation, especially when coupled with the antecedent cumulative rainfall effects (causing constant, gradual deterioration of the soil strength by decreasing the effective stresses).

Fig. 2
figure 2

Antecedent rainfall, from April 15 to May 13, 2014 for Belgrade, Loznica and Valjevo meteorological stations

Full size image

Materials and methods

Data sources

The main input sets for the visual analysis were all available in the cloud-free (less than 5% cloud) and snow-cover-free series of the VHR and HR images from May 23, 2014 to November 04, 2014, provided by the UNDP Office in Serbia through the post-disaster BEWARE Project. The general characteristics of the available satellite images, including their date of acquisition, resolution, number of scenes and name of the mission, are given in Table 1. The VHR and HR satellite images covered 9167.11 km2 or 77.43% of the study area (Table 2 and Fig. 3). For two of the municipalities (Paraćin and Smederevska Palanka), there were no available satellite images (i.e. 23.57% of study area) (Table 2). Pléiades VHR images covered only 19.52% of the study area, while the rest was covered by the HR SPOT 6 and WV2 missions (38.57 and 19.33% of the study area, respectively) during the required time spans (Table 2).

Table 1 Post-disaster VHR and HR satellite data used for visual analysis
Full size table
Table 2 List of municipalities and spatial coverage of municipality areas by VHR and HR satellite data
Full size table
Fig. 3
figure 3

Spatial coverage of municipality areas by satellite data used in the analysis (list of municipalities is given in Table 2)

Full size image

Multitemporal, pre-event (August 9, 2013) and post-event (August 19, 2014) free access Landsat 8 satellite images, number 187/29 (row/path), were analysed for the Krupanj municipality as a test of semi-automatic landslide recognition. These images were also cloud-free, and they had similar vegetation and agricultural conditions. The territory of Krupanj, as indicated earlier, was chosen not only because it provided a diversity and multiplicity of landslide examples but also because it had large coverage in the VHR Pléiades images (87.33% of the territory) (Table 2).

Data processing

The Pléiades, SPOT 6 and WorldView-2 images used in the visual analysis were orthorectified and georeferenced by the vendor. The images have been pan-sharpened and mosaicked. The visual recognition was aided by a 25-m shaded relief derived from the DEM (supplied by the Republic Geodetic Authority of Republic of Serbia) using different transparency settings, as well as a pre-event background image from the Google platform. In all multispectral images, the usual true RGB colour composite and false colour composite were used (Fig. 4).

Fig. 4
figure 4

Flow chart of visual landslide recognition

Full size image

Data pre-processing and Landsat 8 image processing (Fig. 5) were performed in the open source QGIS software. To maximize the usefulness of the data, radiometric calibration and correction procedures were applied (Chavez 1996). Pre-processing included radiometric calibration and atmospheric correction; the derived data were extracted for the surface temperature calculation (Moran et al. 1992; Valor and Caselles 1996; Jiménez-Muñoz et al. 2006; Osiñska-Skotak 2007). The surface temperature is generated using a Split Window algorithm, which includes the emissivity, vapour, temperature profiles, transmittance and reflection (Fig. 5). Pre-processing further included brightness optimization, pan-sharpening and masking procedures (defining the areas covered by clouds and their removal). Processing included Principal Component Analysis (PCA) and semi-automatic classification. PCA was conducted on six pre-processed multispectral bands (from two to seven), and PCA2 was the one that best showed the landslides because PCA2 gives the relationship between bands 5 and 3 (red and infrared), which represents the Vegetation Index. A raster calculator (Change detection) was used to compute the difference between the pre- and post-event PCA2s to compare the differences in their vegetation cover and, thereby, the potential presence of slides and flows. A supervised classification of land cover was completed using a semi-automatic classification plug-in with a Minimum Distance algorithm using four classes (Minu and Bindhu 2016), which will be explained in “Data analysis” section.

Fig. 5
figure 5

Flow chart of applied pre-processing and processing procedures for semi-automatic landslide recognition

Full size image

Data analysis

Visual analysis and interpretation

Visual analysis and interpretation were performed on the available Pléiades, SPOT 6 and WorldView-2 satellite images provided by UNDP. Originally, the high- to low-resolution (1.5–6 m) multispectral and panchromatic SPOT 6 and WorldView-2 satellite images and their acquisition dates had limited usefulness (Table 1) because their visual interpretation was time consuming and constrained only to large slides, according to Metternicht and Hurni (2005). On the other hand, the VHR Pléiades images (0.5 m) were very helpful for fast detection and interpretation of all landslides types. The landslides in these images were detected on the basis of a set of standard criteria: their tone changes and the morphological footprints in the relief. Tone changes (for black-and-white images) or changes in colour (colour images) indicated a change in soil moisture. Visual interpretation allowed a relatively confident discerning of flows, as elongated irregular forms developed down the slope and shallow slides in irregular elliptical forms (Figs. 6 and 7). Some of the flows followed the existing local micro-relief depressions on the slopes, but others were formed in the gullies of the upstream zones. Flash flood zones were also routinely digitized (Fig. 8).

Fig. 6
figure 6

Left: Pléiades satellite image (resolution 0.5 m, RGB 3, 2, 1). Right: Recognition of landslides, Krupanj municipality

Full size image
Fig. 7
figure 7

Left: Pléiades satellite image (resolution 0.5 m, RGB 3, 2, 1). Right: Recognition of earth flow (yellow) and debris flows (pink), Krupanj municipality

Full size image
Fig. 8
figure 8

Left: Pléiades satellite image (resolution 0.5 m, RGB 3, 2, 1). Right: Recognition of flash flood zone Krupanj municipality urban area

Full size image

All landslide occurrences were interpreted with a specified degree of certainty. The level of certainty was determined based on the number of satisfied interpretation criteria, such as the morphological footprints in the relief, tone/colour changes, soil moisture changes, gullies, vegetation, textures (including cracks), surface disturbances and artificial object disturbances (Haugerud et al. 2003; Schulz 2007). Independent of the landslide size, if the phenomenon satisfied all six of these criteria, the highest level of certainty was assigned to it. The highest level of certainty was labelled “1” (890 landslides), medium was labelled “2” (214 landslides, which satisfied a minimum 3/6 criteria), while the lowest level of certainty was marked “3” (179 landslides); see the example in Fig. 9. Approximately 516 of the 890 landslides with the highest interpretation certainty were flow-type landslides, suggesting that this type of mechanism provides the easiest and most reliable interpretation. In contrast, approximately 100 of the 179 occurrences with the lowest interpretation certainties belonged to slide-type landslides, suggesting that this type has the least reliable interpretation variants. Since 70% of all of the interpreted instabilities fell in the “1” category, where all criteria for interpretation were satisfied, it is already clear that VHR-HR images provide an efficient and reliable source of information.

Fig. 9
figure 9

Left: Pléiades satellite image (resolution 0.5 m, RGB 3, 2, 1). Right: Illustration of different level of certainty of landslide (red polygon certainty 1, yellow polygon certainty 2, green polygon certainty 3)

Full size image

Semi-automatic detection of landslides for the test area

The spectral plot, original bands and soil temperature intermediate products described in “Data processing” section were analysed with a semi-automatic supervised classification in order to determine the target classes. The following classes were selected: areas that included clouds and/or shadows—class 0; areas of cultivated land—class 1; areas under water and vegetation—class 2; and areas with bare land (eroded areas)—3 (Fig. 10). The process of automatic classification misclassified alluvial plains and river terraces as eroded areas (that potentially contain landslides); therefore, they were manually removed from class 3. It was found that instabilities that represent a single body were disaggregated into multiple polygons. All such cases were identified and merged into logical entities in class 3. A semi-automatic classification of the PCA2 difference (PCA/before and PCA/after) (Krupanj) singled out more than 4500 occurrences of instabilities in the test area. The obtained polygons were plotted and correlated with a slope map derived from the digital elevation model (25 × 25 m). According to the morphological parameters in the test area (the Krupanj municipality), it was assumed that slopes ranging between 5° and 25° were possible hosts of landslides according to field data and statistical analysis of slope angle range vs. landslides (the average slope angle for all landslides is 23°). All zones outside of this range were excluded from further semi-automatic analysis for simplification, and the number of total polygons was reduced. Many of these were split polygons that belonged to the same landslide and had to be aggregated into single logical entities. The remaining split polygons were visually/manually grouped into logical entities.

Fig. 10
figure 10

Diagram of wavelength for different classes

Full size image

Results and discussion

Results of the analysis of the whole study area

The resolutions of the satellite images that covered the whole research area and included all 23 municipalities varied from 0.5 m for the Pléiades images to 1.5–6 m for the SPOT 6 and WV2 images. This variance was the source of the uneven interpretation qualities and different numbers of recognized landslides throughout the area. The field mapping campaigns, on the other hand, had lesser but more consistent areal coverage (Table 3).

Table 3 List of visually recognized landslides per municipality
Full size table

Based on the visual interpretation and analysis of the VHR and HR satellite images of the study area, a total of 1298 landslides (592 slides and 691 debris/earth flows) were recognized, while 15 instances of movements were not assigned to be slide- or flow-type movements since they represent complex types of movement. The total number of visually recognized landslides per municipality included in the research is given in Table 3. Several municipalities were covered by the relatively low-resolution SPOT 6 and WorldView-2 (MS) satellite images, or they were not covered by any satellite images at all. In these municipalities, subsequent field work has only established slides. Because it is relatively difficult to identify slides with relatively small displacements (below 12 m) and small magnitudes (below 500 m2) without stereo images and considering the resolution of the available satellite images in the territory of these municipalities, it has not been possible to visually interpret slides.

The largest landslide/debris flow was recognized in the municipality of Ljubovija, near the village of Selanac, with an area of 85,222.23 m2 (Table 4). This landslide/debris flow was 1.5 km long and, at its wider part, was 220 m wide (Figs. 11 and 12).

Table 4 Details of visually recognized landslides per municipality
Full size table
Fig. 11
figure 11

Example of a representative visually recognized Selanac landslide/debris flow

Full size image
Fig. 12
figure 12

Photo of the he Selanac landslide/debris flow (from Fig. 11)

Full size image

Results of the analysis of the test area (Krupanj municipality)

Results of the visual interpretation of the satellite images in the Krupanj test area are shown in Fig. 13. A total of 507 landslides (Table 5) were assigned certainty labels of “1” and “2”.

Fig. 13
figure 13

An example of visual recognition of landslides (red colour) in Krupanj area. Pléiades image, June 8, 2014 (transparency 50%)

Full size image
Table 5 Number of recognized landslides in Krupanj municipality
Full size table

A semi-automatic analysis was applied in order to highlight the differences in the relief for the period after a heavy rainfall by comparing the pre-event and post-event images (Fig. 14). The red/orange colour represents a deficit of mass, while the blue colour represents a surplus of mass (accumulation of transported material). Since the resolution of the Landsat 8 images was 30 m, only instabilities longer/wider than 30 m were recognized. White zones in the image represent a mask for clouds and shadows. These results were burdened with undesired interferences, such as alluvial fans and agricultural zones, which were shown as surplus/deficit masses. Given the spectral signature of the target classes (Fig. 10), the minimum distance algorithm had difficulties coping with the classifications in spectral overlap domains, especially for cultivated land, vegetation and bare land.

Fig. 14
figure 14

Difference of PCA on pre-event and post-event Landsat 8 images for Krupanj municipality (red/orange colour represents deficit of mass, blue colour represents surplus of mass)

Full size image

As a result of the semi-automatic classification and reclassification (explained in “Semi-automatic detection of landslides for the test area” section), 490 instances of mass differences were identified. Those polygons were interpreted as landslides without a more precise determination of the mechanism of transport (slides/flows). After their interpretation, it was concluded that most of the identified instances were positioned on slopes with values of approximately 20° and were close to or within gullies.

The total number of recognized landslides in the Krupanj municipality based on the different source data and applied methodology of inventorying is shown in Table 5.

The visual satellite image analysis, supervised classifications and field observations were cross-compared (Fig. 15). Since the lowest common resolution for these three sources was 30 m (Landsat 8 image resolution), only landslides larger than 900 m2 in the area were considered. The field data were loosely considered to be the ground truth but were mostly used as a reference. The reason for this lies in the fact that the flows were recognized with much higher certainty in the satellite images than the slides (i.e. flows were preferred via image analysis), while one of the focuses of the field work was mapping of slides or other types that could not be clearly identified from satellite images (i.e. slides were preferred via field work). The cross-comparison for Krupanj shows some matching between the visual interpretation and the ground truth, i.e. an approximately 26% match, especially for the larger slides. A similar match of approximately 20% was noted between the semi-automatic analysis and ground truth. A cross-comparison between the visual and semi-automatic analyses gives a better match, with an overlap of 41% for larger slides.

Fig. 15
figure 15

Interpretation of instabilities based on three types of data (visual analysis of VHR satellite images, yellow; field observation data, blue; semi-automatic classification of Landsat 8 data, red)

Full size image

Considering the number of identified instabilities and the resolution of the available data, the visual interpretation provided reliable results, especially for the mapping of inaccessible flows. Field data (160 landslides and debris/earth flows) cover only accessible terrain, which is the reason behind the relatively small number of detected instabilities in this dataset compared to the number found via the remote sensing method (507 landslides and flows). The second reason was that the field campaign data were concentrated at predefined locations. These locations were chosen based on complaints obtained by the local administration, wherein damage of objects or infrastructure was cited. Processing and inventorying these complaints were necessary for the subsequent financial restitution from the government. The semi-automatic analysis provided results (490 cases of instability) very similar to those of the visual image analysis but lacked the ability to discern between slides and flows.

Conclusion

This research considered the aftermath of a massive landslide event that occurred in May 2014 in Serbia. This study included the 23 municipalities heavily affected by the landslides, as well as other rainfall-induced calamities. The research objectives focused on two aspects. Firstly, a practical aspect was to provide a reliable landslide inventory that will become a useful support for subsequent field investigations, as well as to observe landslides in inaccessible terrains. Second, a semi-automated analysis was applied over coarser (non-HR/VHR) resolution images that are globally available (Landsat 8 images) to see whether this approach provides any useful information. The idea behind this approach was also a practical one since it entails the use of widely available images to aid field campaigns or directly generates preliminary landslide inventories in a relatively short timeframe, at practically no expense. Given that VHR/HR images are costly and do not have continuous archives (requiring on-demand collection), generating a semi-automated landslide inventory from coarse resolution but globally available images could attract much attention for critical situations or recovery actions, such as this one.

The results highlighted that the remote sensing-aided approach identified a much higher density of landslides. For instance, the field-based landslide inventory of Krupanj has three times fewer identified landslide occurrences than the visually compiled inventory from the VHR satellite images or the semi-automatically compiled inventory. There is a similar ratio in all other heavily affected municipalities (especially those affected by flows in central and western Serbia) that had sufficient VHR Pléiades satellite image coverage. It is further shown that semi-automated and visual analyses can have a relatively good overlap (approximately 40%) even with different sources, which means that the landslide recognition criteria were well tuned in both analyses. The poorer results of the field data are, however, a phenomenological and practical matter. Namely, the field work was first conducted to recheck the preliminary inventories provided by the visual and semi-automated analyses, but, following this, the focus of the field campaign was switched to detect slides (confidently recognized only in situ) that had been reported as damaging or threatening by the local authorities of each municipality.

A successful field campaign that resulted in 1785 mapped landslides within 5 months of limited operational activity proves the efficiency of the presented approach and its role in planning detailed research. Reliable inventories, both preliminary and final, are essential for the quality of all subsequent analyses. It is particularly important for event-based inventories because they provide temporal support that can be put to use in a landslide hazard context and in spatiotemporal landslide frequency analysis if enough event-based time splits are collected in the future.