Landslide Susceptibility Assessment Using AHP, Frequency Ratio, and LSI Models: Understanding Topographical Controls in Hanang District, Tanzania

Abstract

This study evaluates landslide susceptibility in Hanang District, Manyara Region, Tanzania, using three approaches: Analytic Hierarchy Process (AHP), Frequency Ratio (FR), and Landslide Susceptibility Index. A total of 11 environmental and anthropogenic factors were analyzed, with 5879 landslide events identified from satellite imagery to create an inventory map for training and testing. Model performance was assessed using Area Under the Curve (AUC), Consistency Ratio, and Prediction Rate, while multicollinearity among factors was evaluated through Tolerance (TOL) and Variance Inflation Factor (VIF). Results indicate that the Analytic Hierarchy Process model outperformed Frequency Ratio and Landslide Susceptibility Index, achieving an Area Under the Curve of 0.88, demonstrating strong predictive capability. Slope, elevation, and geology were identified as the most influential factors. The susceptibility maps developed in this study aim to support policymakers and disaster management authorities in climate adaptation and risk reduction efforts, contributing to Sustainable Development Goal 13 (Climate Action). Limitations include reliance on remotely sensed data for landslide inventory, which may omit smaller events or introduce classification errors.

1. Introduction

Landslides are among the worst natural disasters, leading to a large number of fatalities, property losses, and environmental deterioration [1,2]. Between 4.3 million and 7 million people reside in high-risk landslide areas, especially in hilly regions like the Andes, the Himalayas, and the tropical parts of Southeast Asia [3]. On a global scale, landslides are frequently caused by both natural phenomena like intense rain, earthquakes, and volcanic eruptions, as well as human activities like mining, infrastructure building, and deforestation [4,5,6]. According to the Centre for Research on the Epidemiology of Disasters (CRED) report, landslides account for 17% of all natural disaster fatal events in the world [7,8], and it is forecast that there will be a subsequent trend in the increase of landslide occurrence due to several factors such as population growth, increased infrastructure constructions, and residential area progressions in high-risk areas, continuous deforestation, and increased precipitation. Landslides cannot be prevented, but with the right techniques and analysis that display the spatial probability of landslide occurrence, disasters can be anticipated and avoided.

In Tanzania, landslide events have increased in frequency and intensity, especially in regions like Hanang District, Manyara Region, due to shifting climate patterns, increased precipitation, and unsustainable land use. Despite this growing risk, landslide susceptibility mapping (LSM) remains underutilized in regional hazard planning and disaster preparedness. Landslide susceptibility mapping (LSM) is one of several landslide prediction techniques, and a useful land use management strategy that can help land managers make better decisions [9,10]. The spatiotemporal mobility and deformation of vulnerable locations can be tracked using remote sensing techniques to assess the risk of regional landslides in complex terrain settings. In landslide monitoring, early warning, and evaluation, landslide susceptibility prediction is a crucial technique that displays the spatial probability of landslide occurrence [11,12,13].

Making landslide susceptibility maps requires several qualitative or quantitative techniques. Expert opinions and assessments are crucial to qualitative approaches, such as geomorphological mapping [14,15], indexing or heuristic approaches, and landslide inventory analysis [12]. It has been acknowledged that the initial phase of landslide hazard reduction measures should involve landslide inventory and susceptibility mapping investigations [16,17]. These maps offer crucial data to help with land use planning and urban development decisions. By using these maps effectively, landslides’ potential for damage and other financial consequences can be significantly decreased. The quantitative approaches analyze the relationship between landslide incidence and affecting factors using mathematical analysis and establish a probability statistical model.

According to [18], the quantitative approaches are mostly statistical. Protection elements of unstable slopes are constructed using deterministic approaches. The majority of the aforementioned research has been carried out utilizing regional landslide inventories that are based on remotely sensed and airborne images [18,19]. In contrast, statistical approaches in mapping landslide susceptibility use binary or multivariate statistical approaches for assessment, including logistic regression and the Weight of Evidence Support Vector Machine, naïve Bayes, and the Analytical Hierarchical Process (AHP) and Frequency Ratio (FR) [8,20]. Considering that landslides are more expected to happen in environments that are comparable to those that have triggered them in the past, data-based approaches adopt a scientific model to forecast the likelihood of landslides based on the spatial distribution of several landslide-influencing factors in places that are prone to them [15,21,22]. The application of the Analytic Hierarchical Process and Frequency Ratio approaches has been used by many scholars and researchers [11,23,24,25]. Each of these models offers ways to map the outputs and integrate different levels of information.

In recent years, scientists and disaster responders have been using satellite-based earth observation technologies and data-based hazard cutting-edge analysis more frequently to quickly assess disaster scenarios globally [25], including deep learning, machine learning, and other approaches. Deep learning methods initially employed in landslide susceptibility mapping (LSM) include convolutional neural networks (CNNs) and recurrent neural networks (RNNs) [26]. Machine learning approaches include XGBoost, random forest [27], artificial neural networks (ANN), decision trees (DT) [19], and support vector machine (SVM) techniques that have gained popularity due to the advancement of artificial intelligence and its broad use, as distinguished by its short sample size, nonlinearity, and high dimension [22,23,28,29,30]. Other approaches include neuro-fuzzy systems [28], weighting factors [16], weighted linear combinations of instability factors, Shannon’s entropy [31], and spatial decision support systems. All of these models have been evaluated based on their performance. Both qualitative and quantitative maps of the landslide hazard zones are produced by these methods, and their spatial results are often pleasing. Comprehending the distinctions among the suggested methods is not always easy. Sometimes it is difficult to understand how the suggested methods differ from one another. The primary variations are in the method utilized to estimate the prior likelihood of landslide occurrence and the level of rigor of the approach [8].

Landslide susceptibility mapping is essential for hazard management in mountainous regions. While statistical approaches such as Frequency Ratio (FR) and Landslide Susceptibility Index (LSI) are widely used, few studies have compared their predictive performance with expert-based methods like the Analytic Hierarchy Process (AHP), particularly in the Hanang region of Tanzania. This lack of comparative analysis limits understanding of the relative strengths and weaknesses of these models for local conditions.

To address this gap, the present study aims to evaluate and compare the performance of AHP, FR, and LSI models for landslide susceptibility mapping in the study region. Specifically, this study focuses on (1) incorporating eleven key conditioning factors: elevation, distance to rivers, distance to roads, slope, aspect, lithology, distance to fault lines, land use/land cover, NDVI, annual rainfall, plan curvature, and topographic wetness index (TWI) to construct susceptibility models; (2) assessing model performance using statistical and expert-based evaluation measures, including the Area Under the Curve (AUC), Consistency Ratio (CR), and Prediction Rate (PR); (3) examining multicollinearity among conditioning factors using the Variance Inflation Factor (VIF) and Tolerance (TOL) to ensure model robustness; and (4) providing a comparative analysis of AHP, FR, and LSI to highlight their relative strengths, weaknesses, and applicability under the environmental and geological conditions of Hanang.

This study attempts to answer the key question of how AHP, FR, and LSI models perform using key environmental and topographic factors, aiming to identify the most accurate and locally applicable approach. Through this comprehensive approach, the study provides a robust framework for understanding landslide susceptibility in the region and informs effective hazard management strategies.

2. Materials and Methods

2.1. Location of the Study Area

The Hanang District, located in the Manyara Region of northern Tanzania between longitudes 34°45″ to 35°48″ E and latitudes 4°25″ to 5°00″ S, covers an area of 3639 km² with an elevation reaching 3399 m and a population of approximately 367,391 [32] (Figure 1). The district is characterized by an equatorial climate with annual rainfall of around 1200 mm and average temperatures up to 37 °C, supporting agriculture and livestock as the primary economic activities, especially the cultivation of wheat, maize, barley, and sunflowers [32,33,34]. Geologically, Hanang consists of low to medium mountain landforms and volcanic soils that are highly susceptible to erosion and slope instability, especially during the rainy season. In 2023, a severe landslide event caused loss of life, property damage, and economic disruption, highlighting the area’s vulnerability due to a combination of steep slopes, intense rainfall, and human activities such as hillside farming.

2.2. Preparation of Landslide Inventory Map

The inventory maps have been crucial in landslide susceptibility assessment, which involves a simple process that records the location, type, and contemporary occurrence of all visible landslides in a given area utilizing several methodologies based on the context, the extent of the research region, base maps scale, aerial images, and quality and details of the existing material, and includes the resources available [35] for Landslide Susceptibility Maps (LSM) to be used for accuracy assessment during the statistical modeling [36]. However, the forthcoming landslide will occur where the previous landslide took place based on the existing environmental conditions [37,38].

The landslide inventory maps can be organized in several approaches, such as field surveying and investigations, and satellite image interpretation [39,40]. The study noted 5879 landslides through satellite image recognition from the land use/land cover map by using the bare land class [40]. Based on this assumption, the impacts of the landslides are in the form of bare soils with erosional features (shallow soils), involving mudflows consisting primarily of fine-grained material (≥50% silt/clay) mixed with water, forming a viscous slurry that typically follows pre-existing drainage channels [41]. These landslides occur due to prolonged rainfall in the loose sediments in the study area. It is then vectorized and overlapped in the boundaries for the statistical data for each criterion. They are branded with a small scale ranging from 10³ m² and the largest slides being 2 × 10⁴ m² [42]. The landslides have been stimulated by the distributions of seasonal heavy rainfalls associated with floods and day-to-day human engineering activities.

Additionally, the landslide inventory may not capture all landslide events accurately. The landslide inventory map was compiled from literature and remote sensing data, with cross-checks to reduce errors. As an indirect method, it may still miss smaller landslides or include classification errors, affecting model confidence. It is important to note that the randomly vectorized samples from the LULC map were categorized into two parts: (i) the training samples and (ii) testing samples for validation and calibration procedures, by splitting them into groups of 70% and 30%, respectively, to obtain the good and excellent agreements (Figure 2).

2.3. Determination of Conditioning Factors

Conditioning factors are the circumstances or criteria that can control the occurrence of landslides and/or absenteeism. The chosen criteria for landslide conditioning are the potential for the Analytic Hierarchy Process (AHP) and the Frequency Ratio (FR) models. According to several works of literature, the conditioning features/criteria should be selected based on the geomorphological characteristics of the area [43,44,45].

In this study, eleven conditioning criteria including, elevation, distance to river, distance to roads, slope, aspect, lithology, distance to fault lines, land use/land cover, normalized difference vegetation index, annual rainfall, plan curvature, and the topographical wetness index (TWI) have been primarily selected, and lately used for further analysis in landslide susceptibility. The criteria should be associated directly or indirectly with the slope of a given study area due to their direct effect on land stability [25,46]. The selected conditioning criteria have been obtained from various data sources, including field investigations, open sources, secondary sources, and remote sensing images (Table 1). The SRTM Digital Elevation Model (DEM) was used to produce other conditioning factors including the slope, curvature, aspect, distance to fault lines, topographical wetness index, distance to river, and elevation [29,47]. However, the Landsat 8 which carries the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS) instruments was used for the computation of the normalized difference vegetation index, and Sentinel-2A from Copernicus was used for the derivation of land use/land cover with the spatial resolutions of 30 m and 10 m, respectively. The main reason for using different resolutions was to upscale the 30 m to 10 m resolution to improve the quality of the data in terms of the resolution [30,48].

It is important to note that the eleven conditioning factors can either be constant or indeterminate variables. For other statistical analytics like the LSM and LR, the continuous variables should be converted into discrete variables to facilitate the easy analysis of decision-making and landslide susceptibility predictions [49]. At this point, there is no uniform standard to attain the variable numbers for the attribute interval [34,50]. The weights have been assigned to each criterion to compute the landslide susceptibility using the AHP and FR models to support the decision-making based on expert knowledge.

The vertical distance from the sea level in the study region ranges between 1357 m and 3402 m as the lowest and highest distance, respectively (Figure 3E). It is linked with the other conditioning factors including slope, which has been classified into four classes using the Natural Breaks (Jenks) approach from 0° to 72° in which flat to gentle (<5°), moderate (5–9°), steep (9–22°), and high slopes (22–75°) [51,52]. From this perspective, the gentle and flat slopes are spatially distributed along the river banks of the study area. In contrast, the steep slopes are distributed along the mountains where human activities are sporadic, and no landslide activities have been noticed due to the hard rocks that are resistant to sliding. The aspect of areas is used to show the facing direction of the slope in a given area [53]. The aspect of an area has been classified into nine classes: flat, north, northeast, east, southeast, south, southwest, west, and northwest. The criteria have been so potent in the landslide susceptibility for microclimate and hydrology characteristics based on its exposure to sunlight, wind, and rainfall [54].

Figure 3K demonstrates the topographical wetness index of an area, which shows the soil moisture on the ground and its distribution concerning the vegetation on the surface of slopes and the annual net primary production that triggers the occurrence of landslide [55]. The TWI is used in the quantification of the topographic controls based on the hydrological approaches. However, the TWI has been calculated from the SRTM Digital Elevation Model raster layer with 30 m resolution using ArcGIS 10.7 software tools to produce a slope, flow direction, flow accumulation, the tangent of the slope, and finally the TWI. It has been classified into five classes according to different reviews of the TWI, (−9.5 to −4.6), (−4.6 to −2.7), (−2.7 to −0.1), (−0.1 to 2.9), and (2.91 to 12.1), showing the spatial distribution of water movements along the areas [15,29,56]. The higher the value, the higher the water trend accumulated, and vice-versa [55]. It was obtained through the formula below:

$$\mathrm{T}\mathrm{W}\mathrm{I}=\mathrm{A}=\ln {\displaystyle \frac{\mathrm{a}}{\tan \left(\phi \right)}}$$

(1)

where a is the upslope (specific catchment) and $\tan \left(\phi \right)$ is the slope.

Geological features, such as faults, joints, weak, sensitive, and shared materials, the existence of fractures, and disparities in absorbency or stiffness of the slope-forming elements, including the uplift, glacial reverberations, and erosion of the hill slope. The area consisted of various geological structures, including the Precambrian Craton, Tertiary-Quaternary unconsolidated, and Tertiary–Quaternary volcanic, which influences the occurrence of landslides due to their weak faults, fracturing structures, and the alluvial sediments along the river valleys and banks (Figure 3). The existing geological features have been associated with several soil groups, including the Ferric Acrisols, which are dominant in the study area and easily eroded with depositional surfaces. The soils are commonly found in mountain areas on stable ridge tops, linked with regosols and cambisols on the steeper and less stable slopes; eutric nitosols, which are red and strongly weathered soils, considered as more fertile and productive than any other tropical red soils; and lithosols, which are found on hills and hill slopes, shallowly associated with gravels close to the surface in mountains areas, with poor water (rainwater) absorption reading to severe erosion [57].

The proximity distances involved both the distance to the main river (m) and the distance to the road (m). The measurements were taken in ArcGIS software between the landslide points and the roads and rivers with equal distances of 100 m and 200 m buffer and hence, classified into four classes. In areas with rich runoff, a reservoir structure is the most common infrastructure development activity to utilize water resources, which significantly affects landslides. On the other hand, the distances to the roads demarcate the influence of human activities on the landslide environment. It also shows the accessibility to their economic events such as mining, livestock farming, and agriculture [58].

The NDVI measures the amount and vigor of vegetation on the land surface. The spatial composite images were easily compiled to distinguish between the green vegetation and bare soil. The NDVI values range from −1 to 1; the negative values indicate the clouds and the water, near-zero values indicate the bare soils, the positive values indicate sparse vegetation (0.1 to 0.5), and 0.6 and above indicate dense vegetation. The NDVI was calculated using the Sentinel-2A, which carries a near-infrared (NIR) Multispectral Instrument (MSI) and the red bands were acquired on 24 November 2023. The areas with less vegetation (bare soils) are easily eroded, which becomes a big influence in landslide occurrence. Thus, the NDVI was derived from the formula below:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4}{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4}}$$

(2)

Land use/land cover maps have been the prejudicing conditioning factors in the landslide occurrence in the study area, as they have been used by many studies. The land use types have abundant effects as they affect the extent and rate of occurrence due to the influence of nature and human activities. It was derived from the Sentinel-2A satellite images acquired during the dry seasons with a low percentage of clouds in the Google Earth Engine. The images were then classified into nine classes: Agriculture, Bare Land, Built-up, Forest, Grassland, Shrubland, Water, Wetland, and Woodland using a Random Forest (RF) classification algorithm. The validation metrics, such as overall accuracy and kappa coefficient, were used to measure the observer agreement between the LULC classifications and reference data, obtaining 0.90 and 0.84, respectively, which confirms that the data effectively represents real-world conditions. However, the value greater than or equal to 0.75 indicates a very good to excellent agreement, while the values between 0.40 and 0.75 indicate only fair agreement, and the values less than 0.40 indicate a poor classification agreement between the categories. Figure 3G indicates that the built-up class is where human settlement takes place. The human activities taking place in this class had a greater effect on the occurrence of the landslide compared to the other classes.

Precipitation (mm) is also an important conditioning criterion influencing the manifestation of landslides in an area. The rainfall data was obtained from the Tanzania Meteorological Agency (TMA), which is responsible for meteorological services, including weather forecasting, climate services, warnings, and advisories information for the country, operating under the Ministry of Works, Transport and Communications of the United Republic of Tanzania. The annual average rainfall of an area has many implications on the circulation features of the average rainfall received, which also have a significant impact on the groundwater level changes and shear strength of the rocks and slopes [25]. The area receives high average annual rainfall from January to April, associated with floods, with the autumn season in late November to January, and the rest of the months being the winter period. The flood seasons in the mentioned months are espoused as a landslide conditioning criterion (Figure 3).

Table 1. Landslide conditioning criteria and their sources.

Conditioning Criteria	Source	Spatial Resolution (m)
Administrative boundaries	https://www.nbs.go.tz/ (assessed on 24 March 2022)
Elevation (m)	USGS Earth Explorer	30
Slope aspect (°)	SRTM Digital Elevation Model	30
Slope degree (°)	SRTM Digital Elevation Model	30
TWI	SRTM Digital Elevation Model	30
Distance to river (m)	https://www.openstreetmap.org/ (assessed on 24 February 2022)	30
Distance to road (m)	https://www.openstreetmap.org/ (assessed on 24 February 2022)	30
Distance to fault (m)	https://www.openstreetmap.org/ (assessed on 24 February 2022)	30
NDVI	from Sentinel-2A	10
Land use	https://dataspace.copernicus.eu/ (assessed on 5 August 2024)	10
Precipitation (mm)	Tanzania Meteorological Agencies (TMA)	30
Geology	Ministry of Minerals	30
Inventory Points	[59,60], and https://dataspace.copernicus.eu/ (assessed on 5 August 2024)	10

Table 1. Landslide conditioning criteria and their sources.

Conditioning Criteria	Source	Spatial Resolution (m)
Administrative boundaries	https://www.nbs.go.tz/ (assessed on 24 March 2022)
Elevation (m)	USGS Earth Explorer	30
Slope aspect (°)	SRTM Digital Elevation Model	30
Slope degree (°)	SRTM Digital Elevation Model	30
TWI	SRTM Digital Elevation Model	30
Distance to river (m)	https://www.openstreetmap.org/ (assessed on 24 February 2022)	30
Distance to road (m)	https://www.openstreetmap.org/ (assessed on 24 February 2022)	30
Distance to fault (m)	https://www.openstreetmap.org/ (assessed on 24 February 2022)	30
NDVI	from Sentinel-2A	10
Land use	https://dataspace.copernicus.eu/ (assessed on 5 August 2024)	10
Precipitation (mm)	Tanzania Meteorological Agencies (TMA)	30
Geology	Ministry of Minerals	30
Inventory Points	[59,60], and https://dataspace.copernicus.eu/ (assessed on 5 August 2024)	10

2.4. Methods

In this study, the landslide inventory map was first prepared using secondary sources, including existing reports and documents from the Disaster Ministry, since no official inventory was available. Eleven conditioning factors were then selected as input variables: elevation, slope degree, slope aspect, geology, land use/land cover, precipitation, distance to rivers, distance to roads, distance to faults, NDVI, and topographic wetness index (TWI) [34,58]. To ensure reliability, multicollinearity analysis was performed using Tolerance (TOL) and Variance Inflation Factor (VIF), followed by significance testing of each factor. The dataset was split into 70% for training and 30% for testing, and three susceptibility mapping approaches were applied: the Analytic Hierarchy Process (AHP), the Frequency Ratio (FR) method, and the Landslide Susceptibility Index (LSI) overlay model. These models were selected because they are widely used in landslide research, require relatively limited data compared to advanced machine learning techniques, and are suitable for regions such as the Hanang District where high-quality datasets are scarce. In AHP, pairwise comparisons and consistency checks were carried out; in FR, factor maps were reclassified and Frequency Ratios calculated; and in LSI, weighted factor maps were normalized and summed [61]. The resulting susceptibility maps were validated using statistical measures, including the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC) index (Equation (14)). In this context, AUC values greater than 0.5 indicate acceptable performance, while values approaching 0.9 denote very strong predictive ability [62]. Sensitivity and specificity curves were further examined to reflect the classification dynamics of the produced models (Equations (11) and (12)). Finally, the validated results were compared across methods, and the most reliable landslide susceptibility maps were produced [25,57], along with additional validation metrics as described in the subsections in Figure 4 below.

2.5. Multicollinearity Analysis of Conditioning Factors

For LSM, it is important to check the multicollinearity between the conditioning criterion as its undistinguishable step. This step points out the relationship between the variables, in which the strong linear correlation among the factors indicates the existence of redundancy in the original conditioning factors [63]. The multicollinear complications become a challenge for landslide susceptibility predictions and pose a risk of errors. However, the two statistical indices (parameters) TOL and VIF are used in the analysis of multicollinearity through a pair of their reciprocals [64]. Thus, when TOL < 0.1 or VIF > 10, it indicates a blunt multicollinearity.

$$\mathrm{V}\mathrm{I}\mathrm{F}={\displaystyle \frac{1}{1-{\mathrm{R}}^2}}$$

(3)

$$\mathrm{T}\mathrm{O}\mathrm{L}={\displaystyle \frac{1}{\mathrm{V}\mathrm{I}\mathrm{F}}}$$

(4)

where R² is the coefficient regression determination among the conditioning criterion

2.6. Importance of Landslide-Related Conditioning Factors

This is the next step after conducting a multicollinearity analysis to scrutinize the most effective landslide conditioning factors. It is because some of the factors are not important in the landslide susceptibility, causing noise and huge errors during the prediction process. Therefore, in this study, the information value approach has been used in the determination of the important conditioning factors to be used for modeling processes [65]. In general, the factors with values greater than zero have much influence on the model, while the factors with values smaller than zero do not contribute to the model and can be omitted before further modeling [25]. The common formula used for this traditional statistical approach is indicated below;

$$\mathrm{B}={\sum }_{\mathrm{k}=1}^{\mathrm{n}}\ln {\displaystyle \frac{L_k/T_k}{L/T}}$$

(5)

where B is the total information value, L_k, indicates the number of landslide grid cells with the presence of conditioning factor, $T_k$ indicates the number of landslide grid cells with the conditioning factors, L indicates the sum number of landslide grid cells, and T is the total number of grid cells in the study area.

2.7. Frequency Ratio Analysis

The landslide susceptibility map used the Frequency Ratio approach based on the eleven conditioning factors. Each factor was classified into four classes and unified with the spatial distribution of the landslide emergency in the study area. There is no common formula or guideline to obtain the number of classes per criteria; instead, it is based on the review of other studies [29,54,64]. It is the ratio in the area where the landslide has occurred under a subclass of the conditioning factor to the total study area of the similar subclass. However, values > 1 indicate a higher landslide occurrence in the respective class than the frequency in the entire zone [42]. On the other hand, the lower values occur when the frequency of the landslide in the respective class is less than that of the entire map, indicating a low potential. The following formula (Equation (6)) was used in the FR manipulation;

$$\mathrm{F}\mathrm{R}={\displaystyle \frac{({AC}_L/{AtOt}_L)}{(AC/AtOt)}}$$

(6)

where ${AC}_L$ is the area of the landslide in each layer class, ${AtOt}_L$ is the total landslide area of the study map, AC is the total area of the layer class, and AtOt is the total area of the study map.

The Prediction Rate (PR) was calculated for the rating of each conditioning factor using the training datasets. This is because of the Frequency Ratio weakness during the normalization process of assigning equal weights to all conditioning criteria [66]. Thus, the whole process has been obtained by the Prediction Rate formula in Equation (7):

$$\mathrm{P}\mathrm{R}={\displaystyle \frac{({\mathrm{R}\mathrm{F}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{R}\mathrm{F}}_{\mathrm{m}\mathrm{i}\mathrm{n}})}{({\mathrm{R}\mathrm{F}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{R}\mathrm{F}}_{\mathrm{m}\mathrm{i}\mathrm{n}})\mathrm{m}\mathrm{i}\mathrm{n}}}$$

(7)

where PR indicates the Prediction Rate, ${\mathrm{R}\mathrm{F}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ indicates the maximum relative Frequency Ratio, and ${\mathrm{R}\mathrm{F}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$ is the minimum relative Frequency Ratio.

Ultimately, multiple maps of the variables that affect landslides are combined in the overlay process. The Landslide Susceptibility Index (LSI) value is obtained by adding the FR values on each controlling factor map, Equation (8).

$$\mathrm{L}\mathrm{S}\mathrm{I}={\mathrm{F}\mathrm{R}}_1+{\mathrm{F}\mathrm{R}}_2+{\mathrm{F}\mathrm{R}}_3+{\mathrm{F}\mathrm{R}}_4+\dots +{\mathrm{F}\mathrm{R}}_{\mathrm{n}}$$

(8)

The landslide map is then classified using the Natural Breaks (Jenks) approach based on its values. The approach produces the map, which is interpreted by showing the spatial distribution of the predicted landslide varying from low to high landslide events.

2.8. Analytic Hierarchy Process (AHP)

Ref. [67] is the creator of the Analytic Hierarchy Process (AHP), one of the multi-criteria decision-making techniques. It is, in essence, a technique for generating ratio scales from paired comparisons. The input can come from objective judgment, such as satisfied emotions and preferences, or objective measurement, such as weight, etc. Since people are not always consistent, AHP permits a little judgmental inconsistency. The consistency index is obtained from the principal Eigenvalue, whereas the ratio scales are derived from the principal [66].

Refs. [68,69] demonstrated that for consistent reciprocal matrices, the greatest Eigenvalue is equal to the size of the comparison matrix, or ${\lambda }_{max}=n$. Then a consistency index was provided, which is a deviation or degree of consistency, using the formula below;

$$\mathrm{Consistency}\mathrm{Index}={\displaystyle \frac{{\lambda }_{max}-n}{n-1}}$$

(9)

According to [56,67,68,69,70,71,72,73,74,75,76,77], this index should be applied by contrasting it with the appropriate one. The appropriate consistency metric is the Random Consistency Index (RI). After bootstrapping, the random consistency index is assessed to see if it is about 10% or less. The average random consistency index of sample size 500 matrices is shown in

Table 2. Random Consistency Index (RI).

n	1	2	3	4	5	6	7	8	9	10
Rn	0	0	0.58	0.9	1.12	1.24	1.32	1.41	1.45	1.49

.

$$\mathrm{Consistency}\mathrm{Ratio}={\displaystyle \frac{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{y}\mathrm{I}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}\left(\mathrm{C}\mathrm{I}\right)}{\mathrm{R}\mathrm{a}\mathrm{n}\mathrm{d}\mathrm{o}\mathrm{m}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{y}\mathrm{I}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}\left(\mathrm{R}\mathrm{I}\right)}}$$

(10)

The inconsistency is acceptable if the Consistency Ratio value is less than or equal to 10%. If the Consistency Ratio exceeds 10%, the subjective assessment requires updating.

2.9. Model Evaluation Procedures

This study utilizes diverse metrices during the evaluation of the model performances such as specificity, sensitivity, AUC, and ROC, as expressed in the following equations:

$$Specificity={\displaystyle \frac{TruePositive}{FalsePositive+TruePositive}}$$

(11)

$$Sensitivity={\displaystyle \frac{TruePositive}{FalseNegative+TruePositive}}$$

(12)

$$Accuracy={\displaystyle \frac{TrueNegative+TruePositive}{FalsePositive+TruePositive+FalseNegative+TrueNegative}}$$

(13)

$$AUC=\left({\displaystyle \frac{\sum TruePositive+\sum TrueNegative}{T_p+T_N}}\right)$$

(14)

where $T_p$ and $T_N$ represent the total number of pixels with and without torrential events, respectively.

3. Results

3.1. Conditioning Factor Determination

Subsequently, each conditioning factor underwent a rigorous pre-processing workflow to ensure spatial and thematic consistency. This process encompassed several critical stages, including geospatial rectification, attribute refinement, and spatial normalization of the study area. A total of 70% of the training datasets were used to calculate the Frequency Ratio (FR) and the Prediction Rate (PR) for each class in the conditioning factors. Further, the susceptibility of the unit grids was extracted, and the value 1 was assigned to the areas with a high probability of landslide occurrence. However, the grid units were randomly selected, and the non-landslide grids were selected and assigned a value of 0, denoting the steady areas.

Table 3. Landslide grid cells and conditioning factors of FR and PR values.

Conditioning Factors	Class	Proportion of Grid in the Whole Area (%)	Grids in Landslide	Percent	FR	PR
	Flat	45.32	32,123	8.13	0.0185
	North	5.89	2305	4.53	0.0102
	Northeast	6.70	2378	4.67	0.0092
	East	8.95	2927	25.75	0.0085
Aspect	Southeast	2.88	825	1.62	0.0074	1
	South	5.75	2064	4.06	0.0093
	Southwest	6.57	2477	4.87	0.0098
	West	6.05	2088	4.10	0.0090
	Northwest	8.97	2875	5.65	0.0083
	North	2.91	820	1.61	0.0073
	0–2000	26.99	3758	7.39	0.0036
	2000–4000	21.86	15,731	30.92	0.0187
Distance 2 Fault (m)	4000–6000	20.14	24,024	47.22	0.0310	3.65
	6000–8000	18.55	4022	7.90	0.0056
	>8000	12.46	3347	6.58	0.0070
	0–200	28.53	4033	7.93	0.0037
	200–500	27.49	3885	7.64	0.0037
Distance to River (m)	500–800	21.29	22,732	44.68	0.0278
	800–1100	14.15	15,949	31.35	0.0293	2.9
	>1100	8.55	4283	8.42	0.0130
	0–300	56.97	27,514	54.07	0.0126
	300–1300	28.18	18,952	37.25	0.0175
Distance to Road (m)	1300–2300	9.94	4416	8.68	0.0116	1.25
	2300–3300	3.67	0	0	0
	>3300	1.24	0	0	0
	0–1509	15.37	19,467	38.26	0.0328
	1509–1609	34.26	25,282	49.69	0.0191
	1609–1710	31.14	1079	2.12	0.0009
Elevation	1710–1894	16.69	5025	9.88	0.0078	4.64
	1894–2203	1.52	29	0.06	0.0005
	2203–2659	0.70	0	0.00	0
	2659–3402	0.32	0	0.00	0
	B-L/M	55.39	33,158	65.17	0.0155
	U-L/H	25.94	16,481	32.39	0.0214	4.48
Geology	I-L/M	18.67	1243	2.44	0.0017
	Wetland	0.41	3792	7.45	0.0204
	Water	1.81	2386	8.20	0.0115
	Forest	2.61	8412	5.53	0.0133
	Agriculture	53.62	12,707	24.97	0.0146
LULC	Woodland	16.40	2103	4.13	0.0210	1.03
	Built-up	20.12	23,868	20.45	0.0282
	Shrub land	2.43	8918	6.15	0.0160
	Grassland	1.13	1034	5.18	0.0156
	Bare land	1.46	12,998	18.36	0.0181
	<0.113	2.44	1252	2.46	0.0133
	0.113–0.403	7.00	6919	13.60	0.0256
NDVI	0.403–0.548	18.29	11,630	22.86	0.0165	3.01
	0.548–0.659	35.13	17,190	25.78	0.0127
	0.659–0.867	37.13	13,891	27.30	0.0097
	0–811	29.65	1709	3.43	0.0016
	811–859	28.14	16,606	33.37	0.0159
Precipitation (mm)	859–915	17.33	16,973	34.11	0.0263	2.58
	915–999	22.11	11,765	23.65	0.0143
	999–1192	2.76	2703	5.43	0.0264
	1–2	54.50	43,245	10	0.0206
	2–7	34.59	7417	14.58	0.0056
Slope	7–14	8.07	196	0.39	0.0506	6.58
	14–27	2.14	24	0.05	0.0003
	27–72	0.70	0	0.00	0
	−8.961 to −4.671	29.54	6199	12.18	0.005
	−4.671 to −3.134	39.11	18,884	37.11	0.013
TWI	−3.134 to −1.110	18.65	10,739	21.11	0.015	3.44
	−1.110 to 1.641	8.22	7283	14.31	0.023
	1.641 to 11.758	4.48	7777	15.28	0.045

shows the percentage obtained per class for the landslide occurrence and the Prediction Rate of each criterion, which was used to provide the weightage of the influencing factor to be used for comparison and production of the Landslide Susceptibility Index. The slope is the most affecting factor with the high Prediction Rate, followed by the elevation of an area, whose PR = 4.64, experiencing a high landslide susceptibility between 1509–1609 m with FR = 0.32, compared to the rest of the factors. Slope influences shear and normal stress on shear surfaces and is a key component of slope instability investigations, serving as the foundation for stability analysis. The way that materials move as a result of gravity is largely determined by slope.

3.2. Multicollinearity Analysis of Conditioning Factors

The VIF and TOF matrices were used in the analysis of the multicollinearity landslide conditioning criterion in the study area. Figure 5 shows that the slope has the highest value of VIF (5.662) with the lowest value of TOL (0.1766). However, the obtained values do not exceed the conforming thresholds, that is VIF > 10 and TOL < 0.1, which denotes that there is no multicollinearity among the selected conditioning criteria. Thus, all eleven conditioning criteria were then used for landslide susceptibility analysis with the AHP and FR models.

3.3. Importance of Conditioning Factors

The spherical plots illustrate the eleven conditioning factors and their contributions to landslide susceptibility (Figure 6). On each axis, the star point plots the normalized contribution (weight) of that factor derived from the informative values in Equation (6) and model-based importance while the shaded (colored) polygon connects these values so that larger radii and broader filled sectors indicate stronger influence and smaller sectors indicate weaker effects. Consistent with geomorphic process understanding, slope gradient exerts the greatest impact because steeper inclinations increase shear stress and gravitational driving forces, elevating failure risk; notable contributions also arise from elevation, geology, distance to faults, TWI, and NDVI. The ranking of factor importance was obtained from the informative values (Equation (6)) within the FR framework and was confirmed by Prediction Rates computed on the training dataset. To address multicollinearity, inputs were adjusted across models: slope was omitted in the FR implementation and precipitation was omitted in the AHP, while RF retained the remaining factors after collinearity checks.

3.4. Landslide Susceptibility Based on Frequency Ratio, AHP, and LSI

The landslide susceptibility results were used to demarcate the spatial distribution of the predicted future events in the study area. The results were obtained using conditioning factors such as aspect, distance to fault, distance to the river, distance to road, distance to road, elevation, geology, land use, normalized difference vegetation index, precipitation, slope, and topographical wetness index. Both models were used to produce the final results based on the conditioning factor importance attained in the Prediction Rates for FR in Equation (6), AHP from Equation (7), and LSI from approach Equation (8). Table 3 illustrates that slope, elevation, geology, distance to fault, topographical wetness index, and the normalized difference vegetation index are the highly affecting factors triggering landslide occurrence in the future, with Prediction Rates of 6.58, 4.64, 4.48, 3.65, 3.44, and 3.01, respectively.

However, the production of the landslide susceptibility map followed the basic guides from Equation (6), in which the final output was then reclassified into four classes; low, moderate, high, and very high using the Natural Break method commonly known as Jenks. Figure 7 shows the landslide susceptibility map based on the PR values in Table 3. In this study, the low susceptibility covered an area of 584.43 km², 1116.45 km² for moderate susceptibility, 884.69 km² for high susceptibility, and 947.21 km² for very high susceptibility of the whole area (

Table 4. Area, percent, and number of pixels in landslide suspected classes.

Landslide Susceptibility Class	Number of Pixels	Area (Km2)	Percentage (%)
Low	612,788	584.43	16.54
Moderate	1,170,629	1116.45	31.60
High	927,626	884.69	25.04
Very high	993,182	947.21	26.81

).

3.5. Model Validation and Comparison

As specified earlier, the Receiver Operating Characteristic curve was applied in this study based on the Area Under the Curve to assess the produced results. From the training datasets, the specificity and sensitivity curves were utilized for validation of the obtained results while the remaining validation datasets were used for the derivation of the prediction curve (Figure 6). Based on the AUC curve, an accuracy of above 0.75 was obtained, which is the highest value in this statistical method. This value is higher and acceptable, indicating the proper selection of the conditioning factors. Similarly, lower values indicate the improper selection of the conditioning factors or the model drawbacks (model limitations).

For the AHP model, all the factors were used for training except precipitation as it ranked with the highest position based on the VIF metric. Meanwhile, in the FR model, the slope was excluded due to the high-ranking VIF during the training processes. Figure 8 shows that the FR model is lower than the AHP model, indicating that the slope has a sufficient effect on the model performances and also denoting the multicollinearity of the omitted factor while the Landslide Susceptibility Index is lower than any other model. This happens after the removal of the two factors (i.e., precipitation and slope) which depicts a highly significant effect on the production of the LSI. Additionally, given that the AHP model obtained a high AUC value of 88.76%, the performance of the models is deemed to be good. Since an AUC value of 80% is considered exceptional in landslide susceptibility mapping, this suggests a strong predicting capacity. In contrast, the FR model, which had an AUC of 78.86% models, and the LSI, which had an AUC of 80.67%, both had a good performance but not as well as AHP.

In comparing the models, both of them seem to have different values of increment and decrement on the landslide susceptibility levels. However, AHP covered 1470.12 km² in the high class compared to the FR model (884.69 km²) while the low class held only 16.54% for FR and 1.84% for the AHP model of the landslide susceptibility (Figure 9). During the AHP calculation procedures, the model showed that the Consistency Index is 0.0347 and the Consistency Ratio is 0.0547, which indicates that the obtained pairwise matrix has a reasonable and acceptable level of consistency. Thus, the results of the AHP were then reclassified using the Natural Break (Jenks) method to produce four classes (

Table 5. Comparison between models per class and LSI.

Landslide Susceptibility Class	Number of Pixels	Area (Km2)	Percentage (%)	Model
Low	612,788	584.43	16.54	FR
Moderate	1,170,629	1116.45	31.6
High	927,626	884.69	25.04
Very high	993,182	947.21	26.81
Low	68,084	64.93	1.84	AHP
Moderate	1,361,488	1298.47	36.73
High	1,541,466	1470.12	41.58
Very high	736,129	702.06	19.86
Low	594,818	567.29	16.06
Moderate	1,263,950	1205.45	34.12	LSI
High	681,004	649.48	18.38
Very high	1,164,453	1110.56	31.44

). Moreover, the final results from the FR model have shown more accuracy than LSI, as the very high class on the landslide susceptibility map has been proportional to the landslide occurrence shown on the inventory map (Figure 2).

The LSI values have been calculated from the PR and FR of the eleven conditioning factors (Table 3). The results were obtained by weighting the susceptibility maps of the conditioning factors to produce a prediction map. The produced map was then reclassified into four classes using the Natural Breaks (Jenks) method i.e., low, moderate, high, and very high, using the ArcGIS software. Based on the LSI, the final prediction map indicated that only 16.06% was for low, 34.12% for moderate, 18.38% for high, and 31.44% for very high class in the spatial distribution of the landslide events (Table 5 and Figure 9).

4. Discussion

Landslide susceptibility mapping has made identifying and managing the landslide-prone areas difficult. Several researchers and scholars have utilized various kinds of approaches for landslide susceptibility mapping and compared their performances to obtain the best technique for their area of interest [15,22,41]. This study compares two approaches and LS index: AHP, FR, and the Landslide Susceptibility Index in the study area, which has different features and geomorphologies (Figure 10). The standard basis for the AHP model is a rating system that is derived from expert judgment. By employing pairwise comparisons to systematically rank components, this approach aids in the creation of an extensive susceptibility map [24,71]. When it comes to resolving complicated issues like landslides, professional input is quite helpful. To a certain degree, each expert’s ideas may vary, making them susceptible to subjectivity and cognitive difficulties. The FR calculates the ratio of landslide occurrences to the total number of occurrences in the area for each factor class, thereby assessing the spatial link between past landslides and causative causes, allowing the quantitative assessment of each contributing factor while the LSI combines several landslide-triggered criteria into a fused index by ranking the area based on the relative susceptibility occurrence [25,60]. The Landslide Susceptibility Mapping analysis has been done considering both environmental and human-induced factors, and the evaluation is measured using the matrices validated from the historical landslide inventory map in the study area. So far, the accuracies of the results have been obtained from splitting the dataset into sampling and testing parts to assess the matrices [16,72]. Selecting various landslide-influencing criterion that plays an important role in susceptibility mapping is the next step after constructing the sample data. It is therefore very important to assess the contribution of each factor in the study area. Initially, the test between the independent factors in the study region was conducted. The findings of the experiment show that the 11 factors that were chosen are independent of one another and that the absolute values of all the correlation coefficients between these factors are below the critical threshold. The second method used to determine the connection between these parameters was multicollinearity analysis. The PR value derived by the slope factor is larger than the other factors, suggesting that this factor is the primary source of landslides in the study area, which is consistent with earlier research [19,31,33].

One of the most interesting findings of the study area’s landslide susceptibility investigation was the absence of landslides in the mountains, despite the widespread belief that mountainous regions are frequently more susceptible to landslides. This finding challenges the commonly accepted belief that steeper slopes and higher elevations are more vulnerable to landslides due to gravity and terrain instability [35,73]. Upon careful examination, many factors unique to the study site were identified as potential causes of this abnormality. First, the soil composition and hard rocks in mountainous areas are more resilient to erosion and degradation, giving the slopes inherent stability. Second, the land cover and land usage in these regions reduce the possibility of landslide occurrences. For instance, a mountain covered with dense vegetation or unaltered forests may serve as a natural barrier, keeping the soil in place and halting landslides caused by surface runoff. Nevertheless, landslides have been observed more commonly in lower-elevation regions, where human activities such as urbanization, deforestation, and agriculture have disrupted the landscape’s natural equilibrium by destroying habitats, altering water cycles, and degrading soil.

To ascertain efficiencies and contrast the outcomes of Landslide Susceptibility Mapping, the success rate and PR curves were employed. The landslide-prone maps, both FR, AHP, and LSI have shown superior results than earlier studies conducted by various Tanzanian researchers [25,49,56]. Furthermore, given that the AHP model obtained a high AUC value of 88.76%, the performance of the models is deemed to be good. Since an AUC value of 80% is typically regarded as outstanding in landslide susceptibility mapping, this suggests a strong predictive capacity. In contrast, the FR model, which had an AUC of 78.86% model, and the LSI, which had an AUC of 80.67%, both did rather well, nonetheless not quite as well as the AHP model. These findings demonstrate that while all models produce insightful forecasts, the AHP model seems to be the most trustworthy. The disparity in performance indicates that AHP provides superior predictions by handling the weighting and integration of different landslide-triggering criteria more skillfully. While the FR and LSI offer reliable but less precise forecasts, the findings show that AHP is a good option for this application.

However, the findings are consistent and recognizable under similar conditions. This approach is consistent with accepted standards, and prior research has produced similar results using the same techniques [25]. While minor deviations may occur due to differences in datasets or environmental circumstances, overarching patterns and results are expected to be similar across research that uses the same methodology and parameters [25,44,66,74]. Table 3 indicates that the study region, a geological structure with an elevation of 1700 m and a slope above 14 degrees, is particularly susceptible to landslides. By taking into consideration the development of various machine learning models, the AHP model still performs better in the area as it involves expert judgments, human preferences, and multi-criteria decision-making in a subjective decision context compared to FR, LSI, and machine learning models, which suffer from overfitting issues, bias, and data quality problems [62,75,76]. Landslides are particularly common along the lineaments. In this study, the landslide susceptibility zonation (Figure S1), slope, elevation, and lithology seemed to be the main influencing factors, followed by distance to fault and topological wetness index (TWI) as the top five influencing factors. Meanwhile, various works and studies have found a high linkage between landslide occurrences and influencing factors such as NDVI, precipitation, distance to river, and LULC [16,60,76,77], but this study did not find any linkage between the landslide occurrence and the influencing factors as mentioned above. Here, intense commercial agriculture is practiced on small terraces. Most settlements are found to be in dangerous (very high) risk areas. Monitoring and reducing landslide incidents in the geographic range and the surrounding area requires site-specific slope control. The type that was found to be the most vulnerable was sandy clay with the highest Frequency Ratio (Figure 9). In addition, the research area’s precipitation patterns might be less concentrated in hilly areas, which reduce the probability of lithology deterioration, slope, and water saturation which are major causes of landslides. The soil type, vegetation cover, and localized precipitation patterns cooperate to maintain the stability of the mountain slopes.

The study is important because it directly contributes to environmental preservation, disaster risk reduction, and sustainable land use planning. Human communities, infrastructure, and ecosystems are seriously threatened by landslides, which are commonly caused by strong rainfall and seismic activity. The creation of precise landslide susceptibility models is essential as climate change is making these risks worse. By showing that the AHP model performs better than FR (78.86%) and LSI (80.67%), this study offers crucial information on which approaches best capture the many processes causing landslides. AHP’s ranking as the best model aids in setting priorities for decisions about catastrophe mitigation and prevention, particularly in susceptible areas.

Through the methodical comparison of several models in the study region, this research enhances comprehension of how heuristic and empirical approaches can be improved to increase prediction accuracy. In this study, factors showing multicollinearity were removed based on VIF and TOL criteria to ensure model stability and interpretability. While this approach reduces redundancy, it may also exclude interactions among variables that can influence landslide occurrence. Alternative statistical approaches such as Principal Component Analysis (PCA), partial least squares regression (PLSR), or regularization methods, Factor Analysis, or hybrid multi-criteria decision-making methods, and high-quality data into hazard mapping could allow correlated factors to be incorporated without compromising model performance.

In general, this study enhances the connection between predictive modeling and practical applications, opening the door for more robust infrastructure building and improved adaptation tactics in the face of natural disasters to prevent high landslide occurrence. Besides causing deaths and financial losses, the collapse of unstable slopes in dangerous locations has catastrophic social and environmental repercussions. Finding locations that are susceptible to landslides is a crucial resource for local authorities, planners, disaster response teams, and decision-makers.

5. Limitations and Future Directions

This study employs the Frequency Ratio (FR), Analytic Hierarchy Process (AHP), and Landslide Susceptibility Index (LSI) models, each with inherent limitations. AHP weights were derived from the literature and the authors’ expertise, rather than broad expert surveys, potentially reducing precision compared to consensus-based approaches. FR and LSI depend heavily on the quality and representativeness of input data and may not fully capture complex nonlinear relationships among landslide-triggering factors. Key factors such as precipitation, NDVI, and distance to rivers showed limited influence, which may restrict generalizability. The choice of buffer zone distances for sampling non-landslide areas critically affects model validation (Figures S1–S4), and the landslide inventory, based on bare land LULC classes, may overlook smaller, vegetated, or slow-moving landslides, introducing uncertainty.

As noted in the discussion, factors with high multicollinearity were excluded based on VIF and TOL criteria. Although this improved model stability, it may have excluded important interactions among variables, which limits the ability to capture complex landslide-triggering processes fully. Future research should incorporate broader expert input, high-resolution remote sensing, field-validated or officially verified landslide inventories, and optimized sampling strategies to enhance the accuracy, reliability, and transferability of landslide susceptibility assessments.

6. Conclusions and Recommendations

This study successfully achieved its objectives by (i) generating landslide susceptibility maps using AHP, FR, and LSI models; (ii) evaluating and comparing their predictive performance using AUC, CR, and PR metrics; and (iii) assessing the statistical independence of twelve conditioning factors through multicollinearity analysis. The results demonstrate that all models are reliable for landslide susceptibility mapping in Hanang District, with AHP showing the highest predictive performance. The comparative analysis provides a robust understanding of each model’s strengths and limitations, forming a scientific basis for spatial hazard management.

To mitigate the impact of landslides on human life and livelihoods, proactive strategies are essential. These include enforcing land use policies that prevent construction and cultivation on unstable slopes, promoting reforestation and terracing in erosion-prone areas, and developing early warning systems that integrate satellite-based data with local reporting networks. Public education programs should raise awareness of slope failure risks and encourage safe settlement practices, while emergency preparedness plans must ensure rapid response capabilities for at-risk communities.

At local and national levels, planners and policymakers should integrate the study’s susceptibility maps into infrastructure development, land use planning, and zoning regulations. Local governments can use these maps to prioritize intervention areas, invest in slope stabilization, and allocate resources to protect high-risk communities. Nationally, incorporating landslide risk analysis into Environmental Impact Assessments (EIAs) and strategic development frameworks will enhance resilience across multiple sectors. Strengthening inter-agency coordination among urban planning, agriculture, and environmental management bodies ensures coherent and effective hazard mitigation.

For international and regional stakeholders, this study offers a replicable, cost-effective framework for landslide risk assessment, particularly in data-scarce regions. Development partners, donor agencies, and research institutions can support these efforts by providing technical expertise, funding high-resolution remote sensing, and building local geospatial capacity. Promoting cross-border collaboration and establishing open-data sharing platforms further enhances the ability of vulnerable nations to manage landslide risks effectively. Ultimately, a multi-scalar, interdisciplinary approach is essential to integrating geospatial science into global disaster risk reduction and sustainable development planning.