Introduction

Landslide research chiefly relies on digital inventories for a multitude of spatial, temporal, and/or process analyses (Damm et al. 2010; Neuhäuser et al. 2012; Van Den Eeckhaut and Hervás 2012). Generally, landslide inventories are populated with data gathered by two strategies that are applied either alone or in combination (Guzzetti et al. 2012). These two strategies are: (i) morphological examination, including field surveys, analysis of remote sensing products, topographic or geological maps (Carrara and Merenda 1976; Ayenew and Barbieri 2005; Meinhardt et al. 2015), and, (ii) archival work, i.e., reviewing of private and public archives for textual information in source types like scientific publications, reports from varying entities (e.g., police, fire department, road construction office, or private contractors), as well as newspaper articles (Guzzetti et al. 1994; Liu et al. 2013; Valenzuela et al. 2017).

Strictly speaking, strategy (i) must be applied before its results can be brought in textual form, and only thus enables strategy (ii). Therefore, in case a landslide inventory is based on sources collected with strategy (ii), the preceding morphological examination is outsourced, which in turn usually makes it a more time and cost-effective approach (Carrara et al. 2003; Klose et al. 2015). However, the preliminary examination may be a lay examination performed by people who lack expertise or incentives for gathering in-depth landslide information, e.g., a newspaper journalist who writes an article about a road block due to a landslide event may be more interested in the blockage rather than the event. Therefore, the question arises whether textual information from a specific source type is generally detailed and comprehensive enough to be considered for landslide inventories and their scientific use cases.

To date, no method exists in the literature that quantifies the usefulness of source types, which provide textual information for landslide inventories. In general, landslide data quality studies concentrate on overall data quality of inventories but lack which source types provide useful information Trigila et al. (2010); Schlögel et al. (2015); Piacentini et al. (2018). Studies that provide an assessment of source types are inconsistent in their methods among each other and exclusively report qualitative results Calcaterra and Parise (2001); Liu et al. (2013); Raska et al. (2015); Calvello and Pecoraro (2018). Additionally, assessments of source types depend on the landslide parameters they are tested against. For example, a specific source type might be useful to acquire a landslide’s location and magnitude but otherwise provides no further landslide information. This makes usefulness a subjective measure regarding information requirements.

The overarching goal of the present work is to provide a comprehensive method for quantitative quality assessment of source types regarding their textual information on landslides. The quantitative measure shall represent the probability that landslide information from a single source type or a combination of those is useful for landslide inventories. Furthermore, the method shall be applicable for sources and parameters of any number and type to enable their comparability.

A further goal of this work is to present the most useful source types regarding three landslide parameters: location, date, and process type. These parameters are considered as prerequisites for a multitude of landslide analyses and are commonly requested for landslide inventories (Aleotti and Chowdhury 1999; Battistini et al. 2017; Reichenbach et al. 2018). In order to reach this goal and to illustrate the proposed method, a dataset from the German landslide database is examined. Moreover, various combinations of the parameters are tested against the dataset, and for each case the most useful source type combination is reported.

Material and methods

The material to be examined is a dataset on landslides provided by the German landslide database (cf. Damm and Klose 2015). The database is organized as a flat file and mainly collects information on landslides from various textual sources, including results from scientific publications of field surveys (cf. Bibus and Terhorst 2001; Hardenbicker and Grunert 2001; Jäger et al. 2018; Klose et al. 2014; Schmidt and Beyer 2003; Von der Heyden 2004) and automated digital data acquisition (cf. Kreuzer and Damm 2020). Moreover, information recorded in the dataset that was found in textual sources was manually validated and, if necessary, corrected in case the author of the text was not an expert—this is generally true for all source types except for scientific publications and expert opinion pieces. The validation and the potentially, following correction process included field surveys, remote sensing analyses, and cross checking with attached photographs. For example, it could be deduced from photos of a news article that the colloquially reported slide was in fact a fall process in this case. Thus, the present study works under the assumption that the information from the database is correct. The timespan for events in the databases ranges from prehistoric to current events. A detailed description of the database’s structure is given in Damm and Klose (2015). For the analysis of the present work, the focus is on information for the following landslide parameters found in the dataset: (a) location, (b) date, (c) process type, and (d) type of source where the aforementioned information was found.

Items (a), (b), and (c) are thereby of varying degree of detail. The detail of the location is specified by three spatial confidence descriptors modified after Calvello and Pecoraro (2018) that describe the accuracy of the position: less than 100m (Sd1); less than 1km (Sd2); equal or greater than 1km (Sd3). The date’s detail is its specified exactness as either of day, month, year, or historic/prehistoric. The landslide’s process type is generally differentiated between “falls” and “slides” (Cruden and Varnes 1996; Hungr et al. 2014; Varnes 1978), and also contains additional details (e.g., grain size of a fall process or depth of a slide’s slip surface) in some cases. Thus, parameter (a) has 3 classes of detail degree (Sd1, Sd2, Sd3), (b) has 4 (day, month, year, historic/prehistoric), and (c) has 2 (general, detailed). To further illustrate this, the following is a practical example text of a news article used for the German database:

Country road 6924 was blocked between Börstingen and Eckenweiler [German villages] due to last week’s landslide. [...] The ground was soaked wet like a sponge from rain and melt water, thus the slope beneath the road started to slide. (Fuchs 2018)

From the text above, the following landslide information can be obtained: (i) the location of the event in form of a roughly 5-km-long road section, (ii) the week-exact date of occurrence relative to the news article’s publishing date (“last week”), and (iii) the general information of the landslide process, in this case a slide.

Item (d) describes a total of seven different source types from which the dataset is compiled: administrative document, archive record, compilation, expert opinion, mission report, news article, and scientific publication. The source types with examples of their content providers are listed in Table 1. No distinction was made between digital and analog forms of source types and their respective acquisition methods. Moreover, the database’s structure is designed in such a way that one record has one source type, i.e., landslide events found in multiple source types consist of multiple records. A shared identifier for all records belonging to the same event prevents confusing multiple records with multiple events. This structure is functional for source-centric analyses and necessary due to the flat file nature of the database (cf. Fig. 1).

Table 1 Source types from the German landslide database and examples of their respective content providers
Full size table
Fig. 1
figure 1

Excerpt of the database schema from the used dataset. For illustration purposes, only the date is shown as a landslide parameter and three special cases are coloured: blue marks a single landslide event that is reported in two different sources of the same type (newspaper article); red marks the same landslide event as blue but also reported from another source type (administrative document); green marks two different landslide events reported by the same source (scientific publication). Blue and red also show how different sources may provide different degrees of parameter detail for the same event

Full size image

The present work proposes the following method to quantify the usefulness of source types according to their content on landslide parameters in three degrees: the first degree describes the usefulness of a single source type regarding a single parameter, the second degree regarding a single source type with arbitrary number of parameters, and the third degree regarding an arbitrary number of source types and parameters.

Initially, preference for a parameter’s detail class is described by a so-called utility function. Such utility functions commonly describe risk-aversion through exponential relations Choi and Ruszczynski (2011); Pratt (1964). In this case, an exponential utility function accounts for the risk of not finding any useful information.

Definition 1

For any parameter (e.g., landslide location) with n detail classes, let \(x_i\) be the ith-detailed class ranked \(x_0 \succ x_1 \succ \ldots \succ x_{n-1}\), i.e., \(x_0\) has the most detail and \(x_{n-1}\) the least. The proposed utility function \(u(x_i)\) that quantifies preference for this detail class is defined as:

$$\begin{aligned} u_i := u(x_i) = e^{-i} \end{aligned}$$
(1)

with \(1 = u_0> u_1> \ldots> u_n > 0\) for \(i = 0, 1, \ldots n-1\).

In prose, Eq. 1 means the utility (or preference) of a parameter decreases exponentially with every step towards a lesser detail class. This assumes that \(x_0\) is the desired degree of detail and any deviation to less detail is an undesired compromise.

Furthermore, for every \(x_i\) there exists a \(p_i\) that gives the probability to find detail class \(x_i\) in the respective source type. Since probability is synonymous to expected frequency of occurrence, probability \(p_i\) is deduced per source type from the share of sources that contain detail class \(x_i\).

Definition 2

The single parameter usefulness U of one source type regarding one specific parameter is defined as the probability to find useful information on that parameter in that source type. This probability can be calculated with an “anticipated utility function” originally described by Quiggin (1985):

$$\begin{aligned} U(\mathbf {p}) = p_0 u_0 + \sum ^{n-1}_{i=1}\left( \sum ^i_{j=0}p_j-\sum ^{i-1}_{k=0}p_k\right) u_i \end{aligned}$$
(2)

with \(\mathbf {p} = \{p_0, p_1, \ldots , p_{n-1}\}\) as the probability vector for detail classes \(x_0, x_1, \ldots , x_{n-1}\) of the specified parameter. The probabilities of \(\mathbf {p}\) are mutually exclusive, thus, \(\sum ^{n-1}_{i=0} p_i \le 1\) and \(0 \le U \le 1\).

Generally, the “anticipated utility function” is a probability weighting function that uses a utility function as weights (Koida 2018). It can be seen from Eqs. 1 and 2 that if a source type contains only the highest detail class \(x_0\) it only has one corresponding probability \(p_0\) and therefore \(U(\mathbf {p}) = p_0\). This normalization results from the definition of Eq. 1 where for any parameter \(u_0 = 1\).

A short example is given to make the described single parameter usefulness from Definition 2 more intuitive. In this example, the source type “news article” is examined about its usefulness regarding the parameter “location”. As mentioned above, the location’s detail is differentiated in Sd1, Sd2, and Sd3 with \(x_0 \equiv \text {Sd1} \succ x_1 \equiv \text {Sd2} \succ x_2 \equiv \text {Sd3}\). In a devised case of \(\mathbf {p} = \{0.5, 0, 0.4\}\) it follows from Eqs. 1 and 2 that \(U = 0.55\). Thus, the usefulness, i.e., the probability to find useful location information in a news article, is 55%. This is more than just \(p_0 = 0.5\) but less than the combined probability to find any location information in news articles, which is 90% (\(\sum ^{n-1}_{i=0} p_i = 0.9\)). In another devised case of \(\mathbf {p} = \{0.5,0.4,0\}\) it follows that \(U = 0.65\), this illustrates the (exponential) influence of higher detail classes for the single parameter usefulness.

Since usefulness is equated to probability (Definition 2), combinations of usefulnesses follow the same mathematical rules as combinations of probabilities Baldi (2017). Thus, the overall usefulness \(\hat{U}\) of one source type is calculated by intersection of all single parameter usefulness’s \(U_k\) (Eq. 2), with \(k = 0, 1, \ldots M-1\) in case of M analysed parameters. Assuming that all \(U_k\) are independent of each other, e.g., the occurrence of a location-parameter does not influence the occurrence of a date-parameter, the intersection is calculated with:

$$\begin{aligned} \hat{U}\left( \bigcap ^M_k U_k\right) = \prod ^M_k U_k \end{aligned}$$
(3)

This corresponds to the probability to find useful information for all analysed parameters in one source type.

Furthermore, Eq. 3 is extended with DeMorgan’s Theorem to allow for combination of different source types to increase the probability to find useful information for landslides. This combined usefulness \(\tilde{U}\) for N source types is calculated with:

$$\begin{aligned} \tilde{U}\left( \bigcap ^M_k\left[ \bigcup ^N_{l} U_{l,k}\right] \right) = \prod ^M_k \left[ 1 - \prod ^N_l (1-U_{l,k})\right] \end{aligned}$$
(4)

Here \(U_{l,k}\) is the single parameter usefulness of the kth-Parameter for the lth-Source type. This corresponds to the probability to find useful information on every examined parameter in, at least, one of the combined source types. Thus, for \(N = 1 \Rightarrow \hat{U} = \tilde{U}\).

The examination of the given dataset’s source types is performed with a customised analysis software written in Python, organized into four different parameter cases. All cases demand the location parameter because it is a requirement for any spatial analysis. The cases are:

  • Case 1 \(\rightarrow\) Location

  • Case 2 \(\rightarrow\) Location and Date

  • Case 3 \(\rightarrow\) Location and Process type

  • Case 4 \(\rightarrow\) Location and Date and Process type

For all four cases, the overall usefulness (Eq. 3) for specific source types is examined, as well as all additional combinations of an increasing number of source types (Eq. 4). This means, first one source type is examined, then all combinations of two source types, then of three, and so forth—the corresponding mathematical relation is determined with a least-square fit of the combinations’ mean values (Guest and Guest 2012). With this analysis, the number of required source types to reach high possibilities to find useful information is determined. In this study, possibilities of greater than 75%, i.e., within the fourth quartile, are considered as high.

Results

All in all, the examined dataset contains 4856 records that describe 4658 events reported by 4556 sources (including multiple mentions, cf. Fig. 1), together with location, date, and process type (Table 2). On average, sources of all source types usually report more than one landslide event (ratio \(> 1\)) with the notable exception of news articles.

Table 2 Distribution of sources and landslide events (including multiple mentions, cf. Fig. 1) for each source type in the dataset from the German landslide database, with the ratio of the number of events to number of sources
Full size table

The three parameters’ (location, date, process type) observed probabilities for their respective detail classes are shown in Fig. 2, differentiated after source type. Notably, scientific publications are the only source type with historic/prehistoric landslides and are the penultimate source type for detailed process type descriptions. In contrast, news articles provide the most sources with detailed process type descriptions. For clarification, it is coincidental that all three parameters have a 100% probability to occur in any source type.

Fig. 2
figure 2

Observed probabilities, i.e., frequency of occurrence, for different degrees of detail (colouring) differentiated after examined source types for parameter (a) Location, (b) Date, and (c) Process type. Here, the maximum of the y-axis is 100%

Full size image

With the probability results plotted in Fig. 2, the single parameter usefulness is calculated (see Fig. 3). Here, scientific publications have the highest probability to find a useful location, and news articles have the highest probability to find a useful date and process type. The average usefulness across all source types is 0.74 for location, 0.38 for date, and 0.55 for process type. “Compilation” is the only source type that is below average for all three parameters and “expert opinion” is the only source type that is above average for all three parameters.

Fig. 3
figure 3

Usefulness of source types for specific landslide parameters after Eq. 2. Here, the single parameter usefulness of location is equivalent to the overall usefulness for Case 1

Full size image

The overall usefulness, i.e., case-by-case intersections of single parameter usefulnesses (Fig. 3), is listed in Fig. 4 for Cases 2-4. Since the overall usefulness is equivalent to the single parameter usefulness for Case 1, it is listed in Fig. 3. Expert opinion (Case 3), news article (Case 2 and 4), and scientific publication (Case 1) display the highest probabilities to contain useful information. Statistically, the overall usefulness is on average 0.74 for Case 1 (see above), 0.29 for Case 2, 0.40 for Case 3, and 0.16 for Case 4.

Fig. 4
figure 4

The examined source type’s overall usefulness after Eq. 3 that correspond to three of the four derived cases: Case 2 (Location and Date), Case 3 (Location and Process type), and Case 4 (Location and Date and Process type)

Full size image

A higher usefulness can be achieved through combination of source types. In Fig. 5, the statistical description (minimum, maximum, mean, and standard deviation band) of all possible combinations for different numbers of source types is shown in a modified box plot. Furthermore, the least-square fit proved a logarithmic relation for the increase of the combined usefulness with the number of combined source types N, i.e., \(\tilde{U} = a \ln (N) + b\). Here, the intercept b corresponds to the initial value of usefulness for 1 source type, and the gradient a represents the logarithmic growth rate for every additional source type (Fig. 5).

Fig. 5
figure 5

Combined usefulness’s (Eq. 4) for (a) Case 1 (Location), (b) Case 2 (Location and Date), (c) Case 3 (Location and Process type), and (d) Case 4 (Location and Date and Process type). The x-axis shows the number of sources that are combined from the pool of all seven sources. The modified box plot corresponding to a specific number of sources shows the minimum and maximum (lower and upper end of line), mean (horizontal bar), and standard deviation band (box) of all possible usefulness-combinations for the given number of sources; the dashed line shows the logarithmic relation

Full size image

The highest combined usefulness for each case and number of combinations seen in Fig. 5 is always the combination of the respective number of source types with the highest overall usefulness for the desired case (cf. Figs. 3 and 4). Thus, the minimum number of source types that need to be combined to get a high usefulness (\(\ge 0.75\)) can be deduced from Fig. 5 when looking at the maximum value bar of the respective box plot and the corresponding value on the x-Axis. After Fig. 5, Case 1 requires one, Cases 2 and 3 require two, and Case 4 three source types to be combined to achieve high usefulness. In Fig. 6 the resulting combined usefulness for each case after combining the minimum number of source types with the highest overall usefulness can be seen. Overall, the results for all four cases are achieved with four source types in total: administrative document, expert opinion, news article, and scientific publication.

Fig. 6
figure 6

Source types with the highest overall usefulness (\(\hat{U}\)) in the respective cases, and thus produce the least number of required source types to obtain a high combined usefulness (\(\tilde{U} \ge 0.75\))

Full size image

Discussion

The present work proposes a method to quantify the usefulness (probability to find useful information) of source types in three degrees:

  1. 1.

    single parameter usefulness (one source type, one parameter, cp. Eq. 2),

  2. 2.

    overall usefulness (one source type, arbitrary number of parameters, cp. Eq. 3), and

  3. 3.

    combined usefulness (arbitrary number of source types and parameters, cp. Eq. 4).

All degrees of usefulness depend on the utility of the examined parameter(s), thus representation of utility is paramount for the method. As mentioned above, the exponential relation is an established method to account for risk-aversion (Choi and Ruszczynski 2011; Pratt 1964). In this case, it is the risk of not finding any useful information. Additionally, the proposed function (Eq. 2) was conceived with an in-built normalization of \(u_0 = 1\) and no additional parameters that need to be determined, thereby making it only dependent on the number of detail classes. Furthermore, the exponential relation was chosen to allow for a strong distinction of detail classes. In case weaker distinctions are required, the detail classes can be adopted accordingly through reallocation of preferred information, either with or without changing the number of detail classes.

It is emphasised that usefulness is a relative concept, however the proposed method adapts its outcome to the user defined detail classes and thus meets the needs of any particular information requirements or preferences. Nevertheless, the results of the present work are produced under the assumption that the more detail an information contains, the more useful it is. The detail classes of the examined material resulted from the given data structure and were not altered. Here, the lowest detail classes from all examined parameters define a wide quality range that includes any parameter information. For example, Sd3 from “location” describes an accuracy of equal or greater than 1km; this means any location information can at least be Sd3. This kind of inclusive detail class is no prerequisite for the method and, in this case, adds to the occurrence of 100% chances in Fig. 2; it is, however, comparable to other works that also define inclusive detail classes (cf. Battistini et al. 2013; Calvello and Pecoraro 2018). In contrast to the inclusive classes for least detail, the used classes for highest detail in this work may not fulfill requirements for all landslides research. For example, the “date” parameter is at best day-exact, this may not be detailed enough to deduce rainfall thresholds, in this case hour-exact information may be necessary Scavia et al. (2020). Another property of the examined material is the similarity of the number of events for each source type, except for scientific publications and compilations (Table 2)—in this respect it is recalled that scientific publications include more than detailed research papers but also maps depicting multiple landslides events and compilations still have a large enough sample size for statistical analysis. Such a similar number of events is relevant to give a comparable and statistical significant estimation on the frequency of occurrence for detailed landslide parameters. In this context, compilations might be underrepresented and findings for scientific publications are likely to be more robust. The calculated outlier-ratio of 0.51 for news articles (ibid.) indicates a strong presence of duplicates in this source type (Kreuzer and Damm 2020). This does not affect frequency of occurrence, it rather shows that the recent overall tempo-spatial frequency of “news-worthy” landslides in Germany is usually low enough to be reported on individually by multiple newspaper articles.

On this basis, the frequency of occurrence for information quality in scientific publications was found to be strongly heterogeneous in this study. On one hand, they provide a location information with the best quality “Sd1” for more than 87% of all events (Fig. 2a). On the other hand, it is the only source type that contains the weakest detail class “historic/prehistoric” for date, and this in 70% of all cases (Fig. 2b). Presumably, this disparity is due to lack of bias in scientific research towards contemporary landslide events, which are less numerous than historic/prehistoric landslides that amassed over time. In contrast, it is likely that all other source types have a bias towards contemporary landslide events as exemplified by the high frequency (68%, Fig. 2b) of day-exact dates in news articles. Additionally, news articles contain the highest frequency of detailed process type descriptions (Fig. 2c), which reinforces the conclusion that they are mainly concerned with the impact of landslide events. To revive the example from the introduction: in case of a road blockage it is important to know the extent of what blocks the road (process type details) and when the road was closed (date of occurrence) in order to have a hint for the duration of the blockage. Conversely, scientific publications do not provide many relevant process descriptions. This is also most likely due to many scientific publications concerning themselves with prehistoric landslide events (cf. Fig. 2b) from which the process is difficult to discern due to continued erosion. Moreover, most scientific publications of the examined material report landslides as a result derived from topographical and geological maps and remote sensing products rather than detailed field surveys (cf. Bibus and Terhorst 2001; Klose et al. 2014; Hardenbicker and Grunert 2001; Von der Heyden 2004). However, such mere visual data acquisitions are generally not capable of identifying process types reliably, e.g., vegetation may obscure visual landslide markers Zhao and Lu (2018). Lastly, in the present analysis, the chosen detail classes for process types do not demand sophisticated interpretations often found in scientific works, thus enabling news articles to compete. From this perspective, finer grained detail classes may reveal a stronger bias in favor of scientific publications.

The aforementioned particularities of scientific publications and news articles are correctly reflected in the single parameter usefulness seen in Fig. 3. Here, scientific publications are the most useful source type for location information, and news articles for date and process type. Moreover, the single parameter usefulness shows expert opinions are above, and compilations below, the average of all source types. Of course, it is expected for expert opinions to deliver detailed information, however, compilations seem to suffer from being a third-hand source type. This can also be stated about archive records, which is the only source type, next to scientific publications, that is below average in two out of the three examined parameters.

In this context, numbers for the overall usefulness seen in Fig. 4 (and Fig. 3 for Case 1) give the consequent qualitative result: scientific publication, expert opinion, and news article are most useful for the four devised cases. Quantitatively, however, the numbers are roughly halved with every additional parameter requirement, i.e., one parameter has an average usefulness of 0.74 (Case 1), two parameters 0.34 (mean of Case 2 and 3), and three parameters 0.16 (Case 4). This means, in case source types are to be examined regarding even more parameters, a single source type is unlikely to produce useful information on all of them.

In contrast to the overall usefulness’s decline with every additional parameter, the opposite is true for the combined usefulness: the more parameters are used the higher the logarithmic growth rate (Fig. 5). Consequently, the more parameters are used the stronger the increase in usefulness for every additional source type combination, this is a weak counterbalance: linear decline vs. logarithmic growth. It follows, any multi-parameter research for landslide information benefits from searching multiple source types. Furthermore, due to the logarithmic behaviour, high values for usefulness could already be achieved with the combination of 2 or 3 source types in this study. Additionally, the range of usefulnesses is wider the less source types are used and therefore makes it more important to select the most productive source types to get a high usefulness. For example, the combination of the two most useful source types for Case 2 have roughly the same usefulness as the average combination of three source types (Fig. 5b). Moreover, the highest combined usefulness is always obtained if source types with the highest overall usefulness are used, and analogously, this is also the case for the lowest combined usefulness.

The final result (Fig. 6), which specifies the most useful combination of source types for the devised cases, is a direct consequence of the probability to find detailed landslide information in the examined material. It is thus, first and foremost, a reflection of detail in German landslide documents and their particularities. For example, while administrative documents are part of the best source types in this study, Calcaterra and Parise (2001) describe them as “most frustrating” for landslide research in the Calabria region, Italy. In contrast, the presence/absence of detailed location/date information in scientific publications, and the reverse in news articles, is generally agreed upon by international publications Calcaterra and Parise (2001); Carrara et al. (2003); Liu et al. (2013); Raska et al. (2015), as well as by this study.

Conclusions

The present work proposes a method to assess a given source type—or a combination of those—on the basis of quality of textual landslide information. Therein, the quantified usefulness describes the probability a researcher is satisfied with the detail of the found information. Moreover, probability is an intuitive measure that is suitable for a wide range of applicants, whose a priori knowledge of source type usefulness helps to use fewer source types during research. More specifically, researchers can adapt the used source types to the information they seek. In case it is not practical to limit research to specific source types, the reverse is also possible: the parameter requirements can be adapted to the source types available. Furthermore, the proposed method is not limited to textual sources on landslides; in theory, the method can be applied to parameters and source types of any natural hazard.

Even though the results presented in this work illustrate the proposed method, they are valid results particular to German documents with content on landslides. Future works should analyse material from different regions and place them into a broader context. Furthermore, there are many more source types and parameters that should be analysed to acquire comprehensive knowledge of textual information on landslides. In this context, social media posts may be of particular interest due to its intensified use over recent years. This way, the increase of effectiveness from the proposed method could encourage the time-consuming and difficult task of creating or extending landslide inventories in more regions of the world.