逻辑回归和人工神经网络模型在滑坡灾害空间预测中的应用

以三峡坝区到巴东段为研究区,选择坡度、坡向、软弱夹层、水系影响范围和土地利用状况等9项评价指标,分别采用逻辑回归和人工神经网络(ANN)模型,在ArcGIS平台上进行滑坡灾害危险性预测。此外,应用受试者工作特征曲线(receiver operating characteristic curve,ROC曲线)分析方法对两种模型的预测结果进行对比,分析结果表明滑坡危险性预测区划图和实际的滑坡发育情况基本吻合,逻辑回归模型和ANN模型的ROC曲线下面积AUC值分别为0.806和0.799,两种模型的预测结果可以相互验证,且逻辑回归模型的预测精度相对较高。     

Comparison of landslide susceptibility mapping methodologies for Koyulhisar,Turkey: conditional probability,logistic regression,artificial neural networks,and support vector machine

This case study presented herein compares the GIS-based landslide susceptibility mapping methods such as conditional probability (CP), logistic regression (LR), artificial neural networks (ANNs) and support vector machine (SVM) applied in Koyulhisar (Sivas, Turkey). Digital elevation model was first constructed using GIS software. Landslide-related factors such as geology, faults, drainage system, topographical elevation, slope angle, slope aspect, topographic wetness index, stream power index, normalized difference vegetation index, distance from settlements and roads were used in the landslide susceptibility analyses. In the last stage of the analyses, landslide susceptibility maps were produced from ANN, CP, LR, SVM models, and they were then compared by means of their validations. However, area under curve values obtained from all four methodologies showed that the map obtained from ANN model looks like more accurate than the other models, accuracies of all models can be evaluated relatively similar. The results also showed that the CP is a simple method in landslide susceptibility mapping and highly compatible with GIS operating features. Susceptibility maps can be easily produced using CP, because input process, calculation and output processes are very simple in CP model when compared with the other methods considered in this study.     

A comparative study of sequential minimal optimization-based support vector machines,vote feature intervals,and logistic regression in landslide susceptibility assessment using GIS

Landslide susceptibility assessment using GIS has been done for part of Uttarakhand region of Himalaya (India) with the objective of comparing the predictive capability of three different machine learning methods, namely sequential minimal optimization-based support vector machines (SMOSVM), vote feature intervals (VFI), and logistic regression (LR) for spatial prediction of landslide occurrence. Out of these three methods, the SMOSVM and VFI are state-of-the-art methods for binary classification problems but have not been applied for landslide prediction, whereas the LR is known as a popular method for landslide susceptibility assessment. In the study, a total of 430 historical landslide polygons and 11 landslide affecting factors such as slope angle, slope aspect, elevation, curvature, lithology, soil, land cover, distance to roads, distance to rivers, distance to lineaments, and rainfall were selected for landslide analysis. For validation and comparison, statistical index-based methods and the receiver operating characteristic curve have been used. Analysis results show that all these models have good performance for landslide spatial prediction but the SMOSVM model has the highest predictive capability, followed by the VFI model, and the LR model, respectively. Thus, SMOSVM is a better model for landslide prediction and can be used for landslide susceptibility mapping of landslide-prone areas.     

A case study using support vector machines, neural networks and logistic regression in a GIS to identify wells contaminated with nitrate-N

Accurate and inexpensive identification of potentially contaminated wells is critical for water resources protection and management. The objectives of this study are to 1) assess the suitability of approximation tools such as neural networks (NN) and support vector machines (SVM) integrated in a geographic information system (GIS) for identifying contaminated wells and 2) use logistic regression and feature selection methods to identify significant variables for transporting contaminants in and through the soil profile to the groundwater. Fourteen GIS derived soil hydrogeologic and landuse parameters were used as initial inputs in this study. Well water quality data (nitrate-N) from 6,917 wells provided by Florida Department of Environmental Protection (USA) were used as an output target class. The use of the logistic regression and feature selection methods reduced the number of input variables to nine. Receiver operating characteristics (ROC) curves were used for evaluation of these approximation tools. Results showed superior performance with the NN as compared to SVM especially on training data while testing results were comparable. Feature selection did not improve accuracy; however, it helped increase the sensitivity or true positive rate (TPR). Thus, a higher TPR was obtainable with fewer variables.     

Machine learning predictive models for mineral prospectivity: An evaluation of neural networks,random forest,regression trees and support vector machines

Machine learning algorithms (MLAs) such us artificial neural networks (ANNs), regression trees (RTs), random forest (RF) and support vector machines (SVMs) are powerful data driven methods that are relatively less widely used in the mapping of mineral prospectivity, and thus have not been comparatively evaluated together thoroughly in this field.The performances of a series of MLAs, namely, artificial neural networks (ANNs), regression trees (RTs), random forest (RF) and support vector machines (SVMs) in mineral prospectivity modelling are compared based on the following criteria: i) the accuracy in the delineation of prospective areas; ii) the sensitivity to the estimation of hyper-parameters; iii) the sensitivity to the size of training data; and iv) the interpretability of model parameters. The results of applying the above algorithms to epithermal Au prospectivity mapping of the Rodalquilar district, Spain, indicate that the RF outperformed the other MLA algorithms (ANNs, RTs and SVMs). The RF algorithm showed higher stability and robustness with varying training parameters and better success rates and ROC analysis results. On the other hand, all MLA algorithms can be used when ore deposit evidences are scarce. Moreover the model parameters of RF and RT can be interpreted to gain insights into the geological controls of mineralization.     

Groundwater-level prediction using multiple linear regression and artificial neural network techniques: a comparative assessment

The potential of multiple linear regression (MLR) and artificial neural network (ANN) techniques in predicting transient water levels over a groundwater basin were compared. MLR and ANN modeling was carried out at 17 sites in Japan, considering all significant inputs: rainfall, ambient temperature, river stage, 11 seasonal dummy variables, and influential lags of rainfall, ambient temperature, river stage and groundwater level. Seventeen site-specific ANN models were developed, using multi-layer feed-forward neural networks trained with Levenberg-Marquardt backpropagation algorithms. The performance of the models was evaluated using statistical and graphical indicators. Comparison of the goodness-of-fit statistics of the MLR models with those of the ANN models indicated that there is better agreement between the ANN-predicted groundwater levels and the observed groundwater levels at all the sites, compared to the MLR. This finding was supported by the graphical indicators and the residual analysis. Thus, it is concluded that the ANN technique is superior to the MLR technique in predicting spatio-temporal distribution of groundwater levels in a basin. However, considering the practical advantages of the MLR technique, it is recommended as an alternative and cost-effective groundwater modeling tool.     

Modelling shallow landslide susceptibility: a new approach in logistic regression by using favourability assessment

A new method for estimating shallow landslide susceptibility by combining Geographical Information System (GIS), nonparametric kernel density estimation and logistic regression is described. Specifically, a logistic regression is applied to predict the spatial distribution by estimating the probability of occurrence of a landslide in a 16 km² area. For this purpose, a GIS is employed to gather the relevant sample information connected with the landslides. The advantages of pre-processing the explanatory variables by nonparametric density estimation (for continuous variables) and a reclassification (for categorical/discrete ones) are discussed. The pre-processing leads to new explanatory variables, namely, some functions which measure the favourability of occurrence of a landslide. The resulting model correctly classifies 98.55% of the inventaried landslides and 89.80% of the landscape surface without instabilities. New data about recent shallow landslides were collected in order to validate the model, and 92.20% of them are also correctly classified. The results support the methodology and the extrapolation of the model to the whole study area (278 km²) in order to obtain susceptibility maps.     

An assessment on the use of logistic regression and artificial neural networks with different sampling strategies for the preparation of landslide susceptibility maps

The main purpose of this study is to highlight the conceptual differences of produced susceptibility models by applying different sampling strategies: from all landslide area with depletion and accumulation zones and from a zone which almost represents pre-failure conditions. Variations on accuracy and precision values of the models constructed considering different algorithms were also investigated. For this purpose, two most popular techniques, logistic regression analysis and back-propagation artificial neural networks were taken into account. The town Ispir and its close vicinity (Northeastern part of Turkey), suffered from landsliding for many years was selected as the application site of this study. As a result, it is revealed that the back-propagation artificial neural network algorithms overreact to the samplings in which the presence (1) data were taken from the landslide masses. When the generalization capacities of the models are taken into consideration, these reactions cause imprecise results, even though the area under curve (AUC) values are very high (0.915 < AUC < 0.949). On the other hand, the susceptibility maps, based on the samplings in which the presence (1) data were taken from a zone which almost represents pre-failure conditions constitute more realistic susceptibility evaluations. However, considering the spatial texture of the final susceptibility values, the maps produced using the outputs of the back-propagation artificial neural networks could be interpreted as highly optimistic, while of those generated using the resultant probabilities of the logistic regression equations might be evaluated as pessimistic. Consequently, it is evident that, there are still some needs for further investigations with more realistic validations and data to find out the appropriate accuracy and precision levels in such kind of landslide susceptibility studies.     

Combining landslide susceptibility maps obtained from frequency ratio,logistic regression,and artificial neural network models using ASTER images and GIS

Landslide-related factors were extracted from Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) images, and integrated techniques were developed, applied, and verified for the analysis of landslide susceptibility in Boun, Korea, using a geographic information system (GIS). Digital elevation model (DEM), lineament, normalized difference vegetation index (NDVI), and land-cover factors were extracted from the ASTER images for analysis. Slope, aspect, and curvature were calculated from a DEM topographic database. Using the constructed spatial database, the relationships between the detected landslide locations and six related factors were identified and quantified using frequency ratio (FR), logistic regression (LR), and artificial neural network (ANN) models. These relationships were used as factor ratings in an overlay analysis to create landslide susceptibility indices and maps. Three landslide susceptibility maps were then combined and applied as new input factors in the FR, LR, and ANN models to make improved susceptibility maps. All of the susceptibility maps were verified by comparison with known landslide locations not used for training the models. The combined landslide susceptibility maps created using three landslide-related input factors showed improved accuracy (87.00% in FR, 88.21% in LR, and 86.51% in ANN models) compared to the individual landslide susceptibility maps (84.34% in FR, 85.40% in LR, and 74.29% in ANN models) generated using the six factors from the ASTER images.     

Delineation of landslide hazard areas on Penang Island, Malaysia, by using frequency ratio, logistic regression, and artificial neural network models

This paper summarizes findings of landslide hazard analysis on Penang Island, Malaysia, using frequency ratio, logistic regression, and artificial neural network models with the aid of GIS tools and remote sensing data. Landslide locations were identified and an inventory map was constructed by trained geomorphologists using photo-interpretation from archived aerial photographs supported by field surveys. A SPOT 5 satellite pan sharpened image acquired in January 2005 was used for land-cover classification supported by a topographic map. The above digitally processed images were subsequently combined in a GIS with ancillary data, for example topographical (slope, aspect, curvature, drainage), geological (litho types and lineaments), soil types, and normalized difference vegetation index (NDVI) data, and used to construct a spatial database using GIS and image processing. Three landslide hazard maps were constructed on the basis of landslide inventories and thematic layers, using frequency ratio, logistic regression, and artificial neural network models. Further, each thematic layer’s weight was determined by the back-propagation training method and landslide hazard indices were calculated using the trained back-propagation weights. The results of the analysis were verified and compared using the landslide location data and the accuracy observed was 86.41, 89.59, and 83.55% for frequency ratio, logistic regression, and artificial neural network models, respectively. On the basis of the higher percentages of landslide bodies predicted in very highly hazardous and highly hazardous zones, the results obtained by use of the logistic regression model were slightly more accurate than those from the other models used for landslide hazard analysis. The results from the neural network model suggest the effect of topographic slope is the highest and most important factor with weightage value (1.0), which is more than twice that of the other factors, followed by the NDVI (0.52), and then precipitation (0.42). Further, the results revealed that distance from lineament has the lowest weightage, with a value of 0. This shows that in the study area, fault lines and structural features do not contribute much to landslide triggering.     

Performance evaluation of a physically based model for shallow landslide prediction

Evaluating the performance of a physically based model for landslide prediction was conducted in this study. The model was developed based on the basis of the infinite slope instability analysis and TOPMODEL for saturated water level estimation, which enabled to predict the location and time of occurrence of shallow landslides. Field data from 2008 to 2013 in two areas vulnerable to landslide in Taiwan were collected to test the applicability of the model for landslide prediction. Three indexes including the probability of detection (POD), false alarm ratio (FAR), and threat score (TS) were adopted to assess the advantages and disadvantages of the model. The results indicated that the POD for the landslide prediction by using the proposed model was 1.00, the FAR was lower than 0.25, and the overall TS value was higher than 0.75. It is promising to apply the proposed model for landslide early warnings to reduce the loss of life and property.     

Spatial prediction of landslide susceptibility in western Serbia using hybrid support vector regression (SVR) with GWO,BAT and COA algorithms

In this study, we developed multiple hybrid machine-learning models to address parameter optimization limitations and enhance the spatial prediction of landslide susceptibility models. We created a geographic information system database, and our analysis results were used to prepare a landslide inventory map containing 359 landslide events identified from Google Earth, aerial photographs, and other validated sources. A support vector regression (SVR) machine-learning model was used to divide the landslide inventory into training (70%) and testing (30%) datasets. The landslide susceptibility map was produced using 14 causative factors. We applied the established gray wolf optimization (GWO) algorithm, bat algorithm (BA), and cuckoo optimization algorithm (COA) to fine-tune the parameters of the SVR model to improve its predictive accuracy. The resultant hybrid models, SVR-GWO, SVR-BA, and SVR-COA, were validated in terms of the area under curve (AUC) and root mean square error (RMSE). The AUC values for the SVR-GWO (0.733), SVR-BA (0.724), and SVR-COA (0.738) models indicate their good prediction rates for landslide susceptibility modeling. SVR-COA had the greatest accuracy, with an RMSE of 0.21687, and SVR-BA had the least accuracy, with an RMSE of 0.23046. The three optimized hybrid models outperformed the SVR model (AUC = 0.704, RMSE = 0.26689), confirming the ability of metaheuristic algorithms to improve model performance.     

Logistic regression versus artificial neural networks: landslide susceptibility evaluation in a sample area of the Serchio River valley,Italy

This article presents a multidisciplinary approach to landslide susceptibility mapping by means of logistic regression, artificial neural network, and geographic information system (GIS) techniques. The methodology applied in ranking slope instability developed through statistical models (conditional analysis and logistic regression), and neural network application, in order to better understand the relationship between the geological/geomorphological landforms and processes and landslide occurrence, and to increase the performance of landslide susceptibility models. The proposed experimental study concerns with a wide research project, promoted by the Tuscany Region Administration and APAT-Italian Geological Survey, aimed at defining the landslide hazard in the area of the Sheet 250 “Castelnuovo di Garfagnana” (1:50,000 scale). The study area is located in the middle part of the Serchio River basin and is characterized by high landslide susceptibility due to its geological, geomorphological, and climatic features, among the most severe in Italy. Terrain susceptibility to slope failure has been approached by means of indirect-quantitative statistical methods and neural network software application. Experimental results from different methods and the potentials and pitfalls of this methodological approach have been presented and discussed. Applying multivariate statistical analyses made it possible a better understanding of the phenomena and quantification of the relationship between the instability factors and landslide occurrence. In particular, the application of a multilayer neural network, equipped for supervised learning and error control, has improved the performance of the model. Finally, a first attempt to evaluate the classification efficiency of the multivariate models has been performed by means of the receiver operating characteristic (ROC) curves analysis approach.     

Evaluation of dynamic regression and artificial neural networks models for real-time hydrological drought forecasting

In this study, application of a class of stochastic dynamic models and a class of artificial intelligence model is reported for the forecasting of real-time hydrological droughts in the Black River basin in the USA. For this purpose, the Standardized Hydrological Drought Index (SHDI) was adopted in different time scales to represent the hydrological drought index. Six probability distribution functions (PDF) were fitted to the discharge time series to obtain the best fit for SHDI calculation. Then, a dynamic linear spatio-temporal model (DLSTM) and artificial neural network (ANN) were used to forecast SHDI. Although results indicated that both models were able to forecast SHDI in different time scales, the DLSTM was far superior in longer lead times. The DLSTM could forecast SHDI up to 6 months ahead while ANN was only capable of forecasting SHDI up to 2 months ahead appropriately. For short lead times (1–6 months), the DLSTM has performed nearly perfect in test phase and CE oscillates between 0.97 and 0.86 while for ANN modeling, CE is between 0.72 and 0.07. However, the performance of DLSTM and ANN reduced considerably in medium lead times (7–12 months). Overall, the DLSTM is a powerful tool for appropriately forecasting SHDI at short time scales; a major advantage required for drought early warning systems.     

Mapping landslide susceptibility with logistic regression, multiple adaptive regression splines, classification and regression trees, and maximum entropy methods: a comparative study

Four statistical techniques for modelling landslide susceptibility were compared: multiple logistic regression (MLR), multivariate adaptive regression splines (MARS), classification and regression trees (CART), and maximum entropy (MAXENT). According to the literature, MARS and MAXENT have never been used in landslide susceptibility modelling, and CART has been used only twice. Twenty independent variables were used as predictors, including lithology as a categorical variable. Two sets of random samples were used, for a total of 90 model replicates (with and without lithology, and with different proportions of positive and negative data). The model performance was evaluated using the area under the receiver operating characteristic curve (AUC) statistic. The main results are (a) the inclusion of lithology improves the model performance; (b) the best AUC values for single models are MLR (0.76), MARS (0.76), CART (0.77), and MAXENT (0.78); (c) a smaller amount of negative data provides better results; (d) the models with the highest prediction capability are obtained with MAXENT and CART; and (e) the combination of different models is a way to evaluate the model reliability. We further discuss some key issues in landslide modelling, including the influence of the various methods that we used, the sample size, and the random replicate procedures.     

基于正交设计下SVM滑坡变形时序回归预测的超参数选择

超参数的选择直接影响着支持向量机（SVM）的泛化性能和回归效验,是确保SVM优秀性能的关键。针对超参数穷举搜索方法的难点,从试验设计的角度,提出了正交设计超参选择方法,并分析了基于混合核函数（比单一核函数具有更好的收敛性和模型适应性）SVM各个超参数的取值范围,选定了每个参数的试验水平。通过考虑参数间的正交性和交互性,选取最优超参数组合下的SVM模型。应用该方法,对两种典型滑坡位移时序的SVM建模进行了超参数组合正交优化设计,获得了精度高且泛化性能良好的滑坡预测模型,其试验结果验证了方法的可靠性。正交设计超参选择方法较之其他超参选择法简单实用,其高时效的特点更有助于SVM在实践工程中的良好应用。