基于机器学习的滑坡易发性预测建模及其主控因子识别

不同机器学习预测滑坡易发性的建模过程及其不确定性有所差异, 另外如何有效识别滑坡易发性的主控因子意义重大。针对上述问题, 以支持向量机(support vector machine, 简称SVM)和随机森林(random forest, 简称RF)为例探讨了基于机器学习的滑坡易发性预测及其不确定性, 创新地提出了"权重均值法"来综合计算出更准确的滑坡主控因子。首先获取陕西省延长县滑坡编录和10类基础环境因子, 将因子频率比值作为SVM和RF的输入变量; 再将滑坡与随机选择的非滑坡样本划分为训练集和测试集, 用训练好的机器学习预测出滑坡易发性并制图; 最后用受试者工作曲线、均值和标准差等来评估建模不确定性, 并计算滑坡主控因子。结果表明: ①机器学习能有效预测出区域滑坡易发性, RF预测的滑坡易发性精度高于SVM, 而其不确定性低于SVM, 但两者的易发性分布规律整体相似; ②权重均值法计算出延长县滑坡主控因子依次是坡度、高程和岩性。实例分析和文献综述显示RF模型相较于其他机器学习模型属于可靠性较高的易发性模型。      

考虑线状环境因子适宜性和不同机器学习模型的滑坡易发性预测建模规律

对于滑坡易发性预测中的水系、公路和断层等线状环境因子, 现有研究大多采用缓冲分析提取距离线状因子的距离。但缓冲分析得到的线距离属于离散型变量, 带有大小不等的随机波动性且对点或线要素的误差较为敏感, 导致滑坡易发性建模精度下降。提出了使用水系和公路的空间密度等连续型变量改进线状环境因子的适宜性。以江西省安远县为例, 选取高程、地形起伏度、距水系和公路距离等14个环境因子(原始因子), 再将距水系和公路距离2个线状因子改进为水系密度和公路密度(改进因子); 之后采用逻辑回归、多层感知器、支持向量机和C5.0决策树等机器学习模型, 分别构建了基于原始因子和改进因子的机器学习模型以预测滑坡易发性; 最后利用ROC曲线和易发性指数分布特征等来研究建模规律。结果表明: ①改进因子机器学习预测精度均高于原始因子机器学习模型, 表明空间密度对于易发性预测的适宜性更好; ②在4类机器学习模型中C5.0模型对于滑坡易发性预测性能最好, 其次是SVM、MLP和LR; ③水系和公路两类环境因子的重要性较高且使用改进因子机器学习后这两类环境因子重要性排名依然非常靠前。      

基于贝叶斯优化机器学习超参数的滑坡易发性评价

利用机器学习模型进行滑坡易发性评价时, 不同的超参数设置往往会导致评价结果的不同。采用贝叶斯算法对4种常见机器学习模型(逻辑回归LR、支持向量机SVM、人工神经网络ANN和随机森林RF)的超参数进行了优化, 探索了该算法对滑坡易发性机器学习模型的优化效果。以湘中地区4县(安化县、新华县、桃江县和桃源县)滑坡易发性评价为例说明该算法的可行性与适用性。基于滑坡历史编录, 确定研究区内1 017个滑坡点, 并选定15个滑坡影响因子, 以此构建滑坡易发性模型的训练集和测试集。利用贝叶斯优化算法对4种机器学习模型的主要超参数进行了优化, 依据优化后的超参数建立了4种优化模型, 并使用AUC值等指标来比较其预测能力。结果表明: 经超参数优化后的4种机器学习模型预测性能均有所提高, 且基于贝叶斯优化的随机森林模型表现最好。      

基于空间多尺度分析的滑坡易发性评价——以三峡库区秭归-巴东段为例

以三峡库区秭归-巴东段为例,将地理加权回归(GWR)模型引入到研究区的空间尺度分割方法中,利用粒子群优化(PSO)算法对支持向量机(SVM)模型参数进行优化,构建GWR-PSO-SVM耦合模型,完成研究区滑坡易发性评价,并与传统的PSO-SVM耦合模型结果进行对比。结果表明,在特定类别精度分析、总体预测精度分析和曲线下面积分析中,本文方法评价效果均优于传统方法。     

确定性系数与随机森林模型在云南芒市滑坡易发性评价中的应用

编制科学的滑坡易发性分区图,可以有效降低灾害带来的损失。以云南省芒市为研究区,利用确定性系数模型（certainty factor,简称CF）方法计算各个因子的敏感值,作为随机森林（random forests,简称RF）的分类数据,选取合适的训练数据和最优化的模型参数进行模型预测,从而对研究区进行滑坡易发性评价分区。采用频率比方法将连续性因子离散化,从而通过确定性系数计算因子不同区间的滑坡易发性,同时利用CF先验模型,对研究区负样本进行选取。通过计算袋外误差得到最优化的RF参数,随后利用RF模型对研究区模型进行训练及预测。绘制ROC曲线和三维遥感影像对预测模型结果分别进行定量和定性评价,结果表明,所得到的模型精度为91%,优于随机抽样得到的结果。最后,采用平均基尼不纯度减少和平均准确度下降两种计算方法计算、评价了研究区各个因子的重要性。基于以上对研究区进行的滑坡易发性评价结果,可以为该区灾害风险评估和管理提供依据。      

基于逻辑回归模型和3S技术的思南县滑坡易发性评价

区域滑坡易发性评价对滑坡灾害防治具有重要意义，贵州省思南县由于其特殊的自然地理和地质条件，受滑坡地质灾害的影响非常严重，因此，非常有必要对思南县的滑坡易发性进行评价。在滑坡编录的基础上，采用由RS、GIS和GPS组成的3S技术，获取了思南县的数字高程模型、坡度、坡向、剖面曲率、坡长、岩土类型、地表湿度指数、距离水系的距离、植被覆盖度和地表建筑物指数10个滑坡影响因子；再在频率比和相关性分析的基础上，利用逻辑回归模型对思南县的滑坡易发性进行了评价并绘制了易发性分布图。结果表明:利用逻辑回归模型预测思南县滑坡易发性的准确率(AUC值)达到0.797，较为准确地预测出了思南县滑坡分布规律；极高和高滑坡易发区主要分布在高程低于600 m、地表坡度较大且以软质岩类为主的区域；而极低和低滑坡易发区主要分布在高程较高、地表坡度较小且以硬质岩类为主的区域。      

综合非滑坡样本选取指数与异质集成机器学习的区域滑坡易发性建模

为解决基于机器学习的滑坡易发性建模存在的单模型分类能力弱和传统随机抽取非滑坡样本准确性不高的问题，本研究以三峡库区奉节县为例，应用优化的非滑坡样本和Stacking异质集成机器学习模型进行滑坡易发性建模研究。首先，基于地形、地质和遥感影像等数据提取16个评价指标并进行相关性分析，剔除高相关指标，构建易发性评价指标体系；其次，基于信息量模型提出非滑坡样本选取（Non-Landslide Sampling, NLS）指数；最后，应用NLS指数选取更高质量的非滑坡样本，并与滑坡样本组成训练集；采用随机森林(Random Forest, RF)，轻量级梯度提升树（Light Gradient Boosting Machine, LGBM），梯度提升决策树（Gradient Boosting Decision Tree, GBDT），以及以三者为基模型的同质（Boosting）和异质（Stacking）集成方法进行易发性建模。结果表明：应用NLS指数能选取得到质量更高的非滑坡样本，提升了易发性建模精度；Stacking异质集成机器学习模型的精度最高，为0.941，优于3个同质集成模型和3个单模型...     

不同日降雨工况下万州区滑坡灾害危险性分析

以三峡库区万州区为例，选择具有代表性的地质环境指标，分析各指标等级，利用逻辑回归、支持向量机和决策树3种数理统计模型，计算全区滑坡灾害易发性程度，分析3种日降雨工况下滑坡的发生概率，得到各日降雨工况下万州区滑坡灾害危险性分布图。确定了支持向量机模型为万州区滑坡灾害易发性分析的最优模型；万州区滑坡灾害高易发区和高危险区主要表现出沿河道水系呈带状分布、沿高程垂直分布、在城镇区集中分布的特点；特定工况下，万州区滑坡灾害危险性随着日降雨量增大而增大。      

基于信息量模型和数据标准化的滑坡易发性评价

本文以北川曲山-擂鼓片区为研究区,将坡度、坡向、高程、地层、距断层的距离、距水系的距离和距道路的距离作为该区域滑坡易发性评价因子。采用信息量模型计算了各项评价因子的信息量值,并运用4种标准化模型对信息量值进行标准化处理。各评价因子的权重由层次分析法（AHP）确定。在GIS中将权重值和各评价因子的标准化信息量值,进行叠加计算得到区域滑坡总信息量值,并基于自然断点法对其进行重分类,将研究区划分为极高易发区、高易发区、中易发区、低易发区和极低易发区5级易发区。将基于4种标准化模型和信息量模型得到的滑坡易发性评价结果进行了对比分析,结果表明：基于最值标准化信息量模型的滑坡易发性评价结果的ROC曲线下面积AUC值为0.807,高于其余模型的AUC值,说明最值标准化信息量模型的滑坡易发性评价效果最好。极高易发区面积占研究区面积的20.03%,离断层和水系较近,主要分布地层为寒武系、志留系和三迭系。研究结果可为区内滑坡风险评价和灾害防治提供参考。     

采用多模型融合方法评价滑坡灾害易发性:以湖北省五峰县为例

不同的易发性评价模型可以得到有差异的滑坡空间预测结果，选取最优模型甚至综合各模型的优势是提高易发性评价精度的有效方法。为检验模型融合思路的有效性，以鄂西地区五峰县渔洋关镇为研究区，提取坡度、地层、断层、河流、公路等7个滑坡成因条件，分别采用信息量模型、证据权模型和频率比模型进行滑坡易发性评价；并将3种模型分别进行归一化、主成分分析（PCA，Principal component analysis）和优势融合，得到了6幅易发性分区图。结果表明:优势耦合模型精度最高（90.3%），频率比模型次之（89.7%），归一化融合模型和PCA融合模型分别为89.3%和89.1%，以上4种结果的精度均高于证据权模型（87.7%）和信息量模型（87.6%）；6幅预测图对应的评价结论与历史滑坡空间分布的实际情况相符。空间一致性对比结论表明，主成分融合模型与优势耦合模型的同格率高达68%，其预测结果避免了单个模型预测结论带来的偶然性和片面性，说明多模型融合方法与优势耦合模型在提高滑坡易发性预测精度上是可行性的，该思路对其他地区滑坡灾害易发性评价具有借鉴意义。      

GIS-based landslide susceptibility mapping using hybrid integration approaches of fractal dimension with index of entropy and support vector machine

The loess area in the northern part of Baoji City, Shaanxi Province, China is a region with frequently landslide occurrences. The main aim of this study is to quantitatively predict the extent of landslides using the index of entropy model(IOE), the support vector machine model(SVM) and two hybrid models namely the F-IOE model and the F-SVM model constructed by fractal dimension. First, a total of 179 landslides were identified and landslide inventory map was produced, with 70%(125) of the landslides which was optimized by 10-fold crossvalidation being used for training purpose and the remaining 30%(54) of landslides being used for validation purpose. Subsequently, slope angle, slope aspect, altitude, rainfall, plan curvature, distance to rivers, land use, distance to roads, distance to faults, normalized difference vegetation index(NDVI), lithology, and profile curvature were considered as landslide conditioning factors and all factor layers were resampled to a uniform resolution. Then the information gain ratio of each conditioning factors was evaluated. Next, the fractal dimension for each conditioning factors was calculated and the training dataset was used to build four landslide susceptibility models. In the end, the receiver operating characteristic(ROC) curves and three statistical indexes involving positive predictive rate(PPR), negative predictive rate(NPR) and accuracy(ACC) were applied to validate and compare the performance of these four models. The results showed that the F-SVM model had the highest PPR, NPR, ACC and AUC values for training and validation datasets, respectively, followed by the F-IOE model.Finally, it is concluded that the F-SVM model performed best in all models, the hybrid model built by fractal dimension has advantages than original model, and can provide reference for local landslide prevention and decision making.     

GIS-based Frequency Ratio and Logistic Regression Modelling for Landslide Susceptibility Mapping of Debre Sina Area in Central Ethiopia

Ethiopia has a mountainous landscape which can be divided into the Northwestern and Southeastern plateaus by the Main Ethiopian Rift and Afar Depression. Debre Sina area is located in Central Ethiopia along the escarpment where landslide problem is frequent due to steep slope, complex geology, rift tectonics, heavy rainfall and seismicity. In order to tackle this problem, preparing a landslide susceptibility map is very important. For this, GISbased frequency ratio(FR) and logistic regression(LR) models have been applied using landslide inventory and the nine landslide factors(i.e. lithology, land use, distance from river fault, slope, aspect, elevation, curvature and annual rainfall). Database construction, weighting each factor classes or factors, preparing susceptibility map and validation were the major steps to be undertaken. Both models require a rasterized landslide inventory and landslide factor maps. The former was classified into training and validation landslides. Using FR model, weights for each factor classes were calculated and assigned so that all the weighted factor maps can be added to produce a landslide susceptibility map. In the case of LR model, the entire study area is firstly divided into landslide and non-landslide areas using the training landslides. Then, these areas are changed into landslide and non-landslide points so as to extract the FR maps of the nine landslide factors. Then a linear relationship is established between training landslides and landslide factors in SPSS. Based on this relationship, the final landslide susceptibility map is prepared using LR equation. The success-rate and prediction-rate of FR model were 74.8% and 73.5%, while in case of LR model these were 75.7% and 74.5% respectively. A close similarity in the prediction and validation rates showed that the model is acceptable. Accuracy of LR model is slightly better in predicting the landslide susceptibility of the area compared to FR model.     

Rapid susceptibility mapping of co-seismic landslides triggered by the 2013 Lushan Earthquake using the regression model developed for the 2008 Wenchuan Earthquake

The primary objective of landslide susceptibility mapping is the prediction of potential landslides in landslide-prone areas.The predictive power of a landslide susceptibility mapping model could be tested in an adjacent area of similar geoenvironmental conditions to find out the reliability.Both the 2008 Wenchuan Earthquake and the 2013 Lushan Earthquake occurred in the Longmen Mountain seismic zone,with similar topographical and geological conditions.The two earthquakes are both featured by thrust fault and similar seismic mechanism.This paper adopted the susceptibility mapping model of co-seismic landslides triggered by Wenchuan earthquake to predict the spatial distribution of landslides induced by Lushan earthquake.Six influencing parameters were taken into consideration: distance from the seismic fault,slope gradient,lithology,distance from drainage,elevation and Peak Ground Acceleration(PGA).The preliminary results suggested that the zones with high susceptibility of coseismic landslides were mainly distributed in the mountainous areas of Lushan,Baoxing and Tianquan counties.The co-seismic landslide susceptibility map was completed in two days after the quake and sent to the field investigators to provide guidance for rescue and relief work.The predictive power of the susceptibility map was validated by ROC curve analysis method using 2037 co-seismic landslides in the epicenter area.The AUC value of 0.710 indicated that the susceptibility model derived from Wenchuan Earthquake landslides showed good accuracy in predicting the landslides triggered by Lushan earthquake.     

Analysing post-earthquake landslide susceptibility using multi-temporal landslide inventories——a case study in Miansi Town of China

The earthquake that occurred on May 12, 2008, in Wenchuan County aroused a great deal of research on co-seismic landslide susceptibility assessment, but there is still a lack of an evaluation method that considers the activity state of the landslide itself. Therefore, this paper establishes a new susceptibility evaluation model that superimposes the active landslide state based on previous susceptibility evaluation models. Based on a multi-phase landslide database, the probabilistic approach was used to evaluate landslide susceptibility in the Miansi town over many years. We chose the elevation, slope, aspect, and distance from the channel as trigger factors and then used the probability comprehensive discrimination method to calculate the probability of landslide occurrence. Then, the susceptibility results of each period were calculated by superposition with the activity rate. The results show that between 2008 and 2014, the proportion of areas with low landslide susceptibility in the study area was the largest, and the proportionof areas with the highest susceptibility was minimal. The landslide area with highest susceptibility gradually decreased from 2014 to 2017. However, in 2017, 15.06% of the area was still with high susceptibility, and relevant disaster prevention and reduction measures should be taken in these areas. The larger area under the receiver operating characteristic curve(AUC) indicates that the results of the landslide susceptibility assessment in this study are more objective and reliable than those of previous models. The difference in the AUC values over many years shows that the accuracy of the evaluation results of this model is not constant, and a greater number of landslides or higher landslide activity corresponds to a higher accuracy of the evaluation results.     

Optimization of causative factors using logistic regression and artificial neural network models for landslide susceptibility assessment in Ujung Loe Watershed,South Sulawesi Indonesia

Landslide susceptibility maps(LSMs) play a vital role in assisting land use planning and risk mitigation. This study aims to optimize causative factors using logistic regression(LR) and an artificial neural network(ANN) to produce a LSM. The LSM is produced with 11 causative factors and then optimized using forward-stepwise LR(FSLR), ANN, and their combination(FSLR-ANN) until eight causative factors were found for each method. The ANN method produced superior validation results compared with LR. The ROC values for the training data set ranges between 0.8 and 0.9. On the other hand, validation with the percentage of landslide fall into LSM class high and very high, ANN method was higher(92.59%) than LR(82.12%). FSLR-ANN with nine causative factors gave the best validation results with respect to area under curve(AUC) values, and validation with the percentage of landslide fall into LSM class high and very high. In conclusion, ANN was found to be better than LR when producing LSMs. The best Optimization was combination of FSLR-ANN with nine causative factors and AUC success rate 0.847, predictive rate 0.844 and validation with landslide fall into high and very high class with 91.30%. It is an encouraging preliminary model towards a systematic introduction of FSLR-ANN model for optimization causative factors in landslide susceptibility assessment in the mountainous area of Ujung Loe Watershed.     

Landslide Susceptibility Mapping along Bhalubang–Shiwapur Area of Mid-Western Nepal Using Frequency Ratio and Conditional Probability Models

Roads constructed in fragile Siwaliks are prone to large number of instabilities. Bhalubang–Shiwapur section of Mahendra Highway lying in Western Nepal is one of them. To understand the landslide causative factor and to predict future occurrence of the landslides, landslide susceptibility mapping(LSM) of this region was carried out using frequency ratio(FR) and weights-of-evidence(W of E) models. These models are easy to apply and give good results. For this, landslide inventory map of the area was prepared based on the aerial photo interpretation, from previously published/unpublished reposts, and detailed field survey using GPS. About 332 landslides were identified and mapped, among which 226(70%) were randomly selected for model training and the remaining 106(30%) were used for validation purpose. A spatial database was constructed from topographic, geological, and land cover maps. The reclassified maps based on the weight values of frequency ratio and weights-of-evidence were applied to get final susceptibility maps. The resultant landslide susceptibility maps were verified andcompared with the training data, as well as with the validation data. From the analysis, it is seen that both the models were equally capable of predicting landslide susceptibility of the region(W of E model(success rate = 83.39%, prediction rate = 79.59%); FR model(success rate = 83.31%, prediction rate = 78.58%)). In addition, it was observed that the distance from highway and lithology, followed by distance from drainage, slope curvature, and slope gradient played major role in the formation of landsides. The landslide susceptibility maps thus produced can serve as basic tools for planners and engineers to carry out further development works in this landslide prone area.     

Regional scale rainfall- and earthquake-triggered landslide susceptibility assessment in Wudu County, China

Wudu County in northwestern China frequently experiences large-scale landslide events.High-magnitude earthquakes and heavy rainfall events are the major triggering factors in the region.The aim of this research is to compare and combine landslide susceptibility assessments of rainfalltriggered and earthquake-triggered landslide events in the study area using Geographical Information System(GIS) and a logistic regression model.Two separate susceptibility maps were produced using inventories reflecting single landslide-triggering events,i.e.,earthquakes and heavy rain storms.Two groups of landslides were utilized: one group containing all landslides triggered by extreme rainfall events between 1995 and 2003 and the other group containing slope failures caused by the 2008 Wenchuan earthquake.Subsequently,the individual maps were combined to illustrate the locations of maximum landslide probability.The use of the resulting three landslide susceptibility maps for landslide forecasting,spatial planning and for developing emergency response actions are discussed.The combined susceptibility map illustrates the total landslide susceptibility in the study area.     

Landslide inventory and susceptibility modelling using geospatial tools,in Hunza-Nagar valley,northern Pakistan

A comprehensive landslide inventory and susceptibility maps are prerequisite for developing and implementing landslide mitigation strategies. Landslide susceptibility maps for the landslides prone regions in northern Pakistan are rarely available. The Hunza-Nagar valley in northern Pakistan is known for its frequent and devastating landslides. In this paper, we have developed a landslide inventory map for Hunza-Nagar valley by using the visual interpretation of the SPOT-5 satellite imagery and mapped a total of 172 landslides. The landslide inventory was subsequently divided into modelling and validation data sets. For the development of landslide susceptibility map seven discrete landslide causative factors were correlated with the landslide inventory map using weight of evidence and frequency ratio statistical models. Four different models of conditional independence were used for the selection of landslide causative factors. The produced landslides susceptibility maps were validated by the success rate and area under curves criteria. The prediction power of the models was also validated with the prediction rate curve. The validation results shows that the success rate curves of the weight of evidence and the frequency models are 82% and 79%, respectively. The prediction accuracy results obtained from this study are 84% for weight of evidence model and 80% for the frequency ratio model. Finally, the landslide susceptibility index maps were classified into five different varying susceptibility zones. The validation and prediction result indicates that the weight of evidence and frequency ratio model are reliable to produce an accurate landslide susceptibility map, which may be helpful for landslides management strategies.     

Co-seismic landslide inventory and susceptibility mapping in the 2008 Wenchuan earthquake disaster area, China

The Ms 8.0 May 12,2008 Wenchuan earthquake triggered tens of thousands of landslides.The widespread landslides have caused serious casualties and property losses,and posed a great threat to post-earthquake reconstruction.A spatial database,inventoried 43,842 landslides with a total area of 632 km 2,was developed by interpretation of multi-resolution remote sensing images.The landslides can be classified into three categories:swallow,disrupted slides and falls;deep-seated slides and falls,and rock avalanches.The correlation between landslides distribution and the influencing parameters including distance from co-seismic fault,lithology,slope gradient,elevation,peak ground acceleration(PGA) and distance from drainage were analyzed.The distance from co-seismic fault was the most significant parameter followed by slope gradient and PGA was the least significant one.A logistic regression model combined with bivariate statistical analysis(BSA) was adopted for landslide susceptibility mapping.The study area was classified into five categories of landslide susceptibility:very low,low,medium,high and very high.92.0% of the study area belongs to low and very low categories with corresponding 9.0% of the total inventoried landslides.Medium susceptible zones make up 4.2% of the area with 17.7% of the total landslides.The rest of the area was classified into high and very high categories,which makes up 3.9% of the area with corresponding 73.3% of the total landslides.Although the susceptibility map can reveal the likelihood of future landslides and debris flows,and it is helpful for the rebuilding process and future zoning issues.