1. Introduction
Landslide disasters have increased significantly in the last decades, leading to economic disruption, damage to properties, and loss of lives. According to the EM-DAT database, landslides claimed the lives of 26,750 people worldwide from 1991 to 2020. In Brazil, 629 and 2502 fatalities were reported due to landslides and due to the combination of landslides and floods, respectively. Hence, Brazil is regarded as one of the countries with the highest annual mortality levels due to landslides [
1]. To reduce the existing risk, public managers have long been adopting tools to subsidize decision-making by optimizing regional planning [
2]. To this end, delineating landslide-prone areas is considered essential, as these assessments provide decision-makers with relevant information regarding landslide occurrences [
3,
4].
Overall, existing methods used for modeling landslide susceptibility can be classified into physically based models, statistical models, knowledge-driven (i.e., heuristic) models [
4,
5], and, more recently, machine learning models [
2,
6]. Among these approaches, physically based models such as the Shallow Landslide Stability Model (SHALSTAB) [
7], Stability Index Mapping (SINMAP) [
8], and Transient Rainfall Infiltration and Grid-based Regional (TRIGRS) [
9,
10] have been commonly adopted as effective tools for predicting areas susceptible to slope instability. These models currently deliver the highest prediction accuracy among all existing methods [
6]. They are based on the Safety Factor (FS), which reduces the subjectivity of the model outcomes.
Among the aforementioned models, SHALSTAB stands out for its efficiency. It is worth mentioning that the SHALSTAB was conceptualized as a simple model to map shallow landslide potential [
7]. Some studies demonstrate the better performance of SHALTAB, even when compared with accurate models such as TRIGRS [
11,
12,
13], which incorporates varying hydrological parameters over time, and SINMAP, which follows a probabilistic approach [
12,
14]. Therefore, the frequent use of SHALTAB is associated with its high efficiency due to its deterministic aspect, easy access, availability of free software, and the small amount of data it requires.
Given its advantages, the SHALSTAB model has been applied in several countries: the USA [
7,
15]; Canada [
16]; Argentina [
15]; Portugal [
17]; Italy [
18,
19,
20]; Guatemala [
21]; and Australia [
22]. In Brazil, SHALSTAB has also been widely used [
13,
14,
23,
24,
25,
26,
27,
28,
29,
30,
31]. In a recent systematic review, Melo et al. [
32] highlighted aspects that should be considered to advance the SHALSTAB studies. Among the findings, the authors emphasized the importance of determining input parameters with local sampling. Indeed, it was found that, due to lack of resources, many studies in Brazil did not carry out field sampling to determine soil mechanical properties. Studies often use data obtained in other study areas which are pedologically similar [
18,
23,
24,
27,
31,
33,
34]. Even when the studies sample data, these tend to have few data points. For instance, even though Kim et al. [
35] characterized the soil by sampling it, they considered that the soil mechanical properties were homogeneous over the study area. This limitation is linked to the lack of resources and/or research time for collecting relevant information. This is the case in many developing countries such as Brazil, where data scarcity prevails.
The data required for applying SHALSTAB are related to topographic and physical characteristics of the soil for each terrain unit (i.e., pixel). These include variables such as the soil friction angle, cohesion, specific mass, and layer depth. In terms of topographic features, the counter length, the upstream contribution area, and the slope are variables often included. SHALSTAB also allows for customization. In Brazil, Michel et al. [
25,
26], inserted new variables (cohesion and root weight) and considered the spatially discretized characterization of the values of the geotechnical units into the model. Here, the geotechnical unit is defined as the area which has common soil type and lithology. Within one geotechnical unit, mechanical and hydraulic properties of soil are considered uniform. Therefore, a geotechnical unit can be determined with soil type and geological maps [
36].
Concerning the model uncertainty and sensitivity, several studies highlight the influence of spatial resolution of digital terrain models (DTM) on the estimation of instability areas [
37]. These studies demonstrated the importance of adopting more detailed spatial scales. This is relevant because, with more detail on the slope relief, it is possible to verify the degree of water concentration. The slope relief can also be used as an indication of failure predisposition. Based on a sensitivity analysis of SHALSTAB model outputs, Borga et al. [
38] and Michel et al. [
25,
26] verified the most relevant parameters and demonstrated the strong sensitivity of SHALSTAB to the cohesion value.
The original SHALSTAB algorithm considers a single set of geotechnical values for a whole basin. However, for better performance, Michel et al. [
25] modified the algorithm, inserting a spatial discretization of the geotechnical values of each unit into this model. Using the algorithm established by Michel et al. [
25], Sbroglia et al. [
28] considered spatially discretized geotechnical unit values. Although the consideration of these aspects enabled the generation of more realistic predictive maps, there are still many uncertainties of both the actual benefits as a consequence of the geotechnical-unit characterization and the inherent processes related to the spatialization as well as the amount of soil sample data required for this purpose. Hence, Michel et al. [
26] and Sbroglia et al. [
28] proposed a new technique for the discretization of geotechnical-unit characteristics. However, the effect of such discretization was not well investigated in detail.
Therefore, this study aimed at investigating the advantages of using the geotechnical-unit characterization as a variable in the slope stability calculation using SHALSTAB. To this end, several scenarios were proposed to represent distinct contexts of distribution and quantification of soil sampling points. To analyze the model performance, we used state-of-the-art techniques, such as the ROC (Receiver Operating Characteristic) curve [
39], and the Success Index and Error Index [
20]. Besides this, we also proposed a new metric which is termed Detective Performance Index of Unstable Areas (DPIUA).
3. Results
By considering 13 scenarios, 360 different simulations were executed, encompassing seven stability classes proposed by Montgomery and Dietrich [
7]. To better visualize the outcomes, these classes are expressed in logarithmic scale. Two classes (unconditionally unstable and unconditionally stable) resulted from the topographical and pedological characteristics, being independent of the parameter
q/
T. The other classes are linked to the hydrological parameters relating to the parameter
q/
T. The lower values of log
q/
T (red areas in
Figure 4,
Figure 5 and
Figure 6) represent areas that need less magnitude in the hydrological conditions for the area to become unstable. Therefore, they present higher instability.
The simulation results show that, in all scenarios, the instability classes increased as the soil layer depth increased. A reduction in unstable areas was also observed in places of lower soil depths, although they are located on steep slopes where the rate of surface erosion and the propensity for landslides are very high. When the instability threshold log
q/
T = −2.5 was adhered to, the best performance was obtained with Scenario 1 by considering
z = 2 m, with an IA = 87% and IE = 10% (
Figure 4 and
Figure 7A).
By relating the stability classes and the cumulative percentage of the area and landslides in each class, 89% of the landslides fall into only 9% of the unstable areas in Scenario 1 when considering the threshold log
q/
T = −2.5. For this scenario, an area calculated between the curves of DPIUA = 55 was obtained (
Figure 8A). The elaborated graphical presentation (
Figure 8) presents two curves with values for each log
q/
T category: the first one is the “percentage of cumulative area”, and the second one is the “percentage of cumulative area of landslides”. The further these curves are, the higher the DPIUA is, and the better the performance of the SHALSTAB model.
When the instability threshold log
q/
T = −2.8 is adopted, the best performance was obtained with Scenario 4, where the values of the parameters of sample point 8, located in the geotechnical unit 5 and
z = 1 m, were used (
Figure 5). This result is observed in
Figure 7B, where the values of SI and IE are 75% and 14%, respectively.
Figure 8B shows the stability classes of the model and the cumulative percentage of the area and landslides in each class in Scenario 4 and the resulting area between such curves (DPIUA = 71). Thus, 93% of the landslides were classified in 3% of the unstable areas.
For the instability threshold log
q/
T = −3.1, the best performance was obtained with Scenario 13, which covered 13 sampling points and considered
z = 3 m (
Figure 6). The evaluation results corresponding to such a condition is shown in
Figure 7C and
Figure 8C, with SI and EI values of 90% and 26%, respectively and a DPIUA of 66. In this case, 93% of the landslides were encountered in 7% of the unstable area in the basin.
When considering instability thresholds of log
q/
T = −2.5 and log
q/
T = −3.1, the best results were obtained from the Scenarios that used the values of the discretized soil parameters per geotechnical unit. The DPIUA area in
Figure 8 was larger in Scenario 13, which presented better performance when a threshold log
q/
T = −3.1 was adopted (DPIUA = 66). However, this result was obtained from the large value (82%) with the unconditionally unstable class, independent of the hydrologic parameters. Since rainfall is responsible for the landslides, this result can be considered incoherent.
Scenario 1, with z = 1 m, presented the best performance with the threshold q/T = −2.5, most parts of the scars from landslides (75%) were captured by this category, while 14% of landslides were captured by the unconditionally unstable category (q/T = −3.5). Observing the performances of Scenarios 13 and 1, the latter presents better coherence in the results due to a large part of the scars captured by the category associated with hydrological parameters. It is conjectured that the reduction in the c value of geotechnical unit 2, slightly more than 50%, caused an overestimation of unstable areas in the basin. Allied to this, the high soil thickness contributed to this result.
On the other hand, when the instability threshold log q/T = −2.8 is considered, the best performance was presented by Scenario 4, which adopted constant soil parameters for the whole basin. Thus, from the values of point 8 located in the geotechnical unit 5 and z = 1, 61% of the landslides were predominantly framed in the unconditionally unstable class. Meanwhile, 18% of the landslides were classified into the log q/T < −3.1 class. Similar to the aforementioned analysis associated with the log q/T = −3.1 threshold, we can assume that these results were influenced by values that are not consistent with the local reality. The low ϕ value (10°) for the whole basin probably contributed to the overestimation of unstable classes, setting the predominant classification of landslides in the unconditionally unstable class.
4. Discussion
Despite being widely used, the SHALSTAB model is often applied without considering the importance of in situ determination of geotechnical parameters and the necessity of spatial discretization of soil characteristics. Indeed, several applications tend to adopt values from other studies which investigated similar soils [
18,
23,
27,
33,
34]. Michel et al. [
25,
26] and Sbroglia et al. [
28], which used the modified algorithm of Michel et al. [
25], proposed a useful procedure for discretizing soil characteristics. However, none of the previous studies quantified the effect of such discretization. To address this gap, here we analyzed it by considering different performance metrics: (1) SI and EI [
20], (2) the ROC curve [
39], and (3) a new index, the DPIUA (
Figure 8).
Even though there are several uncertainties related to the effects of geotechnical-unit characterization, the evaluation of different instability thresholds, together with the different scenarios, showed that considering 20 sampling points to characterize the geotechnical units presented a better performance when compared to the original SHALSTAB model (Scenario 11). When considering the thresholds log q/T = −2.8 and log q/T = −3.1, the best performances had polarized results due to abrupt variation in soil resistance parameters. The importance of discretization of soil characteristics by using the geotechnical units confirms the necessity to carry out the field sampling at every unit. Hence, for improved performance, future studies should divide the study area (e.g., basins) into different geotechnical units through the use of existing data generated by other researchers or with similar soils, and, when possible, characterize them with in situ soil sampling.
The simulation results confirm the important role of soil depth in hillslope stability analysis. As demonstrated by Borga et al. [
38] and Michel et al. [
26], the increase in the soil depth reduces the safety factor of the hillslope. The results obtained here confirm this tendency. By analyzing the values of DPIUA, it can be said that soil depth of 2 m supported better performance in general. These results might imply that the mean depth in the basin could be 2 m. A condition with a steep hillslope and shallow soil layer can sustain the soil even in its saturated situation. Therefore, the soil layer thickness can be thought to influence hillslope failures significantly.
Although the present study showed the effect of geotechnical unit consideration on mapping performance with SHALSTAB, it did not investigate how uncertainties in parameter determination affect the reliability of the results. This theme should be investigated in more detail in future studies. How uncertainty affects results’ reliability is among the most central challenges in geosciences in general, including hydrology, geomorphology, and hydrogeomorphology [
47,
48]. The uncertainty depends on (i) the nature of each parameter (model input data), (ii) the methodological procedures used to collect data in the field and to analyze in the laboratory, (iii) the technical level of the involved researchers, and (iv) the adopted models, among other variables. Therefore, when feasible, uncertainty evaluations should be carried out together with sensitive analysis such as that in Michel et al. [
26]. The results presented here advance this direction, showing how results vary when different assumptions are considered in each scenario.
5. Conclusions
The SHALSTAB model has been frequently applied to assess shallow-landslide susceptibility. In developing countries, where geotechnical-unit characterization is difficult due to lack of financial support, many studies measure the soil mechanical properties at few points or do not measure at all and use the data available in the SHALSTAB tutorial [
32]. Therefore, the present study demonstrated the importance of considering the geotechnical-unit characterization on landslide mapping performance by considering the case of the Jaguar creek basin, southern Brazil.
In general, results showed that the use of geotechnical units characterized by several sampling points provides better results compared to the original SHALSTAB model. To evaluate the model performance, we proposed a new index called DPIUA, which allows evaluating the model robustness according to different instability thresholds (
Table 6 and
Figure 8). Through using several performance evaluation methods, the benefits generated by the characterization of the geotechnical units were verified. These advantages were evident as a result of the better performance of the scenarios in which the soil parameters were discretized into various geotechnical units.
Thus, future assessments should consider greater cartographic precision to obtain the variability of the geotechnical units in the application of the SHALSTAB model and also a sufficient amount of in situ soil sampling points to characterize these units. In this regard, it should be noted that the choice of the minimum number of sampling points requires careful consideration. This is one of the most critical issues for field workers who want to be time-efficient while at the same time achieving precise results. This number may depend on (i) soil types within an adopted soil classification, (ii) soil parameters for analysis, and (iii) other physical conditions such as soil covers (vegetation) and soil depth. In any case, adopting geotechnical units in future assessments will surely enable the generation of more reliable landslide-susceptibility maps.