Application of Artificial Intelligence and Remote Sensing for Landslide Detection and Prediction: Systematic Review

Abstract

This paper systematically reviews remote sensing technology and learning algorithms in exploring landslides. The work is categorized into four key components: (1) literature search characteristics, (2) geographical distribution and research publication trends, (3) progress of remote sensing and learning algorithms, and (4) application of remote sensing techniques and learning models for landslide susceptibility mapping, detections, prediction, inventory and deformation monitoring, assessment, and extraction and management. The literature selections were based on keyword searches using title/abstract and keywords from Web of Science and Scopus. A total of 186 research articles published between 2011 and 2024 were critically reviewed to provide answers to research questions related to the recent advances in the use of remote sensing technologies combined with artificial intelligence (AI), machine learning (ML), and deep learning (DL) algorithms. The review revealed that these methods have high efficiency in landslide detection, prediction, monitoring, and hazard mapping. A few current issues were also identified and discussed.

1. Introduction

Landslides are natural disasters which are caused by erosion, earthquakes, heavy downpours, deforestation, floods, and poor human-planned development activities in many parts of the world [1,2]. Landslides rank among the top seven geological natural hazards [3,4] and may result in casualties and enormous damage to infrastructure and socio-economic conditions [5,6].

Developing landslide susceptibility maps to identify landslide-prone areas and monitoring are the key strategic management of risk associated with this natural phenomenon. Several studies have mapped landslide susceptibility areas and developed models to predict or forecast landslides based on previous events and certain probabilistic assumptions [1,7,8]. Landslide susceptibility can be grouped according to heuristics, deterministic, and statistics [9,10]. The knowledge-driven approach to categorizing landslide susceptibility levels by examining the relative influence of several factors is preliminary and relies on subjective observations [2,11,12].

Remote sensing is an effective and cost-effective tool for landslide susceptibility mapping because it can cover a wide range of territorial datasets with different degrees of geospatial and temporal resolution. The commonly used remote sensing techniques include Sentinel-1, Sentinel-2, Landsat-8, Synthetic Aperture Radar (SAR), Informetric Synthetic Aperture Radar (InSAR), Google Earth Engine (GEE), and Unmanned aerial vehicle (UAV). Over the past years, landslide susceptibility mapping, detection, and prediction have been exploited with different remote sensing data and imagery with the application of various sophisticated AI, DL, and ML algorithms and models in many cases [13,14,15,16,17,18,19,20,21,22,23,24,25,26]. They provide fast, accurate predictions and can handle several non-linear and linear variables simultaneously compared to traditional methods [13,27,28]. To create landslide susceptibility maps, AI, ML, and DL algorithms and datasets must first be ascertained and prepared from past and present landslide condition factors. These datasets are obtainable from remote-sensing imagery to develop the landslide-conditional factors.

The AI, DL, and ML modeling techniques, namely boosted regression tree (BRT), random forest (RF), logic regression (LR) and support vector machines (SVM) [13,29,30,31], convolution neural network (CNN), deep convolution neural network (DCNN), back-propagation neural network (BPNN) and U-Net model [22,32,33,34], fuzzy expert system (FES), extreme learning machine (ELM), artificial neural network (ANN), explainable artificial intelligence (XAI), and certainty factor–random (CF–RF) hybrid [23,24,35,36] have been successfully used with remote sensing datasets from different sources to map the susceptibility of landslides and predict and monitor landslide deformation zones. However, traditional algorithms or models can apply different input factors without considering a specific statistical rule [37]. Some studies reported that ML models are intrinsically non-linear [2]. Additionally, Choi et al. [38] reported that ANN models have limitations in the local optimum and over-fitting in terms of landslide susceptibility mapping, detection, and forecasting. Hence, selecting an appropriate model for landslide susceptibility mapping, detection, and forecasting is necessary.

Recent studies have focused on improving the traditional AI, ML, and DL by developing new algorithms and models to increase the accuracy of landslide hazard mapping [39,40,41,42]. Learning algorithms have been developed to predict and detect landslides, including LR, SVM, and RF, with Certainty Factor-based (CF-based) hybrid models like CF–LR, CF–SVM, and CF–RF; ANN, XAI, and CNN [22,35,38,39] with prediction and detection accuracy ratios above 0.9. These models greatly enhance the computational efficiency of landslide susceptibility, prediction, and deformation monitoring to provide vital information for government agencies, engineers, and decision-making bodies. In addition, advancements have recently been made in AI, ML, and DL studies with remote sensing data and imagery to map landslide susceptibility, prediction, detection, and deformation monitoring. However, despite the significant interest in landslide studies involving remote sensing and learning algorithms, no systematic literature review has been provided to identify the major trends and advances in the field.

This systematic review aims to analyze the existing knowledge by providing an in-depth systematic review of the research articles, focusing on the advantages and disadvantages of AI, DL, and ML algorithms with remote sensing data and images on landslide susceptibility mapping, detection, prediction, and deformation monitoring. The research objective of this paper is to provide answers to the following questions:

• Over the years, what are the commonly used AI, ML, and DL methods for landslide hazard mapping?
• What are the commonly used remote sensing techniques in landslide hazard studies?
• How accurate are the AI, ML, and DL, as well as the remote sensing techniques, capable of evaluating landslides?

This systematic review provides the basis for further research on artificial intelligence, machine learning, and deep learning with remote sensing datasets and imagery to map and evaluate landslide hazards. Additionally, an analysis based on learning machines and various remote sensing techniques could be widely utilized in landslide evaluations.

2. Materials and Methods

2.1. Literature Search, Exclusion, and Inclusion Strategy

The studies included in this systematic review were retrieved from an extensive search for peer-reviewed journal research articles in Scopus and Web of Science databases. The following combinations were utilized in all the two databases through the Griffith University library; “landslide susceptibility AND remote sensing OR artificial intelligence”, “landslide detection AND machine learning OR remote sensing”, “deep learning AND remote sensing OR landslide deformation monitoring”, “landslide prediction AND remote sensing AND susceptibility OR different machine learning”. The search for research articles was conducted without restriction to the year of publication.

A total of 1328 references were pulled from the abovementioned databases. Following the literature search from Scopus and Web of Science, the retrieved peer-reviewed articles were imported into EndNote-21 for initial screening. The number of research articles identified, excluded, or included was recorded following the “Preferred reporting items for systematic reviews and meta-analysis statement” by Page et al. [43], as shown in Figure 1. The eligibility criteria of articles used in this review were the following:

• The study focuses on landslide susceptibility mapping, detection, prediction, and deformation monitoring with remote sensing data, imagery, satellite imagery or data, AI, ML, and DL, and no other type of landslide study.
• The articles are written only in English.
• The research articles were published in reliable journals.

Figure 1. The PRISMA flow diagram is used for the identification, screening, and inclusion process.

Figure 1. The PRISMA flow diagram is used for the identification, screening, and inclusion process.

After the initial screening in EndNote-21, 903 articles were excluded from the total record, which included articles focused on only “landslide and remote sensing”, without any of the learning machine algorithms and models, and “landslide and machine learning”, without remote sensing and all the articles not available in portable document format (PDF). In total, 425 records were exported from the EndNote-21 as an “XML” file and imported into the “Covidence software version 2.0” to carry out a detailed screening of the papers. The systematic review processes in Covidence included identification, screening, inclusion, and extraction of final research articles. In the identification process, 125 articles were excluded as duplicates in the first phase. The essential bibliographic information (abstract and title) of the remaining 300 articles were examined to check whether the articles contained remote sensing, landslide susceptibility, detection, prediction, deformation monitoring with the machine and deep learning, and artificial intelligence algorithms or models. Upon abstract and title screening, 75 articles were excluded as they were less relevant to this study. The remaining 225 articles were assessed for eligibility by further screening with the complete text, and 39 records were excluded as non-articles, non-English articles, and studies with partial information that is not relevant to the studies. In total, 186 articles met the selection criteria and were extracted and included in this review.

2.2. Data Extraction

The data from Covidence was exported into an Excel spreadsheet. The data were imported into EndNote to set the bibliographic information, including the author’s name, publication year, title, journal name, abstract, keywords, digital object identifier (DOI), and uniform resource locator (URL). In addition, information on the study area locations, remote sensing techniques and AI, ML, and DL algorithms or models with special attention to landslides susceptibility mapping, prediction, detection, and deformation monitoring being investigated were extracted after critically reading through each research article. The prediction models were compared with the traditional ones, and newly developed ones using machine learning algorithms and remote sensing data and imagery on landslides were extracted for the review analysis and discussions. The landslide susceptibility mapping, prediction, and detection accuracies were presented in percentages (i.e., 0–100%) and numbers (0–1). All the results were converted into percentages to compare the results from various research articles.

2.3. Data Analysis

The articles extracted were subjected to qualitative synthesis and quantitative analysis. The bibliometric analysis was first conducted to visualize the occurrence of and co-occurrence networks regarding keywords from the articles utilized in this study. The bibliometric analysis is a meta-analytical tool widely used to identify the interconnected keywords related to a given research topic from published journal articles [44]. The VOSviewer software version 1.6.20 was used for the bibliometric analysis [45]. The VOSviewer presents a network visualization of keywords in the form of linked clusters, and developing a cluster map in VOSviewer, four steps need to be followed: (1) selecting a counting method (i.e., whole or binary counting); (2) selecting the minimum number of occurrence for the keywords, thus a calculation of similarity index; (3) the relevance of the co-occurrence keywords are calculated, and based on the calculated results, display the essential keywords; and (4) the map displayed based on the selected keywords.

The VOSviewer functionality for bibliometric analysis and mapping is detailed by Eck et al. [45]. The abstract, title, and keywords of the 186 research articles were utilized as input text in VOSviewer to present the graphic visualizations of the keywords. Basic statistical analysis was performed with MS-Excel to assess the progress of landslide susceptibility, prediction, and detection with the learning paradigms such as artificial intelligence, deep learning, and machine learning [46]. The review was divided into sections to address the study objectives. Section 1 explores the progress of landslide susceptibility, prediction, and detection with the learning paradigms with remote sensing technologies studied and employed in the field. The outcome of Section 1 was utilized to articulate existing research gaps on remote sensing and learning paradigms related to landslide susceptibility, prediction, and detection.

3. Results

3.1. Literature Search Characteristics

The keywords co-occurrence in VOSviewer shows the relationship between the essential items frequently used in scientific research articles and publications, the necessary research topics, and the cognitive structure of the field of study [47]. In analyzing the 186 research articles retrieved from Scopus and WoS databases for this study, the VOSviewer network map is presented in Figure 2. The map categorized the identified keywords from the articles into four clusters of concepts. The red clusters included keywords such as “susceptibility”, “hazard assessment”, “landslide”, “remote sensing”, “u-net”, “deep learning”, “earthquake”, “change detection”, “supervised vector machine”, “mark r-nn”, “machine learning models”, “change detection”, and “landslide mapping”. This cluster links accurate evaluation of performance with valuing landslide characteristics with remote sensing and learning paradigms. The inclusion of keywords like “remote sensing and deep learning” in this red cluster presents the linkages between remote sensing images and data, and deep learning algorithms or models to detect, predict, and for the susceptibility mapping and deformation monitoring of landslides.

The green cluster had “natural hazard”, machine learning”, “landslides”, “remote sensing images”, and “landslide susceptibility” as the major keywords, which implies the potential of machine learning models in evaluating landslide characteristics with remote sensing datasets or images. The blue cluster connected key items such as “landslide susceptibility mapping and detection”, “remote sensing”, and “convolutional neural networks (CNN)”. This articulates the vast use of remote sensing data derived from imagery as a proxy for studying landslide susceptibility mapping and detection as a significant component of landslide characteristics, with most studies conducted with CNN, a part of deep learning models. Lastly, the yellow cluster keywords include “support vector machine”, “back-propagation neural network”, “machine learning”, “gis”, “uncertainty analysis”, and “susceptibility evaluation”. This may be attributed to the influence of scale settings on machine learning and remote sensing imagery’s performance in landslide mapping.

3.2. Geographic Distribution and Research Publication Trends

Of the 186 articles in this study, the spatial distribution involved in the meta-analysis was conducted in 27 countries, as illustrated in Figure 3. In analyzing the frequency of research publications by country, it was found that most of the studies on landslides with remote sensing and machine learning, artificial intelligence, and deep learning were conducted in Asian Pacific nations, with China having the highest number of publications. However, some studies have been conducted in Africa, with approximately 2% on landslide evaluation with remote sensing and learning paradigms, whereas all the studies were conducted in Egypt. About 2% of the studies were conducted in Australia, 9% in Europe, and 7% in North and South America. None of the studies in this study were conducted on a global scale. From Figure 3, it can be seen that there is a considerable gap in the geographic distribution of published research articles, especially in a significant part of Africa, the Eastern part of European countries, and some countries in South America. This calls for more research work to extend the application of remote sensing technologies and learning algorithms in assessing landslide susceptibility mapping, detection, prediction, and deformation monitoring across the global south and north.

Figure 4 shows research articles on landslides, remote sensing, and the application of artificial intelligence, deep and machine learning algorithms, or models over the past years. No studies dated 2012 or 2015 were retrieved from the literature as they did not satisfy the search criteria. It was noticed that a few research articles (about 3.23%) were published between 2011 and 2016. This small number of articles could be attributed to (1) the specific keywords and databases used for the article search and (2) researchers not giving much attention to utilizing learning algorithms and remote sensing data in investigating landslides and associated geohazards. Since then, the use of learning algorithms and models in landslide studies has significantly increased, reaching a total of 186 articles in 2024, amounting to about 97% of all published research articles. This surge in studies on utilizing learning algorithms or models with remote sensing datasets demonstrates the researchers’ interest in using AI, ML, and DL in landslide studies since these learning algorithms can deal with large datasets obtained from remote sensors with high accuracy in mapping landslide susceptibility.

3.3. Major Topics in the Reviewed Research Article

The results of this study show that nine major topics were mentioned in the research articles retrieved from the literature, as shown in Figure 5. The topics include susceptibility mapping, detection, monitoring, inventory, stability, the influence of train data, and management. Landslide susceptibility mapping is the most widely studied, followed by landslide detection and assessment, with few other studies. For instance, 88 studies use learning algorithms like AI, DL, and ML models for landslide susceptibility mapping or zoning. A total of 42 studies investigated landslide detection, 17 focused on landslide assessment, 7 on landslide monitoring, and 5 studies each studied landslide extractions and inventory. The results show that there was only one study each for stability [25], the influence of spatial heterogeneity on landslide susceptibility [48], and landslide management [49] utilizing remote sensing data and learning algorithms to investigate landslide characteristics. The application of remote sensing and learning algorithms related to these topics are discussed in detail in Section 4.

3.4. Commonly Used Remote Sensors

The utilization of remote sensors and data and imagery for landslide investigations has recently increased. Forty sensor types were observed in the literature, as illustrated in Figure 6. It can be observed that the digital elevation model (DEM) is the most widely used in landslide susceptibility, prediction, detection, and deformation monitoring with learning algorithms or models amounting to 27% of all the studies. The study’s findings also show that 12% of the studies employed Landsat-8 OLI and Google Earth Engine (GEE or GEI) in landslide investigations. Of these studies, approximately 7% specifically used high-resolution advanced land-observed satellite-phase array L-band synthetic aperture radar (ALOS–PALSAR) to derive digital elevation models.

Multispectral instruments such as Sentinel-1 and Sentinel-2 sensors have potential in landslide studies with learning models (7%). High spatial satellites, including light detecting range (LiDAR), information synthetic aperture radar (InSAR), advanced spaceborne thermal emission and reflection radiometer (ASTER), unmanned aerial vehicle (UAV) based sensors, and shuttle radar topography mission (STRM) have also been utilized in 18% of studies. The results also show that many sensors have not been used enough to investigate landslide susceptibility mapping, detection, prediction, and deformation monitoring with the learning paradigms in Figure 6. This is because remote sensors or satellite datasets from WorldView-1, 2, and 3, SuperView-1, and Zeiss Aerial Survey are commercial. For example, Zeiss Aerial Survey sensors require airplanes or helicopters to be able to acquire surface data, which makes them expensive to operate. Although, some remote sensors are not widely used, due to the availability of high-resolution remote sensor datasets online from Landsat-7 and 8, Sentinel-1 and 2, DEM, InSAR, ALOS–PALSAR, and GEI. The UAV and LiDAR are also commercial but have become popular because they can be manned quickly and cover a large area. They can be operated at shallow heights, in access areas that are impossible to access from the ground and are complex enough to be flown over by airplanes [17,32,50].

Figure 7 presents the progress of high-resolution remote sensor datasets employed in landslide studies with learning algorithms from 2011 to 2024. The use of DEM and Landsat-8 OLI sensors can be noticed from 2011, and they have been utilized almost every years since then, except the fiscal years of 2012 and 2015. Additionally, the UAV and Google Earth Images (GEI) in this study were observed trending from 2014 until now. The last five years have seen the ALOS–PALSAR being utilized, from 2019 to 2023 and counting; from 2020 to 2024, there were some shifts in the frequent use of ASTER, Sentinel-1, Sentinel-2, InSAR, and other sensors (i.e., RapidEye, WorldView-2 and 3, and MODIS) in the studies of landslides with machine learning models. It should be noted that no records of studies were found in the 2012 and 2015 fiscal years. Figure 7 also demonstrates a shift in leading remote sensing technologies, data sources, and methods with high resolution that are freely available online for research purposes.

3.5. Learning Algorithms with Remote Sensor Techniques

The applications of learning algorithms in landslide studies with remote sensing data and imagery from 2011 to 2024 are shown in Figure 8. The stack columns represent the annual research article journal publications. Except for 2012 and 2015, the number of research publications has increased steadily from the 2016 to 2022 fiscal years and decreased after the same year. It is evident from Figure 8 that researchers have given much attention to the use of learning algorithms and models and remote sensing in investigating landslides with supportive vector machines (SVMs) and random forests (RFs). It should be noted that a single research article can use multiple learning algorithms, resulting in more algorithms or models employed than the number of articles retrieved for this study. Figure 9 shows the specific models for artificial intelligence, the deep and machine learning algorithms, or the models used in this literature review. Machine learning (ML) has been widely used with remote sensing data for assessing landslides, especially the SVM, followed by deep learning with CNN, which is widely utilized, and artificial intelligence with ANN, which is primarily used.

The learning algorithms, remote sensors, and their prediction, detection, susceptibility zoning, and deformation monitoring of landslides are summarized and presented in

Table 1. Summary of some data sources and learning paradigms.

Learning Algorithms	Prediction Accuracy	Data Sources	Reference
ANN	0.97	TERRASAR-X; ALOS–PALSAR	Masruroh et al. [51]
CF–RF	0.93	ALOS–PALSAR; DEM; LISS-III; Radarsat	Peethambaran et al. [23]
CF–RF, U-Net	0.94; 0.96	Landsat-8; DEM; ALOS–PALSAR	Sharma et al. [35]
DCN	0.97	ALOS–PALSAR; Sentinel-2	Aladhi et al. [36]
ANN, RF	0.97; 1.0	ASTER-GDEM; Landsat-8	Sun et al. [52]
ANN, RF; SVM	0.98; 0.95; 0.97	PlanetScope; RapidEye; Sentinel-2	Das et al. [53]
MLP	0.88	DEM	Lee et al. [24]
ANN, SVM, RF; ELM	0.96; 0.96; 0.99; 0.99	MODIS, DEM	Sharma et al. [41]
RF, ANN, SVM	0.94; 0.90; 0.90	DEM; Landsat-8	Hu et al. [54]
FES, ELM	0.84; 0.80	ALOS–PALSAR; DEM; LISS-III; Radarsat	Peethambaran et al. [23]
ANN, BN, LR, SVM	0.92; 0.86; 0.93; 0.93	ASTER-DEM	Xie et al. [55]
DNN–HBA	0.95	DEM; Landsat-8	Nguyen et al. [56]
DADRCNN	0.90	DEM; OpenStreetMap	Ma et al. [57]
CNN	0.96	SBAS–InSAR	Sheng et al. [58]
FC–SAE	0.85	DEM; Landsat-8	Huang et al. [33]
ResNet-50 + SCAM	0.99	DEM	Ji et al. [59]
ACO–DBN	0.97	DEM; GDCP	Xiong et al. [60]
DNN; LSTM; CNN; RNN	0.87; 0.87; 0.86; 0.83	DEM; GEE	Habumugisha et al. [61]
LD–BiLSTM	0.90	DEM	Wang et al. [62]
GA–SVM	0.96	UAV; DEM	Niu et al. [63]
RF, BRT, CART, GLM	0.78; 0.96; 0.82; 0.82	SRTM; DEM	Youssef et al. [29]
LSSVM–BC	0.90	SPOT satellite images	Tien et al. [64]
LR, LB, NB	0.84; 0.71; 0.85	DEM	Pourghasemi et al. [65]
LNRF–BRT, LNRFLMR	0.91; 0.90	PALSAR; DEM	Arabameri et al. [66]
RF	0.97	Sentinel-1 and -2	Kocama et al. [67]
ANN, RF, DT, SVM	0.93; 0.92; 0.89; 0.91	ALOS–PALSAR; DEM	AI-Najjar et al. [68]
RF, XGBoost, NB, ANN, KNN, FR	0.99; 0.99; 0.97; 0.92; 0.88; 0.67	ALOSPALSAR; GDEM; Landsat-8	Khalil et al. [69]

. As seen in Table 1, the accuracy of landslide susceptibility mapping, prediction, detection, and monitoring from the various models and remote sensors shows an excellent prospect for landslide investigations. For instance, the prediction accuracy of 0.97 was observed by Masruroh et al. [51] using TERRASAR-X, ALOS–PALSAR high-resolution sensors imagery, and ANN algorithms. Similarly, Sharma et al. [41] utilized MODIS and DEM-sensing-derived data with four different algorithms and recorded an average accuracy of 0.98 for landslide susceptibility mapping.

Table 2. Model prediction accuracy measurements used in previous works.

Metric Measures	Parameters	Reference
$AUC={\displaystyle \frac{\left({\displaystyle \sum TP+TN}\right)}{\left(P+N\right)}}$ $Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$ $Specificity={\displaystyle \frac{TN}{TN+FP}}$ $Recall={\displaystyle \frac{TP}{TP+TN}}$ $MCC={\displaystyle \frac{TP\times TN-FN\times FP}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}}$	True positive (TP), false negative (FN): (P) is the total number of landslides; and (N) is the total number of non-landslides. True negative (TN), false positive (FP), and Mathew coefficient correlation (MCC)	Hong et al. [15], Huang et al. [70], Nhu et al. [11], Hussain et al. [71], Sun et al. [52], Yuan et al. [39]
$Kappa\text{-}index={\displaystyle \frac{p_{observed}-p_{expected}}{1-p_{expected}}}$	pobserved is identical to accuracy and pexpected	Lee et al. [24]
$F1\text{-}score={\displaystyle \frac{2\times PRE\times Recall}{PRE+Recall}}$	Precision (PRE)	Chowdhuri et al. [72], Zhang et al. [73]
$MAE={\displaystyle \frac{\left|p_1-a_1\right|+\left|p_2-a_2\right|+\mathrm{\dots }+\left|p_n-a_n\right|}{n}}$	Mean absolute error (MAE), p1 is the predicted value of landslide sensitivity, a1 is the actual value of landslide sensitivity, and n is the number of sample instances.	Deng et al. [74]
$mIoU={\displaystyle \frac{TP}{FP+TP+FN}}$	Mean Intersection over Union (mIoU), True positive (TP), false positive (TN), false negative (FN)	Liu et al. [34]

shows some of the metric measurements employed to predict the learning algorithms and models available in the research articles, including area under curve (AUC), F-score (F1-score), mean intersection over union (mIoU), recall, accuracy, Kappa index, and sensitivity. It can be seen from Table 2 that the AUC, accuracy, specificity, recall, and MCC metric measurements have been widely used in predicting the accuracy of learning algorithms with the actual landslide remote sensing datasets in terms of statistical analysis. The F1-score, MAE, and mIoU have also been utilized with reasonable prediction accuracy in the literature, demonstrating their capabilities for mapping, detecting, and predicting landslides.

4. Remote Sensing and Learning Algorithms Application for Landslides

4.1. A Brief Overview of the Application of Learning Models and Remote Sensing

Recently, artificial intelligence models (fuzzy expert system (FES), artificial neural network (ANN), extreme learning machine (ELM), fuzzy set procedure (FSP)) [53,75,76,77], deep learning models (deep convolutional network (DCN), convolutional neural network (CNN), U-Net, YOLO–SA, Deeplabv3+) [78,79,80] and machine learning models (supportive vector machine (SVM), random forest (RF), logistics regression (LR), extreme gradient boosting (XGBoost)) [81,82,83,84] with high-resolution-sensor remote sensors include UAV, ASTER-GDEM, SRTM, DEM, Sentinel-1, Sentinel-2, Landsat-8 OLI, InSAR, Google Earth Engine, MT-InSAR, and many others, as well as geospatial data like topography, hydrology, and environmental factors have been used to map, detect, extract, predict, and monitor landslide susceptibility successfully [22,85,86,87,88,89,90].

4.2. Landslide Mapping with AI, DL, ML, and Remote Sensing Data

Landslide susceptibility mapping, detecting, extracting, predicting, inventory, and monitoring are crucial for geo-hazard management, control, and mitigation. The complexity of geological terrains makes it challenging to map, predict, and monitor landslides using traditional analytical methods. The traditional methods also involved too much human interference and assumptions, significantly influencing landslide susceptibility prediction, mapping, and detection. For the past decades, remote sensing data and imagery and AI, ML, and DL have been used to reduce human interference and assumption to accurately predict, detect, and monitor landslide susceptibility to provide accurate information to governmental authorities and agencies to manage and control landslide disasters. In this section, we explore the application of remote sensing data and imagery with the learning algorithms grouped into landslide susceptibility mapping, decisions, predictions, inventory and deformation, extraction, and management.

4.2.1. Landslide Susceptibility Mapping

Landslide susceptibility mapping is an indispensable prerequisite for reliable geo-hazard risk analysis, management, and control [85]. The application of learning algorithms, models, remote sensing datasets, imagery, and landslide environmental factors, including climatic, geological, tectonic, and human engineering activities, have been studied [23,85,91,92].

Table 3. Summary of studies on landslides susceptibility mapping.

No. of Landslide Factors	No. of Landslide Maps	Source	Algorithms	Results/Metric	Reference
9	226	UAV, DEM	GA–SVM (best) and SVM	OA = 0.959, Kappa = 0.957	Niu et al. [63]
14	82	Surface area ratio (SAR), DEM, Landsat/ETM+ satellite	SVM (best)	Polynomial degree-3 with AUC of 0.88	Pourghasemi et al. [93]
12	1	Landsat, DEM	SVM15% (best), DT, LR	Kappa-index15% = 0.57, AUC15% = 0.79, FP15% = 0.4	Marjanović et al. [13]
11	125	QuickBird, SRTM, DEM, Enhanced thematic mapper plus ETM+	RF, BRT (best), CART, and GLM	AUC = 0.958, Precision rate = 0.86	Youssef et al. [29]
9	639	DEM, Landsat-8	SOM–ELM (best), SOM–SVM, and single ELM	Prediction rate 0.867, FPR = 10.2%	Huang et al. [2]
8	N/M	ASTER–GDEM	SVM, PSVM, and L2–SVM–MFN (best)	AUC = 0.83, AA = 0.72, TPR = 0.93, TNR = 0.72	Kumar et al. [30]
17	N/M	Ultracam-X, DEM	RF and Boosted Tree (best)	Landslide susceptibility index rank-regression = 84.87 and classification = 85.98%	Kim et al. [14]
16	249	GEE, ALOS–PALSAR, DEM, Sentinal-2, Landsat-8	ADTree, FLDA, and RF (best)	Susceptibility zone of = 11.35%, AUC = 0.97	Arabameri et al. [49]
20	151	DEM	Multi-layer perceptron (MLP)	Sensitivity = 82.61%, accuracy = 80.43%, PPV = 79.17%, NPV = 81.82%, Kappa = 0.61 and AUC = 0.88	Lee et al. [24]
17	152	InSAR, GEI, Sentinel-2, Sentinel-1, and Landsat-8	AB (best), ADTree, and AB–ADTree	AUC = 0.96	Nhu et al. [11]
17	544	DEM, GEE	Support Vector Classification (SVC)	Accuracy of 0.91 and AUC = 0.96	Qing et al. [94]
13	210	SRTM, Landsat-7 and Landsat-8, GEE	LR (best), SE, WoE, and FR	AUC = 85.3%, precision = 83.1%	Ahmad et al. [95]
11	35	LiDAR, RIEGL, DEM	ANN with Ohe-X	AUC = 0.93 with an improvement rate of 37.15%	Al-Najjar et al. [96]
19	616	Resources satellite three (ZY-3), ASTER GDEM	ANN, LR, and SVM (best)	AUC = 0.93 with accuracy = 83.86%	Xie et al. [55]
16	7184	DEM, Geospatial Data Cloud Platform	ACO–DBN	Accuracy = 93%, precision = 95%, sensitivity = 91% and F1-score = 0.93	Xiong et al. [60]
12	7556	Landsat-8, DEM, ALOS, GEE	SE–CapNet (best), CapNet, CNN, and RF	Accuracy = 96%, precision = 97%, sensitive = 95% and AUC = 0.97	Chang et al. [97]
10	13,886	ALOS	BO_GBT, BO_RF, SVM, and CNN_BO_GBT (best)	Accuracy = 0.79, F1-score = 0.84, Kappa = 0.67 and precision = 0.82	Gao et al. [26]
13	1290	GEE, DEM	CNN, DNN (best), LSTM, and RNN	F1-score = 0.82 and AUC = 0.873	Habumugisha et al. [61]
12	413	DEM, Landsat-8 OLI	DRNN	AUC = 0.96 (train and tested), Kappa = 0.829	Huang et al. [98]
12	235	ASTER GDEM, Landsat-8, DEM	SVM and RF (best)	AUC > 0.80 for landslide aggregate, neighborhood factor-based	Huang et al. [70]
12	400	SRTM, Landsat-8 OLI, DEM	RF, SVM, maxENT, GBM, and LR	SVM, RF, and GBM demonstrated the best results, but the AUC of SVM = 0.969 and performance overall (POA = 2669)	Shahzad et al. [99]
9	213	ALOS–PALSAR, DEM	FSP, FES, and ANN	AUC > 0.80	Sweta et al. [89]
19	59	SBAS–InSAR, Setinel-2, Setinel-1, SRTM DEM, GEE	XGBoost	AUC = 0.996, accuracy = 97.98%, TPR = 98.77%, F1-score = 0.98, and Kappa = 0.96	Wang et al. [100]
13	1081	ASTER GDEM, Landsat-8 OLI, DEM	CF–RF	POA = 257.046, AUC = 0.946	Yuan et al. [39]
12	233	GEE, Landsat-8 satellite, ASTER GDEM	WLR, WLightGBM, and WRF (best)	Balance accuracy = 0.84, G-mean = 0.84, recall = 0.90, accuracy = 0.79 and AUC = 0.91	Zhang et al. [101]
11	405	GDEMV2, Landsat-8 OLI, DEM	CNN	Precision = 0.99, recall = 0.91, F1-score = 0.94, MCC = 0.77, SRC-AUC = 93.14% and PRC-AUC = 91.81%	Zhang et al. [102]
16	256	DEM, Geographic Remote sensing Ecological network	SVM, LR, RF, XGBoost (best), and LDA	AUC = 0.876	Cao et al. [103]
15	235	ASTER GDEM, Landsat-8 TM	Slope-RF and Slope-MLP	Both models performed equally with AUC = 0.895	Chang et al. [104]
14	293	Landsat-8, SRTM DEM, GEE	SVM, RF, and GAMI-net (best)	AUC = 0.94, accuracy = 0.87, precision = 0.86, F1-score = 0.87, and recall = 0.87	Fang et al. [76]
17	85	UAV, GEI, ALOS–PALSAR, TERRASAR-X	ANN	AUC = 0.965 and precision-recall = 0.976	Masruroh et al. [51]
12	30	Landsat-8, DEM	ANN, boosting-C5.0 DT, and SVM	AUC = 0.99	Miao et al. [105]
22	86	GEE, Landsat-9, DEM, JAXA ALOSWorld 3D	RF and XGBoost	Both models achieved AUC = 0.96	Parra et al. [106]
13	2308	DEM, Landsat-8, ALOS	LR, RF (best), NB, and MLP	AUC = 0.92, susceptibility = 12.24%	Wang et al. [107]
13	302	Sentinel-2, Landsat-8, SRTM, DEM, GEI	RF, EXT–XGBoost (best), and KNN	AUC = 0.97, and all models have an accuracy rate > 90%	Abbas et al. [108]
12	1215	DEM, Maps	SVM, LR, LDA, RF, and XGBoost	AUC > 90% for all the models	Agboola et al. [83]
15	811	DEM	ABC–ANFIS	Precision = 0.89, AUC = 0.96, RMSE = 0.21, MAE = 0.21, and Kappa = 0.70	Chen et al. [109]
16	337	ASTER GDEM, Landsat-8	MLP and RF	AUC > 0.9 for both models	Huang et al. [110]
16	370	DEM, Landsat-8 TM	SVM, MLP, and RF (best)	AUC > 0.90, but the accuracy decreases with increasing random error	Huang et al. [111]
11	407	ASTER V2 and V3, COP-DEM, ALOS, FABDEM, SRTM	KNN and RF (best)	AUC > 0.80 and 0.95 COP	Lu et al. [112]
14	945	DEM, Landsat-8	DNN and DNN-hybrid models (DNN–MPA) (best)	AUC = 0.96, with accuracy > 90%	Nguyen et al. [56]
14	170	GEI, DEM, Sentinel-2	SVM	AUC = 0.88, recall = 0.73, precision = 0.86, accuracy = 0.81, and F1-score = 0.79	Patil et al. [84]
16	489	MERIT DEM, SRTM DEM, GEE, MODIS	SVMSMOTE	Accuracy = 96%, sensitivity = 97% and MCC = 0.92	Sharma et al. [41]
12	12	Landsat-8 OLI, DEM, GEE, ASTER GDEM	LGBM (best), GRU, LSTM, RNN, RF, and ET	F1-score = 0.56, recall = 0.52, precision = 0.60, AUC = 0.84, and accuracy = 0.82	Song et al. [42]
15	1718	Landsat-8 OLI, DEM	RF (best), and XGBoost	AUC > 0.80	Sun et al. [113]
13	78,529	MT–InSAR, DEM	Stacking RF-XGBoost	AUC = 0.96	Zeng et al. [114]
9	202	GEE, Landsat-5 TM, Landsat-8 OLI	CTLGNet	AUC = 0.97 and recall = 0.98	Zhao et al. [115]

Note: The learning algorithm with the highest susceptibility mapping is indicated as “best” in this Table.

summarizes critical studies conducted on landslide susceptibility mapping using different remote sensing techniques, AI, ML, and DL algorithms.

Figure 10 shows various learning algorithms and the accuracy of remote sensing data sources. The learning algorithms provide a high accuracy (0.76 to 1.0) level for landslide susceptibility mapping. It is also clear from this figure that the data source and landslide environmental factors play a crucial role in obtaining good accuracy for landslide susceptibility mapping. For example, using ASTER GDEM and Landsat-7 datasets, Hong et al. [15] achieved an accuracy of 0.87, while Hussain et al. [71] achieved an accuracy of 0.84 utilizing AOS–PALSAR, DEM, Sentinel-2, and Landsat-8 and Zhou et al. [116] also used MT-InSAR to be obtained prediction accuracy of 0.92. In contrast, Sun et al. [52] achieved an accuracy of 1.0 using DEM and ASTER GDEM datasets, demonstrating the significance of data sources. Overall, considering all the learning algorithms, there is not much difference in the accuracy level. However, they give a reasonable accuracy for landslide susceptibility mapping in geological terrains prone to landslides.

A few key studies that utilized different techniques and learning algorithms conducted for different landslide locations are briefly discussed below, with a focus on the effectiveness and accuracy of each method.

(1)

Marjanovic et al. [13] assessed the landslide susceptibility in Fruška Gora Mountain, Serbia, using remote sensing imagery derived from the digital elevation model (DEM); landslide environmental factors such as elevation, aspect, lithology, topographic wetness index; and three different machine learning algorithms, support vector machine (SVM), decision tree (DT), and logistics regression (LR). The results from unbalanced and balanced training modeling, the area under the curve (AUC) for SVM, provide the best results for landslide susceptibility mapping.

(2)

Stumpf et al. [83] combined the object orientation image analysis with varieties of remote sensing datasets obtained from Geoeye-1, IKONOS, and Quickbird and obtained landslide susceptibility mapping accuracies ranging from 73% to 87% for four different landslide sites in China, France, Haiti, and Italy.

(3)

Peethambaran et al. [23] compared the fuzzy expert system (FES) and extreme learning machine (ELM) using remote sensing images from Alos Pulsar DEM (15 m resolution), satellite images (LISS-III and Radarsat), and eight causative factors of several landslides in Uttarakhand, India. It was found that the FES outperformed the ELM with an AUC of 0.84 compared to 0.80.

(4)

Qi et al. [17] analyzed 14,397 shallow landslides in Southern Tianshui, China, using 18 geomorphic concentration factors, including slope aspect, geological conditions, and human activities. Their results indicated that out of 21 ML models developed, the ExtraTrees model produced better outcomes, with an AUC of 0.91. It was also found that the slope terrain aspect was the most significant contributor to landslide susceptibility compared to the other factors.

(5)

Yin et al. [79] explored remote sensing data from the Qinghai-Tibet Plateau to map common landslides based on environmental factors such as terrain and climate. The ML demonstrated that climate conditions, such as summer temperature and rainfall, were the main contributing factors causing landslides.

(6)

Zeng and Chen [74] investigated an unstable slope using DEM-derived data and established artificial neural networks (ANNs) and GIS-based techniques with relevant environmental factors to map landslide susceptibility in Enshi, China. Their proposed models successfully predicted the unstable slope location based on AI theory and GIS. Later, Zeng et al. [136] proposed a new model for environmental heterogeneity for Fengjie County and Fulin District in Chongqing, China, which was based on a graph neural network with environmental consistency (GNN–EC) and compared it with three different learning models such as SVM, ANN, and random forest (RF). The investigators reported a landslide susceptibility accuracy higher than 0.80, a precision of 0.80, a recall of 0.91, an F-measure of 0.81, and an AUC of 0.87, indicating a strong capability of mapping landslide susceptibility compared to the other three learning models.

(7)

An innovative susceptibility-guided (SG) landslide-mapping method based on a fully convolutional neural network (FCNN) and mean changing the magnitude of objects (MCMO) was proposed by Chen et al. [137] to improve the landslide hazard mapping in Lantau Island, Hong Kong, China. This method produced better outcomes with a precision, recall, and F1-score between 80% and 95% and an intersection over union (IoU) value between 70% and 85%, indicating a significant reduction in false and landslide detections compared to the traditional FCNN.

(8)

Garcia et al. [92] proposed a semi-automatic detection of relict landslides in the Serra do mountain range, Southeastern Brazil, based on a deep learning CNN framework (Unet, FPN, and Linknet). The k-means cluster algorithm was used for pre-training, and weights were computed to fine-tune the training process in the CNN. Forty-two CNN tests of two datasets were performed, achieving recall values higher than 75% and a precision of less than 20%. It was found that the model predicted relict landslides accurately but was limited to detecting landslides in the terrains covered with rainforests.

(9)

Zhang et al. [138] proposed a prototype-guided domain-aware progressive representation learning (PG-DPRL) technique to map cross-domain landslides based on the multitarget domain adaptation (MTDA) method. The model adopted a near-to-far adaptation strategy, reliable domain-specific pseudo-label, and cross-domain sharing boundary decision. A category-wise representative alignment was performed to enhance the discriminative capacity of the model by using the Wasserstein distance metric and cross-domain prototype consistency loss. Using landslide datasets from Hokkaido, Japan, the study demonstrated that cross-domain landslides could be successfully mapped using the PG-DPRL model and global high-resolution landslide datasets, and the proposed model outperformed many traditional deep learning algorithms in comparison.

(10)

Ouyang et al. [139] developed a model (PU-pullbggingDT) algorithm to cater for the limitations of the positive unlabeled (PU) algorithm in mapping landslide susceptibility and prediction. The model uses the global shuttle radar topography mission (SRTM), Google Earth Engine (GEE) datasets, and 25 environmental factors associated with landslides in Zigui County, China. The proposed PU-pullbggingDT model outperformed the existing SVM, DT, LR, AdaBoost, XGBoost, and PU-learning, indicating a strong capability of accurately predicting landslides in complex geological terrains.

(11)

Zhao et al. [20] combined the CNN and transformer model to develop CNN–transformer local–global extraction network features (CTLGNet) to investigate landslide hazards in the Three Gorges Reservoir and Jiuzhaigou areas, China. Considering nine landslide environmental factors, the CTLGNet performed better than other methods such as the CNN, residual neural network (ResNet), densely connected convolutional network (DenseNet), vision transformer (ViT), and fractional Fourier image transformer (FrIT).

In summary, remote sensing techniques combined with traditional and relatively newly developed learning algorithms have been effective in landslide hazard mapping. Based on the metrics measurement and statistical analysis tools such as the Kappa index, F1-score, and area under the curve (AUC), the models demonstrated a high level of accuracy, ranging from 0.56 to almost 1. The support vector machine (SVM), convolutional neural network (CNN), and random forest (RF) have been widely used with remote sensing techniques such as UAV, Sentinel-1 and Sentinel-2, Landsat-8 ILO, and ALOS–PALSAR GDEM in landslide susceptibility mapping. The extreme learning model (ELM) could not map landslides, especially when data were insufficient for training and testing. A semi-automatic model based on CNN was reported to predict relict landslides accurately but was limited to detecting landslides in forest areas.

4.2.2. Landslide Detection

AI, DL, and ML have changed the field of remote sensing in dealing with landslide detection challenges in complex geological terrain [12,22,32,140,141].

Table 4. Summary of studies on landslide detection using AI, DL, and ML with remote sensing.

Data Used	Algorithms Evaluated	Some Key Results	Reference
DEM	Deep convolution Neural Network (DCNN)	The DCNN model demonstrates a high capacity for detecting landslides in mountainous terrain.	Chen et al. [22]
Unmanned aerial vehicles (UAVs), DEM	CNN models	The CNN model shows a precision equivalent to 0.90, an F-score of 0.85, and mIoU of 0.74; the effectiveness of CNN for slope or landslide detection depends strongly on its design. CNN proves to be a powerful tool, but it depends on trial and error.	Ghorbanzadeh et al. [78]
DEM, satellite images	ResNet-50	The ResNet-50 achieved a precision as high as 0.987, a recall of 0.9455, an accuracy of 0.9816, and an F1-score of 0.9662, indicating a successful prospect of landslide detection with satellite images.	Ji et al. [59]
GEE	You only look once-small attention (YOLO-SA)	Comparing the YOLO-SA model to the other 11 models, the YOLO-SA has an accuracy of 0.94 and an F1-score of 0.91, showing a potential landslide detection.	Cheng et al. [142]
Sentinel-2A, DEM, ALOS–PALSAR, GEE	convolutional auto-encoder (CAE)	The results show that CAE with a mini-batch K-means clustering algorithm can be applied to primary landslide mapping.	Shahabi et al. [87]
Zeiss RMK TOP 15 aerial survey, DEM	Object-oriented change detection CNN (CDCNN)	The CDCNN shows a better and more robust detection of landslides in vast areas, even with a considerable volume of datasets and complex land cover, requiring less human interaction.	Shi et al. [143]
Landsat-8 and JAXA ALOS DSM and (GEE)	SVM, CART, minimum distance, RF and NB	The SVM and RF models achieved the best detection results with a true positive ratio of 87.5% compared to others.	Singh et al. [144]
UAV images	ResNet-50 and ResNet-101 by exploring Mask R-CNN	With the incorporation of Mask R-CNN, the ResNet-101 performed better than ResNet-50 with a precision of 1.0, recall of 0.93, and F1-score of 0.97, indicating good landslide detection.	Ullo et al. [145]
PlanetScope, RapidEye, Sentinel-2, ASTER, GDEM and DEM	ANN, RF, and SVM	From the studies, the results from all three models using remote sensing datasets provided high accuracies of 0.94, sensitivity of 0.96, and specificity of 0.92, demonstrating the models’ robustness in accurately detecting landslides.	Das et al. [53]
Sentinel-2, ALOS DEM	ResU-Net, OBIA, and ResU-Net-OBIA	According to the proposed model ResU-Net-OBIA, the model detects landslide with a precision of 0.73, recall of 0.80, and F1-score of 0.77	Ghorbanzdeh et al. [18]
GEE, Sentinel-2, Landsat-8	Transfer learning Mask R-CNN (TL-Mask R-CNN)	The result shows that the TL-Mask R-CNN model can detect landslides with a recall of 0.78 and an F1-score of 0.80, demonstrating a solid detection and segmentation along Sichuan–Tibet transportation and landslide recognition along the Sichuan–Tibet transportation corridors.	Jiang et al. [146]
Sentinel-2, GEE, DEM and ALOS PALSAR	DeepLab-v2, DeepLab-v3+, FCN-8s, LinkNet, FRRN-A, FRRN-B, SQNet, U-Net, and ResU-Net.	The ResU-Net performed better for detecting landslides than all the other models after all the models were trained from scratch on patches.	Ghorbanzdeh et al. [147]
RapidEye, DEM and ALOS–PALSAR	U-Net, SVM, KNN and RF	The U-Net model achieved slightly better results than the other learning algorithms; however, it is still at the preliminary stage of detecting landslides.	Meena et al. [148]
Airborne remote sensing image	HRNet, DeepLabV3, Attention-UNet, U2Net, FastSCNN and SegFormer	Comparing the experimental results, the SegFormer model enhances landslide detection accuracy of the mIoU by 2.2%, IoU by 5%, and F1-score by 3% and reduces the pixel-wise classification error ratio by 14%	Tang et al. [149]
UAV	Transfer learning (TL) model	The result from UAV remote sensing images with a proposed TL model demonstrated a better detection of landslides when accurate landslide data are used.	Yang et al. [50]
TripleSat satellite and the RGB images	U-Net, DeepLab v3+ and PSPNet	After experimenting with the three models with different backbone networks, PSNet with ResNet-50 demonstrated landslide detection with an mIoU of 91%, a recall of 97%, and a precision of 94%	Yang et al. [150]
RapidEye, Geoeye-1, GEE, TripleSat, Sentinel-2 and GaoFen-2	HADeenNet	With ResNet50 as the backbone, HADeenNet modes built with the DL framework increase the F1-score by about 21% and 10% higher than six traditional deep learning models.	Yu et al. [151]
TripleSat	U-Net with ResNet-50, ResNet-101, VGG-19, and DenseNet-121 as backbone	Comparing the U-Net model with the four backbones, the U-Net + ResNet50 demonstrates better landslide detection by recording the highest precision value of 0.98, a recall of 0.98, an F1-score of 0.98, an overall accuracy (OA) of 1.0, and an MCC of 0.88. Although, the other three models show measured accuracy values > 0.90.	Chandra et al. [152]
Google Earth Pro	SG-FCNN + MCMO	The results from the study show that the SG-FCNN+MCMO can detect landslides and reduce the false and miss detection of landslides compared to the traditional FCNN model.	Chen et al. [137]
Sentinel-2, and ALOS PALSAR DEMs	U-Net	The U-Net model demonstrates a promise for case event inventory but shows lower detection accuracy for geomorphology inventory.	Das et al. [80]
CBERS-04A and remote sensing	CNN, Unet, FPN and Linknet	The models show recall values similar for all (>75%) with precision values < 20%, while fewer precision values were attributed to the false positive sample. Additionally, the models cannot detect relict landslides in terrains with rainforests.	Garcia et al. [92]
Optically sensing	Faster R-CNN with VGG16 and ResNet50 as backbone	The Faster R-CNN+ResNet50 performed better than the Faster R-CNN+VGG16 with an AP of about 91%, an F1-score of 94%, a recall of 91%, and a precision of 98% and can detect slope failure and landslides, while R-CNN+VGG16 can detect small-slope failure and landslides.	Guan et al. [153]
LiDAR and optical images, DEM	ResUNet, LandsNet, HRNet, MLP, SegFormer, and proposed (DemDet)	The DemDet model has a mean accuracy of 0.95, mIoU of 0.61, and F1-score of 0.78, demonstrating the capability of the model to detect landslide in a forest cover compared to the ResUNet, LandsNet, HRNet, MLP, and SegFormer models	Li et al. [154]
Shipborne images	Transfer learning models and VGG19, DenseNet121, EfficientNetB0, SEResNet50, and ViT	The result shows that the decision-level fusion and transfer learning integration can classify landslide-based shipborne images.	Li et al. [19]
TripleSat satellite	Vision transformer (ViT) models	Comparing the ViT model with the traditional deep learning models, the ViT model has the benefit of detecting landslides with remote-sensing images.	Lv et al. [155]
Landsat-8, GEE	Mask RCNN + weight transfer function + Mask IoU model	The proposed model precision is enhanced by 20% compared to Mask R-CNN, and the average accuracy higher than 0.75 achieved for the proposed mask IoU threshold is 0.5.	Wang et al. [156]
Sentinel-2 Level 1C, DEM and ALOS DEM	Semantic segmentation network (EGCN)	The EGCN model outperformed traditional major deep learning models by obtaining an OA of 0.998, an mIoU of 0.997, a Kappa of 0.973, an F1-score of 0.974, a precision of 0.973, and a recall of 0.974, indicating the model’s capability of recognized landslides.	Yang et al. [157]
Synthetic aperture radar (SAR), Sentinel-1A, Sentinel-2A and DEM	Multi-input channel U-Net	The U-Net with multi-input channel (16) demonstrates a good prospect for detecting landslide with a recall value of 69.74%, precision of 62.96%, F1-score of 66.18, and an mIoU of 74.67%	Chen et al. [158]
Sentinel-2	U-Net, U2-Net, and U-Net3+	The U-Net models applied provided the best results for detecting traces of landslides with remote sensing data.	Dang et al. [159]

summarizes some studies on landslide detection utilizing remote sensing datasets and imagery and learning algorithms. A few key studies are discussed below to demonstrate recent advances in this field.

(1)

Ghorbanzadeh et al. [32] used high-resolution optical satellite data from RapidEye, 20 landslide environmental condition factors, and maps combined with the the ANN, different CNN, SVM, and RF models to detect landslide-hazard-prone areas along the Ganga River in the Himalayas, northern India. This study showed that the CNN model achieved a mean intersection over union (mIoU) of 0.782 with insignificant window size among other models, indicating the model’s ability to detect landslides.

(2)

Similarly, Hu et al. [54] utilized a cross-validation method and nighttime imagery, Landsat-8, ASTER DEM, and Visible Infrared Imaging Radiometer Suite (VIIRS) from the disaster-prone Jiuzhaigou region with ANN, RF, and SVM considering 30 landslide condition factors. The investigators noticed that the precision and accuracy of landslide identification were reduced for the three models, especially for ANN and SVM. They attributed the accuracy reduction to the “users” and “producers” imbalance accuracy. However, the models can be improved to detect landslides through the cross-application probability.

(3)

Sameen et al. [140] designed a residual network (ResNet) for landslide detection in the Cameron Highlands, Malaysia, utilizing altitude, slope, aspect, and curvature. The DEM and light detection and ranging (LiDAR) were used to create spatial data. The ResNet was compared to traditional deep learning models like CNN, and the results showed that the CNN model enhanced landslide detection with the fusion method. However, the ResNet model outperformed the CNN with the same fusion method and showed convergence to the test area with an overall training and validation accuracy and F1-score between 0.80 and 0. 95 and an mIoU of 90.24%.

(4)

Nhu et al. [12] adopted LR, logistic model tree (LMT), and RF models to analyze 152 landslide sites in the Cameron Highlands, Malaysia, utilizing Google Earth Images (GEI), interferometry synthetic aperture radar (InSAR), DEM, and field datasets. Seventeen landslide environmental condition factors were considered to detect landslides using the machine learning algorithms with different accuracy measurement statistical tools, including sensitivity, AUC, accuracy, and root-mean-square error (RMSE). This study demonstrated an AUC of 0.92 for LMT, 0.90 for LR, and 0.88 for RF, showing that the datasets from remote sensors and learning algorithms such as LMT could detect landslides.

(5)

Cai et al. [160] reported that the DenseNet model reached a Kappa value of about 0.95 and an F1-score of 0.95 after the data augmentation and fine-tuning technique, indicating the model’s applicability to landslide detection. This research around the Three Gorges Reservoir, China, utilized datasets from the ZY-3 satellite, global digital elevation (GDEM), and Landsat-8 with 12 topographic, geological, hydrological, and land cover factors. However, it was noted that the model had poor micro-landslide detection and needed improvement.

(6)

Huang et al. [161] demonstrated that incorporating Landsat-8 OLI, advanced spaceborne thermal emission and reflection radiometer, and global digital elevation model (ASTER GDEM) data with nine geospatial landslide factors could benefit a landslide detection analysis. The investigators developed a Distilled Swin–Transformer (DST) model with Swin–Transformer as a backbone to eliminate the long runtime and insufficient model challenges facing ML and DL. They tested this model to detect landslides in Zigui County, Hubei Province, China. In a comparison of the quantitative results with ResNet, data-efficient image transformers (DeiT), and the proposed DST model, the DST model achieved a higher landslide detection accuracy of 98.17%, a precision of 98.17%, a recall of 98.16%, an F1-score of 98%, and a Kappa index of 0.977, with less floating-point operation (FLOP) for landslide detection.

(7)

Li et al. [88] indicated that an image-based data-driven framework with satellite images could benefit landslide detection. The proposed approach comprised two essential phases: firstly, an object detection algorithm, thus (Faster-RNN) trained within large-scale satellite imagery and proposed bounding box for each landslide visualization; and secondly, the bounding box location information was utilized to crop the satellite images and boundary detection algorithms proposed to identify each loess landslide dissection performance. A total of 150 loess landslides from northern China were reviewed to validate the effectiveness of the two frameworks. The results showed that the frameworks could accurately detect the segmentation of loess landslides.

(8)

To eliminate the anomalous deformation area in InSAR data and human interactions with acceptable criteria, Zhang et al. [73] developed two-stage deep learning networks (InSARNet). The model was utilized to identify anomalous deformation terrains in Maoxian County, China. From the quantitative analysis with different measurement metrics, the InSARNet model performed better than the commonly used learning algorithms in detecting landslide anomalous deformation areas. However, the model was recommended to have future improvements.

(9)

Sreelakshmi et al. [162] tackled the challenges of detecting landslides in rainforest vegetation in Bijie City, China, using deep-learning frameworks with visual saliency and high-resolution remote sensing imagery. The saliency feature maps were employed to fine-tune the U-Net model on the landslide data. The results showed that the proposed model achieved a landslide detection accuracy of 94% compared to the commonly used deep-learning algorithms.

In summary, remote sensing datasets combined with learning algorithms like SVM, RF, ANN, Mask R-CNN, and ResNet-101 could effectively detect landslides, producing an accuracy ranging from 70% to almost 100%. A few challenges were reported as well: (a) the CNN model strongly depended on its design framework, and (b) the accuracy of SVM and ANN models depended on the user and producer interactions.

4.2.3. Prediction of Landslide Offset

The offset of landslide forecasting and early warning is essential to geological hazard analysis. Many studies have recently utilized remote-sensing data and imagery with AI, ML, and DL algorithms to predict landslide deformation [27,64,163,164].

Table 5. Summary of studies on the prediction of landslide deformation.

Landslide Factors	Data Used	Algorithms Evaluated	Results/Metric	Reference
Topography, hydrological, geological, land cover	Landsat TM8, DEM	FC–SAE, SVM and BPNN	Prediction rate and total accuracy of FC–SAE (0.85), SVM (0.82 and 0.81), and BPNN (0.82 and 0.81); the FC–SAE algorithms proposed predict the landslide susceptibility better compared to the other models.	Huang et al. [33]
Slope, aspect, elevation, lithology, rainfall, curvature	GEI, Landsat-8, DEM	MBNBT, MLPN, SVM, and NBT	AUC of MBNBT (0.82), MLPN (0.80), SVM (0.80), and NBT (0.80); MBNBT offers better prediction than the other learning algorithms	Nguyen et al. [27]
TWI, NDVI, elevation, slope, aspect, curvature	GEE, Landsat-8, ASTER-GDEM	LR–GBDT–VFI	The proposed LR–GBDT–VFI model (AUC: 0.98) best predicts landslide potential after comparing the model with nine traditional models.	Sachdeva et al. [165]
Slope, lithology, aspect, total curvature, STI	ALOS PALSAR, DEM	GAN and SMOTE for correcting data imbalanced with ANN, RF, DT, KNN and SVM	The results from the study show that landslide data balancing could affect the prediction capacity of learning algorithms	AI-Najjar et al. [68]
Landslide, terrain, lithology, land cover, NDVI, landform and precipitation	DEM, Landsat, STS Endeavour OV-105, Crowdsourcing	Multi-graded Cascade Forest, thus, (GCF) algorithm	The GCF model shows a precision efficiency of 0.93 for landslide prediction, and the workflow alleviated poor prediction problems from limited landslide data.	Chen et al. [166]
Rainfall, slope, curvature, TWI, soil texture, lineament density, NDVI	ALOS PALSAR, DEM, Sentinel-2 MSI,	RF and DNN	The RF and DNN models achieved higher accuracy of AUC greater than 0.90, which indicates a high prediction of landslide susceptibility, but DNN outperformed the RF with an AUC of 0.96. Rainfall was noticed to be a significant contributing factor to landslides, according to the results	Dahim et al. [40]
Aspect, elevation, lithology, distance to a river	ASTER, DEM, Landsat-8	MLP, RF, and proposed semi-supervised machine learning (SSML); MLP, RF, SVM, and low-pass filter method	The proposed SSML model reduced landslide position errors and significantly improved landslide accuracy compared to MLP and RF. The random errors in landslide condition factors result in higher uncertainty, and the low-pass filter method reduced random error significantly.	Huang et al. [110], Huang et al. [111]
Slope, aspect, elevation, rainfall, TWI, NDVI, land cover	Global Digital Elevation Model (GDEM), ALOSPALSAR and Landsat 8	RF, XGBoost, NB, ANN, KNN and FR	The AUC of RF (0.992), XGBoost (0.991), NB (0.970), ANN (0.922), KNN (0.877), and FR (0.674), indicating that the RF model provided effective means of predicting landslide on a global scale, followed by XGBoost and ANN comparing the models.	Khalil et al. [69]
Slope, aspect, elevation, sediment transport index, lithology, fault proximity, TWI, NDVI	Resources-1 multispectral LISS-IV, IRS-P5 Cartosat-1, Landsat-7, ALOS–PALSAR, DEM, TRMM, and Google Earth Engine	Multilayer Perceptron (MLP)	The success and precision rate curves under AUC are 0.94 and 0.92, respectively, indicating high prediction accuracy of MLP for forecasting landslide potentials.	Sundriya et al. [167]
Length-slope, topography, vegetation cover, lithology, rainfall, TWI	DEM	Landslide density-based bidirectional long short-term memory (LD–BiLSTM) model compared with Landslide object model (LO–BiLSTM model)	The precision, recall F1-score, and AUC are 0.90, 0.89, 0.90, and 0.94 for LD–BiLSTM, respectively, and 0.81, 0.82, 0.81, and 0.91 for LO–BiLSTM, respectively, making LD–BiLSTM model superior to predict landslide compared to LO–BiLSTM.	Wang et al. [62]

shows some studies conducted to analyze landslide susceptibility predictions. Other studies have also elaborated on landslide predictions with remote sensing datasets from different landslide sites and learning paradigms.

(1)

Tien et al. [64] employed a hybrid intelligence model of the least square support vector machine (LSSVM) and artificial bee colony (ABC) optimization called LSSVM–BC to study landslides in the Lao Cai area, Vietnam. Using DEM, 10 landslide environmental influence factors, and 340 landslides, the computational and quantitative analysis showed that the hybrid intelligent model LSSVM–BC could predict landslide offset relatively well compared to the SVM learning algorithm.

(2)

Van Natijne et al. [162] used remote sensing data and an ML algorithm to predict landslide deformation. This study investigated the deep-seated Vogelsberg landslide near Innsbruck, Austria, using a long short-term memory (LSTM) model. The investigators pointed out that the LSTM model produced a positive outcome for landslide deformation prediction; however, the ML technique was rather complex.

(3)

Chen et al. [164] developed a neural disaster emergency ontology (NADE) schematic layer based on a constructed knowledge graph and different environmental factors, including geology, landform, soil, climate, vegetation, and transportation. The proposed knowledge graph embedding (KGE)-based model was applied to generate landslide prediction in Xiji County, Ningxia Province, China. The model was trained with 741 landslide records from DEM and SRTM. The findings demonstrated an F1-score improvement of 5% with the complete data and 17% with the reserve data.

(4)

Zhou et al. [168] combined the physics-based and economical landslide displacement prediction framework with a multi-temporal interferometric synthetic aperture radar (MT-InSAR) and ML method such as gate recurrent units (GRU) to predict the non-linear tendency and time displacement of landslides. In the Three Gorges Reservoir area, China. The GRU model outperformed the traditional ML algorithms, including LSTM, Kernel-based extreme learning machine (KELM), and the Adam algorithm.

(5)

He et al. [25] proposed a new method involving photogrammetric and aerial LiDAR surveys and GIS-based kinematic evaluation to analyze the rock slope conditions of the North Coast of Cornwall, UK. The ML models, such as SVM, RF, MLP, and deep learning neural networks (DLNNs), and the discontinuities factors were applied. A frequency analysis (FR) produced better outcomes, suggesting that the inclusion of geological discontinuities into learning algorithms could enhance the prediction accuracy.

(6)

Maxwell et al. [169] proposed an explainable boosting machine (EBM) with a generalized additive model (GAM) as an empirical method to better predict landslides in West Virginia, USA. The EBM model was compared with standard ML models such as LR, RF, KNN, and SVM, producing an overall AUC of 0.90, which indicated that the EBM model provided a relatively high accuracy of landslide prediction

(7)

Huang et al. [170] considered 11 landslide environmental factors and developed four ML models, LR, MLP, SVM, and C5.0 DT, for Huichang County, China. The C5.0 DT model outperformed others, with an AUC of 0.94.

In summary, several studies demonstrated that the remote sensing methods combined with the learning algorithms could predict landslide occurrence with accuracy above the baseline of 0.50. Remote sensing techniques, such as ALOS–PALSAR GDEM, Sentinel-2 MSI, and Landsat-8 with learning algorithm models like XGBoost, gave a landslide prediction accuracy of 0.99, followed by the RF and DNN with an accuracy of 0.96.

4.2.4. Landslide Inventory and Monitoring

Landslide inventory mapping is essential for disaster and emergency rescue management [79,171]. Table 6 lists studies on landslide inventory mapping and monitoring, including the study area, source, and model evaluation. A few representative case studies are discussed below.

(1)

Lei et al. [171] developed a novel approach to detecting landslides in complex geological terrains and geospatial uncertainty based on a fully convolutional network-fusing pyramid pooling (FCN–PP). The proposed model utilized the Zeiss RMK TOP 15 aerial survey camera system to collect landslide data from Hong Kong, China. The results indicated that the FCN–PP model effectively mapped landslides with a precision of 96%, recall of 96.5%, overall error of 93%, F-score of 90%, and accuracy of 81.90%.

(2)

Althuwaynee et al. [172] developed a model incorporating 77 slope deformation factors. They used the t-distribution stochastic neighbor embedding (t-SNE) and the Apriori algorithm to recognize common relationships in the inventory maps with landslide factors. The model was validated using Landsat–TM and provided practical significance for landslide inventory mapping in Pohang state in South Korea.

(3)

Ramos-Bernal et al. [173] developed ASTER DEM-derived datasets with primary landslide factors and ML model KNN, stochastic gradient descent (SGD), linear Kernel–SVM, support vector machine radial basis function (SVM–RBF Kernel), and AdaBoost. A total of 671 landslide sites were analyzed, while 2/3 of them were trained, and 1/3 were used to obtain the inventory maps. The study showed that among the five models, AdaBoost recorded a precision, recall, F1-score, accuracy, and Kappa > 90%, with mean errors below 2%, indicating the model’s capacity to provide a landslide inventory map in Guerrero, Southern Mexico.

(4)

Chen et al. [174] proposed an algorithm based on multi-feature independent components analysis (MICUNet3+) for landslide inventory mapping with 3 landslide elevation factors. The model was compared with the UNet3+ and Unet3+-RFI using the co-seismic landslide datasets from Jiuzhaigou County, Sichuan Province, China. The MICUNet3+ model had a recall of 0.68, an F1-score of 0.70, and an mIoU of 0.76, indicating that the MICUNet3+ model performed better than the other two models and could assist with detecting landslides in complex geological terrains.

(5)

Chen et al. [79] adopted the full convolution networks with focus loss (FCN–FL) ML models such as SVM and RF to map landslides in Bijie City, Guizhou Province, China. The study applied a K-fold cross-validation training model (FCN–FLK) to enhance the data and model sturdiness. The model achieved an accuracy of 0.93, recall of 0.76, F1-score of 0.62, and an mIoU of 0.68, which indicated that it could solve the data imbalance in the landslide inventory mapping. However, the model was limited to the area with large vegetation coverage.

(6)

In recent years, some researchers have used remote sensing data and imagery from Sentinel-2, ALOS–PALSAR, DEM, Gaofen-1, and Gaofen-2 with learning paradigms including ANN, RF, CNN, DCN, and LSTM [58,73,91,175]. Sheng et al. [58] studied the implementation of a dynamic approach to landslide-related hazards in Shiyan City, China by integrating spatial–temporal likelihood analysis considering periodic ground deformation velocity derived from the MT-InSAR technique. The study employed ML and statistical models, including information quantity (IQ), FR, LR, BP–ANN, RBF–ANN, RF, SVM, and CNN, and the result indicated that the distance to a river and from structure, slope angle, and rocks were the main factors controlling landslide development. The ML models performed better than the statistical method. From the effectiveness, F1-score, sensitivity, and AUC calculations, the CNN recorded a high value greater than 0.90 and outperformed the other model in landslide monitoring and prediction.

(7)

Zhou et al. [176] proposed a model based on deep learning (AtmNet) to deal with the effect of topography and climatic conditions in the landslide-prone terrain of Mao County, Sichuan, China. The landslide-monitoring data were obtained from Sentinel-1 SAR. The data analysis showed that AtmNet provided a platform for DInSAR and InSAR to monitor landslides in mountainous terrains.

(8)

Liu et al. [177] utilized an advancing landslide detection method with the multi-period dataset for the landslide-prone area in Tuanjiecun. The small baseline subset interferometric synthetic aperture radar (SBAS–InSAR), GEE, Gaofen-1 (G-1), Gaofen-2 (G-2), Ziyuan-2 (ZY-3), UAV, and DEM were applied for dataset validation. Long short-term memory (LSTM) was used to evaluate and identify the landslide deformation from the SBAS–InSAR quantitatively. The results demonstrated that the LSTM could identify and monitor potential landslide deformations.

(9)

Senogles et al. [174] developed a method called “SlideSim” with an optical flow predictor for 3D landslide deformation utilizing DEM and UAV to monitor a slow-moving landslide. The landslide dataset from Southern Oregon Coast, USA was used to validate the model, and the results indicated that the “SlideSim” model required fewer intuitive parameters with no supervision to monitor landslide displacement successfully.

Table 6. Summary of studies on landslide inventory and monitoring.

Study Area	Source	Model Evaluation	Reference
Xi’an, China	Zeiss RMK TOP 15 Aerial Survey Camera System	Full convolutional network within pyramid pooling (FCN–PP)	Lei et al. [171]
Pohang, South Korea	GEE, Landsat-8, Enhanced thematic mapper (ETM+)	t-distributed stochastic neighbor embedding (t-SNE)	Althuwaynee et al. [172]
Guerrero, Mexico	ASTER, DEM	KNN, stochastic gradient descent (SGD), support vector machine (SVM linear Kernal), support vector machine radial basis function (SVM RBF Kernel), and AdaBoost	Ramos-Bernal et al. [173]
Guizhou, China	GEE, Sentinel-2	Full convolutional networks with focus loss (FCN–FL)	Chen et al. [79]
Sichuan, China	DEM, Sentinel-2, Landsat-8	UNet3+, MICUNet3+ and UNet3+-RFI	Chen et al. [174]
India	Sentinel-2, ALOS–PALSAR, DEM	ANN, RF, and analytical hierarchy process (AHP)	Nath et al. [91]
South Oregon Coast, USA	DEM, UAV	SlideSim	Senogles et al. [175]
Maoxian County, China	SBAS–InSAR, SRTM, DEM	InSARNet	Zhang et al. [73]
Tuanjiecun, China	SBAS–InSAR, GEE, Gaofen-1 (GF-1), Gaofen-2 (GF-2), Ziyuan-3 (ZY-3) and Unmanned Aerial Vehicle (UAV) and DEM	Long short-term memory (LSTM)	Liu et al. [177]
Shiyan, China	Multi-Temporal InSAR (MT–InSAR)	Information quantity (IQ), FR, LR, BP–ANN, RBF–ANN, RF, SVM, and CNN	Sheng et al. [58]
Sichuan, Ghana	InSAR, DEM, GEE, SRTM and Sentinel-1 SAR	AtmNet	Zhou et al. [176]

Table 6. Summary of studies on landslide inventory and monitoring.

Study Area	Source	Model Evaluation	Reference
Xi’an, China	Zeiss RMK TOP 15 Aerial Survey Camera System	Full convolutional network within pyramid pooling (FCN–PP)	Lei et al. [171]
Pohang, South Korea	GEE, Landsat-8, Enhanced thematic mapper (ETM+)	t-distributed stochastic neighbor embedding (t-SNE)	Althuwaynee et al. [172]
Guerrero, Mexico	ASTER, DEM	KNN, stochastic gradient descent (SGD), support vector machine (SVM linear Kernal), support vector machine radial basis function (SVM RBF Kernel), and AdaBoost	Ramos-Bernal et al. [173]
Guizhou, China	GEE, Sentinel-2	Full convolutional networks with focus loss (FCN–FL)	Chen et al. [79]
Sichuan, China	DEM, Sentinel-2, Landsat-8	UNet3+, MICUNet3+ and UNet3+-RFI	Chen et al. [174]
India	Sentinel-2, ALOS–PALSAR, DEM	ANN, RF, and analytical hierarchy process (AHP)	Nath et al. [91]
South Oregon Coast, USA	DEM, UAV	SlideSim	Senogles et al. [175]
Maoxian County, China	SBAS–InSAR, SRTM, DEM	InSARNet	Zhang et al. [73]
Tuanjiecun, China	SBAS–InSAR, GEE, Gaofen-1 (GF-1), Gaofen-2 (GF-2), Ziyuan-3 (ZY-3) and Unmanned Aerial Vehicle (UAV) and DEM	Long short-term memory (LSTM)	Liu et al. [177]
Shiyan, China	Multi-Temporal InSAR (MT–InSAR)	Information quantity (IQ), FR, LR, BP–ANN, RBF–ANN, RF, SVM, and CNN	Sheng et al. [58]
Sichuan, Ghana	InSAR, DEM, GEE, SRTM and Sentinel-1 SAR	AtmNet	Zhou et al. [176]

In summary, the learning models and remote sensing methods have proved effective for landslide inventory mapping and monitoring in mountainous terrain. For example, the full convolution network with focus loss (FCN–FLK) learning model was found to be limited to geological terrains with small vegetation cover, with an accuracy of 0.93, but the recall, F1-score, and mIoU values reduced due to the imbalance of landslide inventory mapping.

4.2.5. Assessment of Landslide Potential

Landslide vulnerability assessment refers to the likelihood of a landslide in a specific terrain under a complex geological environment and triggering conditions, which are critical to preventing and monitoring landslide risks. It is essential for landslide susceptibility assessment and disaster predictions in terrains prone to landslides [97,109,178,179]. More details about the application of learning algorithms in assessing landslide potentials are presented in

Table 7. Summary of studies on assessment of landslides potential.

Algorithms	Source	Objective	Landslide Factors	Accuracy Evaluation Methods	Some Key Result	Reference
LR, NB, and LB	DEM and other multiple-source	Landslide susceptibility assessment	Seventeen factors were used, including aspect, slope, topographic wetness index (TWI), lithology, land used/land cover	Area under curve (AUC)-LR = 84.2%, NB = 70.7% and LB = 85%	LR and LB reveal a reasonable accuracy for landslide susceptibility assessment than NB	Pourghasemi et al. [65]
ANN, GBM, and MaxEnt	InSAR, LiDAR	Slow movement landslide risk assessment	Slope angle, aspect, road density, TWI, TPI and stream density	Area under receiver operating characteristics (AUROC) = 0.96 and true skill statistics (TSS) = 0.82	The models present reasonable results to assess landslide risk with remote-sensing data	Novellino et al. [28]
RF and GradientBoost (GB)	DEM, Gaofen-1 satellite, GEI, etc.	Factor correlation analysis of landslide	Six factors from human activities, geological structures, geomorphology, and mineral composition	Receiver operating characteristics (ROC) for RF and GB = 0.97, and accuracy for RF and GB are 0.92 and 0.96, respectively.	RF and GB offer a better assessment of landslides in a complex geomorphic and tectonic terrain. NDVI was revealed to be closely associated with landslides compared to the other factors.	Qi et al. [17]
NB, DT, SVM, and RF	InASR and another multi-data source	Potential landslide hazards assessment	Seventeen factors, including NDVI, slope, degree of weathering	Precision for NB = 0.66, DT = 0.76, RF = 0.77 and SVM = 0.87, Recall for NB = 0.66, DT = 0.77, RF = 0.77 and SVM = 0.83, F1 for NB = 0.66, DT = 0.75, RF = 0.74 and SVM = 0.82	Among the four different models, the SVM provides the best prediction result	Zheng et al. [180]
Fuzzy set procedure (FSP), fuzzy expert system (FES), and ANN	ALOS–PALSAR, DEM	Landslide susceptibility zoning assessment	Nine factors like slope, aspect, elevation, TWI, land use/landcover (LULC), curvature, lithology	AUC for FSP = 0.78, FES = 0.83 and ANN = 0.90	The ANN was found to provide superior accuracy in landslide mapping	Sweta et al. [89]
Stacking ensemble learning (SEL)	InSAR	Landslide risk evaluation	Topography, human activities, normalized difference vegetation index (NDVI)	ROC curve shows 8% accuracy	The SEL model provides the best evaluation accuracy for landslide mapping	Gao et al. [181]
Logic model (LM)	Shuttle Radar Topography Mission (SRTM)	Rainfall-induced landslide assessment	Hydrology and rainfall, topography	From the ROC curve, the AUC of LM = 0.79	The result indicated that the LM could map landslide-prone terrain based on precipitation and geomorphology.	Maragano et al. [182]
LR, RF, NB, and multi-layer perception (MLP)	DEM, ALOS, and landsat-8	Susceptibility assessment of landslide	Surface cover, human activities, hydrology, lithology, topography	AUC of LR = 0.79, RF = 0.92, NB = 0.79 and MLP = 0.84	All the models show a good result, but the RF has the highest accuracy for landslide susceptibility assessment	Wang et al. [107]
XGBoost	MODIS, DEM, ASTGTM3	Landslide hazard assessment	Slope, climate, landcover, and geological	The AUC > 0.9	The XGBoost model provided better performance and avoided overestimation of landslide susceptibility.	Zhang et al. [90]
Landslide conditioning factor and swin transformer ensemble (LCFSTE)	GEE, Landsat-8 OLI	Landslide susceptibility assessment	Eleven factors, including slope, elevation, lithology, NDVI	AUC of LCFSTE = 0.94 compared to seven evaluation metrics	The LCFSTE model indicates a promising capability for assessing landslide susceptibility	Chen et al. [183]

. Additionally, some studies have been explored and discussed extensively on the use of artificial intelligence (AI), machine learning (ML), and deep learning (DL) algorithms with various remote sensing technologies and datasets to assess landslide vulnerability potentials extensively in this section.

(1)

Lin et al. [125] assessed landslide susceptibility using two case studies of Taiwan Island with landslide data derived from InSAR. The ML models DT, RF, AdaBoost, XGBoost, and NB considered geological and topographic factors for quantitative analysis. The results revealed that the RF model achieved a better precision accuracy of 82.95%, demonstrating its capability of assessing and predicting landslides in Taiwan Island.

(2)

Chang et al. [97] assessed non-landslide data from the Yangtze River in Yichang City, China, along the Three Gorges Reservoir areas prone to landslide. The Landsat-8 OLI, DEM, GEE, and ALOS datasets were analyzed with an unsupervised deep embedding clustering (DEC) algorithm with a deep integration network known as the capsule neural network based on SENet and SE–CapNet. The study results showed that the SE–CapNet had the highest prediction assessment (AUC = 0.97) of landslides, with rainfall being the main driving factor for landslides in the Yangtze River basin.

(3)

Mabdeh et al. [127] used two new genetic algorithms (GAs) based on bagging constructed with DT, KNN, and NB models based on bagging and random sub-space (RS) technique to assess and map landslide susceptibility. Considering 16 landslide environmental factors and Landsat-8 and DEM-derived data from Ajloun and Jerash, Jordan, the RS-based ensemble model produced a high accuracy (0.95) in assessing landslide.

(4)

Arrogate et al. [179] assessed landslide recognition using supervised and unsupervised models based on explainable ML models, as well as continuous change detection and classification (CCDC) models. With landslide data obtained from GEE, Landsat time series, and SRTM in addition to 16 landslide influence factors, the semi-automatic CCDC model showed a strong capability of assessing landslide risk in Guerrero State in Mexico.

(5)

Shao et al. [184] utilized the LR model with remote sensing images from GEE, ALOS–PALSAR, and DEM-derived data, with influencing factors, including elevation, slopes, aspect, lithology, and topography, to assess potential landslide risk in the downstream area of Jinsha River, Tibet Plateau. The data indicated that the LR model effectively identified landslide hazards around Jinsha hydropower station.

(6)

Chen et al. [109] assessed landslide susceptibility in the Tumen River Basin, utilizing a principle component analysis (PCA) and 15 landslide influence evaluation factors. The investigators developed an Adaptive Neuro-Fuzzy Inference System (ANFIS) based on Particle Swarm Optimization (PSO), Artificial Bee Colony algorithm (ABC), Shuffled Frog Leaping Algorithm (SFLA), and Bat algorithm (BAT). The ABC model achieved higher accuracy metrics measurements like AUC, root mean square error (RMSE), and mean absolute error (MAE) compared to the other models.

In summary, the existing landslide hazard studies related to remote sensing datasets from various landslide sites demonstrated that using traditional and newly proposed learning paradigms could make assessing landslide hazards in complex geological settings easy. Based on the accuracy measurement methods, several models (RF, NB, XGBoost, and ANN with remote sensing techniques such as Lansat-8 OLI, SRTM, InSAR, and LiDAR) achieved a relatively high level of landslide assessment accuracy between 0.60 and 0.95.

4.2.6. Landslide Extraction and Management

Extracting landslide information using remote sensing methods provides significant advantages, including disaster prevention and control, while manual extraction could result in low data accuracy.

(1)

Liu et al. [34] proposed a new model based on the DL algorithm (U-Net model) for automatic extraction landslides. The model considered three RGB bands and landslide-influencing factors to obtain the U-Net + six channels + ResNet model to eliminate the traditional U-Net problem. The Jiuzhaigou County earthquake data from Sichuan Province, China, was used to validate the new model, and the obtained results gave a precision accuracy of 91.3%, recall of 95.4%, and mIoU of 87.5%, compared to the U-Net model.

(2)

Using remote sensing images and data from the Gorganroud watershed, Iran, obtained from GEE, ALOS PALSAR, DEM, Sentinal-2, Landsat-8 OLI, and landslide inventory generated with GIS, Arabameri et al. [117] evaluated three ML methods, RF, alternative decision tree (ADTree), and fisher’s linear discriminant function (FLDA). The RF model recorded the highest AUROC of 0.97 and a prediction rate of 0.98, providing important outcomes for managing, predicting, and controlling the current and future landslides.

(3)

He et al. [25] employed aerial LiDAR survey, GEE, UAV, and DEM-derived data together with ML and DL models, RF, SVM, MLP, and DLNN to extract and predict landslide potential vulnerabilities. The validated results from the ML models achieved an accuracy of 87% and an AUC of 0.94.

(4)

Qi et al. [185] developed an ML algorithm with UAV-sensing image technology. The developed model was capable of extracting landslide features and providing reasonable effects.

(5)

Xia et al. [186] derived a landslide extraction model using a full convolution spectral-topographic fusion network (FSTF-Net) based on CNN and geospatial data. The landslide data were obtained from resources satellite-3 and high-resolution remote-sensing technologies such as Beijing-2, SuperView-1, DEM, and Worldview-3 in Mangkam County, Qinghai–Tibet. The proposed FSTF-Net model gave a landslide extraction and recognition precision of 0.85 and an accuracy of 0.89, compared to the current Deeplab_v3+.

(6)

Yang et al. [187] developed a background-enhancement technique that could learn the difference between landslide and background features. The landslide influencing factors were added to further enhance the extraction model’s accuracy. The proposed Mask-R-CNN + background-enhancement + landslide influencing factors were applied to the Ludian County landslide in 2014. Using the data from GEE, GeoEye-01, DEM, and Airbus Maxar Technologies, the proposed model recorded an F1-score of 89%, indicating its ability to accurately assess landslide susceptibility, compared to the traditional DL models.

(7)

A study conducted by Chen et al. [188] in Lanzhou City, China, introduced a new model using a squeeze-and-excitation network (SENet). This model combined the SENet with U-Net and utilized the data from the remote sensing images. The results indicated landslide extraction effectiveness with an F1-score of 87.94%, compared to the U-Net and U-Net backbone.

The above studies on landslide extraction and management also demonstrated significant achievement in utilizing remote-sensing technologies and learning algorithms in evaluating landslides and their associated hazards with a reasonable extraction and management accuracy above 90%.

5. Conclusions

This paper presents a systematic literature review of 186 research articles on the use of remote sensing technologies and artificial intelligence (AI), deep learning (DL), and machine learning (ML) in landslide studies. The following conclusions were obtained to answer the research questions:

• Artificial intelligence models ANN, FES, FSP, and XAI; deep learning models CNN, ResU-Net, DCN, and Deeplab V3+; machine learning models XGBoost, RF, LR, and KNN; and several other algorithms based on the traditional learning algorithms have been developed and proposed in the literature for the susceptibility mapping, prediction, and detection of landslides.
• The remote-sensing techniques with high-resolution data sources commonly used in the literature include DEM, ALOS–PALSAR, Landsat OLI, Sentinel groups, InSAR, SRTM, GEI, and other free remote sensors and satellite-available data sources from online. Commercial remote-sensing technologies such as UAV and LiDAR have been popular due to their ability to manage vast areas rapidly. They can also operate at lower heights and can access complex geological terrains.
• The AI, ML, and DL algorithms and remote-sensing technologies reasonably evaluate landslide susceptibility mapping, prediction, detection, monitoring, and inventory with a prediction accuracy ranging from 56% to almost 100%.

In addition to the answers to the research questions from the 186 articles reviewed in this study, the primary findings and future research directions based on the systematic review are as follows:

• Remote sensing data and imagery have been used with the learning models to study landslide characteristics, especially in susceptibility mapping and detection, followed by assessment and monitoring, with few studies focusing on stability, influence of trained data, and landslide management. Several traditional and newly proposed AI, DL, and ML algorithms available in the literature could be effectively used for landslide hazard assessments and provide local authorities with important information for landslide risk management and control.
• Landsat groups of remote sensing and digital elevation models were widely utilized and found in almost all the research articles reviewed. However, other remote-sensing technologies exist, such as high-resolution unmanned aerial vehicles, sentinel-1 and sentinel-2, advanced spaceborne thermal emission and reflection radiometers, and global digital elevation maps, which better predict and detect landslides.
• Environmental landslide-influencing factors such as topography, geology, and climate were found in almost all the studies reviewed. However, the available learning algorithms and remote-sensing techniques cannot be generalized for predicting landslide hazard studies worldwide, and this deficiency calls for critical attention to provide a generalized and accepted technology and the learning models’ standard range of prediction accuracy metrics.
• Many learning algorithms and models have been developed for a specific case study, which may not apply to other cases. In addition, some models were found to have significant limitations and must be improved.
• There is a lack of studies on landslide susceptibility and learning algorithms, especially in significant parts of Africa, South America, and the Eastern part of Europe. This calls for more research to extend the application of remote-sensing technologies and learning algorithms in assessing landslide susceptibility worldwide.
• Investigating additional remote sensing technologies and learning algorithms for real-time landslide monitoring could help researchers and geoengineers create reliable early warning systems to detect landslides.

One of the primary limitations of this systematic review is that the topic is vast, and we had to narrow the article search to provide a comprehensive overview of this study area. However, we propose conducting additional reviews, including quantitative studies, to develop an in-depth understanding of this topic. Drawing attention to this issue in academic studies could also be a good learning and research opportunity for researchers in this field.