Content area
Background
Landslides pose a significant natural hazard in Northern Thailand, threatening lives, infrastructure, and sustainable watershed management. Developing reliable landslide susceptibility maps for the Northern Chiang Mai Watershed Basins (NCMBs) was therefore crucial for risk reduction and planning.
Objective
This study employed two traditional methods, Analytical Hierarchy Process (AHP) and Frequency Ratio (FR), alongside two advanced deep learning approaches, Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks. The primary objective was to identify areas prone to landslide susceptibility and determine the underlying causative factors that caused landslides in the NCMBs.
Method
A landslide inventory of 1,222 scars collected from 1989 to 2020 was divided into training and validation datasets. The training data and 10 landslide causative factors (grouped as topography, hydrology, geology, and land modification) were used to construct the susceptibility maps. Feature importance, quantified using Mean Decrease Impurity and Information Gain Ratio, revealed that slope, rainfall, lithology, soil texture, and elevation were the dominant causative factors. Susceptibility maps were produced using each model and validated with the historical landslide inventory. Predictive performance was evaluated through the receiver operating characteristic area under the curve (ROC AUC), along with precision, recall, F1-score, and overall accuracy.
Result
The results indicated that 60–70% of the entire watershed areas were classified as moderately to highly susceptible. The deep learning-based models (CNN and LSTM) achieved superior predictive accuracy (AUC > 0.85) compared to FR and AHP. While less accurate, the traditional models remained valuable due to their interpretability and lower data requirements.
Conclusion
The susceptibility maps reveal that highly landslide-susceptible zones are concentrated in steep, lithologically less resistant, and high rainfall-prone areas. Conversely, low-susceptible zones are located in valleys and on gently sloping terrain. This study provides a valuable reference for future research, contributing to the development of more robust landslide early warning systems and informed land-use planning in Northern Thailand and other regions with similar environmental conditions.
Introduction
Landslides are geological hazards defined by the downward movement of materials along a slope (Cruden and Varnes 1996). These devastating disasters are among the primary disasters in mountainous regions worldwide, posing a significant threat to human life and the environment. They cause fatalities, damage property, and disrupt infrastructure (Azarafza et al. 2018, 2021; Dou et al. 2019; Mengistu et al. 2019; Shano et al. 2020). Landslides are typically triggered by a variety of factors, including topography, geology, hydrology, and human activities (Jaafari et al. 2019; Wang et al. 2019). Understanding these causative factors helps create effective predictive models to mitigate the potential consequences of such events. This process, which connects causative factors with landslide occurrences, is known as landslide susceptibility assessment. It indicates the likelihood of a specific area experiencing a landslide (Wu et al. 2015; Mao et al. 2021). Identifying areas with increased susceptibility enables the implementation of proactive strategies in land-use planning, infrastructure development, and disaster preparedness (Mallick et al. 2024; Rihan et al. 2024). Therefore, reliable prediction and efficient management of landslides are crucial for preventing and reducing the associated costs (Wu et al. 2015).
In recent years, researchers have employed both qualitative and quantitative approaches to enhance landslide susceptibility analysis (Guzzetti et al. 1999; Depicker et al. 2020; Cemiloglu et al. 2023). Qualitative methods rely on a knowledge-based strategy, including statistical techniques such as the Analytical Hierarchy Process (AHP) (El Jazouli et al. 2019; Hepdeniz 2020; Panchal and Shrivastava 2022; Gulbet and Getahun 2024; Liu et al. 2024a), Weight of Evidence (WoE) (Saha and Saha 2020; Cao et al. 2021; Getachew and Meten 2021), Frequency Ratio (FR) (Thapa and Bhandari 2019; Zhang et al. 2020; Thongley and Vansarochana 2021; Manopkawee and Mankhemthong 2024, 2025), Evidential Belief Function (EBF) (Mondal and Mandal 2019; Chowdhuri et al. 2020; Roy et al. 2023), Statistical Index Model (SI) (Bui et al. 2011; Berhane and Tadesse 2021), and Information Value (IV) (Sarkar et al. 2013; Farooq and Akram 2021; Niu et al. 2024). These statistical methods estimate the probability of landslides by assigning weights to contributing factors, often based on historical data. They have proven to be reliable and efficient worldwide (Mandal and Mondal 2019; Pasang and Kubíček 2020; Xiao and Zhang 2023).
Although traditional statistical models are widely used, they often struggle with analyzing time-dependent patterns and complex nonlinear relationships between landslides and environmental factors (Wang et al. 2021). Advanced computational techniques, such as Machine Learning (ML) and Deep Learning (DL), offer more accurate and efficient alternatives for spatiotemporal mapping (Guzzetti et al. 1999; Depicker et al. 2020; Cemiloglu et al. 2023). Successful ML applications include Logistic Regression (LR) (Sahana and Sajjad 2017; Basu and Pal 2017; Sujatha and Sridhar 2021), Support Vector Machine (SVM) (Kumar et al. 2017; Huang and Zhao 2018; Roy et al. 2019; Dou et al. 2020), Artificial Neural Network (ANN) (Pham et al. 2017; Bragagnolo et al. 2020; Selamat et al. 2022), Random Forest (RF) (Goetz et al. 2015; Merghadi et al. 2020; Sun et al. 2020), and Decision Tree (DT) (Saito et al. 2009; Pham et al. 2021).
The recent DL models, including Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks, have also proven successful for landslide susceptibility mapping (Wang et al. 2019; Fang et al. 2020; Wei et al. 2021; Zhang et al. 2022; Huang et al. 2023; Zuo et al. 2025). CNNs excel at extracting spatial features and identifying subtle patterns from large geospatial datasets (i.e., satellite imagery, topographic data), which is crucial for precise mapping in complex environments. LSTM networks, a type of recurrent neural network, are designed to learn from and remember long-term dependencies in sequential data, providing a more dynamic and reliable evaluation by analyzing time-varying environmental elements (Huang et al. 2023; Zuo et al. 2025).
Previous studies on tropical basin landslide prediction primarily relied on statistical susceptibility mapping, particularly AHP and FR (Silalahi et al. 2019; Afzal et al. 2022). However, these models often fail to capture both the complex spatial patterns of the landscape and the long-term temporal changes in rainfall and other environmental factors, especially in tropical basins with a scarcity of reliable, detailed information on historical landslide records, surface characteristics under dense vegetation cover, inconsistent or low-resolution topographic data, and limited ground-based data on hydrology and soil properties (Kayitesi et al. 2022). To address these limitations, this study integrates the AHP/FR-derived susceptibility map as an initial spatial weighting layer with both CNN and LSTM (Chiplunkar et al. 2025; Wu et al. 2025). As CNN is effective for extracting spatial features related to slope instability, while LSTM is well-suited for modeling temporal dynamics such as rainfall and soil moisture variations, combining static spatial context with dynamic temporal behavior via the dual use of CNN and LSTM is expected to outperform traditional and single-model methods in landslide susceptibility mapping. Although this integrated approach is novel for improving landslide prediction accuracy in data-limited tropical basins, the comparative studies on evaluating DL models against conventional statistical models in complex Southeast Asian tropical monsoon environments remain underexplored.
A Landslide susceptibility map (LSM) is crucial in Northern Thailand, where the geomorphology, characterized by mountainous terrain, steep slopes, erodible soil, and less-resistant rock, is combined with heavy monsoon rainfall, contributing to varying erosion resistance. Although the Department of Mineral Resources in Thailand has mapped hazard zones, accurate prediction remains hindered by complex natural–human interactions, rapid climatic changes, forecasting uncertainties, low-resolution legacy maps, and delays in early warning systems (Mairaing 2006; Chinkulkijniwat et al. 2022). Chiang Mai Province, the region’s cultural and economic center, is considered the area with the highest landslide risk (Environmental Geology Division 2003). The high likelihood of landslides from May to September is due to the southwest monsoon, intense rainfall, and heavy downpours (Yongsiri et al. 2023). The headwaters of the Ping River are highly susceptible to landslides due to unstable regions resulting from tectonic events, steep hillslopes, underlying less-resistant bedrock, heavy rainfall, and inappropriate land modifications. These landslides damage people living in hilly areas, downstream communities, and infrastructure.
The study aimed to construct LSMs in the Northern Chiang Mai Watershed Basins (NCMBs) using the AHP, FR, and deep learning-based models (CNN and LSTM). The performance and accuracy of the resulting maps will be assessed via the ROC AUC, precision, recall, F1-score, and overall accuracy. These metrics will help evaluate and refine the prediction of landslide susceptibility. This study serves as a reliable reference for future researchers, providing insight into landslide prediction in other parts of Thailand and regions with similar environmental conditions.
Study site
The Northern Chiang Mai Watershed Basins (NCMBs) in Thailand comprise three distinct watersheds: the Mae Taeng River, the Upper Ping River, and the Mae Ngat River (Fig. 1a, b). The NCMBs’ complex landscape has been shaped by tectonic activity, riverine processes, and climatic changes, and the area is already known for its high flash flood susceptibility (Manopkawee et al. 2025). Mae Taeng River originates in the Daen Lao Range (DLR) near the Myanmar border. The river is a major contributor to the Ping River. It flows southeast through the Wiang Haeng Basin (WHB) and Muang Khong Valley (MKV) along the mountainous Thanon Thong Chai Range before merging with the Ping River in the vast, fertile Upper Chiang Mai Basin (UCMB) (Fig. 1c). The river is measured at about 151 km, with a catchment area of approximately 1,900 km2.
[See PDF for image]
Fig. 1
Location map of the North Chiang Mai Watershed Basins (NCMBs). a Location of Chiang Mai Province in Thailand. b The NCMBs are in Chiang Mai Province and include three watersheds: (1) Mae Taeng River, (2) Upper Ping River, and (3) Mae Ngat River. c The figure shows abbreviated names of mountains and valleys with a swath profile (X-X’). d The figure presents the swath profile (x-x’) and the transverse hypsometric integral across the NCMBs. Abbreviations and symbols: DSM-Doi Sam Muen; DLR-Dan Lao Range; WHB-Wiang Haeng Basin; MKV-Mueang Khong Valley; DCD-Doi Chiang Dao; CDP-Chiang Dao Plain; DPD-Doi Pha Daeng; DCH-Doi Chom Hod; PB-Phrao Basin; DJJ-Doi Jik Jong; DML-Doi Mon Lan; UCMB-Upper Chiang Mai Basin
The Upper Ping River is known as the headwaters of the Ping River. Its catchment covers approximately 1,650 km2. This catchment includes the Chiang Dao Plain (CDP) and surrounding mountainous areas with elevations of 1,500 to 2,000 m (i.e., Doi Chiang Dao (DCD), Doi Pha Daeng (DPD), and Doi Chom Hod (DCH)). The river flows south into the UCMB (Fig. 1c). The river joins the Mae Taeng and Mae Ngat rivers to form the larger Ping River, which serves as the central drainage system of the Chiang Mai Basin.
Mae Ngat River is located west of the Upper Ping River. It originates in the highlands of Doi Jik Jong (DJJ) and flows through mountainous terrain and the Phrao Basin (PB) into the Ping River. Its total length is about 95 km, with a catchment area of approximately 1,300 km2. The prominent feature is the Mae Ngat Somboon Chon Dam and reservoir (Fig. 1c). Although its course is mainly natural and surrounded by forests, its hydrology is heavily regulated, which has stabilized the channel, reduced sediment transport, and fostered a fertile alluvial plain downstream.
Topographic variation across the NCMBs is illustrated by the five km-wide swath profile (X-X’) (Fig. 1d), which typically presents a basin and range structure. Two-peaked mountains, Doi Sam Muen (DSM) and DCD, flank the Mae Taeng River watershed, with the V-shaped MKV in the middle. DCD, Thailand’s third-highest mountain, creates a topographic relief of approximately 1,500 m above sea level and bounds the U-shaped CDP. The Mae Ngat River flows through the PB, bounded by DPD and Doi Mon Lan (DML). The rugged, steep terrain, where the rivers originate, is associated with Transverse Hypsometrical Integral (THI) values between 0.35 and 0.6, indicating a mature stage of landscape development (Pérez-Peña et al. 2009).
The lithology across the NCMBs is a complex mix of metamorphic and sedimentary rocks, including Silurian-Devonian carbonaceous phyllite, Carboniferous-Permian sandstone, shale, chert, Middle Permian limestone, and Tertiary/Quaternary unconsolidated sediments. These sedimentary rocks are intruded by large Triassic granite and granodiorite bodies (Braun and Hahn 1976) (Fig. 2a). Moreover, the NCMBs experience a tropical monsoon climate, characterized by significant precipitation variability. The prolonged rainy season runs from approximately May to October, with the heaviest rainfall in August and September, bringing frequent showers and high temperatures (28–32∘C). Conversely, the dry season, from November to April, has little precipitation, with January and February being the driest months (temperatures 24–29∘C). The average annual rainfall ranges from 1,300 to 1,800 mm. The rainfall data had been collected between 1981 and 2020 by the Climate Hazards Group InfraRed Precipitation with Stations (CHIRPS), developed by the Climate Hazards Center at the University of California, Santa Barbara (Funk et al. 2015) (Fig. 2b). The region’s mountains significantly influence localized rainfall as the area often receives more precipitation than the lower areas.
[See PDF for image]
Fig. 2
a Lithologic characteristics of the NCMBs. b Average annual rainfall of the NCMBs. Abbreviations and symbols are Qa, Qt, Qff-Quaternary sediment; Tmm-Tertiary Mae Moh semi-consolidated and unconsolidated sediments; PNg2-Permian Ngao limestone; CPk, CP, C-Carboniferous arkosic sandstone; DC, SD-Lower Paleozoic low-grade metamorphic rocks; O-Ordovician limestone; E-Cambrian quartzite; Trgr-Triassic granite; Trm-Triassic migmatite; PTrv-Permo-Triassic volcanic rocks; Cgr-Carboniferous granite; DCv-Devonian-Carboniferous volcanic rocks; Cb-Carboniferous basalt; PE-Precambrian high-grade metamorphic rocks (Braun and Hahn 1976)
Materials and methods
Four sequential methods were employed to establish LSMs for the NCMBs. (1) Data preparation involved extracting landslide inventory data from historical scars, which were then divided into training and validation datasets. (2) Factor determination was the process of identifying causative factors, which were categorized based on their potential to induce landslide occurrences. (3) Model construction involved integrating inventory data and causative factors to create LSMs using AHP, FR, and deep learning-based (CNN and LSTM) models. (4) Model validation was the final step to validate the models using the area under the receiver operating characteristic curve (ROC AUC) and other validation metrics (Fig. 3).
[See PDF for image]
Fig. 3
Summary flow chart of the landslide susceptibility mapping and validation process: (1) landslide inventory data preparation, (2) determination of causative factors, (3) model construction using statistical and deep learning models, and (4) model validation
Data preparation on landslide inventory data
A landslide inventory map shows locations that have experienced past landslides and are susceptible to future occurrences. In this study, the Department of Mineral Resources, Thailand, collected landslide scars from 1989 to 2020 through remote sensing data extraction (satellite imagery and aerial photographs) and fieldwork. These features were digitized into GIS shapefiles with attribute information (Department of Mineral Resources 2024). To minimize data redundancy and spatial overlap associated with landslide points, an equal number of non-landslide points were randomly selected outside a 500 m buffer zone around known landslide areas. Furthermore, a buffer-based spatial partitioning approach was adopted to split the training and validation datasets, thereby reducing the risk of spatial dependence. The validation set was exclusively selected from regions at least 250 m away from any points in the training dataset (Lucchese et al. 2021; Wang et al. 2024a). This approach ensured a more reliable model evaluation without artificially inflating the area under the curve (AUC) across the study area.
Within the buffered area, the landslide scars were randomly divided into 70% for the training dataset and 30% for the validation dataset using the Subset Features tool in ArcGIS (Fig. 4a). The location of the training and validation datasets near DLR (Fig. 4b) and DJJ (Fig. 4c) was shown on Google Earth Pro 3D. Examples of landslide scars in the NCMBs were illustrated in Fig. 4d and h, including a landslide along road-cut slopes in Wiang Haeng District (Fig. 4d), failure of road networks near MKV (Fig. 4e), a landslide near DML in Phrao District (Fig. 4f), a landslide along the road between Wiang Haeng and Chaing Dao Districts (Fig. 4g), and a landslide at the base of DCD (Fig. 4h).
[See PDF for image]
Fig. 4
a Landslide inventory data from 1989 to 2020 obtained from the Department of Mineral Resources, Thailand. b–c The landslide inventory shows the training dataset (red dots, used for model construction) and validation dataset (yellow dots, used for model evaluation) near DLR and DJJ as visualized on 3D Google Earth Pro, respectively. d–h Examples of landslide scars in the NCMBs
Landslide causative factor (LCF) determination
Landslide Causative Factors (LCFs) are key elements for susceptibility mapping, selected based on landslide type and the distinct features of the study area. In Northern Thailand, landslides are primarily triggered by prolonged heavy rainfall (Chinkulkijniwat et al. 2022). They are also strongly influenced by fractured rock from active faults, steep slopes, high elevations, and land-use changes in mountainous areas (Komori et al. 2018; Chinkulkijniwat et al. 2022; Manopkawee et al. 2023). This study selected 10 LCFs, categorized into four groups: topography, hydrology, geology, and land modification. Together, these factors represented the primary controls on landslide susceptibility in the NCMBs (Table 1).
Table 1. Data sources used for assessing landslide susceptibility in the NCMBs
Data | Categories | Classification scheme | Data Sources (year of data used) | Scale/resolution |
|---|---|---|---|---|
Landslide inventory | – | – | DMR (1989–2020) | – |
Landslide causative factors | Topography | Elevation | Topographic Map, DMR (2000) DEM from ALOS PALSAR (2011) | 1:50,000 12.5 × 12.5 m |
Terrain slope | DEM from ALOS PALSAR (2011) | 12.5 × 12.5 m | ||
Profile curvature | DEM from ALOS PALSAR (2011) | 12.5 × 12.5 m | ||
Hydrology | Average annual rainfall | CHIRPS (1981–2020) | 12.5 × 12.5 m | |
Drainage proximity | DEM from ALOS PALSAR (2011) | 12.5 × 12.5 m | ||
Geology | Lithology | Thailand geologic map, DMR (1995) | 1:250,000 | |
Soil texture | Soil series, LDD (2024) | 1:25,000 | ||
Lineament density | Thailand geologic map, DMR (1995) DEM from ALOS PALSAR (2011) Landsat 8 OLI/TIRS (2024) Sentinel-2 (2024) | 1:250,000 12.5 × 12.5 m 30 × 30 m 10 × 10 m | ||
Land modification | Land use and land cover | Land use series, LDD (2020) ESRI LULC maps (2017–2024) | 1:50,000 10 × 10 m | |
NDVI | Sentinel-2 vegetation index image series (2024) | 10 × 10 m |
For the sources of the LCFs, continuous factors such as elevation, slope, profile curvature, drainage proximity, and lineament density were derived from the 12.5 m ALOS PALSAR digital elevation model (DEM) (Laurencelle et al. 2015). Average annual rainfall was derived from CHIRPS (Funk et al. 2015). The Normalized Difference Vegetation Index (NDVI) was obtained from the Sentinel-2 vegetation index series to capture regional vegetation conditions. These factors were classified using the natural break classification method. Discrete factors such as lithology, soil texture, and land use and land cover classifications were based on expert knowledge and prior studies (Intarawichian and Dasananda 2010; Intarat et al. 2024; Manopkawee and Mankhemthong 2024, 2025) (Table 1).
Topographical-related factors
Elevation
Elevation is a topographic causative factor that directly influences the stability of hills and mountains by affecting vegetation growth (Ayalew and Yamagishi 2005; Reichenbach et al. 2018). It also impacts weathering processes, the downslope movement of soil and rocks, and water runoff paths. Across the NCMBs watershed, elevation ranged from 289 to 2144 m. It was categorized into five classes using the natural break classification method: 289–566 m, 567–793 m, 794–1019 m, 1020–1279 m, and 1280–2144 m (Fig. 5a). Higher elevations, which typically coincide with steeper slopes, are generally more susceptible to landslides (Niu et al. 2018; ÇELLEK 2023).
[See PDF for image]
Fig. 5
Landslide causative factors: a elevation, b terrain slope, c profile curvature, d average annual rainfall, e drainage proximity. Landslide causative factors: f lithology, g soil texture, h lineament density, i land use and land cover, j NDVI
Terrain slope
Terrain slope is a key factor contributing to landslide occurrence (Van Den Eeckhaut et al. 2006; Xie et al. 2017). As the slope gradient steepens, the shear stress that causes material to slide increases (Ayalew and Yamagishi 2005; Nakileza and Nedala 2020). Landslide potential increases when the slope angle exceeds the material’s angle of repose, reducing its stability. The slope’s gradient was calculated using a topographic toolset in ArcGIS with a three-by-three-cell moving window. Due to the significant variation in the NCMBs, the terrain slope was categorized into five classes using the natural break classification method: flat (0–8.8o), gentle (8.9o–16.9o), moderate (17.0o–24.4o), steep (24.5o–33.1o), and very steep (33.2o–82.8o) (Fig. 5b). Steeper slopes significantly increase the likelihood of landslides compared to gentle slopes (Dai et al. 2001; Lee et al. 2004; Sarkar and Kanungo 2004; Rozos et al. 2011; Papathanassiou et al. 2013).
Profile curvature
Profile curvature is the curvature in the downslope direction, measured along the intersection of a vertical plane and the ground surface (Dikau 1989; Moore et al. 1993a, b; Ayalew and Yamagishi 2004; Ohlmacher 2007). Concave slopes (negative curvature) are more susceptible to landslides because they tend to accumulate surface water and debris, which increases pore water pressure and reduces soil strength. Conversely, convex slopes (positive curvature) tend to divert water, reducing the likelihood of a landslide. Planar surfaces may exhibit a higher landslide probability in areas dominated by debris and clayey soils Profile curvatures were derived from the DEM layer using SAGA-GIS software and categorized into three types: flat, convex, and concave (Fig. 5c). For landslide susceptibility, concave slopes are generally the most susceptible, followed by convex and then flat slopes (Yilmaz et al. 2012; Asmare 2023; Tukku et al. 2025).
Hydrological-related factors
Average annual rainfall
Rainfall is a key causative factor in the initiation of landslides, as extensively documented in numerous studies (Semnani et al. 2025; Xie et al. 2025; Zhao et al. 2025). Rainfall triggers slope failure by saturating the soil, thereby increasing soil weight, raising pore water pressure, and reducing soil shear strength (Pham et al. 2015). Consequently, areas receiving higher rainfall are particularly susceptible to landslides. The volume, intensity, and duration of rain are all critical factors in assessing vulnerability (Wang et al. 2024b). The spatial map of rainfall was created using average annual rainfall data collected between 1981 and 2020, provided by CHIRPS (Funk et al. 2015). The spatial rainfall map was created using average annual rainfall data from 1981 to 2020, provided by CHIRPS (Funk et al. 2015). Using the natural break classification method, the data were categorized into five groups: 1,301–1,408 mm, 1,409–1,466 mm, 1,467–1,526 mm, 1,527–1,590 mm, 1,591–1,750 mm (Fig. 5d).
Drainage proximity
A river and its tributaries can cause the erosion of streambanks and the scouring of the base of sloped hills. The stability of these slopes is negatively correlated with their proximity to the drainage system: the closer the slope, the greater the basal erosion and the saturation of the hillside material (Zhao et al. 2019). The drainage network was extracted using a built-in script from the ArcGIS hydrology toolset. A drainage proximity map was then generated by buffering the network at equal intervals of 250 m. Consequently, the spatial variation was categorized into five groups: 0–250 m, 251–500 m, 501–750 m, 751–1000 m, and over 1001 m (Fig. 5e).
Geological-related factors
Lithology
Lithology refers to the physical characteristics of rocks, including color, texture, composition, grain size, strength, porosity, and permeability (Grana and Della Rossa 2010; Liao et al. 2022). Due to these varying physical properties, different lithological units consequently exhibit varying levels of landslide susceptibility (Machay et al. 2023). The lithologic map was derived from a geological map provided by the Department of Mineral Resources, Thailand, and converted to raster format using ArcGIS. For the study area, lithology was divided into five categories: unconsolidated sediment deposits, limestone, clastic sedimentary rocks, low-grade metamorphic rocks, and granite to basic igneous rocks (Fig. 5f).
Soil texture
Soil forms from the disintegration of rock and contains varying proportions of gravel, sand, silt, clay, and organic matter. The composition of these soil materials influences their strength, moisture content, and viscosity, which in turn affects slope stability and potential for sliding (Komadja et al. 2021; Liao et al. 2022). A soil series map was derived from soil characteristics and properties in the NCMBs, which were collected by the Land Development Department (2024) and converted to raster format using ArcGIS. Based on location and drainage characteristics, soil texture was classified into five categories: well-drained sand, moderately-drained sand, poorly-drained sand, poorly-drained silt and clay, and soil on steep slopes (Fig. 5g).
Lineament density
The NCMBs are traversed by two moderately to highly active fault zones: the Mae Tha and Wiang Haeng Fault zones (Department of Mineral Resources 2023). While infrequent seismic events are unlikely to cause significant landslides (Jibson and Tanyas 2020), the rock exposure adjacent to these faults has led to the formation of lineaments such as joints and fractures. Lineaments appear as straight topographic features (Park et al. 2013; Mathew and Ariffin 2018; Murasingh et al. 2018). The presence of these fractured materials weakens the rock mass and provides pathways for water infiltration, leading to slope instability and potential failure. Lineaments within the NCMBs were collected from the geological structure data of the Department of Mineral Resources, Thailand. Finer lineaments were additionally derived using image-filtering techniques and channel flow analysis from Landsat 8, Sentinel-2 imagery, and DEM, utilizing PCI Geomatica (Howard 1967; Mejia and Niemann 2008; Es-Sabbar et al. 2020; Prabowo et al. 2021). Non-geological features were manually removed, broken lines were connected, and the lineaments were validated for geological or geomorphic significance. Lineament density was calculated using the equation provided by Liu et al. (2012) as
1
where is the lineament density in km/km2, is the total length of all lineaments within a given cell in km, and is the area of the cell in km2. Variation in density was categorized into five classes based on the natural break classification method: 0–0.19, 0.20–0.37, 0.38–0.55, 0.56–0.79, and 0.80–1.48 km/km2 (Fig. 5h).Land modification-related factors
Land use and land cover
Land use refers to the modification of natural environments, and land cover refers to the physical material covering the Earth’s surface. Changes in LULC could lead to the degradation of vegetation and soil, negatively impacting slope stability and increasing landslide possibility (Hamedi et al. 2022; Pacheco Quevedo et al. 2023). Areas with reduced vegetation cover are generally more prone to landslides. LULC data for the study area were compiled from surveys by the Land Development Department of Thailand and the 10 m Land Use and Land Cover maps that were collected between 2017 and 2024 (Karra et al. 2021). The spatial distribution of LULC was categorized into five classes: trees and forests, water, shrubland, built-up areas, and agricultural land (Fig. 5i).
Normalized difference vegetation index (NDVI)
The Normalized Difference Vegetation Index (NDVI) evaluates the vitality and health of vegetation across a watershed. NDVI can indicate ground surface disturbance, a common precursor or result of landslides, because slope failure often converts vegetated areas into bare soil, causing a decrease in NDVI values. NDVI was calculated for each pixel in a Sentinel-2 satellite image using the normalized ratio of the difference between the near-infrared (NIR) wavelength (Band 8) and the red (RED) wavelength (Band 4) (Chen et al. 2017). The spatial distribution of NDVI values was categorized into five classes based on the natural break classification method (Aquino et al. 2018): <0 (bare soil), 0–0.2 (sparse vegetation), 0.2–0.4 (moderate vegetation), 0.4–0.6 (high vegetation), and > 0.6 (dense vegetation) (Fig. 5j).
Model construction
Analytical hierarchy process (AHP)
The Analytical Hierarchy Process (AHP) is a semi-quantitative mathematical method used for multiple-criteria decision analysis (MCDA). AHP simplifies complex problems by breaking them into a hierarchy of elements, facilitating comparative assessments, and synthesizing relative significance or ranking (Saaty 1977, 1980, 2008; Yalcin 2008; Ransikarbum and Mason 2016; Zangmene et al. 2023). In this study, the AHP procedure consisted of three steps: (1) construct the pairwise comparison matrix for each landslide triggering factor, (2) determine the weight of each factor, (3) verify accuracy and reliability using the Consistency Ratio (CR) (Saaty 1977). The classification of each factor is evaluated in pairs, and a pairwise matrix is created as shown below:
2
3
where indicate the comparison outcome of the ith factor relative to the jth factor. The values range between 1 and 9 based on the pairwise comparison scale in Table 2.
Table 2. The pairwise comparison scale for the AHP
Importance scale | Definition | Explanation |
|---|---|---|
1 | Equal importance | Two elements contribute equally |
3 | Moderate importance | Judgment and experience slightly favor one element over another |
5 | Strong importance | Judgment and experience strongly favor one element over another |
7 | Very strong importance | One element is favored very strongly over another |
9 | Extreme importance | The evidence favoring one element over another is of the highest possible |
2, 4, 6, 8 | Intermediate values between two adjacent decisions | Used for compromise between two judgements |
The consistency of the weights attributed to relative importance in the pairwise comparison can be evaluated via the consistency ratio (CR) as (Rodriguez 2002)
4
where is the consistency index, and is the random consistency index (Saaty 1980). is computed as:
5
where is the maximum eigenvalue of the matrix, is the order of the matrix (Rodriguez 2002). The Randomness Index (RI) values depend on the number of elements, denoted as . The RI is derived from comprehensive experiments on a large dataset. Table 3 presents the RI corresponding to various values. Consistency in the pairwise comparison is verified using the Consistency Ratio (CR). The comparison is considered consistent if the CR value is 10% or lower. If the CR exceeds 10% the solution is deemed inconsistent, and the weights in the pairwise comparison matrix must be reassigned. The Landslide Susceptibility Index (LSI) is calculated using the following equation:
Table 3. Random consistency index (RI) (Saaty 1980)
n | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
RI | 0 | 0 | 0.58 | 0.89 | 1.12 | 1.24 | 1.32 | 1.41 | 1.45 | 1.49 | 1.52 | 1.54 |
6
where is the rating or score of each factor class, and is the weight assigned to each factor. A map of landslide susceptibility was created using these LSI values. The accuracy and dependability of the produced map were evaluated using the validation method.
Frequency ratio (FR)
The Frequency Ratio (FR) method is a bivariate statistical technique used to assess landslide susceptibility (Lee and Sambath 2006; Yilmaz 2009; Park et al. 2013; Regmi et al. 2014; Yang et al. 2016; Khan et al. 2019). FR quantifies the relationship between landslide occurrence and causative factors by calculating the ratio of the area where landslides have occurred within a specific factor class to the total area of that class (Lee and Sambath 2006; Pradhan and Lee 2010; Manopkawee and Mankhemthong 2025). The training datasets were combined with LCFs and calculated as:
7
where is the frequency ratio of class of parameter . i is the number of pixels with landslides within class of parameter variable , and is the number of pixels within parameter variable , is the number of classes in the parameter variable , and is the number of parameters in the study area (Regmi et al. 2014). The Tabulate Area tool in ArcGIS was used to intersect the landslide causative factors (LCFs) with the landslide training dataset to determine the number of landslide pixels within each factor class. The FR values were then normalized to a probability range of 0 to 1, expressed as the Relative Frequency (RF):8
Since RF still treats all causative factors with equal importance, the Prediction Rate (PR) was introduced to examine the interconnections and relative importance of the independent variables:
9
where and are the maximum and the minimum relative frequencies of the components for each factor. The LSI was then calculated by summing the PR of each factor and the RF of each class:
10
Deep learning-based approaches
Principle of convolutional neural network (CNN)
Multicollinearity analysis of parameters. In deep-learning landslide susceptibility assessments, ensuring the independence of influential factors is essential, especially for the CNN. Feature selection minimizes overfitting and improves model generalization by removing redundant or irrelevant variables (Kumar et al. 2023). Multicollinearity among factors can lead to overfitting or underfitting (Fang et al. 2021; Putriani et al. 2023). Multicollinearity among the selected causative factors was assessed using the variance inflation factor (VIF) and tolerance (TOL) based on the training data (Roy et al. 2019; Mallick et al. 2021). The VIF is computed as:
11
where is the coefficient of determination, which measures how well a linear regression model fits a dataset. Tolerance (TOL) is the reciprocal of VIF as:
12
TOL measures the proportion of variance in a factor that is not explained by the other factors. A factor is considered multicollinear and should be excluded from susceptibility mapping if the TOL value is < 0.1 or the VIF value is > 10 (Kavzoglu et al. 2014; Mallick et al. 2021).
Convolutional neural network (CNN). A Convolutional Neural Network (CNN) is a deep learning architecture commonly used for image analysis (Shin et al. 2016; Wang et al. 2019). It consists of an input layer, hidden layers (convolutional, activation, and pooling), and an output layer. In this study, the CNN performs binary classification to predict landslide (1) or non-landslide (0) categories. The convolutional Layers use learnable filters to extract spatial features. Rectified Linear Unit (ReLU) activation functions introduce non-linearity (Nair and Hinton 2010). Batch normalization (BN) is applied to mitigate overfitting of the training data and improve the model’s generalization ability. Pooling Layers reduce the spatial dimensions of feature maps, thereby decreasing computational complexity. Fully Connected Layers integrate the learned representations for final classification (Sharif Razavian et al. 2014).
The architecture included multiple convolutional and pooling layers designed to extract spatial and hierarchical features relevant to susceptibility. Initial layers detected basic features (i.e., edges), while deeper layers captured complex patterns. The model was trained using the landslide inventory and geospatial raster layers, learning to distinguish landslide-prone areas from stable regions (Chen et al. 2025; Feng et al. 2025). Once the CNN model was trained, the CNN processed new input data to generate an LSM for the NCMBs, highlighting areas with varying probabilities of occurrence (Fig. 6). The qualitative architecture and parameters of the CNN are detailed in Table 4.
[See PDF for image]
Fig. 6
The architecture of the CNN-based landslide susceptibility model
Table 4. CNN’s and lstm’s architecture and parameters
Model | Layer / parameter | Configuration / value | Details |
|---|---|---|---|
CNN | Input layer | 10 thematic layers | One for each Landslide Causative Factor (LCF) |
Convolutional layers | 3 layers, 32 filters each. | Used for automatic spatial feature extraction | |
Kernel size | 3 × 3 | Standard size for capturing local spatial patterns | |
Activation function | Rectified Linear Unit (ReLU) | Introduces non-linearity to the model | |
Pooling layer | Max pooling, 2 × 2 window | Reduces spatial dimensions and computational cost | |
Output layer | Fully connected (dense) | Final classification layer | |
Loss function | Binary cross-entropy | Suitable for binary classification (1 = landslide vs. 0 = non-landslide) | |
Optimizer | Adam | Adaptive moment estimation | |
Learning rate | 0.001 | Standard learning rate | |
Epochs | 50 | Number of training iterations | |
Batch size | 32 | Number of samples processed before the model is updated | |
LSTM | Input layer | 10 thematic layers | Used the 10 LCFs after Information Gain Ratio (IGR) selection |
LSTM layer | 1 layer, 128 hidden units | Enhances capacity for sequential feature representation | |
Output Layer | Fully connected (Sigmoid) | Final classification layer | |
Loss function | Binary cross-entropy | Suitable for binary classification | |
Optimizer | Adam | Adaptive moment estimation | |
Learning rate | 0.001 | Standard learning rate | |
Epochs | 50 | Number of training iterations | |
Batch size | 32 | Number of samples processed before the model is updated |
Principle of long short-term memory (LSTM)
Feature selection of LCFs. The importance of LCFs was assessed using the Mean Decrease Impurity (MDI) and Information Gain Ratio (IGR) methods. MDI, also known as Gini importance, is a robust technique widely used in geoscientific research to quantitatively rank the significance of causative factors, particularly for high-dimensional, non-linear data (Li et al. 2019; Huang et al. 2023). MDI provides a data-driven measure of a variable’s contribution to the model’s predictive power. It is calculated for a single decision tree as:
13
where is the fraction of samples belonging to causative factor, ,and is the number of causative factors (Li et al. 2019; Huang et al. 2023). MDI provides a measure of a variable’s contribution to the model’s predictive power, thereby establishing a data-driven ranking of factor importance. Factors with higher MDI scores are considered more important because they often lead to splits with a substantial reduction in impurity, indicating a strong correlation with landslide occurrence (Yu et al. 2019; Huang et al. 2023; Zuo et al. 2025). IGR is a feature selection and ranking method often used with decision tree algorithms in the DL models. For landslide susceptibility mapping, IGR quantifies a factor’s predictive power by measuring its contribution to classifying landslide versus non-landslide areas, while normalizing for the factor’s intrinsic information (number of possible outcomes). IGR is calculated as:
14
where is the expected reduction in entropy from splitting on factor A, and measures the intrinsic information of the split (Chen et al. 2018; Cheng and Shi 2023; Gu et al. 2023). Factors are ranked by descending IGR value. The highest IGR indicates the most significant predictor, which aids in developing more accurate and efficient susceptibility models (Chen et al. 2018; Fei et al. 2022; Zuo et al. 2025) (Fig. 7). Using both MDI and IGR ensures a robust feature selection process. This dual analysis identifies and removes unnecessary features, which reduces data noise, improves training efficiency, and enhances prediction accuracy. By highlighting the most critical factors, this process makes the resulting model easier to interpret and more reliably applicable to new data (Zuo et al. 2025).
[See PDF for image]
Fig. 7
Weight of each landslide causative factor determined by a MDI and b IGR methods. Abbreviations and symbols are EL-elevation; SL-slope; PC-profile curvature; AAR-average annual rainfall; DP-drainage proximity; LI-lithology; ST-soil texture; LD-lineament density; LULC-land use and land cover; NDVI-normalized difference vegetation index
Long short-term memory (LSTM). The Long Short-Term Memory (LSTM) network is a type of Recurrent Neural Network (RNN) that is particularly effective with time-series data because it can manage long-term dependencies (Hochreiter and Schmidhuber 1997). LSTM enhances the efficiency and accuracy of an LSM by integrating feature selection (Zuo et al. 2025). The core structure of an LSTM includes multiple gated structures (forget, input, and output gates) that selectively retain or discard information across time steps (Ma et al. 2019), thereby preventing the vanishing gradient problem common in standard RNNs (Hochreiter and Schmidhuber 1997).
The LSTM model in this study consisted of a single LSTM layer followed by a fully connected layer. The LSTM layer was designed with 128 hidden units to enhance the model’s capacity for capturing sequential features (Shi et al. 2015). The forget gate decided which previous information to discard. The input gate determined the amount of new information to retain. The output gate produced the final prediction at each time step. This gated design allowed the model to generalize effectively when working with long sequences of data. During the training phase, the output from the LSTM layer was fed into a fully connected layer. Batch normalization (BN) was applied to the output of the fully connected layer to further improve the model’s generalization before final classification (0 for non-landslide, 1 for landslide). After training, the LSTM processed unlabeled cells to predict the probability of a landslide occurring at each location. These predictions were spatially reassembled into an LSM, where higher values indicate a higher likelihood of landslides (Huang et al. 2023; Zuo et al. 2025). The LSTM integrated feature selection (either MDI or IGR) with the LSTM Python module to improve efficiency and accuracy (Fig. 8). The qualitative design and parameters of the LSTM are shown in Table 4.
[See PDF for image]
Fig. 8
The architecture of the LSTM-based landslide susceptibility model. The IGR was selected to be combined with the LSTM Python module for landslide susceptibility mapping
Model validation
Model validation is a crucial step for determining the accuracy and reliability of an LSM. The validation is typically performed using receiver operating characteristic area under the curve (ROC AUC) analysis. The ROC curve illustrates the model’s performance as a probability curve, while the AUC measures the overall performance of the classification model (Wu et al. 2020). The AUC value ranges from 0 to 1; a higher value indicates better predictive capacity to distinguish between landslide-prone and non-landslide areas. The AUC is derived from the ROC curve, which plots the true positive rate (sensitivity) against the false positive rate (1-specificity). The true positive rate (y-axis) represents the accuracy of correctly detecting landslide and non-landslide areas based on field observations. The false positive rate (x-axis) represents the incorrect modeling of non-landslide areas as landslides (Satarzadeh et al. 2022).
The ROC plot demonstrates the model’s reliability and accuracy in classifying landslide occurrences within the LSM classifications (Kalantar et al. 2017). The ROC graph was plotted to show the relationship between the true and false positive rates using the “Calculate ROC curves and AUC values” tool from the ArcSDM5 toolbox (Mas et al. 2013). The model’s accuracy and predictability are assessed by the position of the ROC curve and the AUC value. If the ROC curve is closer to the top-left corner of the graph, the model is considered to have higher accuracy. Conversely, an AUC value of less than 0.50 suggests the model is poorly performing. The ranges of AUC values are interpreted as follows: 0.50–0.60, 0.60–0.70, 0.70–0.80, 0.80–0.90, and > 0.90 are interpreted as representing fair, average, good, very good, and excellent model performance in predicting landslide susceptibility, respectively (Yesilnacar and Topal 2005; Yalcin 2011). Although spatial k-fold cross-validation performed well in validating the model’s performance, the study adopted a rigorous 70/30 random split of the landslide inventory in the buffer-based spatial partitioning approach to ensure the validation dataset was entirely independent and geographically distributed from the training data, mitigating potential sampling bias. The ROC AUC was adequate to validate the performance of the models.
Additionally, particularly for deep learning-based models, a confusion matrix is used to evaluate classification performance (Chen et al. 2018) (Fig. 9). In this study, the following validation metrics were employed as statistical evaluations, including precision, recall, F1-score, and accuracy (Goetz et al. 2015; Sun et al. 2021). These metrics are calculated as:
[See PDF for image]
Fig. 9
The confusion matrix used in the study
13
14
15
16
where is the number of samples correctly predicted as positive, refers to samples incorrectly predicted as positive (actual negative), represents samples incorrectly predicted as negative (actual positive), and indicates samples correctly predicted as negative (Hong et al. 2017; Liu et al. 2024b). Precision, recall, F1-score, and accuracy values range from 0 to 1. Precision, recall, F1-score, and accuracy values range from 0 to 1. Precision represents the accuracy of positive predictions, where a value closer to 1 indicates a higher degree of accuracy in identifying landslides. Recall measures the proportion of actual positive samples correctly identified, where a value near 1 reflects a stronger landslide detection capability (Pham et al. 2016). The F1-score, as the harmonic mean of precision and recall, provides a balanced measure of the model’s performance for both positive and negative samples. Higher values indicate better predictive ability. Accuracy measures the proportion of all correct predictions (both positive and negative) among all samples. A value closer to 1 indicates that the model, measurement, or system is almost always correct (Hong et al. 2016; Merghadi et al. 2020; He 2021).
Results
Landslide inventory map
Using a buffer-based spatial partitioning method across the NCMBs, 1,222 historical landslide sites were compiled into the landslide inventory data, covering an area of almost 5,000 km². The inventory was split, with 855 points designated for the training dataset and the remaining 367 points serving as the validation dataset to assess model accuracy. Most of the detected landslide scars were reported as translational landslides or debris flows. These slides were exposed after the removal of crystalline rock, clastic sedimentary rock, and low-grade metamorphic rocks, and may have been triggered by heavy rainfall. However, determining the exact landslide type was challenging due to dense vegetation on the slopes and limited image resolution.
Influences of landslide causative factors
In this study, 10 landslide causative factors were analyzed and grouped into four types: topographical, hydrological, geological, and land modification-related factors. For topographical factors, the analysis showed that while the 567–793 m elevation range was dominant (26% of the terrain), the higher elevation range (> 794 m) concentrated 70% of landslides, suggesting higher elevations likely corresponded to steeper slopes and an increased frequency of landslides (Table 5). Similarly, steep slope gradients (> 24.5°) were only described as 25% of the terrain but had 53% of the landslides. The lower gradient (< 17°) accounted for 19% of the landslide occurrence (Table 5). The analysis demonstrated that steeply sloped terrain has a higher potential for material downslope movement. Furthermore, as the NCMBs were generally flat (50% of the terrain), concave slopes were generally more susceptible to landslide activity, accounting for 47% of occurrences compared to 15% on convex slopes (Table 5).
Table 5. Information on landslide causative factors and their frequency ratio values
Categories | Factors | Classes | LCFs information | Frequency ratio method | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
Class pixel | % Class pixela | Landslide pixel | % Landslide pixelb | FR (b/a) | RF | PR | Weighted importance | |||
Topography | Elevation (m) | 289–566 | 7,921,125 | 25.38 | 43 | 5.04 | 0.20 | 0.04 | 1.73 | 0.08 |
567–793 | 8,053,788 | 25.80 | 208 | 24.36 | 0.94 | 0.17 | ||||
794-1,019 | 6,982,019 | 22.37 | 234 | 27.40 | 1.22 | 0.22 | ||||
1,020 − 1,279 | 5,383,277 | 17.25 | 255 | 29.86 | 1.73 | 0.31 | ||||
1,280-2,144 | 2,874,331 | 9.21 | 114 | 13.35 | 1.45 | 0.26 | ||||
Slope (o) | 0-8.8 | 6,505,765 | 20.87 | 35 | 4.10 | 0.20 | 0.03 | 2.79 | 0.12 | |
8.9–16.9 | 8,242,495 | 26.44 | 125 | 14.64 | 0.55 | 0.08 | ||||
17.0-24.4 | 8,600,047 | 27.58 | 239 | 27.99 | 1.01 | 0.15 | ||||
24.5–33.1 | 6,029,198 | 19.34 | 297 | 34.78 | 1.80 | 0.27 | ||||
33.2–82.8 | 1,802,588 | 5.78 | 158 | 18.50 | 3.20 | 0.47 | ||||
Profile curvature | Convex | 3,476,446 | 11.14 | 128 | 14.99 | 1.35 | 0.41 | 1.13 | 0.05 | |
Flat | 15,545,637 | 49.80 | 404 | 37.70 | 0.75 | 0.23 | ||||
Concave | 12,192,457 | 39.06 | 322 | 47.31 | 1.21 | 0.37 | ||||
Hydrology | Average annual rainfall (mm/yr) | 1,301-1,408 | 4,199,225 | 13.45 | 43 | 5.03 | 0.37 | 0.07 | 1.94 | 0.08 |
1,409-1,466 | 8,457,526 | 27.09 | 148 | 17.31 | 0.64 | 0.12 | ||||
1,467-1,526 | 8,525,205 | 27.31 | 194 | 22.69 | 0.83 | 0.15 | ||||
1,527-1,590 | 7,040,413 | 22.55 | 301 | 35.20 | 1.56 | 0.29 | ||||
1,591-1,750 | 2,992,319 | 9.59 | 169 | 19.77 | 2.06 | 0.38 | ||||
Drainage proximity (m) | 0-250 | 12,119,810 | 38.83 | 215 | 25.15 | 0.65 | 0.10 | 1.26 | 0.06 | |
251–500 | 9,414,955 | 30.16 | 241 | 28.19 | 0.93 | 0.15 | ||||
501–750 | 6,188,783 | 19.83 | 220 | 25.73 | 1.30 | 0.20 | ||||
751-1,000 | 2,789,901 | 8.94 | 148 | 17.31 | 1.94 | 0.30 | ||||
1,001–2,140 | 701,265 | 2.25 | 31 | 3.63 | 1.61 | 0.25 | ||||
Geology | Lithology | Granite to basic igneous rock | 7,583,425 | 24.33 | 147 | 17.19 | 0.71 | 0.13 | 3.19 | 0.14 |
Low-grade metamorphic rock | 2,077,352 | 6.66 | 166 | 19.42 | 2.91 | 0.54 | ||||
Clastic sedimentary rocks | 13,450,003 | 43.15 | 489 | 57.19 | 1.33 | 0.25 | ||||
Limestone | 4,486,569 | 14.39 | 37 | 4.33 | 0.30 | 0.06 | ||||
Unconsolidated sediment | 3,574,981 | 11.47 | 16 | 1.87 | 0.16 | 0.03 | ||||
Soil texture | Well-drained sand | 1,429,965 | 4.59 | 2 | 0.23 | 0.05 | 0.03 | 4.97 | 0.22 | |
Moderately-drained sand | 2,675,879 | 8.59 | 19 | 2.22 | 0.26 | 0.17 | ||||
Poorly-drained sand | 1,042,996 | 3.35 | 0 | 0.00 | 0.00 | 0.00 | ||||
Poorly-drained silt and clay | 387,790 | 1.24 | 0 | 0.00 | 0.00 | 0.00 | ||||
Steeply sloped soil | 25,611,538 | 82.22 | 834 | 97.54 | 1.19 | 0.79 | ||||
Lineament density (km/km2) | 0-0.19 | 6,090,428 | 19.51 | 157 | 18.36 | 0.94 | 0.21 | 1.00 | 0.04 | |
0.20–0.37 | 10,238,469 | 32.80 | 284 | 33.22 | 1.01 | 0.23 | ||||
0.38–00.55 | 8,745,739 | 28.02 | 254 | 29.71 | 1.06 | 0.24 | ||||
0.56–0.79 | 4,893,027 | 15.68 | 147 | 17.19 | 1.10 | 0.24 | ||||
0.80–1.48 | 1,247,025 | 3.99 | 13 | 1.52 | 0.38 | 0.08 | ||||
Land modification | Land use and land cover | Forest | 25,798,371 | 82.82 | 794 | 92.87 | 1.12 | 0.61 | 3.85 | 0.17 |
Water | 93,864 | 0.30 | 0 | 0.00 | 0.00 | 0.00 | ||||
Shrubland | 93,830 | 0.30 | 0 | 0.00 | 0.00 | 0.00 | ||||
Built area | 422,027 | 1.35 | 3 | 0.35 | 0.26 | 0.14 | ||||
Agricultural land | 4,740,077 | 15.22 | 58 | 6.78 | 0.45 | 0.24 | ||||
NDVI | < 0 | 436,458 | 1.40 | 5 | 0.58 | 0.42 | 0.10 | 1.02 | 0.04 | |
0-0.2 | 6,089,726 | 19.51 | 167 | 19.53 | 1.00 | 0.23 | ||||
0.21–0.40 | 11,244,818 | 36.02 | 308 | 36.02 | 1.00 | 0.23 | ||||
0.41–0.60 | 8,868,661 | 28.41 | 274 | 32.05 | 1.13 | 0.26 | ||||
0.60-1.0 | 4,575,025 | 14.66 | 101 | 11.81 | 0.81 | 0.19 | ||||
Bold numbers are the highest frequency ratio of each landslide causative factor
Regarding hydrological factors, average annual rainfall in the NCMBs ranged between 1,301 and 1,750 mm/yr. Rainfall higher than 1,527 mm accounted for 55% of landslide occurrences, indicating that higher amounts of rainfall corresponded to an increased likelihood of landslides. Additionally, 80% of landslides occurred at distances of less than 750 m from the river system, while only 20% of landslides occurred at distances greater than 750 m (Table 5). This result supported the notion that the river could scour the slope base, leading to slope instability and landslides.
In terms of geological factors, clastic sedimentary rock was the primary lithology, covering 43% of the terrain. This underlying rock corresponds to 57% of landslide occurrences (Table 5). The undifferentiated soil texture was also highly influential for the landslide. It covered 80% of the terrain and accounted for 97% of landslide occurrences (Table 5). The presence of erodible rock and soil on steep slopes contributes to increased slope instability. Although lineament density indicated a slightly higher landslide occurrence in areas of higher density, 63% of landslides were observed in low to moderate density terrain (Table 5).
Finally, for land modification-related factors, trees and forests covered most of the terrain and accounted for 93% of landslide occurrences. The remaining 7% occurred in agricultural and built-up areas (Table 5). A high proportion of landslide occurrences (68%) was also observed within the moderate to high NDVI values, and 20% located in zones of sparse vegetation cover (Table 5). This result suggested that while the vegetation was dense, it may be insufficient to stabilize the susceptible slope fully. Dense vegetation coverage on high elevations and steep slopes likely makes these areas prone to landslides.
Landslide susceptibility map
Landslide susceptibility map from statistical models
The AHP and FR were utilized to create an LSI. The AHP combined 10 landslide causative factors, with numerical weights for each factor and subfactor determined through a pairwise comparison matrix (Tables S1 and S2). For the FR model, the ratio values for selected factors are presented in Table 5. An FR value of 0 indicates low landslide potential, while a value above 1 indicates high susceptibility. The higher FR values show a positive correlation with RF, indicating a more substantial influence of those sub-classes on landslide occurrence.
The AHP-derived LSM was classified into five classes using the natural break classification method: very low, low, moderate, high, and very high. The findings showed that 21.49% and 7.46% of the total area fell into the high and very high susceptibility classes, respectively. Additionally, 31.17% was in the moderate class, while the remaining 25.32% and 14.56% were in the low and very low classes, respectively (Table 6). Based on the AHP’s eigenvalue, elevation (19.6%) and NDVI (14.9%) were the most influential factors, followed by profile curvature (12.8%), lineament density (12.3%), and drainage proximity (11.6%) (Table S2). The high and very high susceptibility zones were concentrated in the central and eastern parts of the NCMBs, corresponding to mountainous or steep-sloped terrains (DCD and DML). Conversely, the very low and low zones were found in valleys and flatter areas (central CDP, PB, and their southern parts) (Fig. 10a).
Table 6. Landslide susceptibility classification with respective area coverage and model-derived percentages for four predictive models
Classes | Models | |||||||
|---|---|---|---|---|---|---|---|---|
AHP | FR | CNN | LSTM | |||||
Area (km2) | % | Area (km2) | % | Area (km2) | % | Area (km2) | % | |
Very low | 707.73 | 14.56 | 566.70 | 11.66 | 599.21 | 12.33 | 691.14 | 14.22 |
Low | 1230.75 | 25.32 | 774.96 | 15.94 | 597.29 | 12.29 | 997.63 | 20.53 |
Moderate | 1515.09 | 31.17 | 1524.70 | 31.37 | 1519.14 | 31.25 | 1499.70 | 30.86 |
High | 1044.45 | 21.49 | 1392.37 | 28.65 | 1551.70 | 31.93 | 1151.62 | 23.69 |
Very high | 362.29 | 7.46 | 601.59 | 12.38 | 592.98 | 12.20 | 520.23 | 10.70 |
[See PDF for image]
Fig. 10
Landslide susceptibility maps and percentages based on four models: a AHP, b FR, c CNN, d LSTM
Using the same natural break classification method, the spatial distribution based on the FR model was also categorized into five susceptibility classes. The results indicated that 28.65% and 12.38% of the area were situated in the high and very high susceptibility regions, respectively. Additionally, 31.37% was in the moderate class, and the remaining 27.60% was in the low and very low classes (Table 6). Based on weighted importance (Table 5), soil texture (22%) was the most significant factor for landslide prediction in the FR model, followed by land use/land cover (17%), lithology (14%), terrain slope (12%), and average annual rainfall (8%). The very high and high susceptibility zones were primarily located on steep slopes and rugged terrain along DSM, DCD, DML, and DJJ. The moderate zones were at the bases of mountain ranges, while low and very low susceptibility zones were in the valleys and lowlands of WHB, CDP, and PB (Fig. 10b).
Landslide susceptibility map from deep learning-based models
Multicollinearity analysis was employed to examine the interdependence of the 10 landslide causative factors. The analytical findings showed that the TOL values ranged from 0.383 to 0.736, and the VIF values ranged from 1.358 to 2.609 (Table 7). Since all TOL values were greater than 0.1 and all VIF values were less than 10, no multicollinearity was detected. Therefore, all 10 causative factors were determined to be independent and suitable for assessing landslide susceptibility.
Table 7. Multicollinearity test among landslide causative factors for the CNN model
Categories | Landslide causative factors | Collinearity statistics | |
|---|---|---|---|
TOL | VIF | ||
Topography | Elevation | 0.519 | 1.928 |
Terrain slope | 0.383 | 2.609 | |
Profile curvature | 0.691 | 1.447 | |
Hydrology | Average annual rainfall | 0.495 | 2.022 |
Drainage proximity | 0.616 | 1.622 | |
Geology | Lithology | 0.599 | 1.668 |
Soil texture | 0.545 | 1.835 | |
Lineament density | 0.724 | 1.380 | |
Land modification | LULC | 0.644 | 1.554 |
NDVI | 0.736 | 1.358 | |
The CNN, a deep learning process involving data processing, model building, and training, was implemented using Python code in Google Colaboratory. The input data included the training dataset and all 10 landslide causative factors. The trained CNN model generated the LSM (Fig. 10c), which was reclassified into five susceptibility levels using the natural break classification method: very low, low, moderate, high, and very high. The classification distributed the study area as follows: 12.33% (very low), 12.29% (low), 31.25% (moderate), 31.93% (high), and 12.20% (very high) (Table 6). The very high susceptibility zones were primarily located along the western (DSM, Doi Yao (DYO), Doi Mon Chia (DMC), and Doi Mon Ngo (DMN), central (Doi Kham Fah (DKF), and DCD), and eastern (DJJ, DML, and DKK) high mountains of the NCMBs. High susceptibility zones were distributed along the hillslopes to the bases of these mountains. Lower elevation areas (i.e., CDP and PB) showed lower landslide susceptibility. Compared to the landslide inventory, most landslides were situated in the very high to high susceptibility zones (Fig. 10c).
The LSTM model was also implemented in Google Colaboratory using Python code to generate the LSM based on the landslide inventory and 10 causative factors. As IGR and MDI ranked the same five most important factors (terrain slope, lithology, elevation, soil texture, and average annual rainfall), the LSTM-based LSMs were developed using a feature selection method based on IGR, a general technique that provides robust estimates of factor importance for LSMs (Quinlan 1986; Yu et al. 2019; Zuo et al. 2025). The LSTM-derived LSM was categorized into five susceptibility levels using the natural break classification method, accounting for 14.22% (very low), 20.53% (low), 30.86% (moderate), 23.69% (high), and 10.70% (very high) (Table 6). Areas near DSM, DYO, DMC, and DMN in the Mae Tang River watershed, as well as DJJ, DML, and DKK in the Mae Ngat River watershed, showed very high susceptibility to landslides. DLR, DKF, and DCD in the middle were highly susceptible. Moderate to low susceptible areas were mainly distributed at the bases of the mountains, while very low-risk regions were located in the CDP and PB (Fig. 10d).
Model comparisons and validation
When comparing the LSMs derived from the four models, AHP, FR, CNN, and LSTM, each model showed the highest proportions in the moderate susceptibility class (approximately 30% of the total area). The CNN showed the most balanced distribution between the moderate and high susceptibility classes. In the very low category, percentages were relatively similar across all models (around 12–15%). For the low class, AHP presented the highest proportion (25%), followed by LSTM, while CNN had a notably smaller proportion (12%). All four models consistently predicted the lowest percentages for the very high susceptibility class, with values ranging from 8% to 12% (Fig. 11).
[See PDF for image]
Fig. 11
Landslide susceptibility classes percentages for all four models
Model performance was evaluated using the ROC AUC by combining the validation dataset with the LSMs (Yesilnacar and Topal 2005; Pradhan et al. 2011). The ROC curves of all models were close to the upper-left corner of the plot, indicating strong performance (Fig. 12). The AUC values of AHP, FR, CNN, and LSTM were 0.794, 0.839, 0.907, and 0.865, respectively (Fig. 12). Based on AUC value ranges, all models performed well in identifying landslide susceptibility within the NCMBs with high accuracy and reliability (Yesilnacar and Topal 2005; Yalcin 2011).
[See PDF for image]
Fig. 12
ROC AUC for LSMs derived from all four models
The CNN and LSTM-based LSMs were further validated using validation metrics including precision, recall, F1-score, and overall accuracy (Goetz et al. 2015; Sun et al. 2021). Using an optimal threshold of 0.5–0.7, the results revealed that the precision for both CNN and LSTM was high at 0.954 and 0.923, respectively, indicating that over 92% of optimistic landslide predictions were correct. The recall was strong for both CNN and LSTM at 0.820 and 0.813, indicating that more than 80% of all potential landslide events were successfully identified. The F1-scores of CNN and LSTM were 0.849 and 0.838, respectively, which demonstrated a highly commendable balance between precision and recall and reflected effective model performance. Moreover, both models achieved a high overall accuracy with values of 0.863 for CNN and 0.860 for LSTM, reflecting strong reliability in identifying landslide susceptibility in the NCMBs (Table 8).
Table 8. Validation metrics, including precision, recall, F1-score, and accuracy for the CNN and LSTM
Model | Precision | Recall | F1-score | Accuracy |
|---|---|---|---|---|
CNN | 0.954 | 0.820 | 0.849 | 0.863 |
LSTM | 0.923 | 0.813 | 0.838 | 0.860 |
Discussion
Influence of landslide causative factors
The LSMs derived from AHP, FR, and DL models (CNN and LSTM) used all 10 LCFs, which were grouped into four categories: topography (elevation, terrain slope, and profile curvature), hydrology (average annual rainfall, and drainage proximity), geology (lithology, soil texture, and lineament density), and land modification (land use and land cover, and NDVI). Despite various DEM resolutions across the study site, these LCFs, which were extracted from the high-resolution 12.5 m DEM, accurately represent the local geomorphology for evaluating areas prone to landslides (Chang et al. 2019).
The importance of LCFs varied across the models. The AHP model ranks LCFs by breaking down the decision process and assigning numerical weights through pairwise comparisons (Tables S1 and S2). The FR model ranks factors by their relative influence: a higher FR ratio indicates a stronger positive correlation with landslide occurrence (Table 5). For the DL models, the importance of LCFs was ranked by using MDI and IGR (Fig. 7). Both MDI and IGR methods identified terrain slope, elevation, rainfall, soil texture, and lithology as the most important factors. A slight difference is highlighted that MDI emphasizes rainfall, while IGR highlights lithology. Although MDI is computationally efficient, it can be biased toward features with more splits, which may lead to an inaccurate ranking. The IGR is considered more reliable because it reduces this bias and provides a balanced assessment of the factors across diverse datasets (Yu et al. 2019; Zuo et al. 2025). Therefore, feature selection using IGR was chosen for integration with the LSTM model. Furthermore, Factors considered less critical across the DL models included drainage proximity, lineament density, profile curvature, land use and land cover, and NDVI.
The top five most influential LCFs include terrain slope, rainfall, lithology, soil texture, and elevation. These factors align with previous findings by Dong et al. (2023), which report a 30-year analysis of geological information and landslide triggering factors in geohazards. Rainfall is the dominant trigger in the world and in the NCMBs. Intense or prolonged rainfall increases pore water pressure, which reduces the slope’s shear strength and induces failure. This factor is the most significant in causing hydrologically controlled landslides (Cui et al. 2025). The dominance of lithology and terrain slope reflects fundamental geomorphological controls. The high concentration of failures in weak, weathered sedimentary rocks on steep slopes corresponds to the notion that landslide occurrence is controlled by a slope’s ability to resist the movement of material strength (lithology) (Keles and Nefeslioglu 2021). The interaction between these static and dynamic factors from landslide and non-landslide samples enhances the prediction accuracy of the DL models (Xiang et al. 2025). Moreover, the importance of soil texture reflects local geomorphological processes. Unlike other studies, such as those in the Himalayas, which report that coarse, gravel-sand-rich materials lead to high-velocity debris flows (Daud et al. 2025), landslides in the NCMBs are dominated by fine-grained soil and weathered overburden. This study confirms that landslides are strongly controlled by cohesion loss and saturation-induced failure, where poor drainage (clay-rich soil) rapidly reduces strength during intense rainfall. Elevation also influences landslide occurrence, as higher altitudes are often associated with steep slopes and strong climates, which intensify the weathering process and slope failure (Savi et al. 2021).
Although drainage proximity, lineament density, profile curvature, land use and land cover, and NDVI are less influential factors, they still contribute to slope instability and landslide initiation. Drainage proximity increases risk through toe erosion, which removes support material, increases soil saturation from nearby water bodies, and reduces shear strength. Land modification can increase landslide susceptibility through anthropogenic activities such as bare ground exposure, slope destabilization, urbanization, and inappropriate construction. However, some human activities decrease landslide susceptibility through land use practices such as reforestation and appropriate management of agricultural land with deep-rooted crops. NDVI also reflects vegetation health: dense vegetation enhances slope stability, while sparse or bare ground increases the area’s susceptibility to landslides. Lineament density reflects structural weaknesses that reduce rock strength and enhance water infiltration. In the study, lineament distribution corresponds to underlying rock types, with the relatively high terrain representing widespread lineaments that partly control local drainage. However, the low correlation between lineament density and landslide occurrences (only 18% of landslides in high-to-very-high-density areas) suggests that tectonic structures were a weaker influence than topographic and hydrological factors in Northern Thailand.
The influence of LCFs varied across the models. Terrain slope, lithology, and rainfall were dominant in all approaches. However, the DL models more effectively captured the complex spatial dependencies of elevation and drainage proximity than the FR model, which assumes independence among variables, or the AHP model, which relies on expert judgment. The AHP and FR models may limit their ability to represent landslide susceptibility in the complex interactions among factors.
Roles of rainfall-lithology-slope interaction in deep learning-based modeling
The dual use of the DL models for landslide susceptibility mapping requires a physical interpretation of the results, moving beyond purely data-driven classification (Pradhan et al. 2023). For example, Daud et al. (2025) used a multi-grid sampling technique in the DL models to identify locations where the interplay between rainfall and lithology caused mass movement. By combining these factors with landslide occurrences derived from the infinite slope model and the effective stress principle (Vahedifard et al. 2016), therefore, the model can be linked to the Factor of Safety (FS), a geotechnical engineering measure for slope stability, that positively relates to slope failure by evaluating material strength (related to lithology), shear stress (related to slope), and pore water pressure (transmitted by rainfall). The hybrid model’s architecture successfully combines these important factors to represent the complex, non-linear interactions of landslide occurrences in nature.
The CNN primarily captured the static, inherent preconditioning of the terrain, focusing on the spatial interaction between lithology and slope. Its hierarchical filters learned complex spatial associations that represent zones of critically low shear strength (Feng et al. 2025). Specifically, the model assigns high susceptibility where steep slopes coincide spatially with weak or weathered lithological units. By performing neighborhood analysis, the CNN identified micro-geomorphological features such as concave slope profiles or intersecting geological layers. These features serve as spatial proxies for preferential water flow and high potential for rainfall-induced pore-pressure accumulation (Hakim et al. 2022). Conversely, the LSTM explicitly modeled the dynamic triggering mechanism related to rainfall (Sham et al. 2025). Its memory cells effectively simulated the non-linear, time-delayed physical process of hydrological response within the slope material (Yu et al. 2024). The LSTM successfully captured the cumulative effect of antecedent rainfall, the time required for water to infiltrate the soil column, and the subsequent build-up of pore-water pressure. Water pressure reduces the effective normal stress by pushing the solid particles apart, which in turn decreases the material’s strength and triggers failure. The model also learns that the site’s lithological properties directly modulate the time-series dependency of rainfall (Ma and Mei 2025).
Therefore, the integration of the CNN’s spatial output with the LSTM’s temporal analysis ensured that the model weighed the rainfall sequence based on the underlying geology. For instance, weak lithology-slope combinations (high CNN output) require a lower cumulative rainfall threshold to predict failure. This coupled mechanism represents the physically coupled rainfall–lithology–slope interaction, advancing the susceptibility assessment beyond statistical correlation and moving toward a process-based, deep learning representation of slope stability.
Units of terrain
Geomorphons are a comprehensive set of landforms identified by a pattern-recognition approach that classifies terrain based on local elevation differences (Jasiewicz and Stepinski 2013). By dividing landforms with unstable slopes and concentrated water flow, Geomorphons help identify regions susceptible to landslides. The NCMBs included 10 common Geomorphon types: flat, summit, ridge, shoulder, spur, slope, hollow, footslope, valley, and depression (Jasiewicz and Stepinski 2013; Ghahraman et al. 2023) (Fig. 13a). The distribution of landforms that was ranked from highest to lowest percentage is: slope, spur, hollow, valley, ridge, depression, summit, footslope, shoulder, and flat (Fig. 13b). The three most distributed landforms across all individual watersheds are slope, spur, and hollow, while flat, shoulder, and footslope are the least common (Fig. 13c).
[See PDF for image]
Fig. 13
Geomorphons in the NCMBs. a The distribution of 10 common types of Geomorphons. b The percentage of 10 common types of Geomorphons. c The percentage of 10 common types of Geomorphons that were classified in the different watersheds
Specific geomorphological features, such as steep slopes, unstable hillslopes on ridges, and large hollows and valleys, characterize high-risk areas for landslides. Geological factors such as faults, joints, and erodible material further contribute to inherent instability. Areas intensified by heavy rainfall, poorly drained soil, or a lack of vegetation can trigger extensive and vulnerable landslides. Conversely, areas that experience more minor landslides are typically flat, with gentler footslopes and depressions, which tend to lack historical landslides or ongoing human-induced destabilization.
In addition, Geomorphons not only distinguish individual landform types but also reveal how terrain geometry controls landslide processes. Concave features, such as hollows, valleys, and depressions, act as natural convergence zones for runoff that promote soil saturation and shallow slope failures. In contrast, convex forms, such as slopes, ridges, and spurs, are more prone to erosion and mass movement due to divergent flow and gravitational stress. Transitional zones such as shoulders and footslopes often represent critical breakpoints where changes in slope curvature initiate instability. In addition, Geomorphons reveal how terrain geometry controls landslide processes. Concave features (i.e., hollows, valleys, and depressions) act as natural convergence zones for runoff, promoting soil saturation and shallow slope failures. Convex forms (i.e., slopes, ridges, and spurs) are more prone to erosion and mass movement due to divergent flow and gravitational stress. Transitional zones (i.e., shoulders and footslopes) represent critical breakpoints where changes in slope curvature can initiate instability. Thus, the spatial arrangement and connectivity of these landforms amplify susceptibility as material and water move downslope. The advantage of Geomorphons is that they provide a morphometric framework that explains hydrology, structural weaknesses, and vegetation in the initiation and propagation of landslides.
Model comparisons
The comparative assessment of AHP, FR, CNN, and LSTM models reveals apparent differences in how these approaches capture geomorphological and environmental controls on landslide occurrences in the NCMBs. The AHP model depends on expert judgment and produces transparent results. However, its subjective weighting process and inability to represent nonlinear interactions among conditioning factors reduce predictive accuracy. The FR model, a statistically data-driven approach, establishes relationships between landslide occurrences and causative factors. However, both AHP’s subjectivity and FR’s inability to represent the complex, nonlinear interplay among variables, such as slope, lithology, and rainfall, limit their utility. In contrast, the CNN model outperforms traditional methods by extracting spatial hierarchies and recognizing subtle geomorphic patterns. It can handle large and heterogeneous datasets, improving representation of steep terrain, diverse lithological units, and rainfall-induced triggering mechanisms. Similarly, the LSTM model performs strongly by integrating temporal dependencies, which are essential in monsoon-dominated regions where rainfall sequences significantly impact slope stability.
The landslide susceptibility maps produced by the four models show general agreement in spatial patterns. High-to-very high susceptibility areas from all models are mainly concentrated in the mountainous zones and their hillslopes. The FR/CNN models presented the largest proportion of highly susceptible zones. Conversely, areas with low landslide susceptibility are consistently located in the lower elevation regions. The AHP/LSTM models identified the most extensive areas classified as low- and very-low-susceptibility zones.
Model comparison underscores key differences in predictive performance and applicability for LSM in the NCMBs. The relatively lower performance of AHP and FR aligns with global observations, indicating that early statistical and opinion-based models are constrained by subjectivity and have limited capacity to represent complex, nonlinear relationships (Dong et al. 2023). Conversely, the DL models exhibited superior accuracy, reflecting the current shift in LSM research toward DL techniques. CNN and LSTM achieved robust predictive capabilities, confirming that DL architectures effectively capture complex feature interactions and internal data representations. Nevertheless, AHP and FR remain useful in data-scarce or resource-limited settings due to their interpretability and low computational demand.
The methodological advancement demonstrated by CNN and LSTM in this study aligns with recent DL-based susceptibility assessments in other tropical mountainous and monsoon-dominated regions. Comparable studies in the Himalayas and Southeast Asia have shown that DL models consistently outperform traditional statistical and machine learning methods (i.e., SVM, Random Forest) in handling nonlinear triggers from intense monsoonal rainfall and heterogeneous lithological conditions (Halder et al. 2025).
CNN’s superior accuracy confirms its capability to automatically extract spatial features from static factors (i.e., lithology, slope). The CNN offers a commendatory balance between accuracy and computational efficiency that supports rapid susceptibility mapping (Dong et al. 2024). LSTM’s robust performance validates the value of recurrent architectures in integrating time-dependent data (i.e., average annual rainfall), which provides a more reliable assessment in a region with high climatic variability (Cui et al. 2025). Furthermore, optimizing non-landslide sample selection using buffer-defined stable zones also improves prediction accuracy (Xiang et al. 2025). By integrating CNN and LSTM models and comparing them directly with AHP and FR within a single tropical watershed, this study demonstrates the potential of deep learning to reduce overfitting, refine sample selection, and enhance landslide susceptibility mapping in Northern Thailand and comparable regions.
Limitations and future research
The combination of landslide inventory and causative factors achieved a high predictive performance on the susceptibility maps, which aligned well with landscape characteristics. However, several limitations must be considered. The nature of the input data constrains the accuracy of the susceptibility maps. (1) Inventory ambiguity from landslide scars that are collected as point-based representations. This data inadequately captures the full dimensions and extent of failure zones. (2) Data quality depends on the resolution and consistency of input datasets that later introduce uncertainties. The shift from traditional methods to DL models necessitates higher-quality data, which requires resources for advanced methods to improve the resolution of geo-environmental information (Dong et al. 2023, 2024). (3) Overfitting risk from the relatively modest sample size of 1,222 landslide occurrences carries an inherent risk of overfitting in the DL models. Despite potential overfitting risks, the advantages of DL models provide a stronger foundation for modern landslide susceptibility assessment, supporting informed land-use planning and the establishment of early warning systems. Nevertheless, AHP and FR remain valuable where computational resources and data availability are constrained.
Each approach for landslide susceptibility mapping presents distinct advantages and limitations. The AHP method is straightforward and useful in areas where expert knowledge is available, but historical landslide data is limited. The FR model provides a more data-driven option, but cannot fully address correlations between parameters. However, Machine learning and deep learning models generally offer higher predictive accuracy. The models’ dependence on large, high-quality datasets and limited interpretability can reduce their practicality for decision-making and non-technical users. Selecting an appropriate method, therefore, requires balancing accuracy, transparency, data availability, and user needs.
Future improvements should focus on several key areas. (1) Validation method should incorporate more robust validation strategies, such as k-fold or spatial cross-validation, to ensure model generalizability across sub-regions and enhance the credibility of reported performance metrics (Abraham et al. 2021). (2) Process-based AI from combining artificial intelligence (AI) with specific terrain features provide a stronger physical geomorphological basis for susceptibility zonation (Dong et al. 2024). (3) The incorporation of dynamic and time-dependent factors, such as rainfall intensity–duration thresholds, soil moisture, and land use change, can better capture evolving landslide triggers (Segoni et al. 2018). (4) Data augmentation from advanced remote sensing techniques (i.e., InSAR and LiDAR) can map detailed landslide boundaries and micro-topographic features instead of single points. This method will increase the adequate training sample size, reduce overfitting, and enhance model robustness (Sekiguchi and Sato 2004; Li et al. 2024). (5) Hybrid models combine the interpretability of traditional approaches with the predictive power of DL to improve both accuracy and usability (Yang et al. 2021; Dong et al. 2024). (6) Forward-looking risk by integrating climate change scenarios into the landslide susceptibility assessment to provide a forward-looking perspective on future risks and ensure strong user engagement for practical adoption by decision-makers (Ciabatta et al. 2016; Gariano and Guzzetti 2016).
Conclusion
The study developed and compared landslide susceptibility maps for the Northern Chiang Mai Watershed Basins (NCMBs) using two traditional methods, Analytical Hierarchy Process (AHP) and Frequency Ratio (FR), and two advanced deep learning models, Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). The analysis, based on 1,222 historical landslide scars divided into training and validation datasets, identified key causative factors, including terrain slope, rainfall, lithology, soil texture, and elevation. Spatially, the high landslide susceptible zones are concentrated in steep, high-elevation mountainous regions. The deep learning models outperformed traditional methods, with CNN achieving the highest accuracy (AUC = 0.907) and LSTM showing robust performance (AUC = 0.865). CNN excelled in precision, recall, F1-score, and overall accuracy. Moreover, the methodological advancement enhanced the ability of the DL models to capture the physically coupled rainfall–lithology–slope interaction: the CNN efficiently extracted static spatial features, while the LSTM explicitly modeled dynamic, time-dependent hydrological triggering effects of average rainfall. These findings align with global trends that advocate DL models as the core knowledge pathway in LSM, as these approaches overcome the accuracy limitations of traditional models in complex tropical monsoon environments. Future research should focus on enhancing model generalizability through spatial cross-validation and incorporating dynamic factors, as well as exploiting remote sensing for detailed landslide mapping. Hybrid models combining the predictive power of deep learning with the interpretability of traditional methods could further improve landslide susceptibility mapping for better hazard assessment and sustainable watershed management.
Acknowledgements
The authors would like to express our sincere gratitude to the Department of Mineral Resources of Thailand for providing landslide inventory data and related geological information. Special thanks are extended to the Geo-Informatics and Space Technology Development Agency (GISTDA) for access to satellite imagery and geospatial datasets. The authors highly appreciate Asst. Prof. Dr. Weerapan Srichan, and Assoc. Prof. Dr. Niti Mankhemthong for the geological background of the mountain ranges in Northern Thailand and the geomorphic characteristics of the Chiang Mai Basin. Appreciation is also extended to Prof. Dr. Saro Lee from the Korea Institute of Geoscience and Mineral Resources (KIGAM), who provided us with Python codes to conduct landslide susceptibility models. We thank Dr. Worawit Tepsan from the International College of Digital Innovation (ICDI), Chiang Mai University, who helped us solve the problem while running Python code. We are also grateful to Chiang Mai University for research support and technical facilities. Finally, we acknowledge the constructive comments and suggestions from anonymous reviewers, which significantly improved the quality of this manuscript.
Author contributions
P.M. conceptualized and designed the research, developed the methods and procedures, performed the analysis, validated the results, wrote the original draft of the manuscript, and revised and edited the manuscript. P.K. and K.I. conducted research and experiments and visualized the results. All authors contributed, read, and approved the final manuscript.
Funding
This work was supported by Chiang Mai University.
Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.
Declarations
Competing interests
The authors declare no competing interests.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
Abraham, MT; Satyam, N; Lokesh, R; Pradhan, B; Alamri, A. Factors affecting landslide susceptibility mapping: assessing the influence of different machine learning approaches, sampling strategies and data splitting. Land; 2021; 10, 989. [DOI: https://dx.doi.org/10.3390/land10090989]
Afzal, N; Ahmad, A; Shirazi, SA; Younes, I; Ha, LTT. GIS-based landslide susceptibility mapping using analytical hierarchy process: a case study of Astore region, Pakistan. Int J Environ Qual; 2022; 48, pp. 27-40. [DOI: https://dx.doi.org/10.6092/issn.2281-4485/12600]
Aquino DN, ROC; Moreira, MA; Santos, TA; Andrade, EM. Use of remote sensing to identify areas at risk of degradation in the semi-arid region. Rev Ciênc Agron; 2018; 49,
Asmare, D. Application and validation of AHP and FR methods for landslide susceptibility mapping around choke mountain, Northwestern Ethiopia. Sci Afr; 2023; 19, e01470.1:CAS:528:DC%2BB38XjtFemt7vJ [DOI: https://dx.doi.org/10.1016/j.sciaf.2022.e01470]
Ayalew, L; Yamagishi, H. Slope failures in the blue nile basin, as seen from landscape evolution perspective. Geomorphology; 2004; 57, pp. 95-116. [DOI: https://dx.doi.org/10.1016/S0169-555X(03)00085-0]
Ayalew, L; Yamagishi, H. The application of GIS-based logistic regression for landslide susceptibility mapping in the Kakuda-Yahiko Mountains, central Japan. Geomorphology; 2005; 65, pp. 15-31. [DOI: https://dx.doi.org/10.1016/j.geomorph.2004.06.010]
Azarafza, M; Ghazifard, A; Akgün, H; Asghari-Kaljahi, E. Landslide susceptibility assessment of South Pars special zone, Southwest Iran. Environ Earth Sci; 2018; 77, 805. [DOI: https://dx.doi.org/10.1007/s12665-018-7978-1]
Azarafza, M; Azarafza, M; Akgün, H; Atkinson, PM; Derakhshani, R. Deep learning-based landslide susceptibility mapping. Sci Rep; 2021; 11,
Basu, T; Pal, S. Identification of landslide susceptibility zones in Gish river basin, West Bengal, India. GEORISK; 2017; 12,
Berhane, G; Tadesse, K. Landslide susceptibility zonation mapping using statistical index and landslide susceptibility analysis methods: a case study from Gindeberet district, oromia regional State, central Ethiopia. J Afr Earth Sci; 2021; 180, 104240. [DOI: https://dx.doi.org/10.1016/j.jafrearsci.2021.104240]
Bragagnolo, L; da Silva, RV; Grzybowski, JMV. Artificial neural network ensembles applied to the mapping of landslide susceptibility. CATENA; 2020; 184, 104240. [DOI: https://dx.doi.org/10.1016/j.catena.2019.104240]
Braun, EV; Hahn, L. Geologic map of Northern Thailand 1:250000, sheet 3 (Chiang Mai); 1976; Germany, Federal Institute for Geosciences and Natural Resources:
Bui, DT; Lofman, O; Revhaug, I; Dick, O. Landslide susceptibility analysis in the Hoa Binh Province of Vietnam using statistical index and logistic regression. Nat Hazards; 2011; 59, pp. 1413-1444. [DOI: https://dx.doi.org/10.1007/s11069-011-9844-2]
Cao, Y; Wei, X; Fan, W; Nan, Y; Xiong, W; Zhang, S. Landslide susceptibility assessment using the weight of evidence method: A case study in Xunyang area, China. PLoS ONE; 2021; 16,
Çellek, S. Morphological parameters causing landslides: a case study of elevation. Bull Min Res Exp; 2023; 2020, 162. [DOI: https://dx.doi.org/10.19111/bulletinofmre.649758]
Cemiloglu, A; Zhu, L; Mohammednour, AB; Azarafza, M; Nanehkaran, YA. Landslide susceptibility assessment for Maragheh County, Iran, using the logistic regression algorithm. Land; 2023; 12,
Chang, KT; Merghadi, A; Yunus, AP; Pham, BT; Dou, J. Evaluating scale effects of topographic variables in landslide susceptibility models using GIS-based machine learning techniques. Sci Rep; 2019; 9, 12296.1:CAS:528:DC%2BC1MXhs1Kkt7%2FM [DOI: https://dx.doi.org/10.1038/s41598-019-48773-2]
Chen, W; Xie, X; Wang, J; Pradhan, B; Hong, H; Bui, DT; Duan, Z; Ma, J. A comparative study of logistic model tree, random forest, and classification and regression tree models for Spatial prediction of landslide susceptibility. CATENA; 2017; 151, pp. 147-160. [DOI: https://dx.doi.org/10.1016/j.catena.2016.11.032]
Chen, W; Peng, J; Hong, H; Shahabi, H; Pradhan, B; Liu, J; Zhu, AX; Pei, X; Duan, Z. Landslide susceptibility modelling using GIS-based machine learning techniques for Chongren County, Jiangxi Province, China. Sci Total Environ; 2018; 626, pp. 1121-1135.1:CAS:528:DC%2BC1cXitFyitrs%3D [DOI: https://dx.doi.org/10.1016/j.scitotenv.2018.01.124]
Chen, X; Li, W; Hsu, CY; Arundel, ST; Higman, B. Harnessing Geospatial artificial intelligence and deep learning for landslide inventory mapping: advances, challenges, and emerging directions. Remote Sens; 2025; 17,
Cheng, Y; Shi, Q. PCMIgr: a fast packet classification method based on information gain ratio. J Supercomput; 2023; 79, pp. 7414-7437. [DOI: https://dx.doi.org/10.1007/s11227-022-04951-0]
Chinkulkijniwat, A; Salee, R; Horpibulsuk, S; Arulrajah, A; Hoy, M. Landslide rainfall threshold for landslide warning in Northern Thailand. Geomat Nat Hazards Risk; 2022; 13,
Chiplunkar, SV; Linkesh, M; Badhe, N; Vasoya, A. A cloud-based hybrid CNN-LSTM system for real-time landslide prediction using Geospatial intelligence. Int J Multidiscip Res; 2025; 7, IJFMR250451408. [DOI: https://dx.doi.org/10.36948/ijfmr.2025.v07i04.51408]
Chowdhuri, I; Pal, SC; Arabameri, A; Ngo, PTT; Chakrabortty, R; Malik, S; Das, B; Roy, P. Ensemble approach to develop landslide susceptibility map in landslide dominated Sikkim Himalayan region, India. Environ Earth Sci; 2020; 79, 476. [DOI: https://dx.doi.org/10.1007/s12665-020-09227-5]
Ciabatta, L; Camici, S; Brocca, L; Ponziani, F; Stelluti, M; Berni, N; Moramarco, T. Assessing the impact of climate-change scenarios on landslide occurrence in Umbria Region, Italy. J Hydrol; 2016; 541, pp. 285-295. [DOI: https://dx.doi.org/10.1016/j.jhydrol.2016.02.007]
Cruden, DM; Varnes, DJ. Landslide types and processes, transportation research board, Us National academy of sciences, special report. Transp Res Board; 1996; 247, pp. 36-57.
Cui, HZ; tong, B; Wang, T; Dou, J; Ji, J. A hybrid data-driven approach for rainfall-induced landslide susceptibility mapping: Physically-based probabilistic model with convolutional neural network. J Rock Mech Geotech; 2025; 17,
Dai, F; Lee, C; Li, J; Xu, Z. Assessment of landslide susceptibility on the natural terrain of Lantau Island, Hong Kong. Environ Geol; 2001; 40, pp. 381-391. [DOI: https://dx.doi.org/10.1007/s002540000163]
Daud, H; Dou, J; Khan, NG; Xu, B; Dong, S; Dong, A; Ma, H. Tree-based machine learning and flow simulation for debris flow susceptibility, runout propagation, and dynamics in the higher Himalayas. Math Geosci; 2025; 57, pp. 1195-1233. [DOI: https://dx.doi.org/10.1007/s11004-025-10196-3]
Department of Mineral Resources (2023) Atlas book of maps of active faults in Thailand. Available via https://online.anyflip.com/kera/wglh/mobile/index.html. Accessed 14 April 2025
Department of Mineral Resources (2024) Landslide Scars. Available via. https://data.dmr.go.th/nl/dataset/landslide_inventory_map. Accessed 14 April 2025
Depicker, A; Jacobs, L; Delvaux, D; Havenith, H-B; Mateso, J-CM; Govers, G; Dewitte, O. The added value of a regional landslide susceptibility assessment: the Western branch of the East African rift. Geomorphology; 2020; 353, 106886. [DOI: https://dx.doi.org/10.1016/j.geomorph.2019.106886]
Dikau, R. Raper, J. The application of a digital relief model to landform analysis in geomorphology. Three-dimensional applications in geographic information systems; 1989; Chichester, Taylor and Francis: pp. 51-77.
Dong, A; Dou, J; Fu, Y; Zhang, R; Xing, K. Unraveling the evolution of landslide susceptibility: a systematic review of 30-years of strategic themes and trends. Geocarto Int; 2023; 38,
Dong, A; Dou, J; Li, C; Chen, Z; Ji, J; Xing, K; Zhang, J; Daud, H. Accelerating cross-scene co-seismic landslide detection through progressive transfer learning and lightweight deep learning strategies. IEEE Trans Geosci Remote Sens; 2024; 62, 4410213. [DOI: https://dx.doi.org/10.1109/TGRS.2024.3424680]
Dou, J; Yunus, AP; Bui, DT; Merghadi, A; Sahana, M; Zhu, Z; Chen, CW; Khosravi, K; Yang, Y; Pham, BT. Assessment of advanced random forest and decision tree algorithms for modeling rainfall-induced landslide susceptibility in the Izu-Oshima volcanic Island, Japan. Sci Total Environ; 2019; 662, pp. 332-346.1:CAS:528:DC%2BC1MXhvF2gsr4%3D [DOI: https://dx.doi.org/10.1016/j.scitotenv.2019.01.221]
Dou, J; Yunus, AP; Bui, DT; Merghadi, A; Sahana, M; Zhu, Z; Chen, CW; Han, Z; Pham, BT. Improved landslide assessment using support vector machine with bagging, boosting, and stacking ensemble machine learning framework in a mountainous watershed. Japan Landslides; 2020; 17, pp. 641-658. [DOI: https://dx.doi.org/10.1007/s10346-019-01286-5]
El Jazouli, A; Barakat, A; Khellouk, R. GIS-multicriteria evaluation using AHP for landslide susceptibility mapping in Oum Er Rbia high basin (Morocco). Geoenviron Disasters; 2019; 6, 3. [DOI: https://dx.doi.org/10.1186/s40677-019-0119-7]
Environmental Geology Division (2003) A manual of landslide geohazard prevention and list of landslide risk villages of the Northern region of Thailand. Department Mineral Resour Bangkok Thail 1, 1–113. (in Thai).
Es-Sabbar, B; Essalhi, M; Essalhi, A; Mhamdi, HS. Lithological and structural lineament mapping from Landsat 8 Oli images in Ras Kammouna arid area (Eastern Anti-Atlas, Morocco). Econ Environ Geol; 2020; 53, pp. 425-440. [DOI: https://dx.doi.org/10.9719/EEG.2020.53.4.425]
Fang, Z; Wang, Y; Peng, L; Hong, H. Integration of convolutional neural network and conventional machine learning classifiers for landslide susceptibility mapping. Comput Geosci; 2020; 139, 104470. [DOI: https://dx.doi.org/10.1016/j.cageo.2020.104470]
Fang, Z; Wang, Y; Peng, L; Hong, H. Comparative study of heterogeneous ensemble-learning techniques for landslide susceptibility mapping. Int J Geogr Inf Sci; 2021; 35, pp. 321-347. [DOI: https://dx.doi.org/10.1080/13658816.2020.1808897]
Farooq S, Akram MS (2021) Landslide susceptibility mapping using information value method in Jhelum Valley of the Himalayas. Arab J Geosci 14:824. doiL10.1007/s12517-021-07147-7
Fei, S; Li, L; Han, Z; Chen, Z; Xiao, Y. Combining novel feature selection strategy and hyperspectral vegetation indices to predict crop yield. Plant Methods; 2022; 18, 119. [DOI: https://dx.doi.org/10.1186/s13007-022-00949-0]
Feng, L; Zhang, M; Mao, Y; Liu, H; Yang, C; Dong, Y; Nanehkaran, YA. Convolutional neural network-based deep learning for landslide susceptibility mapping in the Bakhtegan watershed. Sci Rep; 2025; 15, 13250.1:CAS:528:DC%2BB2MXpt12hsLk%3D [DOI: https://dx.doi.org/10.1038/s41598-025-96748-3]
Funk, C; Peterson, P; Landsfeld, M; Pedreros, D; Verdin, J; Shukla, S; Husak, G; Rowland, J; Harrison, L; Hoell, A; Michaelsen, J. The climate hazards infrared precipitation with stations-a new environmental record for monitoring extremes. Sci Data; 2015; 2, 150066. [DOI: https://dx.doi.org/10.1038/sdata.2015.66]
Gariano, SL; Guzzetti, F. Landslides in a changing climate. Earth-Sci Rev; 2016; 162, pp. 227-252. [DOI: https://dx.doi.org/10.1016/j.earscirev.2016.08.011]
Getachew, N; Meten, M. Weights of evidence modeling for landslide susceptibility mapping of Kabi-Gebro locality, gundomeskel area, central Ethiopia. Geoenviron Disasters; 2021; 8, 6. [DOI: https://dx.doi.org/10.1186/s40677-021-00177-z]
Ghahraman, K; Nagy, B; Nokhandan, FN. Flood-prone zones of meandering rivers: machine learning approach and considering the role of morphology (Kashkan River, Western Iran). Geosci; 2023; 13,
Goetz, J; Brenning, A; Petschko, H; Leopold, P. Evaluating machine learning and statistical prediction techniques for landslide susceptibility modeling. Comput Geosci; 2015; 81, pp. 1-11. [DOI: https://dx.doi.org/10.1016/j.cageo.2015.04.007]
Grana, D; Della Rossa, E. Probabilistic petrophysical-properties Estimation integrating statistical rock physics with seismic inversion. Geophysics; 2010; 75,
Gu, T; Li, J; Wang, M; Duan, P; Zhang, Y; Cheng, L. Study on landslide susceptibility mapping with different factor screening methods and random forest models. PLoS ONE; 2023; 18,
Gulbet, E; Getahun, B. Landslide susceptibility mapping using frequency ratio and analytical hierarchy process method in Awabel Woreda, Ethiopia. Quat Sci Adv; 2024; 15, 100246. [DOI: https://dx.doi.org/10.1016/j.qsa.2024.100246]
Guzzetti, F; Carrara, A; Cardinali, M; Reichenbach, P. Landslide hazard evaluation: a review of current techniques and their application in a multi-scale study, central Italy. Geomorphology; 1999; 31, pp. 181-216. [DOI: https://dx.doi.org/10.1016/S0169-555X(99)00078-1]
Hakim, WL; Rezaie, F; Nur, AS; Panahi, M; Khosravi, K; Lee, CW; Lee, S. Convolutional neural network (CNN) with metaheuristic optimization algorithms for landslide susceptibility mapping in Icheon, South Korea. J Environ Manage; 2022; 305, 114367. [DOI: https://dx.doi.org/10.1016/j.jenvman.2021.114367]
Halder, K; Srivastava, AK; Ghosh, A; Das, S; Banerjee, S; Pal, SC; Chatterjee, U; Bisai, D; Ewert, F; Gaiser, T. Improving landslide susceptibility prediction through ensemble recursive feature elimination and meta-learning framework. Sci Rep; 2025; 15, 5170.1:CAS:528:DC%2BB2MXjvFOgu70%3D [DOI: https://dx.doi.org/10.1038/s41598-025-87587-3hal]
Hamedi, H; Alesheikh, AA; Panahi, M; Lee, S. Landslide susceptibility mapping using deep learning models in ardabil Province, Iran. Stoch Environ Res Risk Assess; 2022; 12, pp. 4287-4310. [DOI: https://dx.doi.org/10.1007/s00477-022-02263-6]
Hepdeniz, K. Using the analytic hierarchy process and frequency ratio methods for landslide susceptibility mapping in Isparta-Antalya highway (D-685), Turkey. Arab J Geosci; 2020; 13, 795. [DOI: https://dx.doi.org/10.1007/s12517-020-05764-2]
Hochreiter, S; Schmidhuber, J. Long short-term memory. Neural Comput; 1997; 9,
Hong, H; Pradhan, B; Bui, DT; Xu, C. Comparison of four kernel functions used in support vector machines for landslide susceptibility mapping: a case study at Suichuan area (China). Geomat Nat Hazards Risk; 2016; 8, pp. 544-569. [DOI: https://dx.doi.org/10.1080/19475705.2016.1250112]
Hong, H; Liu, J; Zhu, AX; Shahabi, H; Pham, BT; Chen, W. A novel hybrid integration model using support vector machines and random subspace for weather-triggered landslide susceptibility assessment in the Wuning area (China). Environ Earth Sci; 2017; 76, 652. [DOI: https://dx.doi.org/10.1007/s12665-017-6981-2]
Howard, AD. Drainage analysis in geologic interpretation: a summation. Am Assoc Petrol Geol Bull; 1967; 51, pp. 2246-2259. [DOI: https://dx.doi.org/10.1306/5D25C26D-16C1-11D7-8645000102C1865D]
Huang, Y; Zhao, L. Review on landslide susceptibility mapping using support vector machines. CATENA; 2018; 165, pp. 520-529. [DOI: https://dx.doi.org/10.1016/j.catena.2018.03.003]
Huang, W; Ding, M; Li, Z; Yu, J; Ge, D; Liu, D; Yang, J. Landslide susceptibility mapping and dynamic response along the Sichuan-Tibet transportation corridor using deep learning algorithms. CATENA; 2023; 222, 106866. [DOI: https://dx.doi.org/10.1016/j.catena.2022.106866]
Intarat, K; Yoomee, P; Hussadin, A; Lamprom, W. Assessment of landslide susceptibility in the intermontane basin area of Northern Thailand. Environ Nat Resour J; 2024; 22,
Intarawichian, N; Dasananda, S. Analytical hierarchy process for landslide susceptibility mapping in lower Mae Cham watershed, Northern Thailand. Suranaree J Sci Technol; 2010; 17,
Jaafari, A; Panahi, M; Pham, BT; Shahabi, H; Bui, DT; Rezaie, F; Lee, S. Meta optimization of an adaptive neuro-fuzzy inference system with grey Wolf optimizer and biogeography-based optimization algorithms for Spatial prediction of landslide susceptibility. CATENA; 2019; 175, pp. 430-445. [DOI: https://dx.doi.org/10.1016/j.catena.2018.12.033]
Jasiewicz, J; Stepinski, TF. Geomorphons-a pattern recognition approach to classification and mapping of landforms. Geomorphology; 2013; 182, pp. 147-156. [DOI: https://dx.doi.org/10.1016/j.geomorph.2012.11.005]
Jibson, RW; Tanyas, H. The influence of frequency and duration of seismic ground motion on the size of triggered landslides—a regional view. Eng Geol; 2020; 273, 105671. [DOI: https://dx.doi.org/10.1016/j.enggeo.2020.105671]
Kalantar, B; Pradhan, B; Naghibi, SA; Motevalli, A; Mansor, S. Assessment of the effects of training data selection on the landslide susceptibility mapping: a comparison between support vector machine (SVM), logistic regression (LR) and artificial neural networks (ANN). Geomatics Nat Hazards Risk; 2017; 9,
Karra K, Kontgis C, Statman-Weil Z, Mazzariello JC, Mathis M, Brumby SP (2021) Global land use/land cover with Sentinel 2 and deep learning. IEEE International Geoscience and Remote Sensing Symposium IGARSS, Brussels, Belgium, pp. 4704–4707. https://doi.org/10.1109/IGARSS47720.2021.9553499
Kavzoglu, T; Sahin, EK; Colkesen, I. Landslide susceptibility mapping using GIS-based multi-criteria decision analysis, support vector machines, and logistic regression. Landslides; 2014; 11,
Kayitesi, NM; Guzha, AC; Mariethoz, G. Impacts of land use land cover change and climate change on river hydro-morphology- a review of research studies in tropical regions. J Hydrol; 2022; 615, 128702. [DOI: https://dx.doi.org/10.1016/j.jhydrol.2022.128702]
Keles, F; Nefeslioglu, HA. Infinite slope stability model and steady-state hydrology-based shallow landslide susceptibility evaluations: the Guneysu catchment area (Rize, Turkey). CATENA; 2021; 200, 105161. [DOI: https://dx.doi.org/10.1016/j.catena.2021.105161]
Khan, H; Shafique, M; Khan, MA; Bacha, MA; Shah, SU; Calligaris, C. Landslide susceptibility assessment using frequency Ratio, a case study of Northern Pakistan. Egypt J Remote Sens Space Sci; 2019; 22,
Komadja, GC; Pradhan, SP; Oluwasegun, AD; Roul, AR; Stanislas, TT; Laibi, RA; Adebayo, B; Onwualu, AP. Geotechnical and geological investigation of slope stability of a section of road cut debris-slopes along NH-7, Uttarakhand, India. Results Eng; 2021; 10, 100277. [DOI: https://dx.doi.org/10.1016/j.rineng.2021.100227]
Komori, D; Rangsiwanichpong, P; Inoue, N; Ono, K; Watanabe, S; Kazama, S. Distributed probability of slope failure in Thailand under climate change. Clim Risk Manag; 2018; 20, pp. 126-137. [DOI: https://dx.doi.org/10.1016/j.crm.2018.03.002]
Kumar, D; Thakur, M; Dubey, CS; Shukla, DP. Landslide susceptibility mapping & prediction using support vector machine for Mandakini river Basin, Garhwal Himalaya, India. Geomorphology; 2017; 295, pp. 115-125. [DOI: https://dx.doi.org/10.1016/j.geomorph.2017.06.013]
Kumar, C; Walton, G; Santi, P; Luza, C. An ensemble approach of feature selection and machine learning models for regional landslide susceptibility mapping in the arid mountainous terrain of Southern Peru. Remote Sens; 2023; 15,
Land Development Department (2024) Soil service. Available via. https://dinonline.ldd.go.th/#header_top_border. Accessed 1 June 2025
Laurencelle J, Logan T, Gens R (2015) ASF radiometrically terrain corrected ALOS PALSAR products. Available via https://asf.alaska.edu/wp-content/uploads/2019/03/rtc_product_guide_v1.2.pdf. Accessed 10 August 2025
Lee, S; Sambath, T. Landslide susceptibility mapping in the damrei Romel Area, Cambodia using frequency ratio and logistic regression models. Environ Geol; 2006; 50, pp. 847-855. [DOI: https://dx.doi.org/10.1007/s00254-006-0256-7]
Lee, S; Ryu, J-H; Won, J-S; Park, H-J. Determination and application of the weights for landslide susceptibility mapping using an artificial neural network. Eng Geol; 2004; 71, pp. 289-302. [DOI: https://dx.doi.org/10.1016/S0013-7952(03)00142-X]
Li X, Wang Y, Basu S, Kumbier K, Yu B (2019) A debiased MDI feature importance measure for random forests. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, Vancouver, Canada, 8–14 December 2019, pp. 8049–8059
Li, N; Feng, G; Zhao, Y; Xiong, Z; He, L; Wang, X; Wang, W; An, Q. A deep-learning-based algorithm for landslide detection over wide areas using InSAR images considering topographic features. Sensor; 2024; 24,
Liao, M; Wen, H; Yang, L. Identifying the essential conditioning factors of landslide susceptibility models under different grid resolutions using hybrid machine learning: a case of Wushan and Wuxi counties, China. CATENA; 2022; 217, 106428. [DOI: https://dx.doi.org/10.1016/j.catena.2022.106428]
Liu, JG; Mason, PJ; Yu, E; Wu, M-C; Tang, C; Huang, R; Liu, R. GIS modeling of earthquake damage zones using satellite remote sensing and DEM data. Geomorphology; 2012; 139–140, pp. 518-535. [DOI: https://dx.doi.org/10.1016/j.geomorph.2011.12.002]
Liu, X; Shao, S; Shao, S. Landslide susceptibility zonation using the analytical hierarchy process (AHP) in the great xi’an Region, China. Sci Rep; 2024; 14, 2941.1:CAS:528:DC%2BB2cXivFChsr0%3D [DOI: https://dx.doi.org/10.1038/s41598-024-53630-y]
Liu, Y; Qiu, H; Kamp, U; Wang, N; Wang, J; Huang, C; Tang, B. Higher temperature sensitivity of retrogressive thaw slump activity in the Arctic compared to the third pole. Sci Total Environ; 2024; 914, 170007.1:CAS:528:DC%2BB2cXht12rtLk%3D [DOI: https://dx.doi.org/10.1016/j.scitotenv.2024.170007]
Lucchese, LV; De Oliveira, GGD; Pedrollo, OC. Investigation of the influence of nonoccurrence sampling on landslide susceptibility assessment using artificial neural networks. CATENA; 2021; 198, 105067. [DOI: https://dx.doi.org/10.1016/j.catena.2020.105067]
Ma, Z; Mei, G. Forecasting landslide deformation by integrating domain knowledge into interpretable deep learning considering Spatiotemporal correlations. J Rock Mech Geotech; 2025; 17, pp. 960-982. [DOI: https://dx.doi.org/10.1016/j.jrmge.2024.02.034]
Ma, J; Ding, Y; Gan, VJ; Lin, C; Wan, Z. Spatiotemporal prediction of PM2.5 concentrations at different time granularities using IDW-BLSTM. IEEE Access; 2019; 7, pp. 107897-107907. [DOI: https://dx.doi.org/10.1109/ACCESS.2019.2932445]
Machay, F; Moussaoui, SE; Talibi, HE. Insights into large landslide mechanisms in tectonically active Agadir, morocco: the significance of lithological, geomorphological, and soil characteristics. Sci Afr; 2023; 22, e01901. [DOI: https://dx.doi.org/10.1016/j.sciaf.2023.e01901]
Mairaing W (2006) Landslide problems and warning by geotechnical methods. In the 5th International Symposium on New Technology for Urban Safety of Mega Cities in Asia organized by Asian Institute of Technology (AIT), Phuket, Thailand, 16–17 November 2006
Mallick, J; Alqadhi, S; Talukdar, S; Alsubih, M; Ahmed, M; Khan, RA; Kahla, NB; Abutayeh, SM. Risk assessment of resources exposed to rainfall induced landslide with the development of GIS and RS based ensemble metaheuristic machine learning algorithms. Remote Sens; 2021; 13, 457. [DOI: https://dx.doi.org/10.3390/su13020457]
Mallick, J; Alkahtani, M; Hang, HT; Singh, CK. Game-theoretic optimization of landslide susceptibility mapping: A comparative study between Bayesian-optimized basic neural network and new generation neural network models. Environ Sci Poll Res; 2024; 31,
Mandal, S; Mondal, S. Statistical approaches for landslide susceptibility assessment and prediction; 2019; Berlin, Springer: [DOI: https://dx.doi.org/10.1007/978-3-319-93897-4]
Manopkawee, P; Mankhemthong, N. Assessment of landslide susceptibility using geoinformatics and a frequency ratio model: a case study of Mae tha river watershed in Northern thailand. Asia Pac J Sci Technol; 2024; 29,
Manopkawee, P; Mankhemthong, N. Landslide susceptibility assessment using the frequency ratio model in the Mae Chan river watershed, Northern Thailand. Quat Sci Adv; 2025; 17, 100263. [DOI: https://dx.doi.org/10.1016/j.qsa.2024.100263]
Manopkawee, P; Mankhemthong, N; Pattarakamolsen, C. Tectonic and lithologic controls on the landscape adjustment along the Eastern terrain of the Mae tha fault, Northern thailand. Geol Acta; 2023; 22,
Manopkawee, P; Mankhemthong, N; Kanthata, S; Sooksabai, C. Quantifying regional-scale flash flood susceptibility using watershed geomorphometry-based approach along the high Range, West of Chiang Mai Basin, Northern Thailand. Nat Hazards Res Doi; 2025; [DOI: https://dx.doi.org/10.1016/j.nhres.2025.07.001]
Mao, Y; Mwakapesa, DS; Wang, G; Nanehkaran, YA; Zhang, M. Landslide susceptibility modelling based on AHC-OLID clustering algorithm. Adv Sp Res; 2021; 68,
Mas, JF; Filho, BS; Pontius, RG; Gutierrez, MF; Rodrigues, H. A suite of tools for ROC analysis of Spatial models. ISPRS Int J Geo-Inf; 2013; 2,
Mathew, TG; Ariffin, KS. Remote sensing technique for lineament extraction in association with mineralization pattern in central belt Peninsular Malaysia. J Phys Conf Ser; 2018; 1082, 012092. [DOI: https://dx.doi.org/10.1088/1742-6596/1082/1/012092]
Mejia, AI; Niemann, JD. Identification and characterization of dendritic, parallel, pinnate, rectangular, and trellis networks based on deviations from planform self- similarity. J Geophys Res; 2008; 113, F02015. [DOI: https://dx.doi.org/10.1029/2007JF000781]
Mengistu, F; Suryabhagavan, K; Raghuvanshi, TK; Lewi, E. Landslide hazard zonation and slope instability assessment using optical and InSAR data: a case study from Gidole town and its surrounding areas, Southern Ethiopia. Remote Sens Land; 2019; 3, pp. 1-14. [DOI: https://dx.doi.org/10.21523/gcj1.19030101]
Merghadi, A; Yunus, AP; Dou, J; Whiteley, J; ThaiPham, B; Bui, DT; Avtar, R; Abderrahmane, B. Machine learning methods for landslide susceptibility studies: a comparative overview of algorithm performance. Earth-Sci Rev; 2020; 207, 103225. [DOI: https://dx.doi.org/10.1016/j.earscirev.2020.103225]
Mondal, S; Mandal, S. Data-driven evidential belief function (EBF) model in exploring landslide susceptibility zones for the Darjeeling Himalaya, India. Geocarto Int; 2019; 35,
Moore, ID; Gessler, PE; Nielsen, GA; Peterson, GA. Soil attribute prediction using terrain analysis. Soil Sci Soc Am J; 1993; 57, pp. 443-452. [DOI: https://dx.doi.org/10.2136/sssaj1993.03615995005700020026x]
Moore, ID; Lewis, A; Gallant, JC. Jakeman, AJ; Beck, MB; McAleer, MJ. Terrain attributes: Estimation methods and scale effects. Modelling change in environmental systems; 1993; Chichester, Wiley: pp. 189-214.
Murasingh, S; Jah, R; Adamala, S. Geospatial technique for delineation of groundwater potential zones in mine and dense forest area using weighted index overlay technique. Groundw Sustain Dev; 2018; 7, pp. 387-399. [DOI: https://dx.doi.org/10.1016/j.gsd.2017.12.001]
Nair V, Hinton GE (2010) Rectified linear units improve restricted Boltzmann machines. In: Proceedings of the 27th International Conference on Machine Learning, Haifa, 21–24 June 2010, pp. 807–814
Nakileza, BR; Nedala, S. Topographic influence on landslides characteristics and implication for risk management in upper Manafwa catchment, Mt Elgon Uganda. Geoenviron Disasters; 2020; 7, 27. [DOI: https://dx.doi.org/10.1186/s40677-020-00160-0]
Niu, Q; Dang, X; Li, Y; Zhang, Y; Lu, X; Gao, W. Suitability analysis for topographic factors in loess landslide research: a case study of Gangu County, China. Environ Earth Sci; 2018; 77, 294. [DOI: https://dx.doi.org/10.1007/s12665-018-7462-y]
Niu, H; Shao, S; Gao, J; Jing, H. Research on GIS-based information value model for landslide geological hazards prediction in soil-rock contact zone in Southern Shaanxi. Phys Chem Earth (Pt AB C); 2024; 133, 103515. [DOI: https://dx.doi.org/10.1016/j.pce.2023.103515]
Ohlmacher, GC. Plan curvature and landslide probability in regions dominated by Earth flows and Earth slides. Eng Geol; 2007; 91, pp. 117-134. [DOI: https://dx.doi.org/10.1016/j.enggeo.2007.01.005]
Pacheco Quevedo, R; Velastegui-Montoya, A; Montalván-Burbano, N; Morante-Carballo, F; Korup, O; Rennó, CD. Land use and land cover as a conditioning factor in landslide susceptibility: a literature review. Landslides; 2023; 20, pp. 967-982. [DOI: https://dx.doi.org/10.1007/s10346-022-02020-4]
Panchal, S; Shrivastava, AK. Landslide hazard assessment using analytic hierarchy process (AHP): a case study of National highway 5 in India. Ain Shams Eng J; 2022; 13,
Papathanassiou, G; Valkaniotis, S; Ganas, A; Pavlides, S. GIS-based statistical analysis of the Spatial distribution of earthquake-induced landslides in the Island of Lefkada, ionian Islands, Greece. Landslides; 2013; 10, pp. 771-783. [DOI: https://dx.doi.org/10.1007/s10346-012-0357-1]
Park, S; Choi, C; Kim, B. Landslide susceptibility mapping using frequency ratio, analytic hierarchy process, logistic regression, and artificial neural network methods at the Inje area, Korea. Environ Earth Sci; 2013; 68, pp. 1443-1464. [DOI: https://dx.doi.org/10.1007/s12665-012-1842-5]
Pasang, S; Kubíček, P. Landslide susceptibility mapping using statistical methods along the Asian highway. Bhutan Geosci; 2020; 10,
Pérez-Peña, JV; Azañón, JM; Booth-Rea, G; Azor, A; Delgado, J. Differentiating geology and tectonics using a Spatial autocorrelation technique for the hypsometric integral. J Geophys Res Earth Surf; 2009; 114, F02018. [DOI: https://dx.doi.org/10.1029/2008JF001092]
Pham, BT; Bui, DT; Indra, P; Dholakia, M. Landslide susceptibility assessment at a part of Uttarakhand Himalaya, India using GIS–based statistical approach of frequency ratio method. Int J Eng Res Technol; 2015; 4,
Pham, BT; Bui, DT; Dholakia, MB; Prakash, I; Pham, HV. A comparative study of least square support vector machines and multiclass alternating decision trees for Spatial prediction of rainfall-induced landslides in a tropical cyclone area. Geotech Geol Eng; 2016; 34,
Pham, BT; Bui, DT; Pourghasemi, HR; Indra, P; Dholakia, MB. Landslide susceptibility assessment in the Uttarakhand area (India) using GIS: a comparison study of prediction capability of naïve bayes, multilayer perceptron neural networks, and functional trees methods. Theor Appl Climatol; 2017; 128, pp. 255-273. [DOI: https://dx.doi.org/10.1007/s00704-015-1702-9]
Pham, QB; Chandra Pal, S; Chakrabortty, R; Saha, A; Janizadeh, S; Ahmadi, K; Khedher, KM; Anh, DT; Tiefenbacher, JP; Bannari, A. Predicting landslide susceptibility based on decision tree machine learning models under climate and land use changes. Geocarto Int; 2021; 37,
Prabowo IA, Verdiansyah O, Prabowo R (2021) Extraction of lineament density analysis from ASTER DEM for determine the vein direction. In: Proceedings of the 2nd international conference on industrial and technology and information design, ICITID 2021, Yogyakarta, Indonesia, 30 August 2021
Pradhan, B; Lee, S. Delineation of landslide hazard areas on Penang Island, Malaysia, by using frequency ratio, logistic regression, and artificial neural network models. Environ Earth Sci; 2010; 60, pp. 1037-1054. [DOI: https://dx.doi.org/10.1007/s12665-009-0245-8]
Pradhan B, Mansor S, Pirasteh S (2011) Landslide susceptibility mapping: an assessment of the use of an advance neural network model with five different training strategies. In: Hiu PCL (ed.) Artificial Neural Networks – Application, InTech, pp. 361–388. https://doi.org/10.5772/2052
Pradhan, B; Dikshit, A; Lee, S; Kim, H. An explainable AI (XAI) model for landslide susceptibility modeling. Appl Soft Comput; 2023; 142, 110324. [DOI: https://dx.doi.org/10.1016/j.asoc.2023.110324]
Putriani, E; Wu, YM; Chen, CW; Ismulhadi, A; Fadli, DI. Development of landslide susceptibility mapping with a multi-variance statistical method approach in Kepahiang, Indonesia. Terr Atmos Ocean Sci; 2023; 34, 18. [DOI: https://dx.doi.org/10.1007/s44195-023-00050-6]
Quinlan, JR. Induction of decision trees. Mach Learn; 1986; 1, pp. 81-106. [DOI: https://dx.doi.org/10.1007/BF00116251]
Ransikarbum, K; Mason, SJ. Multiple-objective analysis of integrated relief supply and network restoration in humanitarian logistics operations. Int J Prod Res; 2016; 54,
Regmi, AD; Devkota, KC; Yoshida, K; Pradhan, B; Pourghasemi, HR; Kumamoto, T; Akgun, A. Application of frequency ratio, statistical index, and weights-of-evidence models and their comparison in landslide susceptibility mapping in central Nepal himalaya. Arab J Geosci; 2014; 7,
Reichenbach, P; Rossi, M; Malamud, BD; Mihir, M; Guzzetti, F. A review of statistically-based landslide susceptibility models. Earth-Sci Rev; 2018; 180, pp. 60-91. [DOI: https://dx.doi.org/10.1016/j.earscirev.2018.03.001]
Rihan M, Talukdar S, Naikoo MW, Shahfahad AR, Rahman A (2024) Improving landslide susceptibility prediction in Uttarakhand through hyper-tuned artificial intelligence and global sensitivity analysis. Earth Syst Environ 1–20. https://doi.org/10.1007/s41748-024-00457-2
Rodriguez, R. Models, methods, concepts and applications of the analytic hierarchy process. Interfaces; 2002; 32, pp. 93-94.
Roy, J; Saha, S; Arabameri, A; Blaschke, T; Bui, DT. A novel ensemble approach for landslide susceptibility mapping (LSM) in Darjeeling and Kalimpong Districts, West Bengal, India. Remote Sen; 2019; 11,
Roy, D; Sarkar, A; Kundu, P; Paul, S; Sarkar, BC. An ensemble of evidence belief function (EBF) with frequency ratio (FR) using Geospatial data for landslide prediction in Darjeeling Himalayan region of India. Quat Sci Adv; 2023; 11, 100092. [DOI: https://dx.doi.org/10.1016/j.qsa.2023.100092]
Rozos, D; Bathrellos, GD; Skillodimou, HD. Comparison of the implementation of rock engineering system and analytic hierarchy process methods, upon landslide susceptibility mapping, using GIS: a case study from the Eastern Achaia County of Peloponnesus, Greece. Environ Earth Sci; 2011; 63, pp. 49-63. [DOI: https://dx.doi.org/10.1007/s12665-010-0687-z]
Saaty, TL. A scaling method for priorities in hierarchical structures. J Math Psychol; 1977; 15, pp. 234-281. [DOI: https://dx.doi.org/10.1016/0022-2496(77)90033-5]
Saaty, TL. The analytic hierarchy process: Planning, priority Setting, resources allocation; 1980; New York, McGraw-Hill:
Saaty, TL. Decision making with the analytic hierarchy process decision making with the analytic hierarchy process. Int J Serv Sci; 2008; 1,
Saha, A; Saha, S. Comparing the efficiency of weight of evidence, support vector machine and their ensemble approaches in landslide susceptibility modelling: A study on Kurseong region of Darjeeling Himalaya, India. Remote Sens Appl Soc Environ; 2020; 19, 100323. [DOI: https://dx.doi.org/10.1016/j.rsase.2020.100323]
Sahana, M; Sajjad, H. Evaluating effectiveness of frequency ratio, fuzzy logic and logistic regression models in assessing landslide susceptibility: a case from Rudraprayag district, India. J Mt Sci; 2017; 14, pp. 2150-2167. [DOI: https://dx.doi.org/10.1007/s11629-017-4404-1]
Saito, H; Nakayama, D; Matsuyama, H. Comparison of landslide susceptibility based on a decision-tree model and actual landslide occurrence: the Akaishi mountains. Japan Geomorphology; 2009; 109, pp. 108-121. [DOI: https://dx.doi.org/10.1016/j.geomorph.2009.02.026]
Sarkar, S; Kanungo, D. An integrated approach for landslide susceptibility mapping using remote sensing and GIS. Photogramm Eng Remote Sens; 2004; 70,
Sarkar, S; Roy, AK; Martha, TR. Landslide susceptibility assessment using information value method in parts of the Darjeeling Himalayas. J Geol Soc India; 2013; 82, pp. 351-362. [DOI: https://dx.doi.org/10.1007/s12594-013-0162-z]
Satarzadeh, E; Sarraf, A; Hajikandi, H; Sadeghian, MS. Flood hazard mapping in Western iran: assessment of deep learning vis-à-vis machine learning models. Nat Hazards; 2022; 111,
Savi, S; Comiti, F; Strecker, MR. Pronounced increase in slope instability linked to global warming: A case study from the Eastern European alps. Earth Surf Process Landf; 2021; 46,
Segoni, S; Tofani, V; Rosi, A; Catani, F; Casagli, N. Combination of rainfall thresholds and susceptibility maps for dynamic landslide hazard assessment at regional scale. Front Earth Sci; 2018; 6, 85. [DOI: https://dx.doi.org/10.3389/feart.2018.00085]
Sekiguchi, T; Sato, HP. Mapping of micro topography using airborne laser scanning. Landslides; 2004; 1, pp. 195-202. [DOI: https://dx.doi.org/10.1007/s10346-004-0021-5]
Selamat, SN; Majid, NA; Taha, MR; Osman, A. Landslide susceptibility model using artificial neural network (ANN) approach in langat river Basin, Selangor, Malaysia. Land; 2022; 11,
Semnani, SJ; Han, Y; Bonfils, CJ; White, JA. Assessing the impact of climate change on rainfall-triggered landslides: a case study in California. Landslides; 2025; 22, pp. 1707-1724. [DOI: https://dx.doi.org/10.1007/s10346-024-02428-0]
Sham, FAF; El-Shafie, A; Jaafar, WZBW; Adarsh, A; Sherif, M; Ahmed, AN. Improving rainfall forecasting using deep learning data fusing model approach for observed and climate change data. Sci Rep; 2025; 15, 27872.1:CAS:528:DC%2BB2MXitVChsrjF [DOI: https://dx.doi.org/10.1038/s41598-025-13567-2]
Shano, L; Raghuvanshi, T; Meten, M. Landslide susceptibility evaluation and hazard zonation techniques—a review. Geoenviron Disasters; 2020; 7, 18. [DOI: https://dx.doi.org/10.1186/s40677-020-00152-0]
Sharif Razavian A, Azizpour H, Sullivan J, Carlsson S (2014) CNN features off-the-shelf: an astounding baseline for recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp 806–813
Shi XJ, Chen ZR, Wang H, Yeung DY, Wong WK, Woo WC (2015) Convolutional LSTM Network: A machine learning approach for precipitation nowcasting. In: Cortes C, Lawrence ND, Lee DD, Sugiyama M, Garnett R (eds.) Advances in neural information processing systems 28. Neural Information Processing Systems (Nips), Cambridge, MA, USA, pp 802–810
Shin, HC; Roth, HR; Gao, M; Lu, L; Xu, Z; Nogues, I; Yao, J; Mollura, D; Summers, RM. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning. IEEE Trans Med Imaging; 2016; 35,
Silalahi, FES; Pamela Arifianti, Y; Hidayat, F. Landslide susceptibility assessment using frequency ratio model in Bogor, West Java, Indonesia. Geosci Lett; 2019; 6, 10. [DOI: https://dx.doi.org/10.1186/s40562-019-0140-4]
Sujatha, ER; Sridhar, V. Landslide susceptibility analysis: a logistic regression model case study in Coonoor, India. Hydrology; 2021; 8,
Sun, D; Wen, H; Wang, D; Xu, J. A random forest model of landslide susceptibility mapping based on hyperparameter optimization using Bayes algorithm. Geomorphology; 2020; 362, 107201. [DOI: https://dx.doi.org/10.1016/j.geomorph.2020.107201]
Sun, D; Shi, S; Wen, H; Xu, J; Zhou, X; Wu, J. Hybrid optimization method of factor screening predicated on geodetectorand random forest for landslide susceptibility mapping. Geomorphology; 2021; 379, 107623. [DOI: https://dx.doi.org/10.1016/j.geomorph.2021.107623]
Thapa, D; Bhandari, B. GIS-based frequency ratio method for identification of potential landslide susceptible area in the Siwalik zone of Chatara-Barahakshetra Section, Nepal. Open J Geol; 2019; 9, pp. 873-896. [DOI: https://dx.doi.org/10.4236/ojg.2019.912096]
Thongley, T; Vansarochana, C. Landslide susceptibility assessment using frequency ratio model at Ossey watershed area in Bhutan. Eng Appl Sci Res; 2021; 48,
Tukku, MA; Karuppannan, S; Meten, M; Garo, T; Fekadu, A. Landslide susceptibility modeling in West Shoa Zone, central ethiopia: an integrated Geospatial and statistical analysis. Sci Afr; 2025; 28, e02772. [DOI: https://dx.doi.org/10.1016/j.sciaf.2025.e02772]
Vahedifard, F; Leshchinsky, D; Mortezaei, K; Lu, N. Effective stress-based limit-equilibrium analysis for homogeneous unsaturated slopes. Int J Geomech; 2016; 16,
Van Den Eeckhaut, M; Vanwalleghem, T; Poesen, J; Govers, G; Verstraeten, G; Vandekerckhove, L. Prediction of landslide susceptibility using rare events logistic regression: a case-study in the Flemish Ardennes (Belgium). Geomorphology; 2006; 76, pp. 392-410. [DOI: https://dx.doi.org/10.1016/j.geomorph.2005.12.003]
Wang, Y; Fang, Z; Hong, H. Comparison of convolutional neural networks for landslide susceptibility mapping in Yanshan County, China. Sci Total Environ; 2019; 666, pp. 975-993.1:CAS:528:DC%2BC1MXjvFOhurY%3D [DOI: https://dx.doi.org/10.1016/j.scitotenv.2019.02.263]
Wang, Q; Zhang, H; Ma, X; Wang, X. A review of landslide susceptibility mapping using machine learning techniques. Geomat Nat Haz Risk; 2021; 12,
Wang, JN; Wang, YQ; Li, C; Li, YM; Qi, HM. Landslide susceptibility evaluation based on landslide classification and ANN-NFR modelling in the three Gorges reservoir area. China Ecol Indic; 2024; 160, 111920. [DOI: https://dx.doi.org/10.1016/j.ecolind.2024.111920]
Wang K, Ge C, Chen Z, Tang Q (2024b) Stability analysis and parameter sensitivity analysis of slopes under rainfall infiltration conditions. In: Duc Long P, Dung NT (eds) Proceedings of the 5th International Conference on Geotechnics for Sustainable Infrastructure Development. GEOTEC 2023. Lecture notes in civil engineering, vol 395. Springer, Singapore. https://doi.org/10.1007/978-981-99-9722-0_113
Wei, X; Zhang, L; Luo, J; Liu, D. A hybrid framework integrating physical model and convolutional neural network for regional landslide susceptibility mapping. Nat Hazards; 2021; 109, pp. 471-497. [DOI: https://dx.doi.org/10.1007/s11069-021-04844-0]
Wu, X; Chen, X; Zhan, FB; Hong, S. Global research trends in landslides during 1991–2014: a bibliometric analysis. Landslides; 2015; 12, pp. 1215-1226. [DOI: https://dx.doi.org/10.1007/s10346-015-0624-z]
Wu, Y; Ke, Y; Chen, Z; Liang, S; Zhao, H; Hong, H. Application of alternating decision tree with adaboost and bagging ensembles for landslide susceptibility mapping. CATENA; 2020; 187, 104396. [DOI: https://dx.doi.org/10.1016/j.catena.2019.104396]
Wu, H; Niu, J; Li, Y; Wang, Y; Qiu, D. Landslide susceptibility prediction based on a CNN–LSTM–SAM–attention hybrid model. Appl Sci; 2025; 15,
Xiang, Z; Dou, J; Zhang, L; Fu, Y; Yao, X; Yang, X; Dong, A; Ma, H. Towards a synergistic progressive ensemble framework for automatic post-earthquake landslide recognition and susceptibility assessment. Math Geosci; 2025; 57, pp. 703-732. [DOI: https://dx.doi.org/10.1007/s11004-024-10168-z]
Xiao, T; Zhang, L-M. Data-driven landslide forecasting: Methods, data completeness, and real-time warning. Eng Geol; 2023; 317, 107068. [DOI: https://dx.doi.org/10.1016/j.enggeo.2023.107068]
Xie, G; Chen, X; Meng, Y; Zhang, L; Qiao, L; Tan, L. A comparative study of landslide susceptibility mapping using weight of evidence, logistic regression and support vector machine and evaluated by SBAS-InSAR monitoring: Zhouqu to Wudu segment in Bailong river Basin, China. Environ Earth Sci; 2017; 76, 313. [DOI: https://dx.doi.org/10.1007/s12665-017-6640-7]
Xie, C; Xu, C; Huang, Y; Liu, J; Jin, J; Xu, X; Cheng, J; Wu, L. Detailed inventory and initial analysis of landslides triggered by extreme rainfall in the Northern Huaiji County, Guangdong Province, China, from June 6 to 9, 2020. Geoenviron Disasters; 2025; 12, 7. [DOI: https://dx.doi.org/10.1186/s40677-025-00311-1]
Yalcin, A. GIS-based landslide susceptibility mapping using analytical hierarchy process and bivariate statistics in Ardesen (Turkey): comparisons of results and confirmations. CATENA; 2008; 72,
Yalcin, A. Environmental impacts of landslides: a case study from the East black sea Region, Turkey. Environ Eng Sci; 2011; 24,
Yang, IT; Acharya, TD; Lee, DH. Landslide susceptibility mapping for 2015 earthquake region of Sindhupalchowk, Nepal using frequency ratio. J Korean Soc Surv Geod Photogramm Cartogr; 2016; 34,
Yang, X; Liu, R; Yang, M; Chen, J; Liu, T; Yang, Y; Chen, W; Wang, Y. Incorporating landslide Spatial information and correlated features among conditioning factors for landslide susceptibility mapping. Remote Sens; 2021; 13, pp. 2166-2190. [DOI: https://dx.doi.org/10.3390/rs13112166]
Yesilnacar, E; Topal, T. Landslide susceptibility mapping: a comparison of logistic regression and neural networks methods in a medium scale study, Hendek region (Turkey). Eng Geol; 2005; 79,
Yilmaz, I. Landslide susceptibility mapping using frequency ratio, logistic regression, artificial neural networks and their comparison: A case study from Kat landslides (Tokat—Turkey). Comput Geosci; 2009; 35,
Yilmaz, C; Topal, T; Süzen, ML. GIS-based landslide susceptibility mapping using bivariate statistical analysis in Devrek (Zonguldak-Turkey). Environ Earth Sci; 2012; 65, pp. 2161-2178. [DOI: https://dx.doi.org/10.1007/s12665-011-1196-4]
Yongsiri, P; Soytong, P; Laongmanee, W; Uthaisri, P. Development of landslide risk area display system by using the Geospatial technology with daily rainfall data via the internet network in the Northern region of Thailand. Int J Agric Sci Tech; 2023; 19,
Yu, L; Cao, Y; Zhou, C; Wang, Y; Huo, Z. Landslide susceptibility mapping combining information gain ratio and support vector machines: a case study from Wushan segment in the three Gorges reservoir Area, China. Appl Sci; 2019; 9, 4756. [DOI: https://dx.doi.org/10.3390/app9224756]
Yu Q, Tolson BA, Shen H, Han M, Mai J, Lin J (2024) Enhancing long short-term memory (LSTM)-based streamflow prediction with a spatially distributed approach. Hydrol Earth Syst Sci 28:2107–2122. https://doi.org/10.5194/hess-28-2107-2024, 2024
Zangmene, FL; Ngapna, MN; Ateba, MCB; Mboudou, GMM; Defo, PLW; Kouo, RT; Dongmo, AK; Owona, S. Landslide susceptibility zonation using the analytical hierarchy process (AHP) in the Bafoussam-Dschang region (West Cameroon). Adv Space Res; 2023; 71, pp. 5282-5301. [DOI: https://dx.doi.org/10.1016/j.asr.2023.02.014]
Zhang, Y-x; Lan, H-x; Li, L-p; Wu, Y-m; Chen, J-h; Tian, N-m. Optimizing the frequency ratio method for landslide susceptibility assessment: a case study of the Caiyuan basin in the Southeast mountainous area of China. J Mountain Sci; 2020; 17, pp. 340-357. [DOI: https://dx.doi.org/10.1007/s11629-019-5702-6]
Zhang, S; Bai, L; Li, Y; Li, W; Xie, M. Comparing convolutional neural network and machine learning models in landslide susceptibility mapping: a case study in Wenchuan County. Front Environ Sci; 2022; 10, 886841. [DOI: https://dx.doi.org/10.3389/fenvs.2022.886841]
Zhao, Y; Wang, R; Jiang, Y; Liu, H; Wei, Z. GIS-based logistic regression for rainfall-induced landslide susceptibility mapping under different grid sizes in Yueqing, southeastern China. Eng Geol; 2019; 259, 105147. [DOI: https://dx.doi.org/10.1016/j.enggeo.2019.105147]
Zhao, B; Zhang, L; Gu, X; Luo, W; Yu, Z; Yuan, L. How is the occurrence of rainfall-triggered landslides related to extreme rainfall?. Geomorphology; 2025; 475, 109666. [DOI: https://dx.doi.org/10.1016/j.geomorph.2025.109666]
Zuo, P; Zhao, W; Yan, W; Jin, J; Yan, C; Wu, B; Shao, X; Wang, W; Zhou, Z; Wang, J. Landslide susceptibility mapping using an LSTM model with feature-selecting for the Yangtze river basin in China. Water; 2025; 17,
© The Author(s) 2025. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.