Intelligent Geoengineering 1 (2024) 1–18

Contents lists available at ScienceDirect

Intelligent Geoengineering
journal homepage: www.keaipublishing.com/en/journals/ige/

Review on the artificial intelligence-based methods in landslide detection

and susceptibility assessment: Current progress and future directions ]]
]]]]]]
]]

Yange Lia,b, Bangjie Fua, , Yueping Yinc, Xiewen Hud, Wenpei Wangc, Weidong Wanga, Xin Lia,
⁎

Guanping Longa
a
School of Civil Engineering, Central South University, Changsha 410075, China
b
Department of Engineering, University of Canterbury, Christchurch, New Zealand
c
China Institute of Geological Environment Monitoring, Beijing 100081, China
d
Southwest Jiaotong University, Chengdu 610031, China

A R T I C L E I N F O A B S T R A C T

Keywords: Landslides pose significant risks to human life, property, and the environment in mountainous regions. Effective de­
Artificial intelligence tection and susceptibility assessment are essential for mitigating these hazards. Recent advancements in artificial in­
Landslide detection telligence (AI), particularly deep convolutional neural networks, have notably enhanced both the efficiency and ac­
Susceptibility prediction curacy in this field. In this review, we provide a comprehensive summary of the up-to-date studies and applications of
Machine learning
AI-integrated methods in two key areas: landslide detection using remote sensing images and data-driven landslide
Deep learning
susceptibility assessment. We summarize the primary AI-based neural network structures and the frameworks em­
ployed for these purposes. Overall, the current body of research indicates that AI-based approaches have reached a
mature stage, with a convergence in methodologies leading to significant improvements in detection and assessment
accuracy. Despite these advancements, challenges remain, particularly regarding data quality and standardization,
robustness in complex environments, and the issue of overfitting. Addressing these limitations is crucial for advancing
the application of AI methods in landslide mitigation. Future research should focus on developing solutions to these
challenges, paving the way for more effective and widespread use of AI technologies in this critical area.

1. Introduction Timely collection of this critical information allows governments and

administrations to evaluate the impact on human lives, the economy,
Landslides are common geological hazards that frequently result in and the natural environment (Pradhan et al., 2016). Moreover, a
casualties, economic losses, and ecological damage, particularly in comprehensive landslide inventory, which includes details like extent,
mountainous regions with complex geological conditions and variable shape, and size, is vital for deepening our understanding of this com­
climates (Donnini et al., 2023; Dente et al., 2023). Annually, landslides plex phenomenon (Galli et al., 2008). Such inventories also play a key
have been reported to cause billions of dollars in economic losses role in hazard preparedness and early warning systems, providing
worldwide and this trend continues to rise (Marrapu and Jakka, 2014; crucial alerts to governments and at-risk populations (Yang et al., 2022;
Schulz et al., 2009). Beyond immediate destruction, landslides pose Noviello et al., 2020).
long-term threats to infrastructure such as roads, railways, and hy­ Building on historical landslide inventories, landslide susceptibility
draulic projects (Zhuang et al., 2018). Secondary hazards, such as prediction (LSP) is another critical component of hazard management
debris flows triggered by landslides, further exacerbate these impacts (Sajadi et al., 2022). LSP refers to the spatial probability of landslides
(Yin et al., 2009; Aleotti, 2004). occurring at specific locations, based on non-linear interactions be­
Rapid detection of landslides, along with systematic recording of tween multiple contributing factors (Huang et al., 2024; Reichenbach
their locations and attributes in a centralized inventory, are essential et al., 2018; Ermini et al., 2005; Guzzetti et al., 1999). These factors
for accurately assessing the resulting damages (Guzzetti et al., 2012). typically include hydrological, geological, topographical, and

Peer review under responsibility of Chinese Society for Rock Mechanics & Engineering
⁎
Corresponding author. School of Civil Engineering, Central South University, Changsha 410075, China.
E-mail address: [email protected] (B. Fu).

https://doi.org/10.1016/j.ige.2024.10.003
Received 3 September 2024; Received in revised form 27 September 2024; Accepted 30 October 2024
Available online 2 November 2024
3050-6190/© 2024 Chinese Society for Rock Mechanics & Engineering. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an
open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

environmental elements (Jiang et al., 2023; Chen et al., 2020). Geo­ learning and deep learning techniques (Fu et al., 2023; Han et al., 2023;
graphic Information System (GIS) tools compile these factors into a Mohan et al., 2021; Han et al., 2019; Li et al., 2014).
database, which can then be used to model and predict future landslides On-site investigations are considered highly accurate (van Westen
based on past events (Merghadi et al., 2020; Nandi and Shakoor, 2010). et al., 2006), but are often challenged by environmental complexity and
This approach is effective because it assumes that future landslides are non-accessibility in mountainous areas, making them inefficient for large-
likely to occur under conditions similar to those that have triggered past scale applications (Gao et al., 2021). Although reliable in small-scale areas,
events (Merghadi et al., 2020; Comert et al., 2018). High and very high such as sub-watersheds, on-situ methods are time-consuming, labor-in­
susceptibility levels identified in LSP results are crucial and require tensive, and limited in coverage. With the advancement of remote sensing
attention for landscape management and engineering planning technology, visual interpretation has become increasingly common in
(Pradhan et al., 2023; Titti et al., 2021; Saleem et al., 2019). landslide identification. This method relies on observing differences in
Traditional methods for landslide detection, such as on-site in­ geometry and texture between landslide areas and their surroundings,
vestigations and visual interpretation of remote sensing images, as well often incorporating topographic information (Guzzetti et al., 2012).
as knowledge-based and physically based approaches to susceptibility However, it is highly dependent on the interpreter’s experience and
prediction, have been verified to be effective, but often struggle to knowledge, which can lead to time-consuming processes and inconsistent
maintain accuracy and efficiency, especially during large-scale events. results (Xu, 2015; Galli et al., 2008). For instance, experienced annotators
Artificial Intelligence (AI), a key driver of the Fourth Paradigm shift in commonly spend several minutes to locate a landslide in an image, and
scientific research, has made significant advancements across various several more minutes to map the landslide polygon.
applications, including speech recognition (Sze et al., 2017; Hinton To improve detection efficiency, some researchers have explored
et al., 2012), autonomous vehicles (Grigorescu et al., 2020; Farabet threshold-based methods for landslide identification (Gabrieli et al.,
et al., 2012), computer vision (Krizhevsky et al., 2012), and natural 2016; Van Den Eeckhaut et al., 2012). These methods recognize landslide
language processing (Zhou et al., 2024; Collobert et al., 2011). The areas from the background by calculating the optimal threshold based on
development of AI technology, combined with increasing computa­ the grayscale distribution of images (Li et al., 2014). Various thresh­
tional power and the availability of large-scale remote sensing and olding algorithms have been proposed, each with different rationales (Im
geographic data, has made more precise and efficient landslide detec­ et al., 2011; Hammouche et al., 2008; Jiang and Mojon, 2003). While
tion and susceptibility prediction a key focus in engineering geology these methods enable rapid landslide detection, they can also produce
research (Dikshit et al., 2021), in particular landslide management noisy pixels, leading to a high rate of false positives as demonstrated in
which belongs to a typical decision-making task. the following studies, e.g., Han et al., (2019). To address these issues,
Recent advancements in artificial intelligence (AI), particularly deep some hybrid algorithms have been developed to enhance thresholding
convolutional neural networks, have notably enhanced both efficiency performance (Chou et al., 2010; Zhang and Yang, 2010; Tong et al.,
and accuracy in landslide hazard mitigation. In this review, we provide 2009). For example, Li et al. (2014) introduced a hybrid method in­
a comprehensive summary of the current state of AI technology in tegrating image differential technology, a multi-level local thresholding
landslide hazard mitigation research. The review is structured to guide strategy, and a directional recursive method, significantly improving
readers through the research fields. Two key areas where AI-integrated detection performance. Similarly, Han et al. (2019) incorporated Monte
methods have been widely reported are focused, i.e., landslide detec­ Carlo iteration with the local thresholding method to address challenges
tion using remote sensing images in Section 2, and data-driven land­ such as complex lighting conditions and block size uncertainty.
slide susceptibility assessment in Section 3. In retrospect, Section 4 With the blooming of Artificial Intelligence (AI), machine learning-
outlines the major challenges and future directions in these fields. based computer vision methods have gained traction in remote sensing
image classification tasks, such as Support Vector Machines (SVM)
2. Landslide detection using remote sensing images (Melgani and Bruzzone, 2004), Random Forests (RF) (Goetz et al.,
2015), and Artificial Neural Networks (ANN) (Wu and Feng, 2018).
2.1. Overview of landslide detection techniques These advancements have significantly impacted landslide research
(Chen et al., 2016; Chen et al., 2015). The computer vision-based ap­
Landslide detection is a crucial task in hazard and risk research, proaches can be primarily divided into pixel-based and object-oriented
focusing on identifying the location, distribution, number, and extent of methods (Gao et al., 2021). The pixel-based approach classifies land­
landslides within a given area (Mondini et al., 2011; Guzzetti et al., slide pixels based solely on the value of individual pixels in remote-
2000). Traditional methods for landslide detection (Table 1) can be sensing images (Lv et al., 2022). In contrast, the object-oriented ap­
categorized into on-site investigations, remote sensing image-based proach relies on initial image segmentation, with results heavily de­
visual interpretation, threshold-based approaches, and machine pendent on the chosen segmentation scale, which is often manually

Table 1
Comparison of landslide detection methods.

Techniques Advantages Disadvantages

On-site Investigation 1) Provides direct, detailed information 1) Time-consuming and labor-intensive

2) High detection accuracy 2) Limited spatial coverage
3) Allows in-site observation of ground conditions 3) Difficult to access remote or hazardous areas
Visual Interpretation 1) Intuitive and easy to implement 1) Highly subjective, dependent on the interpreter's experience
2) Large area coverage 2) May miss subtle features
3) Combines expert knowledge and experience 3) Time-consuming for large datasets
Threshold-based Method 1) Simple and straightforward 1) Relies on preset thresholds with poor generalizability
2) High computational efficiency 2) Difficult to capture complex patterns
3) Partially automated 3) Lower accuracy in heterogeneous environments
Machine Learning-based Method 1) Handles complex relationships between variables 1) Requires labeled training data
2) Higher accuracy than traditional methods 2) Sensitivity to data quality
3) Capable of processing large datasets 3) Requires parameter tuning and expertise
Deep Learning-based Method 1) Automatic feature extraction 1) Requires large amounts of training data
2) High accuracy and generalization ability 2) Computationally intensive
3) Suitable for processing high-dimensional data like images 3) Poor interpretability

2
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

adjusted based on image texture characteristics (Huang et al., 2021). detection performance (Ma and Mei, 2021). Using single-band or RGB
However, some image object segmentation algorithms are limited in bands as features can make it challenging for the model to distinguish
handling complex and large-scale remote sensing data, reducing land­ landslides from ground objects with similar spectral or geometric
slide detection efficiency (Chen et al., 2017). characteristics, such as soil areas. To address this, Ghorbanzadeh et al.
Recently, deep learning networks, particularly Convolutional Neural (2019) used a multi-channel input combining spectral and topographic
Networks (CNNs), have shown exceptional performance in image-based features to build a CNN model, analyzing the accuracy of different
tasks (Ding et al., 2022; Jiang et al., 2022; Liu et al., 2022; Sun et al., convolutional layers and sizes. Similarly, Sameen and Pradhan (2019)
2021). CNNs can automatically capture effective features from images trained a Residual Network (ResNet) on spectral and topographic fea­
through convolutional kernels, which allows for the understanding of tures for landslide recognition. Studies by Gao et al. (2021), Ullo et al.
semantic information without the need for manual feature extraction (2019), and Wang et al. (2019a,2019b,2019c) have also emphasized
(Zhang et al., 2018). According to the differences in landslide detection the importance of multi-source data input.
targets, it can be further classified into object detection, semantic seg­ In terms of network structure improvements, various CNN archi­
mentation, and instance segmentation. Specifically, landslide object tectures, as well as plug-and-play feature extraction and attention
detection aims to locate the target object in the image and output the modules, have been proposed and applied in landslide detection (Zhang
bounding box of the target (Han et al., 2024; Tanatipuknon et al., 2021; et al., 2024; Hacıefendioğlu et al., 2024; Wei et al., 2023; Ji et al., 2020;
Cheng et al., 2021). It shows high computation efficiency, but cannot Van Den Eeckhaut et al., 2012). For instance, Yu et al. (2020) proposed
provide the precise shape and extent of the landslide. As for landslide an end-to-end framework integrating a modified PSPNet for landslide
semantic segmentation, it can accurately outline the boundaries of the detection in Nepal. Prakash et al. (2020) used modified U-Net and
landslide area, which provides detailed geographic information and ResNet models for mapping landslide inventories. Han et al. (2023)
helps to assess the shape, area, and impact range of the landslide. In improved detection performance in complex environments by in­
terms of landslide instance segmentation, it is designed to effectively tegrating the Dynahead module into the YOLOv3 detection model. Fu
differentiate and segment individual landslide instances in complex et al. (2023) conducted landslide detection in the Longtoushan area by
scenarios containing multiple landslides (Ullo et al., 2021; Liu et al., embedding Reverse Image Pyramid Features (RIPFs) into the UNet
2021). However, the application of instance segmentation is still model. These innovative approaches have significantly advanced the
somewhat limited due to the high modeling complexity, such as chal­ application of AI and remote sensing technology in landslide detection.
lenges in labeling data and optimizing multi-loss. Due to the simplicity
in modeling and significant advantages in accurately delineating 2.2. Mainstream AI-based landslide detection structures
landslide areas, the landslide semantic segmentation technique is
widely used in landslide detection. Over the past two decades, several AI-based detection networks fundamentally rely on representation
advanced CNN algorithms have been developed and applied to land­ learning, which involves constructing nonlinear modules across mul­
slide semantic segmentation (Su et al., 2021; Meena et al., 2021; Yu tiple hidden layers. Each module transforms representations at various
et al., 2020; Ghorbanzadeh et al., 2019; Chollet, 2017). levels, enabling the learning of complex functions. This characteristic
These CNN-based methods can be categorized into change detec­ allows these models to automatically learn from data without the need
tion-based approaches, which use pre- and post-landslide images, and for manual feature design and selection. Although the field of deep
object detection-based approaches, which use only post-landslide learning encompasses various structures tailored to a wide range of
images. The change detection-based method identifies landslide areas issues, the dominant architecture for image recognition remains CNN-
by comparing variations across multi-temporal remote sensing images based, and most network structures are extensions of the basic CNN
(Lv et al., 2022; Shi et al., 2021; Lv et al., 2018). For instance, Lei et al., framework (Fig. 1).
(2019a) developed a deep learning model (FCN-PP) that introduced a Therefore, the mainstream CNN structures reported in up-to-date
pyramid pooling module to fuse multi-scale features, improving the studies will be introduced in this section, including FCN, PSPNet,
robustness of pre- and post-landslide image data and enhancing de­ DeepLabv3+, SegNet, UNet, and ResUnet. The following Table 2
tection accuracy. Similarly, Lv et al. (2020) created a novel Dual-Path briefly introduces and explains the underlying principles behind these
Fully Convolutional Network (DP-FCN) that detects landslide areas by commonly used convolutional neural networks.
directly extracting feature differences, avoiding traditional preproces­
sing steps such as binary threshold calculation. Other studies have also 2.2.1. Fully Convolutional Network (FCN)
validated the effectiveness of change detection-based approaches in The FCN is a pioneering structure in semantic segmentation tasks,
landslide detection (Ullo et al., 2021; Shi et al., 2021; Wang et al., specifically designed to handle pixel-level classification (Jonathan
2019a,2019b,2019c; Chen et al., 2018). However, imaging differences et al., 2015). Unlike traditional Convolutional Neural Networks (CNNs)
in multi-temporal sequences can result from factors other than land­ that include fully connected layers, FCNs replace these with convolu­
slides, such as changes in land cover and human activities, complicating tional layers that produce spatial maps of class scores. This approach
accurate detection (Lv et al., 2018). Additionally, access to multi-tem­ allows the network to maintain spatial hierarchies and generate dense
poral data remains challenging. predictions. FCNs utilize an encoder-decoder structure, where the en­
As a result, the commonly used approach is post-landslide image- coder extracts feature maps through successive convolutions, and the
based object detection. This method is easier to implement and less decoder up-samples these maps to match the original image dimen­
dependent on specific datasets. For example, Prakash et al. (2020) ap­ sions. Skip connections between the encoder and decoder layers help
plied U-Net for landslide detection using high-resolution data, while Yu retain fine-grained details, crucial for accurate segmentation (Huang
et al. (2020) employed ResNet-101 and pyramid pooling for landslide et al., 2024; Chen et al., 2022; Ghorbanzadeh et al., 2021; Lei et al.,
detection in Nepal. Qi et al. (2020) utilized ResUnet for landslide de­ 2019b).
tection with GeoEye-1 images, and Gao et al. (2021) compared the
performance of FCN, SegNet, U-Net, ResUNet, and DenseNet in land­ 2.2.2. Pyramid Scene Parsing Network (PSPNet)
slide detection. PSPNet enhances scene parsing by incorporating a pyramid pooling
Recent research in CNN-based landslide detection has increasingly module that captures multi-scale contextual information (Zhao et al.,
focused on multi-source data fusion and network structure improve­ 2017). The network is built on a ResNet backbone, with an additional
ments, such as more sophisticated feature extraction modules and at­ pyramid pooling module that aggregates global information and refines
tention mechanisms. From a data-driven perspective, using multi- features at different scales. This module aids in better understanding
source data for deep learning model training enhances landslide the scene's overall structure and improves segmentation performance

3
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Fig. 1. The development of CNN-based methods for landslide detection.

for objects of various sizes, making it a strong performer in scene par­ (decoder) connected by skip connections. These skip connections
sing tasks (Yong et al., 2023; Yang et al., 2022; Li and Guo, 2020). transfer feature maps from the encoder to the decoder, enabling the
network to combine low-level spatial information with high-level con­
2.2.3. DeepLabv3+ textual features. The UNet architecture is effective in precise boundary
DeepLabv3+ is an advanced network designed for semantic seg­ delineation and accurate segmentation, particularly for segmenting
mentation (Chen et al., 2018). It employs atrous convolution (or dilated small objects and fine structures (Devara et al., 2024; Fu et al., 2023;
convolution) to capture multi-scale contextual information while pre­ Dong et al., 2022; Fang et al., 2022; Bragagnolo et al., 2021).
serving spatial resolution. The Atrous Spatial Pyramid Pooling (ASPP)
module in DeepLabv3+ allows the network to incorporate features 2.2.6. ResUNet
from multiple dilation rates, enhancing its ability to segment objects of ResUNet combines the strengths of UNet and residual learning to
varying scales. The addition of a decoder module further refines the enhance segmentation performance (Zhang et al., 2018). It integrates
segmentation results by restoring spatial details lost during down­ residual blocks into the UNet architecture, allowing for deeper net­
sampling. DeepLabv3+ excels in segmenting complex scenes and small works with improved gradient flow and feature representation. The
objects due to its detailed feature extraction and contextual under­ residual blocks in ResUNet help mitigate the vanishing gradient pro­
standing capabilities (Wan et al., 2023; Gao et al., 2022). blem, enabling the network to learn more complex features. By com­
bining residual connections with UNet’s skip connections, ResUNet
2.2.4. SegNet enhances the network’s ability to capture detailed spatial information
SegNet is a deep convolutional network specifically designed for and achieve accurate segmentation results. This architecture is parti­
pixel-wise semantic segmentation tasks (Badrinarayanan et al., 2017). cularly effective in handling intricate details and variations in images,
It features an encoder-decoder structure with symmetric layers in both making it suitable for applications requiring high-resolution and precise
the encoder and decoder paths. The encoder extracts hierarchical fea­ segmentation (Ghorbanzadeh et al., 2023; Yang et al., 2022).
tures through a series of convolutional layers, while the decoder re­
constructs the image from these features using max-pooling indices to 2.3. Framework of AI-based landslide detection modelling
improve segmentation accuracy. This design allows SegNet to capture
refined details and provide accurate boundary information. The use of Based on previous studies (Hacıefendioğlu et al., 2024; Fu et al.,
max-pooling indices helps preserve spatial information, which is ben­ 2023; Ghorbanzadeh et al., 2022; Shi et al., 2021; Sameen and Pradhan,
eficial for segmenting objects with complex shapes and boundaries (Liu 2019), AI-based landslide detection using the above-mentioned neural
et al., 2021; Yu et al., 2021). network structures generally involves the following steps, i.e., 1) De­
termination of landslide detection strategy. 2) Collection of image data,
2.2.5. UNet detection support factors, and landslide inventory samples. 3) Con­
UNet is a U-shaped network structure initially proposed for bio­ struction of the dataset for the detection model. 4) Selection of the
medical image segmentation (Ronneberger et al., 2015). The network appropriate detection model. 5) Detection model training and accuracy
comprises a contracting path (encoder) and an expansive path evaluation, and 6) Application of the detection model in real scenarios.

Table 2
The underlying principles behind commonly-used CNNs.

Model Model Description Main Advantages

FCN Replaces fully connected layers with convolutional layers to handle images of Allows end-to-end pixel-level classification, suitable for
any size segmentation tasks
PSPNet Introduces pyramid pooling module to extract multi-scale feature information, Handles image details at different scales, suitable for complex
enhancing global context detection scenes
DeepLabv3 + Uses atrous (dilated) convolution and atrous spatial pyramid pooling (ASPP) to Effective in segmenting complex backgrounds, ideal for high-
expand receptive fields while maintaining high-resolution precision segmentation tasks
SegNet Uses an encoder-decoder architecture, progressively downscales spatial Low memory consumption, suitable for memory-constrained
resolution in the encoding stage, and restores it in the decoding stage environments
UNet Adopts a U-shaped structure, combining features from the encoding and Achieves efficient segmentation with limited data, suitable for image
decoding stages, particularly suited for small dataset tasks segmentation and similar fields
ResUnet Introduces residual blocks on top of UNet to address gradient vanishing and Capable of handling deeper networks with high segmentation
improve training efficiency in deep networks accuracy, ideal for tasks requiring precise segmentation

4
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Fig. 2. The detailed network structure of FCN (Jonathan et al., 2015).

Fig. 3. The detailed network structure of PSPNet (Zhao et al., 2017).

2.3.1. Determination of landslide detection strategy 2.3.2. Data collection

As mentioned in Section 2.1, the majority of landslide detection Data collection involves obtaining time series imagery, detection
strategies in up-to-date studies can be categorized into two main ap­ support factors, and landslide inventory samples. Satellite or aerial
proaches based on recognition strategies: change detection-based and images captured at different times are often used to analyze surface
object detection-based. The change detection-based approach involves changes before and after landslide events. Additionally, detection sup­
comparing multi-temporal image data to identify areas that have ex­ port factors such as topographical data and NDVI are crucial for im­
perienced significant changes over time (Shi et al., 2021; Lv et al., proving the accuracy of landslide detection (Wei et al., 2023; Sameen
2018). By analyzing these changes, it becomes possible to detect and Pradhan, 2019). Landslide inventory samples are used to create
landslide regions. On the other hand, the object detection-based ap­ target labels for model training. Ideally, the landslide sample dataset
proach utilizes segmentation algorithms to recognize landslides directly should encompass various types and scales of landslides to ensure the
from post-landslide images (Fu et al., 2023; Meena et al., 2021; model's generalization capability. It is encouraged that some research
Ghorbanzadeh et al., 2019). The selection of a strategy must be tailored has already been proposed (Ishikawa et al., 2021; Ji et al., 2020;
to the specific characteristics of the region and the availability of data. Ghorbanzadeh et al., 2019).

Fig. 4. The detailed network structure of DeepLabv3+ (Chen et al., 2018).

5
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Fig. 5. The detailed network structure of SegNet (Badrinarayanan et al., 2017).

Fig. 6. The detailed network structure of UNet.

2.3.3. Construction dataset for detection model 2.3.5. Detection model training and accuracy evaluation
Dataset construction typically involves defining input and output In this stage, the model is trained using the training dataset, with op­
variables and dividing the data into training and testing sets (Han et al., timization algorithms incrementally introduced to adjust model weights
2023). Input variables usually include data contributing to detection and improve detection performance. Accuracy evaluation is performed
model training, such as pre- and post-landslide images, DEM, and NDVI. using metrics such as Accuracy, Precision, Recall, F1 Score, Kappa, IoU,
These factors can be aligned and stacked to create input data with mIoU, AP, and mAP (Ghorbanzadeh et al., 2022; Lu et al., 2022; Ju et al.,
consistent shapes. The model’s purpose is to determine the spatial lo­ 2022; Zhao et al., 2019). Detection efficiency can be evaluated using
cation and boundary of landslides based on this input data. The dataset metrics like FPs and FLOPs (Zhang et al., 2024; Mo et al., 2023).
is divided into training and testing sets, with the training set used for
model development and the testing set for evaluating performance. 2.3.6. Application of detection model in the real scenario
Testing the model in real scenarios is crucial for assessing its ap­
2.3.4. Selection of the appropriate detection model plicability and reliability (Casagli et al., 2023). This involves evaluating
Selecting an appropriate detection model is a critical step, as it di­ the model's ability to predict landslides in previously unobserved areas
rectly affects the accuracy of detection results. The choice of detection with different environmental conditions. The performance in real-world
model, possibly combined with advanced feature extraction modules applications helps to understand the effectiveness and generalization
and attention mechanisms, can significantly enhance landslide detec­ capabilities of the trained model.
tion performance. The commonly used DL-based detection models are For example, Ghorbanzadeh et al. (2021) applied the DCNN-
described in Section 2.2, each based on different principles. based landslide detection in different study areas (Fig. 8). The

Fig. 7. The detailed network structure of ResUNet.

6
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Fig. 8. Landslide detection results of the ResUNet. (a) Shuzheng Valley in China (b) Western Taitung County in Taiwan, China (c) Eastern Iburi in Japan
(Ghorbanzadeh et al., 2021).

results show that ResUNet exhibits higher generalization (F1-score 3. Data-based landslide susceptibility prediction (LSP)
of 73 %), but the UNet model shows better transferability between
different geographical areas, with F1-score reached 73.32 %. Sree­ 3.1. Overview of LSP techniques
lakshmi et al. (2022) used Hong Kong's Lantau Island as the study
area, and the proposed revised DCNN model obtained a recognition To support the landslide mitigation work, the major task of LSP is to
accuracy of 92.5 %. Fu et al. (2023) trained UNet, UNet++, and answer the question "Where are landslides prone to occur?". The LSP
RIPT-UNet networks using the Bijie dataset and used it for landslide aims to provide reliable prediction of the spatial probability of potential
identification in the Ludian region (Fig. 9). The results showed that landslides using the general information of historical landslides.
the accuracy of landslide detection of the proposed RIPF-UNet Numerous studies have focused on LSP using GIS and remote sensing
model reached 96.63 % with F1-score of 73.76 %. These studies technology (Zhang et al., 2023; Paulin et al., 2022; Razavi-Termeh
have tested the performance of the constructed model for applica­ et al., 2021; Reichenbach et al., 2018).
tion in real scenarios, however, the detection accuracy of the case Following the classifications by Marjanović et al. (2011) and Tien
study is not as good as the training dataset, which may be due to the Bui et al., (2012), the study by Merghadi et al. (2020) categorized the
scenario differences (e.g., image spectra, resolution, and shadows) LSP methods into four major groups, i.e., physical-based models,
between the case study and the open-source dataset (Zou et al., knowledge-based models, data-based models, and more recently, ma­
2018). The survey revealed some typical study areas that are mainly chine learning-based (ML) models.
co-seismic landslide areas (Fu et al., 2023; Ghorbanzadeh et al., Physical-based models assess landslide stability by simulating or
2022; Ghorbanzadeh et al., 2021; Ji et al., 2020), which provide a calculating the slope failure processes, according to the soil mechanics
good basis for landslide identification studies, e.g., (a) Bijie City, and hydrological characteristics that contribute to landslides. These
(b) Longtongshan Town, (c) Shuzheng Valley, (d) Western Taitung models are highly detailed, aiming to replicate the physical conditions
County, (e) Eastern Iburi, (f) Kodagu District of Karnataka, and (g) within a specific area (Whiteley et al., 2019; Piciullo et al., 2018). They
Rasuwa District of Bagmati. allow for a deep understanding of landslide-contributing factors and

7
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Fig. 9. Landslide detection result of the RIPF-UNet. (a) Pre-seismic image (b) Post-seismic image (c) Detection result (Fu et al., 2023).

enable precise predictions for individual landslides or within a small et al., 2020; Guzzetti et al., 2006). Commonly used ML methods in LSP
local area, e.g., sub-catchments. However, their requirement for ex­ include Artificial Neural Networks (ANN) (e.g., Dou et al., 2015; Ermini
tensive and accurate site-specific data, such as soil properties, geolo­ et al., 2005), Support Vector Machines (SVM) (e.g., Huang et al., 2023;
gical conditions, and hydrological parameters, limits their practicality Zhang et al., 2023), Logistic Regression (e.g., Sun et al., 2021), Random
for broader and large-scale analyses. Forests (RF) (e.g., Zhang et al. 2023; Tanyu et al., 2021; Zhou et al.,
Knowledge-based methods play a central role in many LSP studies 2021), K-Nearest Neighbors (KNN) (e.g., Rabby et al., 2022), and
(Guzzetti et al., 2012). The knowledge-based approach relies on avail­ Bayesian Networks (Song et al., 2012). These methods have been ex­
able information about topographic and environmental variables which tensively introduced and applied in LSP. For example, Huang and Zhao
are known as landslide conditioning factors (LCFs), and often involves (2018) reviewed the application of SVM models in landslide suscept­
ranking and weighting these factors based on expert judgment ibility research, while LR, ANN, and RF have been identified as some of
(Merghadi et al., 2020). However, this approach presents challenges in the most used methods (Chen et al., 2017; Felicísimo et al., 2013). Each
objectively quantifying or validating results, as reliance on expert ML-based method is based on different underlying mechanisms. Al­
opinion can lead to subjective biases and inconsistencies. though there is no consensus on the optimal machine learning method
In contrast, data-based models are widely used due to their simpli­ for LSP, effective machine learning techniques remain crucial for im­
city and ability to incorporate various contributing factors in this proving LSP accuracy (Ganerød et al., 2023).
decade (Reichenbach et al., 2018). These models explore statistical Additionally, the impact of hyperparameters on the performance
relationships between historical landslides and environmental variables and generalization ability of machine learning methods has led to the
(e.g., slope, aspect, and land cover) to provide probabilistic estimates of integration of various optimization algorithms to achieve higher pre­
landslide occurrence (Dou et al., 2019; Pham et al., 2016). A key ad­ dictive accuracy in specific study areas. For example, Moayedi et al.
vantage of data-based approaches is their ability to deal with large (2019), Saha et al. (2022), and Nguyen et al. (2019) introduced the
datasets and apply them across different regions with varying geolo­ Particle Swarm Optimization (PSO) algorithm to optimize traditional
gical conditions. However, these models may be limited in capturing ANN models, resulting in higher predictive accuracy. Other frequently
complex non-linear interactions between variables, leading to overly used optimization algorithms include Genetic Algorithm (GA) ( Zhang
simplistic or less accurate predictions. et al., 2023; Hong et al., 2020), Simulated Annealing (SA) (Bajni et al.,
With the advent of Artificial Intelligence (AI), machine learning- 2022), Ant Colony Optimization (ACO) (Jaafari et al., 2022; Razavi-
based methods have emerged as the most efficient and widely adopted Termeh et al., 2021), and Artificial Fish Swarm (AFS) (Tay and Mo­
approaches in LSP. Typically, LSP is treated as a supervised learning hamad-Saleh, 2023).
task, where landslide and non-landslide sample labels serve as outputs, While machine learning has proven invaluable in geological hazard
and landslide-contributing factors as input variables. ML methods are research, shallow learning algorithms face challenges due to complexity
capable of learning a nonlinear function from the training dataset, arising from increased sample size and diversity (Wang et al.,
which is subsequently applied to predict landslide susceptibility (Adnan 2019a,b,c). To address this, researchers have implemented ensemble

8
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

learning or deep learning techniques. Ensemble learning combines mul­ used distance metrics include Euclidean distance, Manhattan distance,
tiple ML algorithms, often leading to superior accuracy compared to single and Cosine distance. For example, the Euclidean distance can be cal­
algorithms (e.g., Sharma et al., 2024; Li et al., 2023; Fang et al., 2021). culated using Eq. 2:
However, the effectiveness and applicability of ensemble models depend n
on input data and the specific ensemble method employed. d (x , x ) = (x i x i )2 (2)
i=1
Deep learning models are increasingly applied in LSP due to their
robust capability for nonlinear representation. For instance, Wang et al. where x and x represent two individual samples, with x i and x i denote
(2019a,2019b,2019c) demonstrated that Convolutional Neural Net­ the values of the samples in the ith feature and n represents the total
works (CNN) outperform optimized SVM models in predicting landslide number of features. After computing the distances, the K training
susceptibility. Similarly, Jiang et al. (2023) applied LR, SVM, LSTM, samples closest to the test sample are selected, and the test sample’s
and Linformer models to predict landslide susceptibility around railway class is determined by aggregating the labels of these K nearest
areas, finding that deep learning models offer superior accuracy. These neighbor samples in training data.
advancements highlight the potential of deep learning in advancing LSP
research. In this section, the mainstream AI-based methods for LSP and 3.2.3. Random Forest
their common frame will be reviewed in detail. Random Forest is an ensemble learning method based on decision
trees (Breiman, 2001). It improves model performance by combining
3.2. Mainstream AI-based methods for LSP multiple decision trees, where each tree is constructed independently.
The bootstrap sampling strategy is used to create different training
As mentioned in Section 3.1, a variety of AI-based models are datasets for each decision tree model, and the final prediction is made
available in the field of LSP, each based on different algorithmic prin­ by aggregating the predictions from all the decision trees (as shown in
ciples. Understanding these foundations is crucial for selecting the most Fig. 11).
appropriate method. Although many AI-based models have been ap­
plied to LSP, machine learning-based approaches remain dominant. 3.2.4. Support Vector Machine (SVM)
This section introduces the mainstream machine learning algorithms in The core of SVM is to map low-dimensional, non-separable data into
a straightforward manner. a higher-dimensional space and find a hyperplane that maximizes the
distance between different classes (Cortes and Vapnik, 1995). In the
3.2.1. Logistic Regression algorithm, support vectors are the data points that are close to the
Logistic Regression is a widely used statistical method for classifi­ hyperplane. In a linear SVM, the hyperplane can be represented as:
cation tasks (Scott et al., 1991). The core idea is to map input features wT x + b = 0 (3)
to probability values between 0 and 1, enabling the classification of
each sample. The model applies a logistic function to convert the output where w represents the weight vector, x is the input vector, and b is the
into a probability value, as shown in Eq.1: bias term. The goal is to find the optimal w and b that maximize the
distance between the classes while correctly classifying the data points
1
p (y = 1 X ) = (as shown in Fig. 12).
1+e ( 0 + 1 X1 + 2 X2 + + n Xn ) (1)
In this formula, p (y = 1|X ) represents the probability of the event y 3.2.5. Artificial Neural Network (ANN)
occurring given the input x conditions. 0 is the intercept term (bias), Artificial Neural Networks (ANNs) is a computational model in­
while 1, 2, …, n are the weight coefficients of the feature, and spired by the human brain's network of neurons. They are designed to
X1 , X2 , …, Xn represent the input features (i.e., landslide contributing recognize patterns through a structure of interconnected nodes, known
factors). The advantages of logistic regression include its simplicity, as neurons, which consist of an input layer, one or more hidden layers,
interpretability, and computational efficiency. However, it assumes a and an output layer, each containing multiple neurons (Zou et al.,
linear relationship between features, which may be overly simplistic 2008).
and struggle to capture complex nonlinear relationships between
landslide conditioning factors (LCFs). (1) Input Layer: This layer receives the initial data inputs.
(2) Hidden Layers: These layers process the inputs by applying weights
3.2.2. K-nearest neighbors (KNN) and activation functions to capture complex patterns and relation­
The KNN algorithm is an intuitive method for both classification and ships within the data.
regression tasks (Cover and Hart, 1967). Its basic principle involves (3) Output Layer: This layer provides the final output or prediction
predicting the label of a test sample based on the labels of the K nearest based on the processed information.
neighbor samples in the training data (as shown in Fig. 10).
In a KNN-based classification task, the first step is to calculate the Information in a neural network flows from the input layer, through
distance between the test sample and all training samples. Commonly the hidden layers, to the output layer. Each connection between

Fig. 10. The K-Nearest Neighbors (KNN) algorithm.

9
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Fig. 11. The Random Forest algorithm (Merghadi et al., 2020).

neurons has an associated weight that adjusts during the learning

process. The weights assigned between neurons and the activation
functions are used to introduce non-linearity, enabling the network to
model complex real-world relationships. Given an input vector
X = [x1, x2, …, xn], where n denotes the number of features. Thus, the
activation value aj of a neuron node j in a hidden layer can be calcu­
lated by Eq.4:

( )
n
aj = wij x i + bj (4)
i=1

In this equation, wij denotes the weight associated with the con­
nection between the input x i and neuron node j , and bj represent the
bias. is the activation function, with common choices being Sigmoid,
ReLU, and Tanh. In the output layer, the sigmoid function is generally
used to convert the network output into a landslide probability (as
shown in Fig. 13).

3.3. Framework of AI-based modeling for LSP

Fig. 13. The Artificial neural network algorithm.
Based on previous studies (Wang et al., 2021a; Napoli et al., 2020;
Huang and Zhao, 2018; Marjanović et al., 2011; Ermini et al., 2005),
the AI-based landslide susceptibility modeling process typically in­
volves the following steps, i.e., 1) Pre-analysis of landslide formation

Fig. 12. The Support vector machine algorithm.

10
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

mechanisms. 2) Collection of landslide inventory samples and landslide 3.3.5. Selection of the appropriate LSP model
conditioning factors (LCFs). 3) Determination of prediction units and The selection of an LSP model is a critical step, as it directly impacts
spatial resolution. 4) Dataset construction for the LSP model. 5) Se­ the accuracy of the results. Choosing a suitable LSP modeling method,
lection of the appropriate LSP model. 6) LSP model training and ac­ or combining it with optimization and ensemble strategies, can sig­
curacy evaluation, and 7) LSP mapping application. Details of the nificantly enhance model performance. The commonly used ML-based
procedure are summarized as follows. LSP methods are described in Section 3.2, each grounded in different
theoretical principles.

3.3.1. Pre-analysis of landslide formation mechanisms

This step involves understanding the potential formative and trig­ 3.3.6. Accuracy evaluation of training and validation
gering conditions of landslides in the research area, including geolo­ In this phase, the model is trained using the training dataset, al­
gical, hydrological, and environmental factors that contribute to land­ lowing it to learn patterns and relationships within the data. Once the
slide occurrences (Huang et al., 2024). This analysis is essential for model has been well-trained, its performance can be evaluated using
selecting appropriate LCFs in the modeling process. the validation dataset, employing various metrics such as Accuracy,
Precision, Recall, and F1 Score. These metrics provide a comprehensive
understanding of the model's effectiveness. Additional metrics such as
3.3.2. Collection of landslide inventory samples and LCFs Specificity, Kappa, ROC, AUC, and PR are also used to further assess
This step involves gathering information on the historical landslide performance (Rabby et al., 2023; Jiang et al., 2023; Othman et al.,
events (landslide inventory), including their location, size, type, and 2018). Moreover, analyzing the impact of various factors on the model's
other characteristics, as well as identifying and compiling the con­ predictions is crucial for determining which factors are most likely to
ditioning factors that influence landslide susceptibility, such as slope, contribute to landslide occurrence.
aspect, soil type, and land cover, in GIS software (Jiang et al., 2024;
Wang et al., 2021b; Dou et al., 2020). As for the selection of landslide
susceptibility factors, a detailed investigation can be referred to the 3.3.7. LSP mapping application
works of (Huang et al., 2024; Reichenbach et al., 2018). The well-trained model should be applied to landslide susceptibility
mapping in the study area, enabling the visualization of the spatial dis­
tribution of landslide susceptibility predictions. Such visualization not only
3.3.3. Determination of prediction units and spatial resolution aids in understanding the spatial patterns of landslide susceptibility but
This step involves defining the prediction units for the landslide also supports decision-making processes in disaster management, land-use
susceptibility model and selecting the appropriate spatial resolution for planning, and risk mitigation efforts (Dikshit et al., 2021; Guzzetti et al.,
LSP mapping. The most used units in LSP include grid units and slope 2012). For example, Li et al. (2023) developed a prediction of the vul­
units (Li et al., 2024). Grid units offer a simple and efficient compu­ nerability of Wenchuan County. The results show that GL-ResNet has
tational process but may not accurately capture the topographic fea­ better accuracy (0.909) compared to machine learning models and CNN
tures of slopes. In contrast, slope units are theoretically more suitable models (Fig. 15(a)). Hong et al. (2020) applied the RF method to LSP in
for LSP due to their ability to capture the interaction between landslides Yanshan County, which achieved an accuracy of 0.8626 (Fig. 15(b)). Jiang
and topographic features (Jacobs et al., 2020; Alvioli et al., 2020; et al. (2023) conducted a study on the susceptibility of railway safety
Huabin et al., 2005). Despite extensive research in this field (Li et al., (Fig. 15(c)), and the results show that the Linformer model has better
2024; Yang et al., 2022; Huang et al., 2021; Alvioli et al., 2020; Turel prediction performance, with an accuracy of 0.8576. By clearly displaying
and Frost, 2011), there is still debate over which strategy yields su­ areas of very high, high, medium, low, and very low susceptibility, these
perior slope units for LSP. maps serve as valuable tools for communicating risks (Reichenbach et al.,
2018; Van Westen et al., 2003).
3.3.4. Dataset construction for LSP model
Dataset construction typically involves defining input and output 4. Discussion on the challenges and future directions
variables and dividing the data into training and testing sets. The input
variables should include some or all of the LCF data for LSP model In recent years, the research methodologies for both landslide de­
training, while the output variable indicates whether landslides occur tection and susceptibility prediction have progressively shifted from
under the given LCF conditions. Input variables should be preprocessed, traditional approaches to a new paradigm, deep learning-based tech­
such as through normalization or standardization, to mitigate scale niques. This transition is largely driven by the superior capability of
effects among LCFs. Additionally, multicollinearity between LCFs deep learning in recognizing complex data patterns. However, these
should be checked to avoid redundant factors (Jiang et al., 2023; fields still face numerous challenges, particularly concerning data lim­
Dormann et al., 2013). The selection of negative samples and the setting itations, the regional specificity of study areas, and the generalization
of the positive-to-negative sample ratio are also critical considerations. ability of methods. The following sections will discuss the primary
Although numerous studies have explored these issues, consensus re­ challenges currently encountered in landslide detection and suscept­
mains elusive due to the non-replicability of study areas. ibility prediction and explore potential future research directions.

Fig. 14. Representation of the different mapping units (a) Standard mesh grid unit (b) Object-oriented slope unit (c) SLIC-based slope unit (Li et al., 2024).

11
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Fig. 15. Landslide susceptibility prediction results in different study areas (a) GL-ResNet model in Wenchuan County (Li et al., 2023) (b) RF model in Yanshan
County (Hong et al., 2020) (c) Linformer model in railway area (Jiang et al., 2023).

4.1. Landslide detection aspect 4.1.1. Data quality and standardization

The effectiveness of landslide detection relies heavily on the avail­
The development of landslide detection has undergone several key ability of high-quality datasets. However, there is still no consensus on
phases, including the field survey phase, visual interpretation phase, the criteria for dataset creation in the field of landslide detection, which
threshold segmentation phase, machine learning phase, and deep makes it difficult to directly compare the results under different studies.
learning phase. Up to date, the mainstream landslide detection methods Although some open-source datasets are available (Gómez et al., 2023;
have become relatively standardized, with the integration of advanced Ghorbanzadeh et al., 2022; Ji et al., 2020), each landslide dataset may
AI technologies. However, significant opportunities remain for further cover different regions, time periods, and topographic features. Such
research, particularly in enhancing data quality and standardization, diversity can definitely provide rich material for research but also in­
addressing the regional specificity of study areas, and improving the creases the difficulty of model generalization. The collection of land­
robustness and practicality of these models in real-world applications. slide datasets usually depends on multiple sources of data, such as

12
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

optical remote sensing imagery, topographic data, and InSAR data, research should focus on enhancing the adaptability of detection
while the quality and accuracy of each data source may potentially models to various regional conditions.
affect the quality of the final dataset (Joshi et al., 2024; Fu et al., 2023;
Sameen et al., 2019). Furthermore, the accuracy of data annotation
remains a critical factor that can significantly impact the model's 4.1.4. Real application and robustness analysis
learning ability. For example, different research teams define and ca­ Generally, landslide detection models are often observed to perform
tegorize landslide events in different ways, and some datasets may better in experimental environments than in real-world applications. In
focus more on large, destructive landslides, while other datasets may the model training stage, high-quality image datasets can be used,
contain more small, hard-to-detect landslide events. Such differences in which are usually rigorously selected and processed with high clarity.
annotation make the model perform ineffectively in specific regions or In this way, the model can learn and recognize the features of landslides
on specific landslide types. more easily and achieve a high accuracy rate. However, in real appli­
Moreover, the severe imbalance between positive and negative sam­ cation scenarios, the image variability faced is far more complex than in
ples often results in a bias toward model prediction. In which, the positive the training dataset. Such performance degradation is primarily due to
samples usually refer to pixel areas where landslides occurred, while the differences in imagery, such as lighting conditions, imaging angles, and
rest are denoted as negative samples. Actually, the area of landslide oc­ surface cover (Asadi et al., 2024; Fu et al., 2023; Li et al., 2023). On the
currence is often only a small part of the total area, so the model is easily one hand, differences in lighting conditions and imaging angles may
over-influenced by the negative samples during training, resulting in a cause blurring of terrain features, texture, and other information, thus
tendency to predict that no landslide has occurred, thus ignoring the real affecting the model's performance of landslide detection. On the other
landslide risk. Such a challenge is also known as “the problem of category hand, the diversity of surface cover also poses a great challenge to
imbalance”. Further investigation and improvement in this area are ne­ landslide detection modeling. In the model training stage, images are
cessary to enhance detection accuracy. often selected for scenes with relatively single or clear surface cover,
such as bare rocks, clear slopes, and so on. However, in real applica­
4.1.2. Improvement of landslide detection efficiency tions, the surface cover in landslide areas can be significantly complex,
As commented in Section 2.1, the AI-based methods are advanced in such as vegetation, snow, sediment, buildings, and other different ele­
pixel-wise semantic segmentation of landslides in the remote sensing ments intertwined together. These complex surface coverings can in­
images, providing a better solution for tasks of landslide inventory map­ terfere with the visual characteristics of landslides, making it difficult
ping as the polygons of each landslide can be delineated. However, this for the model to distinguish the landslides from other terrains, thus
kind of approach has some disadvantages. First, AI-based approaches re­ reducing the accuracy of detection. Therefore, future research should
quire pixel-wise classification, resulting in higher computational resource focus on improving the robustness of these AI models in real applica­
requirements and training costs (Milan et al., 2018; Yu et al., 2018), which tions with complex environment conditions, by developing methods
should be considered when training the models. For example, the YOLO that can handle these variations effectively.
neural network is capable of processing up to 45 frames per second (FPS),
which may be several times greater than UNets and FCNs. Second, the 4.2. Landslide susceptibility prediction aspect
performances of the deep learning models greatly depend on the quality of
training datasets, but the difficulty of labeling training data makes it a Among the various LSP methods, AI-based susceptibility modeling
challenging, exhaustive, and time-consuming task. Nagendra et al. (2022) processes have become relatively standardized. However, considerable
mentioned that experienced annotators commonly use several minutes to research opportunities still exist, particularly in improving data time­
locate a landslide in an image, and several more minutes to map the liness, addressing model overfitting, and enhancing the reliability of
landslide polygon. Currently, such a dataset with the necessary amount of these models in real-world scenarios.
semantic segmentation labels is not as widely available, because it is more
convenient to make bounding box labels for annotators. Integration of
computer vision and AI methods could inspire an alternative solution to 4.2.1. Data timeliness
address this issue, such as in Han et al. (2024), a Dynahead-YOLO-Ostu The credibility of model predictions heavily depends on the quality
model has been proposed, which achieves pixel-wise semantic segmenta­ of the dataset and the relevance of the triggering factors considered
tion of landslide area with 7 times greater processing speed compared to (Jiang et al., 2023). In terms of data quality, the representativeness,
traditional UNet model. However, substantial improvements, in particular completeness, and precision of the landslide data directly impact the
in the structure of AI models should be focused on in future studies. performance of the model. As for data representativeness, if the dataset
of a landslide prediction model fails to cover various types of landslide
4.1.3. Regional specificity of study areas geologic environments, it may cause a lack of universality and accuracy
Currently, most landslide detection studies are focused on co- in prediction results. A complete dataset can also provide more com­
seismic or region-specific landslides, with limited exploration of the prehensive background information, which helps the model to better
models' applicability across diverse geographical and environmental understand the complex relationships between landslide and back­
conditions. Actually, the occurrence of landslides is not limited to ground. In addition, the precision of the data, i.e., whether there are
seismic activity triggering but involves a variety of complex factors, errors or noise in the data, can also lead to biased prediction results.
such as rainfall, geological conditions, climate change, and human ac­ Furthermore, the timeliness of historical data is a crucial factor af­
tivities (Jiang et al., 2023). Such focus on specific types of landslides fecting the accuracy of model predictions (Ren et al., 2024; Dashbold
may lead to significant limitations in the adaptability of models to et al., 2023; Lee et al., 2022). Outdated historical contributing factors
varying geographical and environmental conditions. Furthermore, the and landslide events may no longer reflect current geological and en­
optimal data sources for detecting different types of landslides can vironmental conditions. Moreover, changes in the geological environ­
differ significantly. For instance, in densely vegetated areas, existing ment, such as land use practices and vegetation cover, may cause past
techniques such as optical imagery may struggle to achieve effective landslide events to exhibit different characteristics under current con­
detection. While Interferometric Synthetic Aperture Radar (InSAR) ditions. These issues can introduce biases and inaccuracies when using
technology has shown promise in monitoring landslide deformation historical data to predict current landslide susceptibility. To address
(Joshi et al, 2024; Osmanoğlu et al., 2016; Schlögel et al., 2015), in­ this, advanced remote sensing and landslide detection techniques
tegrating deep learning techniques to improve its application in dif­ should be employed to refine and update existing landslide samples and
ferent regions and complex terrains remains challenging. Future LCFs.

13
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

4.2.2. “Black box” problem in explaining LSP major source that triggers uncertainty in modeling landslide suscept­
Overall, as we reviewed in Sections 3.1 and 3.2, the AI-based ibility. Currently, the methods used for landslide susceptibility mod­
methods have shown their robust capability and satisfactory accuracy eling are diverse, ranging from traditional knowledge-based methods to
in predicting landslide susceptibility. However, the decision process of machine learning methods (e.g., support vector machines, random
the trained AI model is still blind to the users, which is commonly re­ forests, deep learning, etc.). Meanwhile, different models might per­
garded as a “Black box” problem that has been reported in many recent form differently in response to different datasets, and some of them may
studies. This problem is due to the structure of AI models themselves, perform well in some regions while in others there may be large de­
that the model consists of numerous hidden layers, each containing viations. In addition, the sensitivity of different models to parameters
multiple neurons, making it difficult to directly explain their predic­ also varies, as for the same set of data, different parameter settings may
tions. This lack of transparency poses a significant challenge in inter­ lead to very different predictions.
preting features' importance, hindering the balance between accuracy Such uncertainty in model selection can also be reflected in the
and interpretability and limiting the further development of these differences in landslide susceptibility mechanisms. For example, in
models in the field of geological hazards (Alqadhi et al., 2024). While tropical regions with high rainfall, rainfall may be the main trigger for
high accuracy in LSP is crucial for the success of these models, under­ landslides, while in dry mountainous environments, seismic activity
standing the decision-making process is essential to reveal their char­ may become a more dominant factor that triggers landslides. Such di­
acteristics from a mechanistic level. Currently, some studies have versity of environments requires that model selection should not only
highlighted this issue, and suggested the usage of the XAI (explainable consider the performance in a particular region but also focus on its
AI) method for this purpose, such as SHapley Additive exPlanations ability to generalize across different regions.
(SHAP), Partial Dependence Plots (PDP), Local Interpretable Model- Although susceptibility maps can serve as valuable tools for com­
Agnostic Explanations (LIME) in AI field (Nagy and Molontay, 2024; municating risk, a more thorough review and reliability analysis are
Wang et al., 2021; Zhao and Hastie, 2021). However, the research on necessary before LSP mapping can be officially used. To enhance the
this topic is currently limited and requires more attention in the on­ reliability of these models, post-hoc reliability evaluation methods
going work. should be introduced.
For example, a possible post-evaluation approach would be to va­
4.2.3. Model overfitting issue lidate the model in several different geographic regions and assess the
Due to the quantitative limitations and regional specificity of his­ stability and cross-regional applicability of the model, by comparing the
torical landslide data, AI models are prone to overfitting during the predictions with actual landslide occurrences in different regions. Such
training process (Lv et al., 2022). Overfitting not only hampers the methods can improve our understanding of the model's performance
model's predictive effectiveness on new data but can also lead to and help to establish its reliability in real applications.
overestimation or underestimation of landslide susceptibility. Up to
now, several techniques have been proposed to address this issue, such
as cross-validation, regularisation techniques, and data augmentation. 5. Summary
For example, cross-validation methods (such as k -fold cross-validation)
have been widely used to assess the generalization ability of models. By In this review, we summarized the advancements in artificial in­
dividing the dataset into k mutually exclusive subsets, the model can telligent-based technologies for landslide detection and susceptibility
use one subset for validation and the remaining subsets for training in assessment. We highlight the main AI-based detection network struc­
each iteration. This process is repeated k times to ensure stable per­ tures and landslide susceptibility prediction (LSP) modeling methods,
formance across different data splits, which helps reduce the risk of outline general modeling frameworks, and discuss the current chal­
overfitting (Sun et al., 2020). Additionally, regularization techniques lenges and future directions in these fields. The review of the up-to-date
(such as L1 regularization and L2 regularization) add penalty terms to studies and application of AI-based methods indicates that recent re­
the loss function to constrain the magnitude of model parameters, search shows a trend toward using deep learning techniques for both
preventing the model from becoming overly complex (Zhu et al., 2021). landslide detection and susceptibility prediction. While there have been
Data augmentation is also an important strategy to mitigate overfitting. significant advancements in applying AI methods to these areas, several
On the one hand, by calibrating unlabelled data with pseudo-labels, challenges remain that need to be addressed to improve the predictive
more training samples can be generated to increase data diversity. On capabilities and practical value of these models.
the other hand, sampling techniques, such as the Synthetic Minority The performance of the current AI-based methods can be improved
Oversampling Technique (SMOTE), can be used to balance the dataset, by extracting landslide-related information from multi-source input
thus reducing the impact of class imbalance on model training (Lu data. Additionally, incorporating well-designed feature extraction and
et al., 2024). Additionally, the spatially clustered distribution of sam­ attention modules can significantly enhance model performance.
ples further affects the model's generalization ability (Li et al., 2022). However, deep learning methods face challenges such as the limited
To overcome this challenge, distributed learning techniques, such as availability of labeled landslide data and high computational costs for
Federated Learning, which allow for model training using data from model training. Future research will likely focus on overcoming these
multiple regions, should be further explored. The combined application limitations by improving labeled datasets, optimizing computational
of these techniques helps improve the reliability and accuracy of efficiency, and exploring hybrid deep learning models with better
landslide susceptibility prediction, ensuring the robustness of the model generalization capabilities.
across different regions and datasets. Classical machine learning methods, such as Logistic Regression, K-
Nearest Neighbors (KNN), Random Forest, Support Vector Machines
4.2.4. Real application and reliability analysis (SVM), and Artificial Neural Networks (ANN), have become well-es­
The process of landslide susceptibility modeling is fraught with tablished in LSP. The integration of optimization techniques and en­
uncertainties, such as those related to sample selection and model semble strategies has significantly enhanced model performance.
choice (Xing et al., 2023; Liu et al., 2023; Huang et al., 2022). In fact, Building on this foundation, deep learning methods are poised to ad­
both the distribution, size of the sample, and biases in sample selection vance LSP research even further. However, the "black-box" nature of
can affect the performance of the model. For instance, if landslide these models often limits our understanding of their decision-making
samples are concentrated in only some specific geological conditions or processes. The emergence of Explainable AI (XAI) offers a promising
regions, it will result in a degradation of the model's performance when solution to this issue, as it aims to make these complex models more
applied to other regions. On the other hand, model selection is another interpretable and improve our understanding of LSP results.

14
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Declaration of Competing Interest Comert, R., Avdan, U., Gorum, T., 2018. Rapid mapping of forested landslide from ultra-
high resolution unmanned aerial vehicle data. Int. Arch. Photogramm. Remote Sens.
Spatial Inf. Sci. 42, 171–176. https://doi.org/10.5194/isprs-archives-XLII-3-W4-171-
The authors declare that they have no competing interests. 2018
Cortes, C., Vapnik, V., 1995. Support-vector networks. Mach. Learn. 20. https://doi.org/
Acknowledgments 10.1023/A:1022627411411
Cover, T.M., Hart, P.E., 1967. Nearest neighbor pattern classification. IEEE Trans. Inf.
Theory 13 (1), 21–27. https://doi.org/10.1109/TIT.1967.1053964
This study was financially supported by the National Key Research Dashbold, B., Bryson, L.S., Crawford, M.M., 2023. Landslide hazard and susceptibility
maps derived from satellite and remote sensing data using limit equilibrium analysis
and Development Program of China [Grant No. 2018YFD1100401], the
and machine learning model. Nat. Hazards 116 (1), 235–265. https://doi.org/10.
National Natural Science Foundation of China [Grant No. 52078493], 1007/s11069-022-05671-7
the Natural Science Foundation of Hunan Province [Grant No. Dente, E., Katz, O., Crouvi, O., Mushkin, A., 2023. The geomorphic effectiveness of
landslides. J. Geophys. Res. Earth Surf. 128 (12), e2023JF007191. https://doi.org/
2022JJ30700], the Science and Technology Plan Project of Changsha 10.1029/2023JF007191
[Grant No. kq2305006], and the Innovation Driven Program of Central Devara, M., Maurya, V.K., Dwivedi, R., 2024. Landslide extraction using a novel empirical
South University [Grant No. 2023CXQD033]. These financial supports method and binary semantic segmentation U-NET framework using sentinel-2 ima­
gery. Remote Sens. Lett. 15 (3), 326–338. https://doi.org/10.1080/2150704X.2024.
are gratefully acknowledged. 2320178
Di Napoli, M., Carotenuto, F., Cevasco, A., Confuorto, P., Di Martire, D., Firpo, M., Pepe,
G., Raso, E., Calcaterra, D., 2020. Machine learning ensemble modelling as a tool to
References improve landslide susceptibility mapping reliability. Landslides 17 (8), 1897–1914.
https://doi.org/10.1007/s10346-020-01392-9
Adnan, M.S.G., Rahman, M.S., Ahmed, N., Ahmed, B., Rabbi, M.F., Rahman, R.M., 2020. Dikshit, A., Pradhan, B., Alamri, A.M., 2021. Pathways and challenges of the application
Improving spatial agreement in machine learning-based landslide susceptibility of artificial intelligence to geohazards modelling. Gondwana Res 100, 290–301.
mapping. Remote Sens. 12 (20), 3347. https://doi.org/10.3390/rs12203347 https://doi.org/10.1016/j.gr.2020.08.007
Aleotti, P., 2004. A warning system for rainfall-induced shallow failures. Eng. Geol. 73 (3- Ding, Y., Zhang, Z., Zhao, X., Hong, D., Cai, Wei, Yu, C., Yang, N., Cai, Weiwei, 2022.
4), 247–265. https://doi.org/10.1016/j.enggeo.2004.01.007 Multi-feature fusion: Graph neural network and CNN combining for hyperspectral
Alqadhi, S., Mallick, J., Alkahtani, M., Ahmad, I., Alqahtani, D., Hang, H.T., 2024. image classification. Neurocomputing 501, 246–257. https://doi.org/10.1016/j.
Developing a hybrid deep learning model with explainable artificial intelligence neucom.2022.06.031
(XAI) for enhanced landslide susceptibility modeling and management. Nat. Hazards Dong, Z., An, S., Zhang, J., Yu, J., Li, J., Xu, D., 2022. L-unet: A landslide extraction model
120 (4), 3719–3747. https://doi.org/10.1007/s11069-023-06357-4 using multi-scale feature fusion and attention mechanism. Remote Sens 14 (11),
Alvioli, M., Guzzetti, F., Marchesini, I., 2020. Parameter-free delineation of slope units 2552. https://doi.org/10.3390/rs14112552
and terrain subdivision of Italy. Geomorphology 358, 107124. https://doi.org/10. Donnini, M., Santangelo, M., Gariano, S.L., Bucci, F., Peruccacci, S., Alvioli, M.,
1016/j.geomorph.2020.107124 Althuwaynee, O., Ardizzone, F., Bianchi, C., Bornaetxea, T., Brunetti, M.T., Cardinali,
Asadi, A., Baise, L.G., Chatterjee, S., Koch, M., Moaveni, B., 2024. Regional landslide M., Esposito, G., Grita, S., Marchesini, I., Melillo, M., Salvati, P., Yazdani, M.,
mapping model developed by a deep transfer learning framework using post-event Fiorucci, F., 2023. Landslides triggered by an extraordinary rainfall event in Central
optical imagery. Georisk 18 (1), 186–210. https://doi.org/10.1080/17499518.2024. Italy on September 15, 2022. Landslides 20 (10), 2199–2211. https://doi.org/10.
2316265 1007/s10346-023-02109-4
Badrinarayanan, V., Kendall, A., Cipolla, R., 2017. SegNet: a deep convolutional encoder- Dormann, C.F., Elith, J., Bacher, S., Buchmann, C., Carl, G., Carré, G., Marquéz, J.R.G.,
decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell. Gruber, B., Lafourcade, B., Leitão, P.J., Münkemüller, T., Mcclean, C., Osborne, P.E.,
39 (12), 2481–2495. https://doi.org/10.1109/TPAMI.2016.2644615 Reineking, B., Schröder, B., Skidmore, A.K., Zurell, D., Lautenbach, S., 2013.
Bajni, G., Camera, C.A.S., Brenning, A., Apuani, T., 2022. Assessing the utility of re­ Collinearity: a review of methods to deal with it and a simulation study evaluating
gionalized rock-mass geomechanical properties in rockfall susceptibility modelling in their performance. Ecography 36 (1), 27–46. https://doi.org/10.1111/j.1600-0587.
an alpine environment. Geomorphology 415, 108401. https://doi.org/10.1016/j. 2012.07348.x
geomorph.2022.108401 Dou, J., Yamagishi, H., Pourghasemi, H.R., Yunus, A.P., Song, X., Xu, Y., Zhu, Z., 2015. An
Bragagnolo, L., Rezende, L.R., da Silva, R.V., Grzybowski, J.M.V., 2021. Convolutional integrated artificial neural network model for the landslide susceptibility assessment
neural networks applied to semantic segmentation of landslide scars. Catena 201, of Osado Island, Japan. Nat. Hazards 78, 1749–1776. https://doi.org/10.1007/
105189. https://doi.org/10.1016/j.catena.2021.105189 s11069-015-1799-2
Breiman, L., 2001. Random forests. Mach. Learn. 45, 5–32. https://doi.org/10.1023/ Dou, J., Yunus, A.P., Bui, D.T., Sahana, M., Chen, C.W., Zhu, Z., Wang, W., Pham, B.T.,
A:1010933404324 2019. Evaluating gis-based multiple statistical models and data mining for earth­
Casagli, N., Intrieri, E., Tofani, V., Gigli, G., Raspini, F., 2023. Landslide detection, quake and rainfall-induced landslide susceptibility using the lidar dem. Remote Sens
monitoring and prediction with remote-sensing techniques. Nat. Rev. Earth Environ. 11 (6), 638. https://doi.org/10.3390/rs11060638
4 (1), 51–64. https://doi.org/10.1038/s43017-022-00373-x Dou, J., Yunus, A.P., Merghadi, A., Shirzadi, A., Nguyen, H., Hussain, Y., Avtar, R., Chen,
Chen, Y., Irfan, M., Uchimura, T., Meng, Q., Dou, J., 2020. Relationship between water content, Y., Pham, B.T., Yamagishi, H., 2020. Different sampling strategies for predicting
shear deformation, and elastic wave velocity through unsaturated soil slope. Bull. Eng. landslide susceptibilities are deemed less consequential with deep learning. Sci. Total
Geol. Environ. 79, 4107–4121. https://doi.org/10.1007/s10064-020-01841-8 Environ. 720, 137320. https://doi.org/10.1016/j.scitotenv.2020.137320
Chen, T., Niu, R., Jia, X., 2016. A comparison of information value and logistic regression Ermini, L., Catani, F., Casagli, N., 2005. Artificial Neural Networks applied to landslide
models in landslide susceptibility mapping by using GIS. Environ. Earth Sci. 75, 1–16. susceptibility assessment. Geomorphology 66 (1-4), 327–343. https://doi.org/10.
https://doi.org/10.1007/s12665-016-5317-y 1016/j.geomorph.2004.09.025
Chen, T., Niu, R., Du, B., Wang, Y., 2015. Landslide spatial susceptibility mapping by Fang, C., Fan, X., Zhong, H., Lombardo, L., Tanyas, H., Wang, X., 2022. A novel historical
using GIS and remote sensing techniques: a case study in Zigui County, the Three landslide detection approach based on LiDAR and lightweight attention UNet.
Georges reservoir, China. Environ. Earth Sci. 73, 5571–5583. https://doi.org/10. Remote Sens 14 (17), 4357. https://doi.org/10.3390/rs14174357
1007/s12665-014-3811-7 Fang, Z., Wang, Y., Peng, L., Hong, H., 2021. A comparative study of heterogeneous
Chen, T., Trinder, J.C., Niu, R., 2017. Object-oriented landslide mapping using ZY-3 sa­ ensemble-learning techniques for landslide susceptibility mapping. Int. J. Geogr. Inf.
tellite imagery, random forest and mathematical morphology, for the Three-Gorges Sci. 35 (2), 321–347. https://doi.org/10.1080/13658816.2020.1808897
Reservoir, China. Remote Sens 9 (4), 333. https://doi.org/10.3390/rs9040333 Farabet, C., Couprie, C., Najman, L., LeCun, Y., 2012. Scene parsing with multiscale
Chen, W., Xie, X., Wang, J., Pradhan, B., Hong, H., Bui, D.T., Duan, Z., Ma, J., 2017. A feature learning, purity trees, and optimal covers. arXiv preprint arXiv:1202.2160.
comparative study of logistic model tree, random forest, and classification and re­ Felicísimo, Á.M., Cuartero, A., Remondo, J., Quirós, E., 2013. Mapping landslide sus­
gression tree models for spatial prediction of landslide susceptibility. Catena 151, ceptibility with logistic regression, multiple adaptive regression splines, classification
147–160. https://doi.org/10.1016/j.catena.2016.11.032 and regression trees, and maximum entropy methods: a comparative study.
Chen, Z., Zhang, Y., Ouyang, C., Zhang, F., Ma, J., 2018. Automated landslides detection Landslides 10, 175–189. https://doi.org/10.1007/s10346-012-0320-1
for mountain cities using multi-temporal remote sensing imagery. Sensors 18 (3), Fu, B., Li, Y., Han, Z., Fang, Z., Chen, N., Hu, G., Wang, W., 2023. RIPF-Unet for regional
821. https://doi.org/10.3390/s18030821 landslides detection: a novel deep learning model boosted by reversed image pyramid
Chen, X., Zhao, C., Xi, J., Lu, Z., Ji, S., Chen, L., 2022. Deep learning method of landslide features. Nat. Hazards 119 (1), 701–719. https://doi.org/10.1007/s11069-023-
inventory map with imbalanced samples in optical remote sensing. Remote Sens 14 06145-0
(21), 5517. https://doi.org/10.3390/rs14215517 Gabrieli, F., Corain, L., Vettore, L., 2016. A low-cost landslide displacement activity as­
Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H., 2018. Encoder-decoder with sessment from time-lapse photogrammetry and rainfall data: Application to the
atrous separable convolution for semantic image segmentation. Proc. Eur. Conf. Tessina landslide site. Geomorphology 269, 56–74. https://doi.org/10.1016/j.
Comput. Vis. 801–818. https://doi.org/10.1007/978-3-030-01234-2_49 geomorph.2016.06.030
Cheng, L., Li, J., Duan, P., Wang, M., 2021. A small attentional YOLO model for landslide Galli, M., Ardizzone, F., Cardinali, M., Guzzetti, F., Reichenbach, P., 2008. Comparing
detection from satellite remote sensing images. Landslides 18 (8), 2751–2765. landslide inventory maps. Geomorphology 94 (3-4), 268–289. https://doi.org/10.
https://doi.org/10.1007/s10346-021-01694-6 1016/j.geomorph.2006.09.023
Chollet, F., 2017. Xception: Deep learning with depthwise separable convolutions. IEEE Ganerød, A.J., Lindsay, E., Fredin, O., Myrvoll, T.A., Nordal, S., Rød, J.K., 2023. Globally
Conf. Comput. Vis. Pattern Recognit. 1251–1258. https://doi.org/10.1109/CVPR. vs. locally trained machine learning models for landslide detection: a case study of a
2017.195 glacial landscape. Remote Sens 15 (4), 895. https://doi.org/10.3390/rs15040895
Chou, C.H., Lin, W.H., Chang, F., 2010. A binarization method with learning-built rules Gao, X., Chen, T., Niu, R., Plaza, A., 2021. Recognition and mapping of landslide using a
for document images produced by cameras. Pattern Recognit 43 (4), 1518–1530. fully convolutional densenet and influencing factors. IEEE J. Sel. Top. Appl. Earth
https://doi.org/10.1016/j.patcog.2009.10.016 Observ. Remote Sens. 14, 7881–7894. https://doi.org/10.1109/JSTARS.2021.
Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, K., Kuksa, P., 2011. Natural 3101203
language processing (almost) from scratch. J. Mach. Learn. Res 12, 2493–2537

15
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Gao, O., Niu, C., Liu, W., Li, T., Zhang, H., Hu, Q., 2022. E-DeepLabV3+: a landslide Im, J., Lu, Z., Jensen, J.R., 2011. A genetic algorithm approach to moving threshold
detection method for remote sensing images. IEEE Jt. Int. Inf. Technol. Artif. Intell. optimization for binary change detection. Photogramm. Eng. Remote Sens. 77 (2),
Conf. 10, 573–577. https://doi.org/10.1109/ITAIC54216.2022.9836758 167–180. https://doi.org/10.14358/PERS.77.2.167
Ghorbanzadeh, O., Blaschke, T., Gholamnia, K., Meena, S.R., Tiede, D., Aryal, J., 2019. Ishikawa, T., Yoshimi, M., Isobe, K., Yokohama, S., 2021. Reconnaissance report on
Evaluation of different machine learning methods and deep-learning convolutional geotechnical damage caused by 2018 Hokkaido Eastern Iburi earthquake with JMA
neural networks for landslide detection. Remote Sens 11 (2), 196. https://doi.org/10. seismic intensity 7. Soils Found 61 (4), 1151–1171. https://doi.org/10.1016/j.sandf.
3390/rs11020196 2021.06.006
Ghorbanzadeh, O., Crivellari, A., Ghamisi, P., Shahabi, H., Blaschke, T., 2021. A com­ Jaafari, A., Panahi, M., Mafi-Gholami, D., Rahmati, O., Shahabi, H., Shirzadi, A., Lee, S.,
prehensive transferability evaluation of U-Net and ResU-Net for landslide detection Bui, D.T., Pradhan, B., 2022. Swarm intelligence optimization of the group method of
from Sentinel-2 data (case study areas from Taiwan, China, and Japan). Sci. Rep. 11 data handling using the cuckoo search and whale optimization algorithms to model
(1), 14629. https://doi.org/10.1038/s41598-021-94190-9 and predict landslides. Appl. Soft Comput. 116, 108254. https://doi.org/10.1016/j.
Ghorbanzadeh, O., Gholamnia, K., Ghamisi, P., 2023. The application of ResU-net and asoc.2021.108254
OBIA for landslide detection from multi-temporal Sentinel-2 images. Big Earth Data 7 Jacobs, L., Kervyn, M., Reichenbach, P., Rossi, M., Marchesini, I., Alvioli, M., Dewitte, O.,
(4), 961–985. https://doi.org/10.1080/20964471.2022.2031544 2020. Regional susceptibility assessments with heterogeneous landslide information:
Ghorbanzadeh, O., Xu, Y., Ghamisi, P., Kopp, M., Kreil, D., 2022. Landslide4Sense: Slope unit- vs. pixel-based approach. Geomorphology 356, 107084. https://doi.org/
Reference Benchmark Data and Deep Learning Models for Landslide Detection. IEEE 10.1016/j.geomorph.2020.107084
Trans. Geosci. Remote Sens. 60, 1–17. https://doi.org/10.1109/TGRS.2022.3215209 Ji, S., Yu, D., Shen, C., Li, W., Xu, Q., 2020. Landslide detection from an open satellite
Goetz, J.N., Brenning, A., Petschko, H., Leopold, P., 2015. Evaluating machine learning imagery and digital elevation model dataset using attention boosted convolutional
and statistical prediction techniques for landslide susceptibility modeling. Comput. neural networks. Landslides 17, 1337–1352. https://doi.org/10.1007/s10346-020-
Geosci. 81, 1–11. https://doi.org/10.1016/j.cageo.2015.04.007 01353-2
Gómez, D., García, E.F., Aristizábal, E., 2023. Spatial and temporal landslide distributions Jiang, N., Li, Y., Han, Z., Li, J., Fu, B., Yang, J., 2023. A dataset-enhanced Linformer
using global and open landslide databases. Nat. Hazards 117 (1), 25–55. https://doi. model for geo-hazards susceptibility assessment: a case study of the railway in
org/10.1007/s11069-023-05848-8 Southwest China. Environ. Earth Sci. 82 (17), 397. https://doi.org/10.1007/s12665-
Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G., 2020. A survey of deep learning 023-11080-1
techniques for autonomous driving. J. Field Robot. 37 (3), 362–386. https://doi.org/ Jiang, X., Mojon, D., 2003. Adaptive local thresholding by verification-based multithreshold
10.1002/rob.21918 probing with application to vessel detection in retinal images. IEEE Trans. Pattern Anal.
Guzzetti, F., Cardinali, M., Reichenbach, P., Carrara, A., 2000. Comparing landslide maps: Mach. Intell. 25 (1), 131–137. https://doi.org/10.1109/TPAMI.2003.1159954
a case study in the upper Tiber River basin, central Italy. Environ. Manage. 25 (3). Jiang, C., Ren, H., Ye, X., Zhu, J., Zeng, H., Nan, Y., Sun, M., Ren, X., Huo, H., 2022.
https://doi.org/10.1007/s002679910020 Object detection from UAV thermal infrared images and videos using YOLO models.
Guzzetti, F., Carrara, A., Cardinali, M., Reichenbach, P., 1999. Landslide hazard eva­ Int. J. Appl. Earth Obs. Geoinformation 112, 102912. https://doi.org/10.1016/j.jag.
luation: a review of current techniques and their application in a multi-scale study, 2022.102912
Central Italy. Geomorphology 31 (1-4), 181–216. https://doi.org/10.1016/S0169- Jiang, Y., Wang, W., Zou, L., Cao, Y., 2024. Regional landslide susceptibility assessment
555X(99)00078-1 based on improved semi-supervised clustering and deep learning. Acta Geotech 19
Guzzetti, F., Mondini, A.C., Cardinali, M., Fiorucci, F., Santangelo, M., Chang, K.T., 2012. (1), 509–529. https://doi.org/10.1007/s11440-023-01950-0
Landslide inventory maps: new tools for an old problem. Earth Sci. Rev. 112 (1-2), Jonathan, L., Evan, S., Trevor, D., 2015. Fully convolutional networks for semantic seg­
42–66. https://doi.org/10.1016/j.earscirev.2012.02.001 mentation. IEEE Conf. Comput. Vis. Pattern Recognit. 3431–3440. https://doi.org/
Guzzetti, F., Reichenbach, P., Ardizzone, F., Cardinali, M., Galli, M., 2006. Estimating the 10.1109/CVPR.2015.7298965
quality of landslide susceptibility models. Geomorphology 81 (1-2), 166–184. Joshi, M., Kothyari, G.C., Kotlia, B.S., 2024. Landslide detection in Kinnaur Valley, NW
https://doi.org/10.1016/j.geomorph.2006.04.007 India using PS-InSAR technique. Phys. Geogr. 45 (2), 160–174. https://doi.org/10.
Hacıefendioğlu, K., Varol, N., Toğan, V., Bahadır, Ü., Kartal, M.E., 2024. Automatic 1080/02723646.2023.2202932
landslide detection and visualization by using deep ensemble learning method. Ju, Y., Xu, Q., Jin, S., Li, W., Su, Y., Dong, X., Guo, Q., 2022. Loess landslide detection
Neural Comput. Appl. 36 (18), 10761–10776. https://doi.org/10.1007/s00521-024- using object detection algorithms in northwest China. Remote Sens ((Basel)) 14.
09638-6 https://doi.org/10.3390/rs14051182
Hammouche, K., Diaf, M., Siarry, P., 2008. A multilevel automatic thresholding method Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. ImageNet classification with deep
based on a genetic algorithm for a fast image segmentation. Comput. Vis. Image convolutional neural networks. Advances in Neural Information Processing Systems.
Underst. 109 (2), 163–175. https://doi.org/10.1016/j.cviu.2007.09.001 Lee, J.J., Song, M.S., Yun, H.S., Yum, S.G., 2022. Dynamic landslide susceptibility ana­
Han, Z., Fang, Z., Li, Y., Fu, B., 2023. A novel Dynahead-Yolo neural network for the lysis that combines rainfall period, accumulated rainfall, and geospatial information.
detection of landslides with variable proportions using remote sensing images. Front. Sci. Rep. 12 (1), 18429. https://doi.org/10.1038/s41598-022-21795-z
Earth Sci. 10, 1077153. https://doi.org/10.3389/feart.2022.1077153 Lei, T., Zhang, Y., Lv, Z., Li, S., Liu, S., Nandi, A.K., 2019b. Landslide inventory mapping
Han, Z., Fu, B., Fang, Z., Li, Y., Li, J., Jiang, N., Chen, G., 2024. Dynahead-YOLO-otsu: an from bitemporal images using deep convolutional neural networks. IEEE Geosci.
efficient DCNN-based landslide semantic segmentation method using remote sensing Remote Sens. Lett. 16 (6), 982–986. https://doi.org/10.1109/LGRS.2018.2889307
images. Geomatics Nat. Hazards Risk 15 (1), 2398103. https://doi.org/10.1080/ Lei, T., Zhang, Q., Xue, D., Chen, T., Meng, H., Nandi, A.K., 2019a. End-to-end Change
19475705.2024.2398103 Detection Using a Symmetric Fully Convolutional Network for Landslide Mapping, in:
Han, Z., Li, Y., Du, Y., Wang, W., Chen, G., 2019. Noncontact detection of earthquake- ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing,
induced landslides by an enhanced image binarization method incorporating with 3027–3031. https://doi.org/10.1109/ICASSP.2019.8682802.
Monte-Carlo simulation. Geomatics Nat. Hazards Risk 10, 219–241. https://doi.org/ Li, Y., Chen, G., Han, Z., Zheng, L., Zhang, F., 2014. A hybrid automatic thresholding
10.1080/19475705.2018.1520745 approach using panchromatic imagery for rapid mapping of landslides. Giscience
Hinton, G., Deng, L., Yu, D., Dahl, G., Mohamed, A.R., Jaitly, N., Senior, A., Vanhoucke, Remote Sens 51 (6), 710–730. https://doi.org/10.1080/15481603.2014.972867
V., Nguyen, P., Sainath, T., Kingsbury, B., 2012. Deep neural networks for acoustic Li, Y., Fu, B., Han, Z., Fang, Z., Chen, N., Hu, G., Wang, W., Chen, G., 2024. PSO-SLIC
modeling in speech recognition: The shared views of four research groups. IEEE algorithm: A novel automated method for the generation of high-homogeneity slope
Signal Process Mag 29 (6), 82–97. https://doi.org/10.1109/MSP.2012.2205597 units using DEM data. Geomorphology 463, 109367. https://doi.org/10.1016/j.
Hong, H., Tsangaratos, P., Ilia, I., Loupasakis, C., Wang, Y., 2020. Introducing a novel geomorph.2024.109367
multi-layer perceptron network based on stochastic gradient descent optimized by a Li, Z., Guo, Y., 2020. Semantic segmentation of landslide images in Nyingchi region based
meta-heuristic algorithm for landslide susceptibility mapping. Sci. Total Environ. on PSPNet network. in: Proceedings - 2020 7th International Conference on
742, 140549. https://doi.org/10.1016/j.scitotenv.2020.140549 Information Science and Control Engineering, ICISCE 2020 1269–1273. https://doi.
Huabin, W., Gangjun, L., Weiya, X., Gonghui, W., 2005. GIS-based landslide hazard as­ org/10.1109/ICISCE50968.2020.00256
sessment: an overview. Prog. Phys. Geogr. 29 (4), 548–567. https://doi.org/10.1191/ Li, Y., Huang, S., Li, J., Huang, J., Wang, W., 2022. Spatial non-stationarity-based land­
0309133305pp462ra slide susceptibility assessment using PCAMGWR model. Water 14 (6), 881. https://
Huang, W., Ding, M., Li, Z., Yu, J., Ge, D., Liu, Q., Yang, J., 2023. Landslide susceptibility doi.org/10.3390/w14060881
mapping and dynamic response along the Sichuan-Tibet transportation corridor using Li, P., Wang, Y., Si, T., Ullah, K., Han, W., Wang, L., 2023. DSFA: cross-scene domain style
deep learning algorithms. Catena 222, 106866. https://doi.org/10.1016/j.catena. and feature adaptation for landslide detection from high spatial resolution images. Int. J.
2022.106866 Digit. Earth 16 (1), 2426–2447. https://doi.org/10.1080/17538947.2023.2229794
Huang, F., Pan, L., Fan, X., Jiang, S.H., Huang, J., Zhou, C., 2022. The uncertainty of Li, Y., Yang, J., Han, Z., Li, J., Wang, W., Chen, N., Hu, G., Huang, J., 2023. An ensemble
landslide susceptibility prediction modeling: suitability of linear conditioning factors. deep-learning framework for landslide susceptibility assessment using multiple
Bull. Eng. Geol. Environ. 81 (5), 182. https://doi.org/10.1007/s10064-022-02672-5 blocks: a case study of Wenchuan area, China. Geomatics, Nat. Hazards and Risk 14
Huang, F., Tao, S., Chang, Z., Huang, J., Fan, X., Jiang, S.H., Li, W., 2021. Efficient and (1), 2221771. https://doi.org/10.1080/19475705.2023.2221771
automatic extraction of slope units based on multi-scale segmentation method for Liu, W., Li, C., Xu, N., Jiang, T., Rahaman, M.M., Sun, H., Wu, X., Hu, W., Chen, H., Sun,
landslide assessments. Landslides 18, 3715–3731. https://doi.org/10.1007/s10346- C., Yao, Y., Grzegorzek, M., 2022. CVM-Cervix: A hybrid cervical Pap-smear image
021-01756-9 classification framework using CNN, visual transformer and multilayer perceptron.
Huang, F., Xiong, H., Jiang, S.H., Yao, C., Fan, X., Catani, F., Chang, Z., Zhou, X., Huang, Pattern Recognit 130, 108829. https://doi.org/10.1016/j.patcog.2022.108829
J., Liu, K., 2024. Modelling landslide susceptibility prediction: A review and con­ Liu, Q., Tang, A., Huang, D., 2023. Exploring the uncertainty of landslide susceptibility
struction of semi-supervised imbalanced theory. Earth Sci Rev, 104700. https://doi. assessment caused by the number of non–landslides. Catena 227, 107109. https://
org/10.1016/j.earscirev.2024.104700 doi.org/10.1016/j.catena.2023.107109
Huang, Y., Zhang, J., He, H., Jia, Y., Chen, R., Ge, Y., Ming, Z., Zhang, L., Li, H., 2024. Liu, P., Wei, Y., Wang, Q., Xie, J., Chen, Y., Li, Z., Zhou, H., 2021. A research on landslides
MAST: an earthquake-triggered landslides extraction method combining morpholo­ automatic extraction model based on the improved mask R-CNN. ISPRS Int. J. Geo-
gical analysis edge recognition with swin-transformer deep learning model. IEEE J. Inf. 10 (3), 168. https://doi.org/10.3390/ijgi10030168
Sel. Top. Appl. Earth Obs. Remote Sens. 17, 2586–2595. https://doi.org/10.1109/ Liu, Y., Zhang, W., Chen, X., Yu, M., Sun, Y., Meng, F., Fan, X., 2021. Landslide detection
JSTARS.2023.3342989 of high-resolution satellite images using asymmetric dual-channel network, in:
Huang, Y., Zhao, L., 2018. Review on landslide susceptibility mapping using support vector International Geoscience and Remote Sensing Symposium (IGARSS). 4091–4094.
machines. Catena 165, 520–529. https://doi.org/10.1016/j.catena.2018.03.003 https://doi.org/10.1109/IGARSS47720.2021.9553683

16
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Lu, P., Shi, W., Li, Z., 2022. Landslide mapping from planetscope images using improved Pham, B.T., Pradhan, B., Tien Bui, D., Prakash, I., Dholakia, M.B., 2016. A comparative
region-based level set evolution. IEEE Geosci. Remote Sens. Lett. 19, 1–5. https://doi. study of different machine learning methods for landslide susceptibility assessment: A
org/10.1109/LGRS.2021.3122964 case study of Uttarakhand area (India). Environ. Model. Softw. 84, 240–250. https://
Lu, M., Tay, L.T., Mohamad-Saleh, J., 2024. Landslide susceptibility analysis using doi.org/10.1016/j.envsoft.2016.07.005.
random forest model with SMOTE-ENN resampling algorithm. Geomatics, Nat. Piciullo, L., Calvello, M., Cepeda, J.M., 2018. Territorial early warning systems for
Hazards and Risk 15 (1), 2314565. https://doi.org/10.1080/19475705.2024. rainfall-induced landslides. Earth Sci. Rev. 179, 228–247. https://doi.org/10.1016/j.
2314565 earscirev.2018.02.013
Lv, L., Chen, T., Dou, J., Plaza, A., 2022. A hybrid ensemble-based deep-learning fra­ Pradhan, B., Dikshit, A., Lee, S., Kim, H., 2023. An explainable AI (XAI) model for
mework for landslide susceptibility mapping. Int. J. Appl. Earth Obs. Geoinformation landslide susceptibility modeling. Appl. Soft Comput. 142, 110324. https://doi.org/
108, 102713. https://doi.org/10.1016/j.jag.2022.102713 10.1016/j.asoc.2023.110324
Lv, Z., Liu, T., Benediktsson, J.A., Falco, N., 2022. Land cover change detection techni­ Pradhan, B., Jebur, M.N., Shafri, H.Z.M., Tehrany, M.S., 2016. Data fusion technique
ques: very-high-resolution optical images: a review. IEEE Geosci. Remote Sens. Mag. using wavelet transform and taguchi methods for automatic landslide detection from
10 (1), 44–63. https://doi.org/10.1109/MGRS.2021.3088865 airborne laser scanning data and quickbird satellite imagery. IEEE Trans. Geosci.
Lv, Z., Liu, T., Kong, X.B., Shi, C., Benediktsson, J.A., 2020. Landslide inventory mapping Remote Sens. 54 (3), 1610–1622. https://doi.org/10.1109/TGRS.2015.2484325
with bitemporal aerial remote sensing images based on the dual-path fully con­ Prakash, N., Manconi, A., Loew, S., 2020. Mapping landslides on EO data: performance of
volutional network. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 13, 4575–4584. deep learning models vs. Traditional machine learning models. Remote Sens 12 (3),
https://doi.org/10.1109/JSTARS.2020.2980895 346. https://doi.org/10.3390/rs12030346
Lv, Z.Y., Shi, W., Zhang, X., Benediktsson, J.A., 2018. Landslide inventory mapping from Qi, W., Wei, M., Yang, W., Xu, C., Ma, C., 2020. Automatic mapping of landslides by the
bitemporal high-resolution remote sensing images using change detection and mul­ ResU-Net. Remote Sens 12 (15), 2487. https://doi.org/10.3390/RS12152487
tiscale segmentation. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 11 (5), Rabby, Y.W., Hossain, M.B., Abedin, J., 2022. Landslide susceptibility mapping in three
1520–1532. https://doi.org/10.1109/JSTARS.2018.2803784 Upazilas of Rangamati hill district Bangladesh: application and comparison of GIS-
Ma, Z., Mei, G., 2021. Deep learning for geological hazards analysis: data, models, ap­ based machine learning methods. Geocarto Int 37 (12), 3371–3396. https://doi.org/
plications, and opportunities. Earth Sci. Rev. 223, 103858. https://doi.org/10.1016/ 10.1080/10106049.2020.1864026
j.earscirev.2021.103858 Rabby, Y.W., Li, Y., Hilafu, H., 2023. An objective absence data sampling method for
Marjanović, M., Kovačević, M., Bajat, B., Voženílek, V., 2011. Landslide susceptibility landslide susceptibility mapping. Sci. Rep. 13 (1), 1740. https://doi.org/10.1038/
assessment using SVM machine learning algorithm. Eng. Geol. 123 (3), 225–234. s41598-023-28991-5
https://doi.org/10.1016/j.enggeo.2011.09.006 Razavi-Termeh, S.V., Shirani, K., Pasandi, M., 2021. Mapping of landslide susceptibility
Marrapu, B.M., Jakka, R.S., 2014. Landslide hazard zonation methods: a critical review. using the combination of neuro-fuzzy inference system (ANFIS), ant colony (ANFIS-
Int. J. Civ. Eng. Res. 5 (3), 215–220. ACOR), and differential evolution (ANFIS-DE) models. Bull. Eng. Geol. Environ. 80,
Meena, S.R., Ghorbanzadeh, O., van Westen, C.J., Nachappa, T.G., Blaschke, T., Singh, 2045–2067. https://doi.org/10.1007/s10064-020-02048-7
R.P., Sarkar, R., 2021. Rapid mapping of landslides in the Western Ghats (India) Reichenbach, P., Rossi, M., Malamud, B.D., Mihir, M., Guzzetti, F., 2018. A review of
triggered by 2018 extreme monsoon rainfall using a deep learning approach. statistically-based landslide susceptibility models. Earth Sci. Rev. 180, 60–91.
Landslides 18, 1937–1950. https://doi.org/10.1007/s10346-020-01602-4 https://doi.org/10.1016/j.earscirev.2018.03.001
Melgani, F., Bruzzone, L., 2004. Classification of hyperspectral remote sensing images Ren, T., Gao, L., Gong, W., 2024. An ensemble of dynamic rainfall index and machine
with support vector machines. IEEE Trans. Geosci. Remote Sens 42 (8), 1778–1790. learning method for spatiotemporal landslide susceptibility modeling. Landslides 21
https://doi.org/10.1109/TGRS.2004.831865 (2), 257–273. https://doi.org/10.1007/s10346-023-02152-1
Merghadi, A., Yunus, A.P., Dou, J., Whiteley, J., ThaiPham, B., Bui, D.T., Avtar, R., Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for biome­
Abderrahmane, B., 2020. Machine learning methods for landslide susceptibility stu­ dical image segmentation, in: Lecture Notes in Computer Science (Including Subseries
dies: a comparative overview of algorithm performance. Earth Sci. Rev. 207, 103225. Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics).
https://doi.org/10.1016/j.earscirev.2020.103225 234–241. https://doi.org/10.1007/978–3-319–24574-4_28
Milan, A., Pham, T., Vijay, K., Morrison, D., Tow, A.W., Liu, L., Erskine, J., Grinover, R., Saha, S., Saha, A., Roy, B., Sarkar, R., Bhardwaj, D., Kundu, B., 2022. Integrating the
Gurman, A., Hunn, T., Kelly-Boxall, N., Lee, D., McTaggart, M., Rallos, G., Razjigaev, Particle Swarm Optimization (PSO) with machine learning methods for improving
A., Rowntree, T., Shen, T., Smith, R., Wade-McCue, S., Zhuang, Z., Lehnert, C., Lin, the accuracy of the landslide susceptibility model. Earth Sci. Inform. 15 (4),
G., Reid, I., Corke, P., Leitner, J., 2018. Semantic segmentation from limited training 2637–2662. https://doi.org/10.1007/s12145-022-00878-5
data. in: Proceedings - IEEE International Conference on Robotics and Automation Sajadi, P., Sang, Y.F., Gholamnia, M., Bonafoni, S., Mukherjee, S., 2022. Evaluation of the
1908–1915. https://doi.org/10.1109/ICRA.2018.8461082 landslide susceptibility and its spatial difference in the whole Qinghai-Tibetan
Mo, P., Li, D., Liu, M., Jia, J., Chen, X., 2023. A lightweight and partitioned CNN algo­ Plateau region by five learning algorithms. Geosci. Lett. 9 (1), 9. https://doi.org/10.
rithm for multi-landslide detection in remote sensing images. Appl. Sci. 13 (15), 1186/s40562-022-00218-x
8583. https://doi.org/10.3390/app13158583 Saleem, N., Enamul Huq, M., Twumasi, N.Y.D., Javed, A., Sajjad, A., 2019. Parameters
Moayedi, H., Mehrabi, M., Mosallanezhad, M., Rashid, A.S.A., Pradhan, B., 2019. derived from and/or used with digital elevation models (DEMs) for landslide sus­
Modification of landslide susceptibility mapping using optimized PSO-ANN ceptibility mapping and landslide risk assessment: a review. ISPRS Int. J. Geoinf 8
technique. Eng. Comput. 35, 967–984. https://doi.org/10.1007/s00366-018- (12), 545. https://doi.org/10.3390/ijgi8120545
0644-0 Sameen, M.I., Pradhan, B., 2019. Landslide detection using residual networks and the
Mohan, A., Singh, A.K., Kumar, B., Dwivedi, R., 2021. Review on remote sensing methods fusion of spectral and topographic information. IEEE Access 7, 114363–114373.
for landslide detection using machine and deep learning. Trans. Emerg. Telecommun. https://doi.org/10.1109/ACCESS.2019.2935761
Technol. 32 (7), e3998. https://doi.org/10.1002/ett.3998 Schlögel, R., Doubre, C., Malet, J.P., Masson, F., 2015. Landslide deformation monitoring
Mondini, A.C., Guzzetti, F., Reichenbach, P., Rossi, M., Cardinali, M., Ardizzone, F., 2011. with ALOS/PALSAR imagery: a D-InSAR geomorphological interpretation method.
Semi-automatic recognition and mapping of rainfall induced shallow landslides using Geomorphology 231, 314–330. https://doi.org/10.1016/j.geomorph.2014.11.031
optical satellite images. Remote Sens. Environ. 115 (7), 1743–1757. https://doi.org/ Schulz, W.H., Kean, J.W., Wang, G., 2009. Landslide movement in southwest Colorado
10.1016/j.rse.2011.03.006 triggered by atmospheric tides. Nat. Geosci. 2 (12), 863–866. https://doi.org/10.
Nagendra, S., Kifer, D., Mirus, B., Pei, T., Lawson, K., Manjunatha, S.B., Li, W., Nguyen, 1038/ngeo659
H., Qiu, T., Tran, S., Shen, C., 2022. Constructing a large-scale landslide database Scott, A.J., Hosmer, D.W., Lemeshow, S., 1991. Applied logistic regression. Biometrics 47.
across heterogeneous environments using task-specific model updates. IEEE J. Sel. https://doi.org/10.2307/2532419
Top. Appl. Earth Obs. Remote Sens. 15, 4349–4370. https://doi.org/10.1109/ Sharma, N., Saharia, M., Ramana, G.V., 2024. High resolution landslide susceptibility
JSTARS.2022.3177025 mapping using ensemble machine learning and geospatial big data. Catena 235,
Nagy, M., Molontay, R., 2024. Interpretable dropout prediction: towards XAI-based 107653. https://doi.org/10.1016/j.catena.2023.107653
personalized intervention. Int. J. Artif. Intell. Educ. 34 (2), 274–300. https://doi.org/ Shi, W., Zhang, M., Ke, H., Fang, X., Zhan, Z., Chen, S., 2021. Landslide recognition by
10.1007/s40593-023-00331-8 deep convolutional neural network and change detection. IEEE Trans. Geosci. Remote
Nandi, A., Shakoor, A., 2010. A GIS-based landslide susceptibility evaluation using bi­ Sens. 59 (6), 4654–4672. https://doi.org/10.1109/TGRS.2020.3015826
variate and multivariate statistical analyses. Eng. Geol. 110 (1-2), 11–20. https://doi. Song, Y., Gong, J., Gao, S., Wang, D., Cui, T., Li, Y., Wei, B., 2012. Susceptibility assessment
org/10.1016/j.enggeo.2009.10.001 of earthquake-induced landslides using Bayesian network: A case study in Beichuan,
Nguyen, H., Mehrabi, M., Kalantar, B., Moayedi, H., Abdullahi, M.M., 2019. Potential of China. Comput. Geosci. 42, 189–199. https://doi.org/10.1016/j.cageo.2011.09.011
hybrid evolutionary approaches for assessment of geo-hazard landslide susceptibility Su, Z., Chow, J.K., Tan, P.S., Wu, J., Ho, Y.K., Wang, Y.H., 2021. Deep convolutional
mapping. Geomatics, Nat. Hazards and Risk 10 (1), 1667–1693. https://doi.org/10. neural network–based pixel-wise landslide inventory mapping. Landslides 18,
1080/19475705.2019.1607782 1421–1443. https://doi.org/10.1007/s10346-020-01557-6
Noviello, C., Peduto, D., Verde, S., Zamparelli, V., Fornaro, G., Pauciullo, A., Reale, D., Sun, H., Xu, H., Liu, B., He, D., He, J., Zhang, H., Geng, N., 2021. MEAN-SSD: a novel real-
Nicodemo, G., Ferlisi, S., Gulla, G., 2020. Monitoring buildings at landslide risk with time detector for apple leaf diseases using improved light-weight convolutional
SAR: a methodology based on the use of multipass interferometric data. IEEE Geosci. neural networks. Comput. Electron. Agric. 189, 106379. https://doi.org/10.1016/j.
Remote Sens. Mag. 8 (1), 91–119. https://doi.org/10.1109/MGRS.2019.2963140 compag.2021.106379
Osmanoğlu, B., Sunar, F., Wdowinski, S., Cabral-Cano, E., 2016. Time series analysis of Sun, D., Xu, J., Wen, H., Wang, Y., 2020. An optimized random forest model and its
InSAR data: methods and trends. ISPRS J. Photogramm. Remote Sens. 115, 90–102. generalization ability in landslide susceptibility mapping: application in two areas of
https://doi.org/10.1016/j.isprsjprs.2015.10.003 Three Gorges Reservoir, China. J. Earth Sci. 31, 1068–1086. https://doi.org/10.
Othman, A.A., Gloaguen, R., Andreani, L., Rahnama, M., 2018. Improving landslide 1007/s12583-020-1072-9
susceptibility mapping using morphometric features in the Mawat area, Kurdistan Sun, D., Xu, J., Wen, H., Wang, D., 2021. Assessment of landslide susceptibility mapping
Region, NE Iraq: Comparison of different statistical models. Geomorphology 319, based on Bayesian hyperparameter optimization: A comparison between logistic re­
147–160. https://doi.org/10.1016/j.geomorph.2018.07.018 gression and random forest. Eng. Geol. 281, 105972. https://doi.org/10.1016/j.
Paulin, G.L., Mickelson, K.A., Contreras, T.A., Gallin, W., Jacobacci, K.E., Bursik, M., enggeo.2020.105972
2022. Assessing landslide volume using two generic models: application to landslides Sze, V., Chen, Y.H., Yang, T.J., Emer, J.S., 2017. Efficient processing of deep neural
in Whatcom County, Washington, USA. Landslides 19 (4), 901–912. https://doi.org/ networks: a tutorial and survey. Proc. IEEE. 105 (12), 2295–2329. https://doi.org/
10.1007/s10346-021-01825-z 10.1109/JPROC.2017.2761740

17
Y. Li, B. Fu, Y. Yin et al. Intelligent Geoengineering 1 (2024) 1–18

Tanatipuknon, A., Aimmanee, P., Watanabe, Y., Murata, K.T., Wakai, A., Sato, G., Hung, Yang, Zhongkang, Wei, J., Deng, J., Zhao, S., 2022. An improved method for the eva­
H.V., Tungpimolrut, K., Keerativittayanun, S., Karnjana, J., 2021. Study on com­ luation and local multi-scale optimization of the automatic extraction of slope units
bining two faster R-CNN models for landslide detection with a classification decision in complex terrains. Remote Sens 14, 3444.
tree to improve the detection performance. J. Disaster Res. 16 (4), 588–595. https:// Yang, B., Xiao, T., Wang, L., Huang, W., 2022. Using complementary ensemble empirical
doi.org/10.20965/jdr.2021.p0588 mode decomposition and gated recurrent unit to predict landslide displacements in
Tanyu, B.F., Abbaspour, A., Alimohammadlou, Y., Tecuci, G., 2021. Landslide suscept­ dam reservoir. Sensors 22 (4), 1320. https://doi.org/10.3390/s22041320
ibility analyses using Random Forest, C4.5, and C5.0 with balanced and unbalanced Yang, Zhiqiang, Xu, C., Li, L., 2022. Landslide detection based on ResU-Net with trans­
datasets. Catena 203, 105355. https://doi.org/10.1016/j.catena.2021.105355 former and CBAM embedded: two examples with geologically different environ­
Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., Dick, O.B., 2012. Landslide suscept­ ments. Remote Sens 14 (12), 2885. https://doi.org/10.3390/rs14122885
ibility assessment in the Hoa Binh province of Vietnam: a comparison of the Yin, Y., Wang, F., Sun, P., 2009. Landslide hazards triggered by the 2008 Wenchuan
Levenberg-Marquardt and Bayesian regularized neural networks. Geomorphology earthquake, Sichuan, China. Landslides 6, 139–152. https://doi.org/10.1007/
171, 12–29. https://doi.org/10.1016/j.geomorph.2012.04.023 s10346-009-0148-5
Titti, G., van Westen, C., Borgatti, L., Pasuto, A., Lombardo, L., 2021. When enough is Yong, Q., Wu, H., Wang, J., Chen, R., Yu, B., Zuo, J., Du, L., 2023. Automatic identifi­
really enough? On the minimum number of landslides to build reliable susceptibility cation of illegal construction and demolition waste landfills: A computer vision ap­
models. Geosciences 11 (11), 469. https://doi.org/10.3390/geosciences11110469 proach. Waste Manag 172, 267–277. https://doi.org/10.1016/j.wasman.2023.10.
Tong, L.J., Chen, K., Zhang, Y., Fu, X.L., Duan, J.Y., 2009. Document image binarization 023
based on NFCM. in: Proceedings of the 2009 2nd International Congress on Image Yu, B., Chen, F., Xu, C., 2020. Landslide detection based on contour-based deep learning
and Signal Processing 1–5. https://doi.org/10.1109/CISP.2009.5305330 framework in case of national scale of Nepal in 2015. Comput. Geosci. 135, 104388.
Turel, M., Frost, J.D., 2011. Delineation of slope profiles from digital elevation models for https://doi.org/10.1016/j.cageo.2019.104388
landslide hazard analysis. Risk Assess. Manag. 829–836. https://doi.org/10.1061/ Yu, B., Chen, F., Xu, C., Wang, L., Wang, N., 2021. Matrix segnet: a practical deep learning
41183(418)87 framework for landslide mapping from images of different areas with different spatial
Ullo, S.L., Langenkamp, M.S., Oikarinen, T.P., Delrosso, M.P., Sebastianelli, A., Iccirillo, F.P., resolutions. Remote Sens 13 (16), 3158. https://doi.org/10.3390/rs13163158
Sica, S., 2019. Landslide geohazard assessment with convolutional neural networks using Yu, C., Wang, J., Peng, C., Gao, C., Yu, G., Sang, N., 2018. BiSeNet: bilateral segmentation
sentinel-2 imagery data. in: International Geoscience and Remote Sensing Symposium network for real-time semantic segmentation. in: Lecture Notes in Computer Science
(IGARSS 9646–9649. https://doi.org/10.1109/IGARSS.2019.8898632 (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Ullo, S., Mohan, A., Sebastianelli, A., Ahamed, S., Kumar, B., Dwivedi, R., Sinha, G., 2021. Bioinformatics) 325–341. https://doi.org/10.1007/978-3-030-01261-8_20
A new mask R-CNN-based method for improved landslide detection. IEEE J. Sel. Top. Zhang, W., He, Y., Wang, L., Liu, S., Meng, X., 2023. Landslide Susceptibility mapping
Appl. Earth Obs. Remote. Sens. 14, 3799–3810. https://doi.org/10.1109/JSTARS. using random forest and extreme gradient boosting: a case study of Fengjie,
2021.3064981 Chongqing. Geol. J. 58 (6), 2372–2387. https://doi.org/10.1002/gj.4683
Van Den Eeckhaut, M., Kerle, N., Poesen, J., Hervás, J., 2012. Object-oriented identifi­ Zhang, J., Ma, X., Zhang, Jialan, Sun, D., Zhou, X., Mi, C., Wen, H., 2023. Insights into
cation of forested landslides with derivatives of single pulse LiDAR data. geospatial heterogeneity of landslide susceptibility based on the SHAP-XGBoost
Geomorphology 173, 30–42. https://doi.org/10.1016/j.geomorph.2012.05.024 model. J. Environ. Manage. 332, 117357. https://doi.org/10.1016/j.jenvman.2023.
van Westen, C.J., van Asch, T.W.J., Soeters, R., 2006. Landslide hazard and risk zonation - 117357
why is it still so difficult? Bull. Eng. Geol. Environ. 65, 167–184. https://doi.org/10. Zhang, Z., Qingjie, L., Yunhong, W., 2018. Road extraction by deep residual u-net. IEEE
1007/s10064-005-0023-0 Geosci. Remote Sens. Lett. 15 (5), 749–753. https://doi.org/10.1109/LGRS.2018.
Van Westen, C.J., Rengers, N., Soeters, R., 2003. Use of geomorphological information in 2802944
indirect landslide susceptibility assessment. Nat. Hazards 30, 399–419. https://doi. Zhang, C., Sargent, I., Pan, X., Li, H., Gardiner, A., Hare, J., Atkinson, P.M., 2018. An
org/10.1023/B:NHAZ.0000007097.42735.9e object-based convolutional neural network (OCNN) for urban land use classification.
Wan, Y., Huang, J., Ji, Y., Yu, Z., Luo, M., 2023. Combining BotNet and ResNet Feature Remote Sens. Environ. 216, 57–70. https://doi.org/10.1016/j.rse.2018.06.034
Maps for Accurate Landslide Identification Using DeepLabV3+, in: 2023 6th Zhang, Y. bin, Xu, P. yi, Liu, J., He, J. xian, Yang, H. tian, Zeng, Y., He, Y. yong, Yang, C.
International Conference on Artificial Intelligence and Big Data, ICAIBD 2023. feng, 2023. Comparison of LR, 5-CV SVM, GA SVM, and PSO SVM for landslide
777–782. https://doi.org/10.1109/ICAIBD57115.2023.10206334 susceptibility assessment in Tibetan Plateau area, China. J. Mt. Sci. 20 (4), 979–995.
Wang, N., Cheng, W., Zhao, M., Liu, Q., Wang, J., 2019a. Identification of the debris flow https://doi.org/10.1007/s11629-022-7685-y
process types within catchments of Beijing mountainous area. Water 11 (4), 638. Zhang, R.X., Zhu, W., Li, Z.H., Zhang, B.C., Chen, B., 2024. Re-net: multibranch network
https://doi.org/10.3390/w11040638 with structural reparameterization for landslide detection in optical imagery. IEEE J.
Wang, Yi, Fang, Z., Hong, H., 2019b. Comparison of convolutional neural networks for Sel. Top. Appl. Earth Obs. Remote Sens. 17, 2828–2837. https://doi.org/10.1109/
landslide susceptibility mapping in Yanshan County, China. Sci. Total Environ. 666, JSTARS.2023.3344720
975–993. https://doi.org/10.1016/j.scitotenv.2019.02.263 Zhang, C.Y., Yang, J.Y., 2010. Binarization of document images with complex back­
Wang, Yu, Wang, X., Jian, J., 2019c. Remote sensing landslide recognition based on ground, in: 2010 6th International Conference on Wireless Communications,
convolutional neural network. Math. Probl. Eng., 8389368. https://doi.org/10.1155/ Networking and Mobile Computing, WiCOM 2010. 1–4. https://doi.org/10.1109/
2019/8389368 WICOM.2010.5601007.
Wang, L., Wu, J., Zhang, W., Wang, Lin, Cui, W., 2021b. Efficient seismic stability analysis Zhao, Q., Hastie, T., 2021. Causal Interpretations of Black-Box Models. J. Bus. Econ. Stat.
of embankment slopes subjected to water level changes using gradient boosting al­ 39 (1), 272–281. https://doi.org/10.1080/07350015.2019.1624293
gorithms. Front. Earth Sci. 9, 807317. https://doi.org/10.3389/feart.2021.807317 Zhao, Z.Q., Zheng, P., Xu, S.T., Wu, X., 2019. Object detection with deep learning: a
Wang, H., Zhang, L., Luo, H., He, J., Cheung, R.W.M., 2021a. AI-powered landslide review. IEEE Trans. Neural. Netw. Learn. Syst. 30 (11), 3212–3232. https://doi.org/
susceptibility assessment in Hong Kong. Eng. Geol. 288, 106103. https://doi.org/10. 10.1109/TNNLS.2018.2876865
1016/j.enggeo.2021.106103 Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J., 2017. Pyramid scene parsing network, in:
Wei, R., Ye, C., Sui, T., Zhang, H., Ge, Y., Li, Y., 2023. A feature enhancement framework Proceedings - 30th IEEE Conference on Computer Vision and Pattern Recognition,
for landslide detection. Int. J. Appl. Earth Obs. Geoinformation 124, 103521. https:// CVPR 2017. 2881–2890. https://doi.org/10.1109/CVPR.2017.660
doi.org/10.1016/j.jag.2023.103521 Zhou, X., Wen, H., Zhang, Y., Xu, J., Zhang, W., 2021. Landslide susceptibility mapping
Whiteley, J.S., Chambers, J.E., Uhlemann, S., Wilkinson, P.B., Kendall, J.M., 2019. using hybrid random forest with GeoDetector and RFE for factor optimization.
Geophysical monitoring of moisture-induced landslides: a review. Rev. Geophys. 57 Geosci. Front. 12 (5), 101211. https://doi.org/10.1016/j.gsf.2021.101211
(1), 106–145. https://doi.org/10.1029/2018RG000603 Zhou, B., Yang, G., Shi, Z., Ma, S., 2024. Natural language processing for smart health­
Wu, Y. chen, Feng, J. wen, 2018. Development and application of artificial neural network. care. IEEE Rev. Biomed. Eng. 17, 4–18. https://doi.org/10.1109/RBME.2022.
Wirel. Pers. Commun. 102, 1645–1656. https://doi.org/10.1007/s11277-017-5224-x 3210270
Xing, Y., Huang, S., Yue, J., Chen, Y., Xie, W., Wang, P., Xiang, Y., Peng, Y., 2023. Zhu, L., Wang, G., Huang, F., Li, Y., Chen, W., Hong, H., 2021. Landslide susceptibility
Patterns of influence of different landslide boundaries and their spatial shapes on the prediction using sparse feature extraction and machine learning models based on GIS
uncertainty of landslide susceptibility prediction. Nat. Hazards 118 (1), 709–727. and remote sensing. IEEE Geosci. Remote Sens. Lett. 19, 1–5. https://doi.org/10.
https://doi.org/10.1007/s11069-023-06025-7 1109/LGRS.2021.3054029
Xu, C., 2015. Preparation of earthquake-triggered landslide inventory maps using remote Zhuang, J., Peng, J., Wang, G., Javed, I., Wang, Y., Li, W., 2018. Distribution and char­
sensing and GIS technologies: Principles and case studies. Geosci. Front. 6 (6), acteristics of landslide in Loess Plateau: a case study in Shaanxi province. Eng. Geol.
825–836. https://doi.org/10.1016/j.gsf.2014.03.004 236, 89–96. https://doi.org/10.1016/j.enggeo.2017.03.001
Yang, S., Wang, Y., Wang, P., Mu, J., Jiao, S., Zhao, X., Wang, Z., Wang, K., Zhu, Y., 2022. Zou, J., Han, Y., So, S.S., 2008. Overview of artificial neural networks. Artif. Neural
Automatic identification of landslides based on deep learning. Appl. Sci. 12 (16), Netw.: Methods Appl. 14–22. https://doi.org/10.1007/978-1-60327-101-1_2
8153. https://doi.org/10.3390/app12168153

18