In mountain hazard susceptibility mapping and assessment using machine learning models, the choice of model parameters is a significant factor in determining the accuracy of the model prediction. This work provides a novel method for developing a random forest (RF)-based prediction model by embedding hyperparametric optimization ensembles. The performance of the RF model is enhanced by employing a Bayesian optimization (Bayes) method and a Genetic algorithm (GA) and verified in the Wudu section of the Bailong River basin, China, a typical mountainous hazard-prone area. Fourteen influencing factors were selected in conjunction with field measurements to characterize the cascading ''collapse-landslide-debris flow'' hazard chains in the study area, and datasets for training (80%) and validation (20%) models were constructed for 378 hazard sites. Based on multicollinearity analysis and Relief-F two-step evaluation, typical statistical performance metrics such as the confusion matrix, recall, precision, accuracy, F1 and area under the operating characteristic curve (AUC) of individuals were used to evaluate model performance. Our results revealed that all the 3 models (i.e., RF, GA-RF, and Bayes-RF) performed well (AUC: 0.89 ~ 0.92), but the Bayes-RF model performed the best (AUC = 0.92), which can be used as a highly accurate and robust mountain hazard susceptibility assessment model applicable in the study area and other regions. Meanwhile, the generated susceptibility map can guide stakeholders in making appropriate mountain hazard prevention and control measurements in the region.
Research Article
Machine Learning-based Ensemble Modeling for Mountain Hazards Susceptibility Assessment
https://doi.org/10.21203/rs.3.rs-2220814/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Mountain hazard chains
Susceptibility assessment
Machine Learning
Bayes-RF Model
Geographic information system (GIS）
Landslides, collapses, and debris flows are common natural phenomena and land surface processes on the earth (Lee et al. 2004), as well as the primary geomorphic hazard patterns especially in mountainous areas from the human perspective. Landslides and collapses provide solid material supply for debris flows while always developing directly into debris flows in the routing processes (Fig. 1). In other words, those cascading processes ultimately convert to fatal disaster chains under certain conditions (i.e., triggered by precipitation, earthquakes and human activities), which we defined herein as mountain hazards accounting for catastrophic loss to the world economy, production and security of people’s life and property (Huang and Fan 2013; Froude and Petley 2018). According to China's 2019 National Geological Disaster Bulletin, among the 6181 geological disasters in the country, 95.5% of the total were triggered by natural factors, 4.5% were triggered by human factors; the direct economic loss reached RMB 2.77 billion. Mainly due to the unforeseen stochastic characteristics of the conditioning factors as well as the chain and cascading features of the disaster processes, the developing processes of mountain hazards always show a significant and complicated multivariate-nonlinear characteristic. Assessing mountain hazards susceptibility while seeking effective solutions for disaster risk mitigation scientifically, accurately and rapidly, therefore, are not only the key academic issue in the world, but also the scientific guide and urgent need for the government and stakeholders making land use and risk management strategy decisions (Dai et al. 2002; Zhao and Liu 2016).
Conventional geological hazard susceptibility assessment is mainly classified into two categories: qualitative and quantitative. The qualitative evaluation mainly relies on the experience and knowledge of experts in field investigation and the attributes of disaster susceptibility and makes judgments based on the literature and evaluates the disaster susceptibility by the empirical method with strong subjectivity. Quantitative evaluation, in contrast, emphasizes mathematical models as the basis and predicts the occurrence of disasters based on the calculation results of mathematical models and presents them in a quantitative form, which is more objective than qualitative evaluation.
Over the last 3 decades, although many models and methods for disaster risk evaluation have been applied, few models were proved superior for providing effective evaluation modeling solutions (Trigila et al. 2015; Pourghasemi and Rahmati 2018). In recent 10 years, with the technological progress of GIS, remote sensing and GPS, conventional mathematical and statistical models and machine learning models have been widely used in disaster susceptibility assessment, such as weight of evidence, frequency ratio (Huang et al. 2015; Li et al. 2017), even and parallel models (Zuo et al. 2020), and other methods for constructing weighted index systems (Li et al. 2017). Studies have shown that conventional mathematical and statistical models, such as the informative model, are affected by the size and grading of raster image elements and can hardly reflect the nonlinear relationship between disasters and their underlying environmental factors, while machine learning models can accurately detect the data structure and efficiently predict the occurrence of disasters. According to the literature, there are abundant machine learning models in common use, including logistic regression (Aditian et al. 2018; Kalantar et al. 2018), decision trees (Nefeslioglu et al. 2010; Chen et al. 2018), random forests (RF) (Park and Kim 2019; Dou et al. 2019), support vector machines (SVM) (Fallah-Zazuli et al. 2019; Sheng et al. 2021; Wang and Brenning 2021), Bayesian algorithms (Bayes) (Nhu et al. 2020; Chen and Zhang 2021), genetic algorithms (GA) (Chen et al. 2019), artificial neural networks (ANN) (Conforti et al. 2014; Polykretis and Chalkias 2018). However, there is no consensus on which machine learning is more suitable for susceptibility modeling due to the different selection of predisposing factors of geological hazards, the quality of databases and the selection of machine learning models. Bui et al. (2012) conducted a comparative study of machine learning in landslide susceptibility evaluation and concluded that the Bayesian regularized network model seems to be more stable than the Levenberg-Marquardt network model in landslide susceptibility mapping. Sun et al. (2020) used a Bayesian algorithm to establish an optimized logistic regression model and a random forest model for landslide susceptibility evaluation with hyperparameter optimization and concluded that the AUC value of the test dataset in the optimized model increased by 10% in the random forest model, and the AUC value of the test dataset in the logistic regression model increased by 4%. The random forest model had better stability and predictive power than the two cases that required hyper-parameter optimization. All the related research demonstrated that machine learning-based models show an outstanding performance in susceptibility assessment. However, they were seemingly only used for a single hazard type instead of those cascading hazard chains mentioned above. Consequently, it is of great importance for the susceptibility of mountain hazard chains to be studied under changing environmental conditions.
China is one of the countries with the most serious geological disasters in the world. As China is located at the intersection of the mid-latitude global natural disaster belt and the Pacific Rim disaster belt in the northern hemisphere with a mountainous landscape and strong crustal movement, diverse and serious natural disasters occur widely, which have gradually become an obvious obstacle to the construction of ecological security and high-quality economic development in recent years. In this study, based on the sufficient field survey and data collection about the developing features of the mountain hazard chains in Longnan mountainous area, China, while taking the Wudu section of the Bailong river basin (i.e., Wudu District) where mountain hazards are particularly developed as an example, 14 influencing factors such as elevation, slope, aspect, plane curvature, profile curvature, distance to road, distance to river, distance to fault, surface roughness, lithology, normalized difference vegetation index (NDVI), terrain moisture index (TWI), land cover type and precipitation were selected as integrated datasets. Two novel model frameworks were developed with Bayesian algorithm and genetic algorithm embedded into RF model, and the performance analysis of the three different machine learning models was then conducted. Finally, mountain hazard susceptibility mapping was conducted, and model performance and robustness, as well as accuracy, were evaluated. The hybrid RF-Bayes machine learning model and the ensemble modeling approaches established are expected to provide effective theoretical guidance for predictive modeling of vulnerability in other geographically similar areas while putting insight into the application of data-driven models on earth science systems.
2.1 Study area
Wudu District is located in the southeast of Longnan City, Gansu Province, within the Yangtze River basin, between the latitudes 32° 47' ~ 33° 42' N and longitudes 104°34' ~ 105° 38' E. The whole region covers an area of 4683 km2 with a north-south longest span of 100.8 km and east-west widest span of 76.2 km, and is composed of 36 townships, 643 villages and a total population of about 546,600. The mainstream of the Bailong River flows in from the northwest and flows out from the southwest, and most of the G212 highway is arranged along the Bailong River (Fig. 2). As the study area belongs to the north subtropical semi-humid climate zone, the climate is characterized by frequent rainfall from May to September (accounting for 75%~85% of the total 400 ~ 900 annual precipitation), which is exactly the season of the high occurrence of collapse, landslide and debris flow disasters. The high and middle mountain erosion structures in the study area are mainly distributed in the north-central and southern parts and the southern mountainous area of the Bailong River, consisting of metamorphic sandstone, sand slate and tuff, with an elevation of 1000 ~ 3600 m and a relative elevation difference of 1000 ~ 1500 m, where landslides and collapses occur frequently, especially in the steep mountains with deep gullies and valleys. The mountains in the north-central area are mainly composed of phyllite rock, metamorphic sandstone, chalk and tertiary sandstones dated from Devonian together with sandstone, shale and mudstones in Cretaceous and Tertiary, which are prone to landslides, collapses and debris flows due to the weak and loose geotechnical body, sparse vegetation and fragmentation. In recent years, climate warming-induced extreme rainfall events, population growth, urbanization and other human activities have been leading to increasing mountain disasters, causing serious impacts on the sustainable development of the mountain ecosystem. The Bailong River basin in the Longnan mountainous area, to which the Wudu District belongs, is one of the earthquake-prone areas in China, providing internal geological dynamics for the occurrence of disasters, while precipitation, weathering and human engineering activities provide external geological dynamics. Thus, in this work, based on the historical data of 378 typical mountain hazards in the study area, we conducted the prediction analysis of mountain hazard susceptibility assessment based on GIS and machine learning models.
2.2 Data preparation
2.2.1 Mountain hazard inventory
A hazard inventory records the location and time of occurrence of a geological hazard body, as well as the hazard types that left identifiable traces in an area (J. Pašek 1975). The mountain hazard sites covered in this paper were obtained from historical geohazard literature up to 2015 (Chen et al. 2006; Ma et al. 2015), field surveys and the government open resource website (http://www.ngac.cn/), and a total of 378 geohazards of various types were counted, including 110 landslides, 260 debris flows and 8 collapses. Field observations indicated that landslides and collapses are generally associated cascading processes in the study area, and collapses and landslides also transformed gradually into debris flows in the routing processes or entrained by heavy rainfall-induced flooding after channel fill and ultimately developed into debris flows and even disasters. In the formation of these hazard chains, landslides often act as a link between the other two (Fig. 1). According to the implementation rules of the Basic Requirements for Geological Hazard Investigation and Zoning in Counties (Cities), the huge landslides in the Wudu District account for 27%, the large landslides for 47%, the medium landslides for 6% and the small landslides for 20%, with the huge and large landslides in dominant and the medium and small landslides sporadically distributed. According to their formation era, the old landslides (early Holocene) and the ancient landslides (early Late Pleistocene) account for 69% of the total number of landslides and the new landslides (or modern landslides) for 31% (Li 1997). In this study, the 378 hazard sites were used as positive samples (PS), and 378 non-geological hazard sites were randomly generated as negative samples (NS) outside the 500 m buffer zone of the hazard sites. Ultimately, 605 sample points (80%) were used as the training dataset and 151 sample points (20%) were used as the validation dataset. The collected historical hazard data revealed that the geological hazards here show both a characteristic of a large number and high distribution density, with landslides in all 43 townships in the region, making it the center of hazards in the Longnan mountains area (Li 1995). In addition, statistics along the G212 highway show an average of one landslide every 2 km on the roadside, with one landslide developed every 1 km in the serious sections.
| No. | Name | Type | Time | Size(104 m2) | Economic loss (RMB) | Main conditioning factors |
|---|---|---|---|---|---|---|
| 1 | Dujia gully | Landslide | 2015.8 | 590 | 200,000 | Lithology, Fluvial erosion |
| 2 | Taoshu gully | Debris flow | 2020.8 | 85 | 250,000 | Topography, Rainfall |
| 3 | Xiaoshui gully | Collapse | 2008.5 | 68 | 200,000 | Lithology, Human activity |
In order to ensure the accuracy of the statistics of historical disaster sites as far as possible, and to provide data accuracy and reliability for the subsequent attributes analysis, impact factors selection, dataset classification testing, model validation, algorithm optimization and even disaster susceptibility assessment based on hazard inventory, we carried out field surveys on three of the representative historical disaster sites that have long been active under the funding of "Research on the prevention and control techniques of landslides and debris flows from Longnan section of the G212 highway (Lanzhou ~ Chongqing)" (supported by the western transportation construction science and technology project of Ministry of Transport, China).
As shown in Table 1, the Xiaoshui gully is representative of typical gully-debris flows where flash floods in the gully channel entrain the collapsed accumulation. The new tectonic movement in this region is characterized by differential vertical movement, with fragmented and easily movable rocks and crevice water development as well as intense human engineering activities. Under the direct influence of the Wenchuan earthquake in May 2008, a collapse occurred. According to the on-site measurement, the collapse volume was 680,000 m3, resulting in various economic losses of RMB 2 million, which is the largest typical collapse event in recent years. The Dujia gully is a typical historical landslide hazard site, and the landslide body is located on the left bank of the Bailong River, with the free face directly towards the Bailong River. Field investigation in August 2015 demonstrated that the exposed bedrock of the landslide body is strongly-fully weathered, with severe extrusion and deformation due to long-term tectonic movement. The loose cover of the landslide is mainly composed of loess-like clay, fully weathered phyllite rock rubble soil with the soil stone ratio is 3:7. The surface of the landslide body is seriously disintegrated and the foot of the slope is scoured to form a 20 ~ 40 m-high scrape with an obvious shear outlet, which is a large landslide. The Taoshu gully is a typical valley-type dilute debris flow channel, with steep terrain on both sides and an average longitudinal slope of 21.3%. It is a mid-mountain area characterized by large topographic relief with a relative elevation difference of 673 m, providing sufficient potential energy for the initiation of debris flows. The intense rainfall events in this region are mainly characterized by high-intensity, short-duration storms, which can provide sufficient hydrodynamic conditions for the formation of debris flows. Additionally, the wide distribution of loose solid materials in the channel is the abundant natural supply for the formation of debris flows. Following a landslide disaster in August 2017, a serious debris flow disaster occurred in August 2020. The site investigation revealed that the event led to serious channel damming and the destruction and collapse of part of the drainage guide dike, posing greater threats to the lives and property of 57 households and 236 people downstream.
The fieldwork not only further demonstrates the chain and cascading nature of mountain hazards in the study area, but also indicates some of the main influencing factors.
2.2.2 Mountain hazard influencing factors
As hazard susceptibility assessment involves a wide range of factors, the reliability of field data collection is crucial. To accurately evaluate the background of the study area, a systematic and scientific index system needs to be established through the analysis of the historical occurrence of geological hazards as well as regional natural conditions, geo-environmental conditions and human activities (Qiao and Zhao 2001; Liu 2003). Therefore, in this work, 14 evaluation factors, including elevation, slope, aspect, plan curvature, profile curvature, distance to road, distance to river, distance to fault, roughness, lithology, NDVI, TWI, ground cover and precipitation, were selected from three perspectives (i.e., topography and tectonics, external dynamic geological environment, and engineering geological rock group) for susceptibility assessment, respectively. The analysis of Ning et al. (2013) on the conditioning factors of debris flows and the analysis of Qi et al. (2014) on the integration of indicators for geological hazard assessment in the study area further confirmed that the above-mentioned influencing factors have a significant influence on the development of landslides, mudslides and collapses in the Wudu district.
According to the field survey and the analysis of the underlying geological conditions in the study area, seven factors were selected in terms of topography and tectonics (Fig. 3a ~ g): elevation, slope, aspect, plan curvature, profile curvature, distance to fault, and roughness. The Wudu mountainous area controlled by regional tectonics extends in the east-west direction, with the highest elevation of 3600 m in the northwest and the lowest 660 m in the southeast. The overall topography is characterized by strong gully-dissected, steep mountains, large topography and steep slopes. In this study, a 30 m×30 m digital elevation model DEM of the study area was derived from ALOS multispectral remote sensing image data (http://www.usgs.gov/). Analysis of the dataset by ArcGIS 10.8 software revealed that most of the mountain hazards occurred between 918 m ~ 1214 m, with a distribution rate of 34% (Fig. 3a). Most of the hazards are located within the slope range of 20°~30°, with nearly no occurrence recorded in slopes exceeding 50° (Fig. 3b). Although there are many hills in the study area with different slope aspects, the distribution of hazards is generally even and shows no significant difference (Fig. 3c), and few hazards occurred in flat areas. Ground curvature is a quantitative measure of the degree of distortion at a point on the surface of the terrain, and in terms of plan curvature and profile curvature, the majority of hazard sites occur in areas where the curvature of the ground contours is relatively gentle (Fig. 3d and 3e). The geotechnical fragmentation, complex structural deformation and poor stability in the vicinity of faults and fracture zones are obvious and the number of hazard sites decreases with the increasing of the distance from the fracture (Fig. 3f). The value of slope roughness reflects the ability of runoff diversion and accumulation. Previous results once indicated that landslides, collapses and debris flows are mainly developed in locations with high slope roughness, in that higher slope roughness is not conducive to the diversion of precipitation, thus accelerating the alteration of the physical and mechanical characteristics of the rock by groundwater, which in turn leads to landslides, mudslides and collapses (Santacana et al. 2003), whereas the occurrence of mountain hazards in the study area decreases significantly with increasing slope roughness (Fig. 3g, also see Appendix A for detail).
The development of landslides, collapses and debris flows is controlled by the internal dynamics of the environment and is also influenced by the external dynamics. External dynamic factors are often the direct cause of the above-mentioned mountain hazards. In this study, 6 factors were selected as the external dynamic geoenvironmental factors (Fig. 3h ~ m): distance to road, distance to river, NDVI, TWI, ground cover, and precipitation. Since the Holocene, human activities have gradually become the main trigger for geological hazards. We applied ArcGIS 10.8 to extract the Euclidean distances of roads and rivers based on the road and river data already collected in the study area (http://www.tianditu.gov.cn/) and concluded that 62.43% of the hazards occurred within 600 m of the roads and 46.3% of the hazards occurred within 600 m of the rivers (Fig. 3h and 3i). The vegetation cover within the study area was calculated according to the NDVI proposed by Kaufman and Tanre (1992). If there is good vegetation cover on the ground, the rate of erosion is very slow due to the shading of the plants and the consolidation of the roots, and most geological hazards occur in areas with sparse vegetation cover (Fig. 3j). Based on the TWI concept proposed by Beven and Kirkby (1979), the influence of topography and soil characteristics on soil moisture distribution is taken into account. The equations provided by Obled et al. (1994) and Tarboton (1997) were used to calculate TWI and to assess the development of geohazards in the context of the spatial distribution characteristics of soil moisture (Fig. 3k). The surface cover types provided by Globeland 30 (http://www.globallandcover.com/) were also used to count 9 different land use types in the study area, with 15.24% of the land used for human activities and 46.83% of the total number of hazard occurrences. Compared to the 84.06% vegetation cover and the corresponding 49.74% hazard occurrence, it is clear that human activities have a significant impact on hazard occurrence (Fig. 3l). In addition, rainfall is quite important for the occurrence of mountain hazards in the study area. The heavy infiltration of precipitation leads to saturation of the soil and rock layers on the slopes and even water accumulation on the lower part of the slopes in the water-resisting layer, thus increasing the weight of the sliding body and reducing the shear strength of the soil and rock layers, leading to disasters. We produced rainfall contour maps for the study area based on the annual average rainfall data recorded at 19 stations at the National Meteorological Science Data in China (http://data.cma.cn/) (Fig. 3m).
Lithology and engineering geological rock formations are closely linked to the development and formation of mountain hazards, and lithology influences and even determines the characteristics of mountain hazards. Lithology (Fig. 3n) was selected to assess the geology of the region based on statistical data and hazard susceptibility assessments conducted by Ning et al. (2013). In order to rapidly and accurately recognize the lithological factor grouping, on the basis of the 1:500000 geological map in the study area and with reference to the Engineering Rock Classification Standard(2014.GB/T 50218 − 2014.), the Code for the Design of Building Foundations (2011.GB 50007 − 2011.), and the Code for Geotechnical Investigation (2002.GB 50021 − 2001.), the rock groups in the study area are divided into four categories according to their softness and hardness: hard rocks, harder rocks, softer rocks, soft rocks and loose rocks
The dataset was created in the study area and each subset of influence factors was treated as a point attribute in the model. We used 5,164,435 grids of 30 m × 30 m in size and raster aligned using 'Alignment Points' in the projection tool. The raster data was processed to obtain the attributes of all points from a raster map of the study area and all data were normalized before the model was implemented. Finally, the dataset was analyzed and evaluated using ArcGIS 10.8 software (see Appendix A for the frequency ratio of the influencing factors).
In this paper, we use a hyperparametric optimized machine learning model for mountain hazard susceptibility mapping with 5 main steps (Fig. 4). 1) generation of a list of mountain hazards and mountain hazard influencing factors; 2) influencing factor evaluation and dataset classification; 3) benchmark RF model construction; 4) optimization of RF model using Bayesian and genetic algorithms, susceptibility mapping;5) mountain hazard susceptibility assessment through model validation and comparison.
3.1 Influencing factor evaluation
The occurrence of disasters is controlled by a variety of influencing factors and there are differences in the impacts of individual factors on the occurrence of disasters (Wu et al. 2020). Therefore, it is necessary to study the importance and mechanism of each influencing factor to guide the stability and accuracy of prediction models and disaster prediction and prevention. As mentioned above, although we have initially selected 14 possible gestation factors related to mountain hazards in the study area through literature research and field surveys, it is still necessary to identify the sequence of the most significant of these influencing factors for modeling purposes. Given the complex non-linear characteristics of the development of hazard chains such as collapse, landslide and mudslide, we use multicollinearity analysis and relief-F two-step analysis to quantitatively assess and measure the predictive ability of the 14 influencing factors, in order to establish a targeted and highly accurate evaluation and prediction model for the vulnerability of chain and cascading mountain hazards in the study area.
3.1.1 Multicollinearity analysis
When performing linear regression model analysis, the independent variables are prone to be correlated with each other, and when serious co-linearity problems occur, it will lead to unstable model prediction results. In this case, it is necessary to eliminate the independent variables with significant co-linearity to eliminate the effects of multiple co-linearity. (Tien Bui et al. 2016). The Pearson correlation coefficient method (Alimohammadlou et al. 2014) is usually used to quantify multicollinearity. Its value ranges from − 1 to 1. -1 indicates that the two variables are perfectly negatively correlated, 1 indicates that the two variables are perfectly positively correlated, and 0 indicates that they are not correlated. In this study, Pearson's correlation coefficient was used to detect the correlation between two influencing factors of mountain hazards. If the value is greater than 0.5, the correlation is high. Meanwhile, the discriminant variance inflation factor (VIF) and tolerance level (TOL) were introduced in this study. It is generally considered that when VIF > 10, TOL < 0.1, a serious covariance problem occurs in the model, and the two indicators have logical correspondence. Either one of them can be chosen. In this study, multiple cointegration tests were conducted for the 14 disaster impact factors using SPSS 26 software.
3.1.2 Relief-F
Relief-F is an effective method for quantitative analysis to evaluating the importance of factors (Li et al. 2017). It measures the importance of each feature by the correlation statistic and assigns different weight values, and negative importance is eliminated as a redundancy factor. In this study, using Relief-F on MATLAB 2016b platform, we randomly take one sample R from the training sample set at a time, find k nearest-neighbor samples of R from the R sample set, and then find k nearest-neighbor samples from each sample set of different classes of R. Then, we update the weight of each feature as shown in (1):
3.2 Random Forest (RF)
RF is a "Leatherman" in the field of machine learning, not only having the best accuracy of any current algorithm, but also being able to run large data sets efficiently. RF is an integration of decision tree, which extracts multiple training sets from the original sample set by Bagging method, and performs decision tree modeling for each training set separately. During the growth of the decision tree, a random sampling of the feature set is performed every time the internal nodes split, and the optimal selection is made among the extracted feature sets. Finally, the final result is obtained by averaging or voting using the results of all decision tree calculations (Zhu and Zhang 2014). The method of classification prediction using RF model is as follows (Breiman 1996, 2001; Wu and Li 2006; Zhang and Ni 2007; Xiao et al. 2010; Fang et al. 2011).
(1) A training set is formed by sampling N times from N original samples in the form of sample bagging. The unsampled samples are called Out-Of-Bag (OOB) data and can be used to evaluate the model performance, named as OOB estimation.
(2) For each training set, a decision tree is generated. Assuming that the sample has M features and a number of features F≤M is specified, F features are randomly selected from the M features as the split feature set at each internal node of the decision tree, and the node is split by the best split in the split feature set. The value of F is generally kept constant and is usually taken as F=M/3 for regression and as (3) for classification.
(3) Decision trees are generated using classification and regression tree algorithms, with each tree growing freely without pruning.
(4) Repeat the above steps k times to obtain a total of k training sets, forming k decision trees, and each tree corresponding to the unselected sample set forms a total of k out-of-bag data.
(5) The generated k decision trees are formed into a RF to perform regression analysis or classification prediction on the new data. When used for regression, the final result is the mean of the computed results of each tree. When used for classification, the final result is generated by voting the results of each tree.
In this study, Python 3.5 was used to build a RF model using the Scikit-learn module. The basic parameters of the running device are: 11th Gen Intel(R) Core (TM) i7-11800H @2.30 GHz, running memory of 16 GB.
3.3. Optimization Algorithm
3.3.1 Bayesian hyper-parameter optimization
In machine learning, hyperparameters are parameters that need to be set in advance during the learning process of a machine learning model, and algorithms are rarely hyperparameter free. For all algorithms, some hyperparameters have a significant impact on the predictive power and some hyperparameters have a significant impact on the predictive accuracy or on the accuracy of the model. Therefore, rational adjustment of these hyperparameters, i.e., hyperparameter optimization, is particularly important. However, hyperparameter optimization is a combinatorial optimization problem. It is not optimized by gradient descent like normal parameters. The computation used to evaluate a set of hyperparameter configurations is very difficult because adjusting individual hyperparameters requires retraining to evaluate their effects.
Nowadays, more and more hyperparameter optimization processes are done by automated methods, which aim to find the best hyperparameters in a shorter time using policy-based informed search. Moreover, no initial setup, as well as additional manual operations, are required, except for manual operations during the process. Bayesian optimization is the primary choice for optimizing the objective function (Snoek et al. 2012; Ghahramani 2015; Xia et al. 2017). For most optimization algorithms, it is necessary to assume that the objective function is known in order to solve the derivative, and the function must be convex, but for some problems, the objective function position is a non-convex function in the process of parameter tuning, which leads to a huge amount of computation and an insignificant solution effect. To solve such problems, Bayesian optimization algorithms are born. The posterior distribution of the unknown objective function is approximated by using prior knowledge, and then the next combination of hyperparameters for sampling is selected according to the optimal partial objective function. To get as close as possible to the true objective function, the Bayesian algorithm uses an agent model. Moreover, Bayesian optimization algorithm is now widely used in hyperparameter tuning for machine learning.
3.3.2 Genetic Algorithm (GA) hyper-parameter optimization
As a classical algorithm, GA uses a stochastic global search algorithm to simulate the physical process of natural biological evolution. The basic principle lies in simulating a population of potential solutions, where the initial individuals in this population are randomly generated. Then, the GA algorithm performs selection, crossover and mutation operations on each individual to search for the optimal solution, corresponding to the principles of survival of the fittest, recombination of genetic material and random ach tree. When used for classificati mutation observed in nature, respectively. By iterating these three operations, GA can continuously regenerate solutions that approximate the true value until a certain number of new generations of individuals are generated, and then the objective function is recalculated for all individuals of the new generation. The best-performing individuals in the new generation are retained in the next generation according to their fitness. Thus, as each generation reproduces, the fitness function of the entire population decreases, and the process is repeated until the results can no longer be improved.
In this study, the idea of genetic function is invoked to optimize the parameters, and the process of parameter optimization can be defined as a multi-objective optimization problem, i.e., using (4):
where ∨ — denotes vector minimization, i.e., each sub-objective function in the vector objective f (\(x\)) is minimized as much as possible.
The parameter optimization of the genetic function needs firstly to be encoded, and the encoding is done in binary. The length of the chromosome is selected according to the type of the model parameters and the number of decimal numbers corresponding to binary is calculated (Mo et al. 2001), using (5):
The initial population is thus generated and the fitness is calculated, i.e., the value of the evaluation criterion for measuring cross-validation, and it enters the reproduction iteration loop, and the reproduction replication is performed according to different cross-validation and variation ratios. If the set number of iterations is reached, the reproduction will be terminated, otherwise, the reproduction will be continued and the optimal solution will be found in the final population.
3.4 Model validation and comparison
3.4.1 Model Performance Evaluation Metrics
In this paper, five commonly used statistical metrics are used to validate the performance of the prediction model: Accuracy (ACC), Recall, Precision, F1, and ROC-AUC. The first four are statistically based metrics. ACC is the ratio of the number of samples correctly classified by the model to the total number of samples. Recall is the ratio of the number of positive samples classified by the model into positive classes to the number correctly classified by the model. Precision is the ratio of the number of positive samples predicted as positive classes by the classification model to the sum of samples predicted as positive classes, and F1 is a composite measure of the recall and precision of the model. ROC and AUC are metrics for evaluating classifiers, and ROC is the abbreviation for the subject operating characteristic curve, also known as the perceptivity curve. The points on the curve respond to the same perceptivity, and they are all responses to the same signal stimulus, differing in the results obtained under several judgment criteria. AUC is the abbreviation for area under the ROC curve. Usually, the larger the AUC is, the higher the diagnostic accuracy. They are commonly used and representative indicators.
3.4.2 Mountain hazard susceptibility mapping
Spatial prediction of geological hazards is a binary classification process (Bennett et al. 2016). An equal number of positive samples (hazard points) and negative samples (non-hazard points) need to be prepared before model training (Kornejady et al. 2017). In this study, the negative sample data were randomly generated by ArcGIS 10.8 software outside the 500 m range of positive samples, after which the positive and negative data were divided into training and test sets in the ratio of 8/2. The established hazard evaluation model was used to calculate a unique probability value for each raster cell in the whole study area, which was expressed as the hazard susceptibility index of the area. The classification method of geometric intervals was used to classify the hazard susceptibility index into five susceptibility classes: very low, low, medium, high, and very high, which is an effective classification method for classifying hazard susceptibility values (Hong et al. 2019; Pham et al. 2019).
4.1 Importance of Influencing Factor
| Variables | TOL | VIF |
|---|---|---|
| Elevation | 0.423 | 2.363 |
| Slope | 0.169 | 5.914 |
| Aspect | 0.912 | 1.097 |
| Plan curvature | 0.768 | 1.302 |
| Profile curvature | 0.744 | 1.345 |
| Distance from roads | 0.574 | 1.741 |
| Distance form rivers | 0.655 | 1.528 |
| Distance form faults | 0.792 | 1.263 |
| Roughness | 0.191 | 5.229 |
| Lithology | 0.870 | 1.150 |
| NDVI | 0.441 | 2.266 |
| TWI | 0.732 | 1.367 |
| Surface coverage | 0.947 | 1.056 |
| Precipitation | 0.540 | 1.851 |
According to the results of the multicollinearity analysis of the evaluation factors (Table 2), the maximum VIF and the minimum TOL values were 5.914 and 0.169, respectively. Therefore, there was no multicollinearity in the factors used in this study.
The Pearson correlation coefficient method was used to correlate the 14 factors in the study area according to the two-step analysis of multiple covariance and Relief-F significance analysis, and the results are shown in Fig. 5.
As shown in Fig. 5, the correlation coefficient between slope and roughness is 0.90, which is greater than 0.50, indicating a high correlation. Therefore, in this paper, the roughness factor is excluded from the mountain disaster susceptibility modeling process, and the remaining 13 factors are selected for hazard susceptibility assessment.
Based on the Relief-F model, the contributions of the 13 influencing factors to the occurrence of disasters in the study area were obtained. As shown in Fig. 6, lithology, precipitation, and NDVI contribute more to the occurrence of hazards compared to other evaluation factors, which is consistent with the common perceptions about the causes of geological hazards (Zhang et al. 2005; Rne et al. 2008) and is also highly consistent with the results of field survey investigations in this paper (Table 1). The statistics on the importance of plane curvature and profile curvature factors were negative and should be considered redundant factors to be eliminated, and this result was also similar to the results of (Hong et al. 2019). Therefore, 11 assessment factors of elevation, slope, aspect, distance to road, distance to river, distance to fault, lithology, NDVI, TWI, ground cover and precipitation will be used for subsequent model construction.
4.2 Hyperparametric Optimized RF Model
The RF model contains six hyperparameters: n_estimators, max_features, max_depths, min_samples_split, min_samples_leaf, and bootstrap, among which n_estimators represents the number of trees in the random forest, max_features is the number of features at each split, max_depths is the maximum number of splits each tree can have, min_samples_split is the minimum number of observations required before a node in the tree splits, min_samples_leaf is the minimum number of observations required at the end of each tree's leaf nodes at the end of each tree, bootstrap indicates whether bootstrapping is used to provide data for each tree in the random forest (bootstrapping is random sampling with replacement from the dataset). In this study, n_estimators and max_depths, which have a large impact on the results, are selected for hyperparameter optimization. The samples are selected using repeated sampling with replacement sampling, so the bootstrap value is true.
Bayesian algorithm is used to optimize the hyperparameters of n_estimators, max_depths and to obtain the hyperparameter values for each iteration. In the hyperparameter optimization process, the accuracy is chosen as a measure of the accuracy of the model. In this study, after several hyperparameter optimization training, the optimal parameters of the model appeared before the 20th iteration and the accuracy of the model could not be improved after 30 iterations. Therefore, for hyperparameter optimization, the number of iterations was chosen to be 50, and the accuracy of the 10th iteration was 0.942. The optimal hyperparameters were [n_estimators:96.76385895424721, max_depths: 32].
The RF model was modeled and trained using optimized hyperparameter values for 80% of the training samples in the dataset. The selected 11 gestation factor datasets were input into the optimized RF model, and the mountain hazard susceptibility index of the study area was taken to be in the range of 0 ~ 1.The natural break method was used to classify them into five classes (Fig. 7c), i.e., very low (0 ~ 0.19), low (0.19 ~ 0.36), medium (0.36 ~ 0.55), high (0.55 ~ 0.75), and very high (0.75 ~ 1). The area percentages of each susceptibility class were 34.27%, 29.13%, 17.19&, 11.37% and 8.03%, respectively. From the results of the RF model (Fig. 7a), it can be seen that the very high and high mountain disaster-prone areas in Wudu district are mainly distributed along rivers and river valleys, where human activities are frequent and soil erosion is serious. The low susceptibility areas are mainly distributed in areas with less human activities, higher elevation, and harder lithology, which is consistent with the historical distribution of mountain hazards.
The GA model was optimized with hyperparameters and performed for 50 iterations with an accuracy of 0.915. The accuracy reached the maximum at the 41st iteration and the model performed best. The hyperparameters were [n_estimators: 42, max_depths: 27].
The RF model was modeled and trained using the above optimized hyperparameters and 80% of the training sample set. The 11 influencing factors were input into the RF model to be optimized, and the mountain hazard susceptibility index of the study area was taken as 0 ~ 1. It was classified into five classes using the natural breakpoint method (Fig. 7b), which were very low (0 ~ 0.14), low (0.15 ~ 0.33), medium (0.33 ~ 0.52), high (0.52 ~ 0.74), and very high (0.74 ~ 1). The percentage area of each susceptibility zone class was 30.35%, 31.64%, 17.35%, 12.54% and 8.12%%. Among them, the distribution results of very high and high susceptibility areas were in good agreement with the results predicted by the random forest model (Fig. 7a), which were consistent with the distribution pattern of historical mountain hazards.
4.3 Model Validation and Comparison
To measure the overall performance of the models, we utilize the ROC-AUC metric to validate the training success and prediction ability of two optimization models (GA-RF and Bayes-RF). The AUC value of the training dataset indicates the success ability of the model, while the AUC of the validation dataset represents the predictive ability of the model.
As shown in Fig. 8, the AUC values for the training dataset of the GA-RF and Bayes-RF models are 0.99 and 0.99, respectively. The AUC values for the validation dataset are 0.91 and 0.92, while the AUC values for the training and validation datasets of the unoptimized RF model are 0.97 and 0.89, respectively. The difference in the AUC values of the training dataset indicates that the GA-RF and Bayes-RF models exhibit good fit (success capability) and outperform the unoptimized RF model in terms of overall performance. However, the Bayes-RF model had a higher success capability for the training data. Similarly, both optimized models have higher prediction ability, while the Bayes-RF model predicts with higher accuracy. In addition, to evaluate the merit of the training dataset, four metrics (i.e., ACC, recall, precision and F1) calculated from the confusion matrix, were applied in the training phase of this study. The results are shown in Table 3.
| Model | TP | TN | FP | FN | ACC (%) | Recall (%) | Precision (%) | F1 |
|---|---|---|---|---|---|---|---|---|
| RF | 302 | 721 | 76 | 41 | 89.74 | 88.05 | 79.89 | 83.78 |
| GA-RF | 354 | 751 | 24 | 11 | 96.93 | 96.99 | 93.65 | 95.29 |
| Bayes-RF | 355 | 752 | 23 | 10 | 97.11 | 97.26 | 93.92 | 95.56 |
As demonstrated from the metric assignments, both the Bayes-RF model and the GA-RF model exhibit better fit and higher prediction accuracy compared to the RF model. However, overall, the training and prediction results of the Bayes-RF model are better than those of the GA-RF model. Therefore, in the model application of this study, the Bayes-RF model has higher predictive power than the GA-RF model.
Based on the predicted mountain hazard susceptibility map of the study area and historical mountain hazard occurrences, the model of this study was quantitatively analyzed and compared in terms of frequency ratios. When the frequency ratio is greater than 1, the mountain hazards are prone to occur, and the larger the frequency ratio, the higher the likelihood of mountain hazards. The results of hazard points in Table 4 show that there are 251 hazard points in the very high susceptibility area of the Bayes-RF model, accounting for 66.40% of the total hazards, and 254 hazard points in the very high susceptibility area of the GA-RF model, accounting for 67.20% of the total hazards. There are 1 and 2 hazard points in the very low susceptibility area of the Bayes-RF and GA-RF models, accounting for 0.26% and 0.53% of the total hazards, respectively. However, the very low susceptibility area of the Bayes-RF model (0.26%) was smaller than that of the GA-RF model (0.53%). The frequency ratio (FR) results in Table 4 showed that the high susceptibility region (1.977) and low susceptibility region (0.136) of the Bayes-RF model were larger than the high susceptibility region (1.519) and low susceptibility region (0.117) of the GA-RF model. The very high susceptibility region (8.266) and very low susceptibility region (0.008) of the Bayes-RF model were smaller than the very high susceptibility region of the GA-RF model (8.273) and very low susceptibility (0.017) of the GA-RF model. The above statistical results together indicate that the susceptibility assessment results and the mountain hazard distribution pattern of the Bayes-RF model are more reasonable than those of the GA-RF model.
| Model | Class | Area of zones (%) | Number of hazards | Hazards percentage (%) | FR |
|---|---|---|---|---|---|
| RF | Extremely high | 9.36 | 227 | 60.05 | 6.415 |
| High | 10.57 | 73 | 19.31 | 1.827 | |
| Moderate | 16.38 | 45 | 11.90 | 0.727 | |
| Low | 29.30 | 23 | 6.08 | 0.208 | |
| Extremely low | 34.39 | 10 | 2.65 | 0.077 | |
| GA-RF | Extremely high | 8.12 | 254 | 67.20 | 8.273 |
| High | 12.54 | 72 | 19.05 | 1.519 | |
| Moderate | 17.35 | 36 | 9.52 | 0.549 | |
| Low | 31.64 | 14 | 3.70 | 0.117 | |
| Extremely low | 30.35 | 2 | 0.53 | 0.017 | |
| Bayes-RF | Extremely high | 8.03 | 251 | 66.40 | 8.266 |
| High | 11.37 | 85 | 22.49 | 1.977 | |
| Moderate | 17.19 | 26 | 6.88 | 0.400 | |
| Low | 29.13 | 15 | 3.97 | 0.136 | |
| Extremely low | 34.27 | 1 | 0.26 | 0.008 |
This paper combines machine learning models (data-driven) with field survey (knowledge-driven) research, which is a beneficial attempt to explore human-machine complementarity toward the urgent need to improve disaster prediction accuracy in the future. Based on the physical mechanism of mutual transformation and cascading of various types of geological hazards in the study area, a multi-factor dataset was established from the perspective of collapse, landslide and mudslide hazard chains, and the obtained results of mountain hazard susceptibility assessment were generally consistent with the results of historical geological hazard surveys. The team's long-term and continuous analysis of the causes of mountain hazards showed that human activities are increasingly becoming the primary driver of disaster occurrence. Therefore, in this section, we will further discuss the model performance through an in-depth analysis, based on exploring how the results of data analysis in this paper reflect the impact of human activities, and then propose the future research objective of the team to improve the prediction accuracy of the model.
5.1 Susceptibility responses to human activities
The importance of the influencing factors obtained from the Relief-F model further reveals that elevation and NDVI are important factors that characterize the association between disaster occurrence and human activities in the study area. For this purpose, the frequency ratios of disaster distribution at different elevations and vegetation ranges are plotted as shown in Fig. 9.
As shown in Fig. 9a, the frequency ratio at elevation 918 ~ 1214 m is the highest, reaching 4.62, and gradually decreases as the elevation rises. Also as shown in Appendix A, 65.35% of the hazards occurred in the low elevation area of elevation 918 ~ 1510 m. Frequent human activities in low altitude area such as large area of reclamation led to low vegetation cover, which in turn caused serious impacts on slope stability and frequent hazard development. Relatively speaking, human activities are less in the high-altitude areas and fewer hazards occur in the mountains. The results in Fig. 9b show that 84 hazards occurred in the range of vegetation cover 0.18–0.27 with a frequency ratio as high as 5.79. The frequency ratio decreases with the increase of vegetation cover. Meanwhile, 81.2% of the hazards occurred in the range of low vegetation cover (Appendix A), indicating that human activities-triggered destruction of vegetation is positively related to hazard frequency.
5.2 Model Optimization and performance improvement
Hyperparameters of machine learning algorithms can affect the accuracy of model prediction (Sameen et al. 2020; Rong et al. 2020), and previous studies have typically used trial-and-error and grid search methods to determine hyperparameters (Chen et al. 2017; Pham et al. 2019). However, using such methods may miss the best parameters and consume a lot of time, which greatly reduces the accuracy and stability of model prediction. Therefore, the Bayesian optimization algorithm was created to find the best hyperparameters rapidly. The algorithm fits the posterior distribution of the objective function by increasing the data volume of samples, input and output data, and finally achieves hyperparameter optimization of the model to obtain the best parameters, while ensuring the accuracy and stability of the model. GA is the most popular evolutionary algorithm, highlighting the properties in the biological evolutionary process. Most of the previous algorithms are heuristic models built up based on different ideal assumptions, which cannot guarantee the basic local optimum. However, GA basically has no prerequisite assumptions and only needs to design the value of the objective function, which is very less relying on the nature of the optimization problem itself and does not depend on a large set of conditions. From the results of the RF model before and after optimization shown in Fig. 7, it can be seen that the model after using both optimization algorithms has a good fit on training and the accuracy has been improved. In comparison, the Bayesian algorithm hyperparametric optimization model is more capable of preventing the overfitting problem. Therefore, the accuracy of fitting and prediction can be improved to a certain extent, and the results are more robust and reliable in the hazards susceptibility assessment.
In addition to the parameters of the model itself, which can affect the final accuracy, the strategy of collecting disaster samples is equally important. There are numerous sampling methods for hazard data, such as seed cell, Point, Scarp, and fuzzy C-means clustering (Yilmaz 2010; Che et al. 2012; Alimohammadlou et al. 2014). Comparison of the effects of scarp, seed cell, and point sampling strategies on landslide susceptibility partition maps concluded that the scarp sampling strategy gave better results than the point sampling strategy, while the scarp and seed cell methods gave similar results. Dou et al. (2020) derived the order of predictive power by comparing landslide slope surface samples, slope centroids, landslide body samples and landslide body centroids as landslide surface > landslide body > slope centroids > body centroids, which further demonstrated that the sampling strategy of disasters also has a large impact on the prediction ability of the model. Therefore, in future research, not only the influence of model parameters but also the sampling strategy of hazards needs to be taken into consideration. In addition, the sample number of the hazards also affects the prediction accuracy of the model. Adequate disaster samples can make the trained model more stable and the prediction results more accurate. The sample data of mountain hazards used in this study were obtained from field geological hazard surveys, geological hazard data collection library and remote sensing image verification. However, individual hazard samples may be missed in the collection process, resulting in incomplete hazard lists. Therefore, the interferometric synthetic aperture radar (InSAR) technique can be considered in future studies to identify hazards (Zhao et al. 2019) to compensate for the missing hazard data and make the evaluation results more reliable.
5.3 Assessment on susceptibility mapping accuracy
In disaster susceptibility studies, the model performance of different models is generally compared by AUC value, but the differences in susceptibility maps cannot be identified distinctly in visualization. Therefore, the susceptibility maps were compared before and after optimization according to the research method of Xiao et al. (2020). The optimized susceptibility maps (GA-RF and Bayes-RF) were used to subtract the susceptibility map without optimization (benchmark RF), and the results are shown in Fig. 10.
The values of RF_Bayes-RF are within the range of -0.42 to 0.42 (Fig. 10a), and the values of RF_GA-RF are within the range of -0.52 to 0.69 (Fig. 10b). The comparison indicated that the overall difference of RF_Bayes-RF is small and shows no obvious distribution law, while the difference of RF_GA-RF is obvious. The high values of RF_GA-RF (0.25 ~ 0.69) mainly distributed in river valleys (elevation: 1214 ~ 1806 m), and the low values (-0.52 ~ 0.25) distributed in mountain ridges (elevation: 1806 ~ 2649 m) and areas of high human activity. In this paper, we argue that there are two reasons for the large discrepancy in RF_GA-RF. On the one hand, it may be due to the overfitting of the unoptimized RF model that occurred during the training process, making the error of the model-predicted susceptibility map larger. Another reason may be that factors such as altitude, human activities, and distance to rivers are associated with extremely high and low values of hazard susceptibility, implying the importance of geomorphological meaning and other underlying patterns in susceptibility assessments. Accordingly, it further revealed that the Bayes-RF model is quite robust, while the GA-RF seems to still have much room for optimization, so when conducting susceptibility prediction studies, different state-of-the-art models and metrics besides AUC are suggested to be selected for comparison as much as possible to further guarantee the reliability of the susceptibility mapping.
This study is based on GIS technology and machine learning methods for mountain hazard susceptibility prediction and assessment. Two integrated machine learning models, GA-RF model and Bayes-RF model, were developed using RF model as the benchmark classifier. The results showed that the Bayes-RF model (AUC = 0.92) was significantly better than the GA-RF model (AUC = 0.91) and the RF model (AUC = 0.89) when used to hazard susceptibility prediction and mapping. The machine learning-based ensemble modeling method of Baye-RF operated robustly and performed well, which can be acted as an adopted model for spatial prediction of mountain hazard susceptibility in the study area. The results of the susceptibility mapping indicate that the densely populated areas along the Bailong River and the G212 highway in the northwest part of the study area should be the focus of disaster prevention and mitigation planning.
Based on the physical mechanism of mountain hazards mainly characterized as cascading landslides, collapses and debris flows chains and the complex nonlinear characteristics of hazard occurrence, this study established a more reliable dataset by selecting 14 hazard impact factors with the aid of field survey analysis and consequently enhanced the accuracy of Bayes-RF model by multicollinearity analysis and Relief-F two-step evaluation. The data preparation method, the corresponding hazard susceptibility assessment and the prediction method can be extended and applied to other mountainous areas susceptible to hazards. Moreover, this study proposes the future developing objective of exploring the establishment of a hybrid model that integrates physical mechanism model and data-driven model, in order to further expand the application of machine learning and even deep learning technologies in multiple fields of scientific assessment and prediction on earth systems.
Confict of Interest The authors declare no competing interests.
Acknowledgments This work was supported by the National Natural Science Foundation of China (Nos. 42167043 and 51269009), Gansu Province Key Research and Development Program (No. 20YF8ND141) and the Western Transportation Construction Science and Technology Project of the Ministry of Transport, China (No. 2002 318 000 36). The funding body does not play any role in the design of the study, collection and analysis of data and decision to publish.
Author Contributions Zhijun Wang: Conceptualization, Data curation, Methodology, Writing – original draft preparation, Writing – review & editing, Funding acquisition. Zhuofan Chen: Modeling & computation, Methodology, Writing – original draft preparation. Ke Ma: Mapping. Zuoxiong Zhang: Field survey. All authors read and approved the fnal manuscript.
Funding This study was supported by the National Natural Science Foundation of China (Nos. 42167043 and 51269009), Gansu Province Key Research and Development Program (No. 20YF8ND141) and the Western Transportation Construction Science and Technology Project of the Ministry of Transport, China (No. 2002 318 000 36).
Availability of Data and Code The datasets used or analysed during the current study are available from the corresponding author on reasonable request.
- Aditian A, Kubota T, Shinohara Y (2018) Comparison of GIS-based landslide susceptibility models using frequency ratio, logistic regression, and artificial neural network in a tertiary region of Ambon, Indonesia. Geomorphology 318:101–111. https://doi.org/doi.org/10.1016/j.geomorph.2018.06.006
- Alimohammadlou Y, Najafi A, Gokceoglu C (2014) Estimation of rainfall-induced landslides using ANN and fuzzy clustering methods: A case study in Saeen Slope, Azerbaijan province, Iran. Catena 120:149–162. https://doi.org/doi.org/10.1016/j.catena.2014.04.009
- Bennett GL, Miller SR, Roering JJ, Schmidt DA (2016) Landslides, threshold slopes, and the survival of relict terrain in the wake of the Mendocino Triple Junction. Geology 18:363–366. https://doi.org/doi.org/10.1130/g37530.1
- Beven KJ, Kirby M (1979) A Physically Based, Variable Contributing Area Model of Basin Hydrology. https://doi.org/doi.org/10.1080/02626667909491834
- Breiman L (2001) Random Forests. Mach Learn. https://doi.org/10.1023/A:1010933404324
- Breiman L (1996) Bagging Predictors. Mach Learn. https://doi.org/doi.org/10.1023/A:1018054314350
- Bui DT, Pradhan B, Lofman O, et al (2012) Landslide susceptibility mapping at Hoa Binh province (Vietnam) using an adaptive neuro-fuzzy inference system and GIS. Comput Geosci. https://doi.org/10.1016/j.cageo.2011.10.031
- Che VB, Kervyn M, Suh CE, et al (2012) Landslide susceptibility assessment in Limbe (SW Cameroon): A field calibrated seed cell and information value method. CATENA 92:83–98. https://doi.org/10.1016/j.catena.2011.11.014
- Chen W, Panahi M, Tsangaratos P, et al (2019) Applying population-based evolutionary algorithms and a neuro-fuzzy system for modeling landslide susceptibility. CATENA 172:212–231. https://doi.org/10.1016/j.catena.2018.08.025
- Chen W, Xie X, Wang J, et al (2017) A comparative study of logistic model tree, random forest, and classification and regression tree models for spatial prediction of landslide susceptibility. CATENA 151:147–160. https://doi.org/10.1016/j.catena.2016.11.032
- Chen W, Zhang S (2021) GIS-based comparative study of Bayes network, Hoeffding tree and logistic model tree for landslide susceptibility modeling. CATENA 203:105344. https://doi.org/10.1016/j.catena.2021.105344
- Chen W, Zhang S, Li R, Shahabi H (2018) Performance evaluation of the GIS-based data mining techniques of best-first decision tree, random forest, and naïve Bayes tree for landslide susceptibility modeling. Sci Total Environ 644:1006–1018. https://doi.org/10.1016/j.scitotenv.2018.06.389
- Chen W, Zhao Z, Liu G, et al (2006) Research on engineering geology of gansu section of lanzhou-haikou expressway. Lanzhou University Press, Lanzhou
- Conforti M, Pascale S, Robustelli G, Sdao F (2014) Evaluation of prediction capability of the artificial neural networks for mapping landslide susceptibility in the Turbolo River catchment (northern Calabria, Italy). CATENA 113:236–250. https://doi.org/10.1016/j.catena.2013.08.006
- Dai FC, Lee CF, Ngai YY (2002) Landslide risk assessment and management: an overview. Eng Geol 64:65–87. https://doi.org/10.1016/S0013-7952(01)00093-X
- Dou J, Yunus AP, Merghadi A, et al (2020) Different sampling strategies for predicting landslide susceptibilities are deemed less consequential with deep learning. Sci Total Environ 720:137320. https://doi.org/10.1016/j.scitotenv.2020.137320
- Dou J, Yunus AP, Tien Bui D, et al (2019) Assessment of advanced random forest and decision tree algorithms for modeling rainfall-induced landslide susceptibility in the Izu-Oshima Volcanic Island, Japan. Sci Total Environ 662:332–346. https://doi.org/10.1016/j.scitotenv.2019.01.221
- Fallah-Zazuli M, Vafaeinejad A, Alesheykh AA, et al (2019) Mapping landslide susceptibility in the Zagros Mountains, Iran: a comparative study of different data mining models. Earth Sci Inform 12:615–628. https://doi.org/10.1007/s12145-019-00389-w
- Fang K, Wu J, Zhu J, Xie B (2011) A review of technologies on random forests. Stat Inf Forum 26:32–38. https://doi.org/10.3969/j.issn.1007-3116.2011.03.006
- Froude MJ, Petley DN (2018) Global fatal landslide occurrence from 2004 to 2016. Nat Hazards Earth Syst Sci 18:2161–2181. https://doi.org/10.5194/nhess-18-2161-2018
- Ghahramani Z (2015) Probabilistic machine learning and artificial intelligence. Nature 521:452–459. https://doi.org/10.1038/nature14541
- Hong H, Jaafari A, Zenner EK (2019) Predicting spatial patterns of wildfire susceptibility in the Huichang County, China: An integrated model to analysis of landscape indicators. Ecol Indic 101:878–891. https://doi.org/10.1016/j.ecolind.2019.01.056
- Huang J, Zhou Q, Wang F (2015) Mapping the landslide susceptibility in Lantau Island, Hong Kong, by frequency ratio and logistic regression model. Ann GIS 21:191–208. https://doi.org/10.1080/19475683.2014.992373
- Huang R, Fan X (2013) The landslide story. Nat Geosci 6:325–326. https://doi.org/10.1038/ngeo1806
- J. Pašek (1975) Inventaire Des Glissements de Terrain. Bull Int Assoc Eng Geol 12:73–74. https://doi.org/10.1007/BF02635432
- Kalantar B, Pradhan B, Naghibi SA, et al (2018) Assessment of the effects of training data selection on the landslide susceptibility mapping: a comparison between support vector machine (SVM), logistic regression (LR) and artificial neural networks (ANN). Geomat Nat Hazards Risk 9:49–69. https://doi.org/10.1080/19475705.2017.1407368
- Kornejady A, Ownegh M, Bahremand A (2017) Landslide susceptibility assessment using maximum entropy model with two different data sampling methods. CATENA 152:144–162. https://doi.org/10.1016/j.catena.2017.01.010
- Lee S, Ryu J-H, Won J-S, Park H-J (2004) Determination and application of the weights for landslide susceptibility mapping using an artificial neural network. Eng Geol 71:289–302. https://doi.org/10.1016/S0013-7952(03)00142-X
- Li L, Lan H, Guo C, et al (2017) A modified frequency ratio method for landslide susceptibility assessment. Landslides 14:727–741. https://doi.org/10.1007/s10346-016-0771-x
- Li S (1997) Discussion on landslide activities in bailong river basin of wudu. Bull Soil Water Conserv 6:28–32. https://doi.org/doi.org/10.13961/j.cnki.stbctb.1997.06.005
- Li S (1995) Preliminary study of the landslides in middle reaches of Bailong River. Hydrogeol Eng Geol 6:13–15. https://doi.org/doi.org/10.1007/s10346-016-0771-x
- Liu Y (2003) Risk analysis and zoning of geological hazards (chiefly landslide,rock fall and debris flow) in China. Chin J Geol Hazard Control 14:95–99. https://doi.org/doi.org/10.3969/j.issn.1003-8035.2003.01.020
- Ma J, Wang X, Zhang P, Qi S (2015) Geological hazard and risk analysis of landslides and debris flows in the bailong river basin. Lanzhou University Press, Lanzhou
- Mo H, Luo F, Hou X (2001) An analysis and improvement of the shortcomings of the common real number coding scheme in genetic multilayer feedforward neural networks. Control Theory Appl 18:281–284
- Nefeslioglu HA, Sezer E, Gokceoglu C, et al (2010) Assessment of Landslide Susceptibility by Decision Trees in the Metropolitan Area of Istanbul, Turkey. Math Probl Eng-01-10 2010:242–256. https://doi.org/10.1155/2010/901095
- Nhu V-H, Shirzadi A, Shahabi H, et al (2020) Shallow Landslide Susceptibility Mapping by Random Forest Base Classifier and Its Ensembles in a Semi-Arid Region of Iran. Forests 11:421. https://doi.org/10.3390/f11040421
- Ning N, Tian L, Zhang P, et al (2013) The hazards assessment of debris flow in wudu of southern gansu, china. https://doi.org/10.16089/j.cnki.1008-2786.2013.05.013
- Obled Ch, Wendling J, Beven K (1994) The sensitivity of hydrological models to spatial rainfall patterns: an evaluation using observed data. J Hydrol 159:305–333. https://doi.org/10.1016/0022-1694(94)90263-1
- Park S, Kim J (2019) Landslide Susceptibility Mapping Based on Random Forest and Boosted Regression Tree Models, and a Comparison of Their Performance. Appl Sci 9:942. https://doi.org/10.3390/app9050942
- Pham BT, Jaafari A, Prakash I, Bui DT (2019) A novel hybrid intelligent model of support vector machines and the MultiBoost ensemble for landslide susceptibility modeling. Bull Eng Geol Environ 78:2865–2886. https://doi.org/10.1007/s10064-018-1281-y
- Polykretis C, Chalkias C (2018) Comparison and evaluation of landslide susceptibility maps obtained from weight of evidence, logistic regression, and artificial neural network models. Nat Hazards 93:249–274. https://doi.org/10.1007/s11069-018-3299-7
- Pourghasemi HR, Rahmati O (2018) Prediction of the landslide susceptibility: Which algorithm, which precision? CATENA 162:177–192. https://doi.org/10.1016/j.catena.2017.11.022
- Qi S, Zhang Y, Zhang P (2014) An assessment index system for landslide risk in bailong river basin. J Yangtze River Sci Res Inst 31:23–28. https://doi.org/10.3969/j.issn.1001-5485.2014.01.005
- Qiao J, Zhao Y (2001) Review on rick degree regionalization of landslide. J Mt Sci 19:157–160. https://doi.org/10.3969/j.issn.1008-2786.2001.02.012
- Rne F, Chen W, Han W (2008) Study on reason and spatial-temporal distribution characteristics of debris flow in longnan area along g212. Chin J Rock Mech Eng 3237–3243
- Rong G, Alu S, Li K, et al (2020) Rainfall Induced Landslide Susceptibility Mapping Based on Bayesian Optimized Random Forest and Gradient Boosting Decision Tree Models—A Case Study of Shuicheng County, China. Water 12:3066. https://doi.org/10.3390/w12113066
- Sameen MI, Pradhan B, Lee S (2020) Application of convolutional neural networks featuring Bayesian optimization for landslide susceptibility assessment. CATENA 186:104249. https://doi.org/10.1016/j.catena.2019.104249
- Santacana N, Baeza B, Corominas J, et al (2003) A GIS-Based Multivariate Statistical Analysis for Shallow Landslide Susceptibility Mapping in La Pobla de Lillet Area (Eastern Pyrenees, Spain). Nat Hazards 30:281–295. https://doi.org/10.1023/B:NHAZ.0000007169.28860.80
- Sheng M, Liu Z, Zhang X, et al (2021) Landslide susceptibility prediction based on frequency ratio analysis and support vector machine. Sci Technol Eng 21:10620–10628. https://doi.org/10.3969/j.issn.1671-1815.2021.25.009
- Snoek J, Larochelle H, Adams RP (2012) Practical Bayesian Optimization of Machine Learning Algorithms. In: Advances in Neural Information Processing Systems 25, vol. 4: 26th annual conference on Neural Information Processing Systems 2012, December 3-6, 2012, Lake Tahoe, Nevada, USA. Deren of Couer Science Universiy of Torono, Lake Tahoe, NV(US), pp 2951–2959
- Sun D, Wen H, Wang D, Xu J (2020) A random forest model of landslide susceptibility mapping based on hyperparameter optimization using Bayes algorithm. Geomorphology 362:107201. https://doi.org/10.1016/j.geomorph.2020.107201
- Tarboton DG (1997) A new method for the determination of flow directions and upslope areas in grid digital elevation models. Water Resour Res 33:309–319. https://doi.org/10.1029/96WR03137
- Tien Bui D, Tuan TA, Klempe H, et al (2016) Spatial prediction models for shallow landslide hazards: a comparative assessment of the efficacy of support vector machines, artificial neural networks, kernel logistic regression, and logistic model tree. Landslides 13:361–378. https://doi.org/10.1007/s10346-015-0557-6
- Trigila A, Iadanza C, Esposito C, Scarascia-Mugnozza G (2015) Comparison of Logistic Regression and Random Forests techniques for shallow landslide susceptibility assessment in Giampilieri (NE Sicily, Italy). Geomorphology 249:119–136. https://doi.org/10.1016/j.geomorph.2015.06.001
- Wang Z, Brenning A (2021) Active-Learning Approaches for Landslide Mapping Using Support Vector Machines. Remote Sens 13:2588. https://doi.org/10.3390/rs13132588
- Wu X, Li K (2006) The application of random forests for the classification of gene expression data. Chin J Health Stat 491–494. https://doi.org/10.3969/j.issn.1002-3674.2006.06.004
- Wu Y, Ke Y, Chen Z, et al (2020) Application of alternating decision tree with AdaBoost and bagging ensembles for landslide susceptibility mapping. CATENA 187:104396. https://doi.org/10.1016/j.catena.2019.104396
- Xia Y, Liu C, Li Y, Liu N (2017) A boosted decision tree approach using Bayesian hyper-parameter optimization for credit scoring. Expert Syst Appl 78:225–241. https://doi.org/10.1016/j.eswa.2017.02.017
- Xiao C, Tian Y, Shi W, et al (2010) A new method of pseudo absence data generation in landslide susceptibility mapping with a case study of Shenzhen. Sci China Technol Sci 53:75–84. https://doi.org/10.1007/s11431-010-3219-x
- Xiao T, Segoni S, Chen L, et al (2020) A step beyond landslide susceptibility maps: a simple method to investigate and explain the different outcomes obtained by different approaches. Landslides 17:627–640. https://doi.org/10.1007/s10346-019-01299-0
- Yilmaz I (2010) The effect of the sampling strategies on the landslide susceptibility mapping by conditional probability and artificial neural networks. Environ Earth Sci 60:505–519. https://doi.org/10.1007/s12665-009-0191-5
- Y.J K, D T (1992) Atmospherically resistant vegetation index (ARVI) for EOS-MODIS. IEEE Trans Geosci Remote Sens 30:261–270. https://doi.org/doi.org/10.1109/36.134076
- Zhang C, Ni Z (2007) Comparative study for ensemble learning algorithms. J Hebei Univ 27:551–554. https://doi.org/10.3969/j.issn.1000-1565.2007.05.025
- Zhang Y, Qu Y, He F, et al (2005) Research on the formation mechanism of debris flow on slope around ningshan county in the hinterland of the qinling orogen. Hydrogeol Eng Geol 32:84–88. https://doi.org/10.3969/j.issn.1000-3665.2005.05.022
- Zhao F, Meng X, Zhang Y, et al (2019) Landslide Susceptibility Mapping of Karakorum Highway Combined with the Application of SBAS-InSAR Technology. Sensors 19:2685. https://doi.org/10.3390/s19122685
- Zhao M, Liu X (2016) Regional risk assessment for urban major hazards based on GIS geoprocessing to improve public safety. Saf Sci 87:18–24. https://doi.org/10.1016/j.ssci.2016.03.016
- Zhu C, Zhang S (2014) Gis based analysis of accessibility impact of brt and metro line in hefei, china. Geogr Geo-Inf Sci 30:21-24,30,封3. https://doi.org/10.3969/j.issn.1672-0504.2014.06.005
- Zuo X, Zhang Z, Su Y, et al (2020) Extraction algorithm of ndvi based on gpu multi-stream parallel model. Comput Sci 47:25–29
No competing interests reported.