1. Introduction
Landslide susceptibility assessment and mapping are essential tools for landslide risk management and serve to support decision-makers in developing territorial planning strategies, risk mitigation measures, and the implementation of monitoring and warning systems [
1,
2,
3,
4].
Landslide susceptibility refers to the probability of landslides occurring in an area based on local terrain conditions [
5], considering morphological, geological, land use, land cover, and local environmental and climatic characteristics, thereby providing an estimate of where such events are most likely to occur [
6]. Note that landslide susceptibility does not consider the temporal occurrence or magnitude of expected landslides [
7].
Several predictive models have been developed to assess and map landslide susceptibility, considering factors such as scale, study objectives, modeling approaches, and evaluation criteria [
8,
9]. These models include both qualitative and quantitative methods [
10]. Qualitative methods typically provide a zonation of landslide susceptibility using weighted indices and relative categories (e.g., low, medium, high), which are suitable for local-scale and site-specific studies. In contrast, quantitative methods offer numerical estimates of landslide occurrence.
According to the United States Geological Survey (USGS), a landslide is defined as the downward movement of earth materials (such as rocks, debris, and soil) at velocities ranging from millimeters per year to meters per hour [
11]. These landslides may be triggered by various natural factors, such as slope, geological faults, heavy rainfall, and earthquakes, or result from human activities [
12], causing significant impacts in the affected areas.
Landslides vary in their characteristics, and understanding their causes and mechanisms is essential for effective mitigation [
13]. In this context, the inventory of critical sites or event occurrence points, combined with the assessment of the resulting damage, plays a crucial role in landslide analysis [
14,
15]. These inventories also contribute to a more comprehensive spatio-temporal analysis of landslides; consequently, zoning—which is defined as the delineation of geographic areas where specific events occur—enables the identification of regions susceptible to landslides due to the interaction of various land instability factors [
16].
Landslide Susceptibility Zoning classifies the territory based on the probability of landslide occurrence, utilizing different methods (including direct, heuristic, statistical, and process-based models) and tools [
7]. International guidelines for landslide susceptibility zoning recommend that landslide mapping should include the following components: (a) landslide inventory mapping, which indicates the points or areas identified as having failed due to landslide processes; and (b) landslide susceptibility zoning, which involves the classification, extent, and spatial distribution of existing and potential landslides in the study area [
3,
17].
Landslide susceptibility mapping begins with the creation of a landslide inventory map [
18]. Landslide inventory maps provide information on the location and distribution of landslides that have left noticeable features in an area [
7]. Most landslide inventory maps are generated by visually integrating the interpretation of one or more sets of satellite imagery or aerial photographs, along with varying degrees of direct on-site observations. An inventory is essential for preparing a susceptibility map [
19]. In addition, successful mapping requires the identification of optimal causal factors, which are generally selected based on an analysis of the types of landslides and the characteristics of the study area [
20]. Consequently, landslide occurrence is influenced by the interaction of topographic, hydrologic, and geologic factors [
21,
22].
Identifying the factors that influence landslide occurrence is a complex task due to the lack of standardized criteria for selecting which factors to include or exclude [
23]. The causal factors used for assessing and zoning landslide susceptibility are determined based on a comprehensive literature review and detailed field observations [
18,
24]. These causal factors are often effectively derived from satellite imagery and Digital Elevation Models (DEMs) [
25]. Based on this, thirteen landslide causal factors were selected for this study: elevation, slope, aspect, flow length, flow accumulation, insolation, Sen2p, distance to faults, distance to folds, distance to rivers, ground cover, geology, and geomorphology.
As noted previously, various methods have been employed for landslide susceptibility assessment and zoning, typically focusing on empirical, statistical, and deterministic approaches [
26,
27]. Statistical approaches are the most commonly used for both landslide susceptibility and hazard zoning [
8,
28], although they have a general limitation: they do not propose mechanisms for controlling slope failure. Instead, it is assumed that the prediction of future landslide areas can be achieved by evaluating combinations of variables that contributed to past landslide occurrences [
29]. The Weight of Evidence method (WoE) is widely used to predict areas that are potentially susceptible to landslides by employing the Bayesian probability model in a log-linear form. In this approach, each factor is linearly superimposed in the GIS environment based on its independent characteristics [
30,
31].
The WoE method calculates the weight of each landslide causal factor based on the presence or absence of landslides within the area. Its fundamental assumption is that future landslides will occur under conditions similar to those that contributed to past landslides [
32].
By overlaying the landslide inventory map with each causal factor map, spatial statistical relationships can be derived to quantify the impact of various landslide-causing factors by assigning them positive or negative weights. This process evaluates the potential for future landslides based on the presence or absence of each causal factor class, Ei, using a pair of likelihood ratios as outlined in Equation (1) [
33].
For calculating the weight of each causal factor contributing to landslide occurrence,
Equations (3) and (4) have been employed [
34]:
In these equations, Npix1 denotes the number of pixels where both landslides and landslide-contributing factors are present; Npix2 represents the number of pixels where landslides are present but the contributing factors are absent. Conversely, Npix3 indicates the presence of landslide-contributing factors without a landslide, and Npix4 signifies the absence of both landslides and contributing factors.
The final weight, expressed as Wc, is calculated using Equation (4):
Here, W
c is defined as the difference between W
+ and W
−. This formulation refines
the spatial relationship between all contributing factors and landslides.
where
i is the class of a factor whose values vary from 1 to
n.
Another method that employs a bivariate statistical approach—widely applied by numerous authors [
28,
30,
35,
36,
37,
38] is the Information Value (IV) method. Based on information theory, the IV method measures the amount of information carried by an event [
39] and is primarily used as a statistical predictive approach in environmental geological research, particularly for spatial predictions of landslides and slope stability [
40]. It converts measured data of landslide triggers, such as elevation and slope, into information values that reflect landslide susceptibility, which can then be used to determine potential landslide occurrence areas [
41]. Based on the presence or absence of the classes of causal factors within past landslides, information values can be determined [
42].
where
Nslpix is a number of landslide pixels in a given class,
Ncpix is the number of pixels in a given class,
Ntspix is a total number of landslide pixels in the study area, and
Ntapix is a total number of pixels in the entire study area.
Regardless of the zoning method used, it is recommended to utilize descriptors or parameters associated with nominal scales to differentiate the magnitude and intensity of landslides [
3] and to describe the degree of susceptibility to landslides [
43]. When interpreting a landslide susceptibility map, different shades (colors) are employed, with the nomenclature reflecting actual field observations and descriptions of common scenarios [
30] (see
Table 1).
Landslide susceptibility assessment is essential in mountainous regions, where geological and environmental conditions pose significant risks to communities. Bivariate statistical methods such as Weight of Evidence (WoE) and Information Value (IV) have been widely used for their simplicity and efficiency, even in tropical and South American studies. Research such as Rahman et al. [
34] in the Hindu Kush and Achu et al. [
44] in the Western Ghats have demonstrated their ability to generate reliable susceptibility maps from key geographical variables.
However, these methods also have limitations. Several authors have pointed out that bivariate models fail to capture complex or non-linear relationships between causal factors [
45]. For example, Maharani et al. [
46], Getachew and Meten [
24] and Xie et al. [
47] highlight that, while WoE is effective in open data and limited access scenarios, the incorporation of machine learning models could substantially improve predictive capacity by addressing more sophisticated patterns. In the same vein, Batar and Watanabe [
48] propose hybrid approaches that integrate classical statistical methods with modern geospatial analysis tools and artificial intelligence.
Overall, the recent literature suggests that, while WoE and IV remain valuable tools, their true potential may be best harnessed when used as part of a combined approach. This is particularly relevant in regions such as the Cundiboyacense highlands, where geographic complexity and variable data availability require flexible and adaptive modelling strategies.
The purpose of this research is to develop a reliable landslide susceptibility zoning map for the Zipaquirá-Pacho road corridor, an area where emergencies have been reported, due to the presence of various triggering factors that decrease slope stability and increase the risk of landslides, especially in areas with steep slopes. This process will be generated through the application of bivariate statistical approaches, specifically the Weights of Evidence (WoE) method and the Information Value (IV) method, supported by specialized GIS software.
Figure 1 shows the location of the Zipaquirá–Pacho road corridor within the department of Cundinamarca, Colombia, highlighting its geographical relevance and the susceptibility to mass movements that motivate this study
This map is intended to evaluate and establish mitigation strategies for the impacted area. Effective assessment of landslide occurrence in an area that is already failing or susceptible to failure requires the identification of existing and past landslides (inventory), the determination of the contribution of predominant causal factors, and the generation of landslide susceptibility mapping and zoning [
49,
50].
2. Materials and Methods
The selection of an appropriate method for landslide assessment and zoning depends on the objective and scope of the investigation [
33]. In this study, two bivariate statistical models, namely, the Information Value (IV) model and the Weight of Evidence (WoE) model, were used to generate a landslide susceptibility zoning map for the Zipaquirá-Pacho Road corridor. The methodology comprises a workflow that integrates data capture and collection; the identification and determination of causal factors; the creation of a raster geodatabase; normalization of these factors; extraction of weights through WoE analysis and the Information Value method; generation of the susceptibility map; and, finally, evaluation through the proposal of mitigation strategies based on observed events, affected areas, and incidence zones. From these steps, the landslide inventory map and the landslide susceptibility maps are produced, and the model is subsequently validated (see
Figure 2).
The choice of the bivariate methods Weight of Evidence (WoE) and Information Value (IV) is based on their demonstrated effectiveness in modelling landslide susceptibility across diverse geographical contexts, particularly when reliable inventories and categorical geospatial data are available. Moreover, these methods are highly interpretable and require minimal computational resources, making them especially suitable for local-scale applications. Their combined use in this study aims not only to validate their performance in tropical mountainous regions but also to compare their outputs as a decision-making tool for landslide risk management.
Although the WoE and IV methods fall within the same category of bivariate models, their differing approaches to weight calculation reveal important methodological distinctions. WoE estimates the probability of occurrence based on the contrast between the presence and absence of landslides within each subclass, whereas IV computes the relative proportion of occurrences with respect to the total study area. As a result, IV tends to smooth out differences in smaller classes, while WoE can amplify positive or negative weights when sharp contrasts are present. This distinction may significantly influence model interpretation, particularly when working with categorical input layers or underrepresented subclasses.
2.1. Study Area
This study focuses on the Zipaquirá–Pacho road corridor, which connects the municipality of Zipaquirá (5.024107°, −74.014034°)—located 29 km from Bogotá, Colombia’s capital and part of its metropolitan area—with the municipality of Pacho (5.133582°, −74.154024°), situated in the department of Cundinamarca, 88 km from Bogotá and 35.93 km from Zipaquirá. The region where the road corridor is located is an area where emergencies have been reported due to several factors that reduce slope stability and increase the risk of landslides, especially in areas with steep slopes.
The Zipaquirá–Pacho road corridor is located on the eastern slopes of the Eastern Cordillera of Colombia, an area characterised by mountainous terrain with steep slopes resulting from tectonic and erosional activity. Geologically, the region is composed of Cretaceous and Paleogene sedimentary formations, including the Guaduas Formation, as well as Quaternary alluvial deposits. From a geomorphological perspective, the landscape features hills, structural escarpments, alluvial fans, and fluvial terraces—elements that contribute to slope instability processes.
Climatically, the area experiences a temperate-humid climate. In Zipaquirá, the average annual rainfall is approximately 1493 mm, with the highest precipitation recorded in April (255 mm), October (230 mm), and November (266 mm), and the driest month being July (61 mm). In Pacho, the average annual rainfall reaches 1692.63 mm, with November showing the highest monthly average of 266 mm. These climatic conditions, combined with the geological and geomorphological characteristics, significantly contribute to the region’s susceptibility to landslides.
2.2. Landslide Inventory and Mapping
The data source and accuracy directly affect the results of landslide susceptibility evaluation. Primary data were collected through field observations using tools such as GPS and mobile applications with smart forms, which allowed for the collection of location data and the classification of events based on type. Additionally, data from satellite platforms, specifically Google Earth Pro, were utilized.
The recorded events correspond primarily to rotational landslides and debris flows, both commonly observed on slopes composed of unconsolidated or highly weathered materials—particularly in geological units such as shales and clays of the Guaduas Formation. This classification was based on field observations and the morphodynamic interpretation of scarp features identified through satellite imagery, following the criteria proposed by Cruden and Varnes [
51] and later updated by Hungr et al. [
52].
This information can then be used to estimate the frequency and potential severity of future landslides [
53]. In the study area, 20 landslide events were identified through field inspections along the road corridor. These data were further supplemented by 81 points obtained from visual interpretation and digitization of landslides using images available on Google Earth Pro, captured between 2014 and 2022, bringing the total inventory to 101 critical points. The locations of these events were recorded, imported, and analyzed using the specialized software ArcGIS Pro 3.2.0, and are presented as georeferenced points using the Magna Sirgas—National Origin projection.
The inventory and distribution of landslides along the Zipaquirá–Pacho road corridor were developed through visual interpretation of Google Earth Pro images, detailed field assessments, and the implementation of a mobile application that faciliated the capture of both event locations and their typology based on the type of mass movement (landslide). This process resulted in a database with 101 landslide event records collected along the road corridor.
In this study, 71 (70%) of the landslides were used to train landslide susceptibility models, while the remaining 30 (30%) were used for model validation, as illustrated in
Figure 3.
2.3. Causal Factors of Landslides
The identification of these factors is a critical step that significantly influences the application of the model. In this process, 13 conditioning factors were considered, allowing us to determine landslide susceptibility around the road corridor. These factors are based specifically on geo-environmental conditions and the Digital Elevation Model (DEM). They were converted to a raster format with a cell size of 12.5 × 12.5 m. The datasets and their sub-classes are presented in
Table 2.
The slope variable was derived from a 12.5 m spatial resolution Digital Elevation Model (DEM) from the ALOS PALSAR mission. In addition to the standard slope (in degrees), a terrain-transformed slope factor known as Sen2p slope was also calculated. This factor applies a sinusoidal transformation to the slope angle, aiming to better represent how terrain inclination influences the initiation of shallow landslides. The transformation enhances the relevance of mid-to-steep slopes (typically 15–35°), where mass movements are more likely to occur, as supported by Londoño-Linares (2016) [
27]. This transformed slope was then normalized and reclassified into five ranges (0–15°, 15–25°, etc.) according to morphodynamic thresholds observed in field data and previous studies.
The selection of causal factors is directly linked to the failure mechanisms observed in the field. Rotational landslides were frequently identified on steep slopes composed of fine-grained, clay-rich materials, supporting the inclusion of variables such as slope and lithology. On the other hand, debris flows were more commonly associated with areas near roads and streams, where runoff concentration and surface disturbance are significant. Therefore, flow accumulation, land cover, and distance to roads were incorporated as key factors. Although distance to geological faults was considered, its influence is likely indirect, possibly associated with structural discontinuities or inherited terrain weaknesses.
These conditioning factors are often effectively derived from satellite images and the DEM [
23]. In this study, a DEM derived from Alos Palsar—one of the many cartographic resources available from the Japanese Aerospace Exploration Agency’s ALOS satellite—with a spatial resolution of 12.5 m was utilized.
Although hydrometeorological variables such as cumulative precipitation are widely recognised as key drivers of mass movements, they were excluded from this study due to the lack of accurate and high-resolution climate data for the study area. Instead, indirect variables closely related to water dynamics were incorporated, such as flow accumulation and insolation, which serve as proxies for runoff and soil saturation. It is important to note, however, that this represents a limitation of the present study, and future research is expected to address it in greater depth.
The selection of causal factors was informed by field observations, a review of specialized literature, and their direct relationship with the landslide mechanisms identified along the corridor, such as rotational slides and debris flows. Variables such as slope, land cover type, and flow accumulation are closely linked to the onset of instability caused by overloading, loss of shear strength, and soil saturation—typical conditions on disturbed or highly weathered slopes. This connection between the physical processes and the modelled factors is intended to ensure a coherent and physically meaningful representation of the phenomenon.
2.4. Raster Geodatabase
The construction of geodatabases is a fundamental element in landslide susceptibility mapping [
50]. As part of this process, a database was created encompassing all causal factors such as slope, aspect, curvature, and land cover (see
Figure 3), along with field data collected via a mobile application supported by Survey 123 (ArcGIS Sur-vey123 is part of Esri’s geospatial cloud, based on smart XLSForm-based forms). This resulted in a landslide inventory map within the ArcGIS Pro GIS environment. The in-formation layers were referenced with the same projection system (Magna Sirgas, National Origin projection) and pixel size (12.5 × 12.5 m) for the corresponding data analysis (see
Figure 4).
3. Results
Using both methods, namely, the Weights of Evidence (WoE) method and the Information Value (IV) method, weight values were determined for each subclass, indicating their influence on landslide generation in the study area. Thirteen causal factors were considered for predicting and generating the zoning map.
Table 1 outlines the parameters for each type of causal factor. The obtained weight values reflect the number of pixels associated with landslides and their distribution across each causal factor class. In this context, the calculated weights reveal that most factors exhibit ranges where landslide occurrence is probable, with a generally random distribution and specific concentrations within certain ranges (see
Table 3).
3.1. Landslide Suceptibility Zoning Map
The objective of this mapping is to identify areas or zones that are susceptible to landslides. To create the map, a range of factors was considered to understand where landslides occur and to assess the extent of the problem. Additionally, potential causal factors, such as elevation, slope, and geology, were identified. Finally, both past and recent events were documented using available data sources. The map was generated using Esri’s ArcGIS Pro tool, with the spatial reference system Magna Sirgas—National Origin (see
Figure 5).
The zoning map was classified into five susceptibility zones: very low (rare), low (unlikely), moderate (possible), high (likely), and very high (almost certain). This classification was achieved by reclassifying the data using natural breaks, which established the aforementioned ranges (see
Table 4).
The map aims to identify and visualize areas of the terrain exhibiting different levels of landslide susceptibility, with a particular focus on comparing zones classified as ‘very low’ and ‘very high’, and evaluating the similarity of their spatial distribution. These classes are derived from statistical models (IV and WoE) applied to geographical, topographical, and geological variables.
The graphical representation uses light green shades to depict ‘very low’ susceptibility and deep red for ‘very high’. According to the total number of cells estimated for each class (1,104,941), approximately 20,539 spatial units were simultaneously classified in the ‘very low’ and ‘very high’ categories by model IV, and around 44,778 units by WoE. Using this information, the percentage of similarity between datasets was estimated using Jaccard’s Index, a statistical method for measuring the similarity between finite sample sets [
54]. The calculated Jaccard values were approximately 0.0707 (≈7.07%) for IV and 0.1136 (≈11.36%) for WoE.
These values suggest a moderate overlap between zones classified as opposite susceptibility levels in both models. This could indicate ambiguity in classification, possibly due to variables that do not effectively discriminate between stable and unstable terrain conditions. For instance, if certain variable values (e.g., slope or land use) do not present strong contrasts, or if transition zones exist where terrain behaviour is uncertain, the model might assign extreme categories to spatially similar areas, leading to confusion.
Alternatively, the overlap may point to unbalanced weighting of input variables. If a particular variable is overemphasized or highly correlated with another, the model could misclassify regions as ‘very stable’ or ‘very unstable’ despite having similar physical characteristics. This inconsistency is reflected in increased overlap between opposing classes.
In addition, technical limitations such as medium spatial resolution may contribute to this issue, as it can blend terrain conditions and reduce classification precision. Under such circumstances, extreme classes may be assigned to the same cells, artificially increasing the similarity between areas that should otherwise be well separated.
Regardless of the cause, identifying this type of ambiguity is valuable for diagnosing and improving model quality. It can help inform variable reweighting, enhance image resolution, or guide field validation in areas with high uncertainty.
As shown in the landslide susceptibility zoning map, the high and very high susceptibility zones, based on the raster units, represent 19.71% and 19.62% of the area, respectively, for the WoE model, and 26.55% and 17.57% for the IV model. Of the 101 landslide event records captured along the road corridor, the density distribution shows a gradual increase from the very low to the very high susceptibility zones. Only five event points fall within the low to very low susceptibility ranges for the WoE model, and only one for the IV model. Most events are concentrated in the high and very high susceptibility zones, with a total of 88 events for the WoE model and 90 for the IV model.
3.2. Mitigation and Risk Management Strategies in Landslide Suceptibility Zones
The zoning map highlights the areas affected by these events, indicating zones that are highly dangerous for users of the road corridor. In such areas, the establishment of settlements and urban development should be avoided. In zones where low or very low susceptibility is detected, the slopes are gentle, reducing the risk, however, mitigation measures, such as proper surface water management, are still required. In areas with moderate susceptibility, the slopes are more pronounced, increasing the risk of ground instability. Therefore, it is essential not only to manage surface water effectively but also to implement additional measures, such as slope stabilization, drainage works, revegetation, and bioengineering projects, to reduce the likelihood of failure. In zones marked with high and very high susceptibility, the steeper slopes further increase the risk of instability, necessitating the aforementioned measures along with specific protection and stabilization works, such as shotcrete and anchored walls, which should be designed based on prior geotechnical and geological studies.
The landslide susceptibility map was developed for the Zipaquirá–Pacho road corridor, which is located in a geologically complex mountainous region characteristic of the Eastern Cordillera of the Colombian Andes. This corridor features steep slopes, deep canyons, and ravines, making it highly susceptible to landslides due to unstable and weathered soils and the prevalence of seasonal rainfall. The susceptibility map was produced using the statistical methods Weight of Evidence (WoE) and Information Value (IV), incorporating multiple causal variables such as slope, land use, lithology, distance to geological faults, and flow accumulation.
Both methods identified similar critical zones, particularly on slopes steeper than 16° and in areas with unconsolidated soils. However, the IV method exhibited greater sensitivity in capturing spatial variations among variable classes, resulting in a higher pixel count, greater surface coverage, and more recorded events per category. This suggests that IV may better detect complex spatial relationships in the dataset.
Nevertheless, because IV relies on the frequency of landslide occurrence within each class, it may overestimate susceptibility in areas with few historical events. In contrast, WoE incorporates conditional probabilities, offering a more balanced distribution of susceptibility levels.
Regarding model validation, both methods demonstrated comparable predictive performance. The IV model achieved an AUC of 81.79%, while the WoE model yielded 81.21%, indicating a slightly higher predictive capacity for IV. These results support the conclusion that both WoE and IV are suitable for modelling in contexts with multiple interrelated variables.
Zones classified as ‘high’ or ‘very high’ susceptibility by both models strongly coincide with segments of the road corridor that have historically experienced frequent landslides. This spatial correlation provides empirical validation of the models and highlights priority areas for slope stabilisation and maintenance interventions to improve trafficability along this critical transport route.
It is important to note that neither model incorporated real-time variables such as rainfall or soil moisture, which limits their application for early warning purposes. Moreover, the spatial and temporal quality of the landslide inventory plays a key role in determining the overall accuracy of the models.
4. Discussion
A landslide susceptibility map without validation is meaningless in the scientific community [
55], making the validation process crucial for assessing model accuracy using various techniques [
50]. In this study, 70% of the landslide events (71 events) were used to train the susceptibility model, while the remaining 30% (30 events) were reserved for validation, maintaining a random spatial distribution.
A limitation of the present study lies in its synchronic approach, as it relies on a dataset corresponding to a time window between 2014 and 2022. As a result, temporal variations in the conditioning factors and the occurrence of landslides were not considered. Although the 70/30 split between training and validation data was performed randomly, full spatial representativeness may not have been achieved. This could affect the model’s ability to generalise the results and apply them to other areas within the study region. It is recommended that future research address these aspects through the integration of multi-temporal analyses and more robust validation strategies.
While WoE and IV methods have been widely used for landslide susceptibility zonation, it is important to acknowledge that, as bivariate approaches, they do not fully capture the complex and non-linear interactions between causal factors. Consequently, more advanced models, such as Random Forest, Support Vector Machines (SVM), or artificial neural networks, may yield better results by simultaneously considering multiple variables. Moreover, although the AUC is a useful metric for evaluating the predictive accuracy of a model, it does not account for issues such as spatial autocorrelation or potential overfitting. These limitations should be taken into account and present valuable opportunities for further research.
Validation of the landslide susceptible zones produced by the IV and WoE models was performed using the receiver operating characteristic (ROC) curve method—specifically, by calculating the area under the curve (AUC). This statistical tool is widely used to estimate the accuracy of presence/absence predictive models [
56]. An AUC value between 0.9 and 1 indicates excellent performance; between 0.8 and 0.9, very good performance; between 0.7 and 0.8, good performance; between 0.6 and 0.7, average performance; and between 0.5 and 0.6—or equal to or less than 0.5—poor performance [
50,
57]. In this study, based on ROC curve analysis, the AUC values for both models exceeded 81%, indicating very good performance for landslide prediction (see
Figure 6).
The landslide susceptibility zoning map for the Zipaquirá–Pacho road corridor in the department of Cundinamarca was prepared using both the Weight of Evidence (WoE) and the Information Value (IV) methods. This map incorporates a variety of causal factors related to the study area’s specific characteristics, such as geomorphology, geology, and elevation. When comparing these two GIS-based statistical models, it is evident that both are widely used for landslide susceptibility mapping and zoning, although each has limitations. For instance, Wubalem and Meten (2020 [
55]) note that the Information Value method cannot determine the direct relationship between landslide factors and occurrences; while it provides insights into the influence of each factor class on landslide occurrence and predicts the probability of each class contributing to a landslide, it does not indicate which factor is most influential.
The accuracy of the two models demonstrates nearly equal predictive ability, with values of 81.21% for the WoE method and 81.79% for the IV method, suggesting that the IV method is slightly more effective. In this study, the success rate curve value for the IV method was 82.23%, while its prediction rate curve value was 81.79%. In comparison, the WoE method yielded a prediction rate curve value of 81.51% and a success rate curve value of 82.67%. The close similarity between the curves for both methods may be attributed to the use of the same dataset for both analyses. Ultimately, while our primary focus was on generating the susceptibility zoning map for the road corridor rather than solely on model accuracy, we recognize the importance of validation. Therefore, the Information Value model outperforms the logistic regression model in predicting the probability of landslide occurrences.
This study provides evidence on the usefulness of bivariate methods such as WoE and IV for landslide susceptibility assessment in mountain road corridors. In contrast to more complex multivariate approaches, these methods allow the construction of interpretable and reproducible models with accessible geospatial information, which is particularly valuable in regions with limited technical capabilities. The comparative application of both models in the same context provides objective criteria on their predictive behaviour and can serve as a methodological basis for other similar areas.
The validation of both models showed AUC values above 81%, with success and prediction curves of 82.67% and 81.51% for the WoE method, and 82.23% and 81.79% for the IV method, respectively. Although the results are similar, differences were identified in the spatial distribution of areas classified as ‘high’ and ‘very high’ susceptibility, highlighting nuances in the sensitivity of each approach to certain causal factors. This reflects that, while both methods are useful, they may vary in the way they interpret the influence of certain factors. This direct comparison between WoE and IV represents a relevant contribution, as previous studies often use only one of them without contrasting them under controlled conditions.
The results obtained in this study are consistent with findings reported in previous research. For instance, Rahman et al. applied the WoE method in the Shahpur Valley, Pakistan, obtaining an AUC value of 80.65% [
44], while Achu et al. applied the IV method in the Western Ghats, India, reporting values close to 83% [
44]. These accuracy levels are comparable to those achieved in the present study, supporting the applicability of both methods in topographically complex contexts using remotely sensed geospatial information [
45].
Similarly, Getachew and Meten found that although bivariate methods may be outperformed by machine learning techniques in terms of statistical performance, their simplicity and low computational cost make them ideal for baseline studies in remote or data-scarce areas [
47]. In this regard, the direct comparison of WoE and IV along the Zipaquirá–Pacho corridor provides locally grounded evidence that complements existing literature and reinforces their potential application in territorial planning and disaster risk management in tropical mountainous regions.
The approach adopted in this study is replicable in other Andean regions or tropical mountainous areas with comparable geological and climatic conditions. The model can be readily adapted to other rural road corridors, which extends its potential to support territorial planning and the development of resilient infrastructure. Its implementation enables the prioritisation of intervention zones, guides the formulation of mitigation strategies, and informs decision-making processes by local authorities, even in resource-constrained settings.
A potential limitation of this study lies in the spatial distribution of the landslide inventory, which is primarily concentrated in areas adjacent to the road corridor. Although the study area was expanded to capture a broader range of topographic and environmental conditions, this may introduce spatial bias into the models by limiting the representativeness of more remote zones. Future research is encouraged to incorporate more comprehensive inventories or implement systematic sampling strategies that include areas beyond the road axis, in order to enhance the model’s robustness and generalizability.
Although the inventory includes landslides classified as natural, in several cases their occurrence may have been influenced by anthropogenic interventions, such as slope cuts, infrastructure overload, or inadequate drainage systems. These factors were not explicitly incorporated as independent variables in the model due to the lack of detailed spatial data on roadworks and engineering structures. However, it is acknowledged that such interactions may be significant, and future studies are encouraged to integrate variables related to anthropogenic pressure and infrastructure in order to improve the model’s explanatory capacity.
In the context of landslide susceptibility mapping, statistical resampling techniques such as the Jackknife test have been widely applied to assess the relative importance of causal factors [
8,
9,
10]. However, this study did not apply such a technique for several reasons. First, the methods used (WoE and IV) already allow for analysing the contribution of each subclass of factor through the direct assignment of weights. Second, the Jackknife test involves the sequential exclusion of variables to assess their effect on model performance, which may induce biases when variables are correlated or the dataset is unbalanced.
Moreover, this study—like other recent works [
11,
12,
13]—adopts a conceptual approach in which all selected factors are considered relevant to landslide occurrence. Given the qualitative nature of the model, aimed at supporting risk management and land-use planning, the integration of variables was prioritised over ranking them by statistical influence. This approach aims to more comprehensively represent the complexity of the phenomenon, avoiding the dominance of variables with greater statistical variance that could obscure less evident but significant causal relationships.
Finally, considering that the available data have medium spatial and temporal resolution, it was deemed that the application of techniques such as the Jackknife test might yield unstable or unreliable results. Instead, a simple, interpretable, and reproducible model was preferred—one that is suitable for institutional contexts with limited technical capacity.
5. Conclusions
The landslide susceptibility zoning of the road corridor enabled the identification of areas potentially prone to landslides—an essential step for developing disaster prevention and mitigation plans in the study area. In this study, using two methods—the Weight of Evidence (WoE) method and the Information Value (IV) method—a GIS-based landslide susceptibility zoning map was developed for the Zipaquirá–Pacho road corridor in the department of Cundinamarca. The results indicate a direct relationship between landslide susceptibility and certain causal factors, such as soil conditions.
The landslide susceptibility zoning map was classified into five levels: very low, low, moderate, high, and very high. Overlay analysis of the test data revealed that the high and very high susceptibility zones encompassed more than 88% of the analyzed landslides for both methods. Validation using the area under the ROC curve (AUC) indicated that both models achieved AUC values above 81%, demonstrating very good predictive performance for landslide occurrence.
Overall, the study’s findings show that landslide susceptibility analysis and zoning have significant implications for disaster prevention and mitigation in the affected area. By predicting and identifying zones of high and very high susceptibility, local authorities can prioritize resource allocation to these critical areas and enhance monitoring and early warning systems for landslides. The high predictive accuracy demonstrated by both models provides robust scientific support for developing effective prevention strategies and for strengthening control and mitigation measures, particularly in high-risk zones.
In future studies, it would be valuable to explore the integration of both methods, combining the simplicity of the IV model with the robustness of WoE, in order to enhance classification performance and reliability.
It is also essential that landslide susceptibility maps be updated periodically to reflect new landslide occurrences and, where possible, incorporate recent changes in land use and rainfall patterns. This continuous updating process would significantly improve their relevance and accuracy over time.
While these maps provide critical support for land-use planning, their effectiveness in early warning systems requires integration with additional technologies, such as real-time monitoring networks and environmental sensors. Such integration would strengthen disaster preparedness and enable more proactive risk management.
Additionally, it is recommended that the weighting schemes of input variables in both models be reassessed to improve the differentiation between susceptibility classes. Field-based validation should also be prioritised, particularly in zones where overlapping or ambiguous classifications occur, to ensure a closer alignment between model outputs and on-the-ground conditions.