LANDSLIDE SUSCEPTIBILITY ASSESSMENT OF THE 2014 LUDIAN EARTHQUAKE: COMPARATIVE STUDY BASED ON CONVOLUTIONAL NEURAL NETWORK AND RANDOM FOREST

doi:10.3969/j.issn.0253-4967.20240139

Abstract

Abstract:

Landslides are among the most common geological hazards worldwide, causing severe environmental damage and significant socio-economic losses in affected areas. In China, geological disasters are characterized by wide spatial distribution, high frequency, strong destructiveness, and substantial losses. In particular, landslides triggered by strong earthquakes have profoundly impacted the socio-economic development of southwestern China. Ludian County in Yunnan Province is located within the Sichuan-Yunnan region, where both seismic activity and intense rainfall pose serious hazards. The region’s steep topography, fractured rock masses, abundant precipitation, and active faulting contribute to frequent landslides and debris flows. Therefore, establishing a rapid and accurate landslide susceptibility assessment system is of great importance for improving disaster prevention and risk management and for reducing losses associated with earthquake-induced landslides.
Using the 2014 Ludian earthquake as a case study, a database of coseismic landslides was established through visual interpretation of ultra-high-resolution satellite imagery from the Google Earth platform. Three-dimensional landslide and non-landslide samples were extracted based on multiple influencing factors. A rapid landslide susceptibility assessment model, AlexNetCBAM, was developed based on a convolutional neural network(CNN), and a Random Forest(RF)model was employed for comparison. Model performance was quantitatively evaluated using statistical indicators, including receiver operating characteristic(ROC)curves, area under the curve(AUC), and F1-score, to verify the reliability and applicability of the models in the Ludian area.
Ten influencing factors, including elevation, slope angle, slope aspect, seismic intensity, distance to faults, and lithology, were selected. The Gini index was used to quantify the contribution of each factor to landslide occurrence. The results indicate that distance to faults, seismic intensity, and peak ground acceleration(PGA)exhibit the strongest predictive capability, suggesting that seismic factors exert a dominant control on landslide occurrence in the Ludian earthquake area compared with topographic and geological conditions. Removing highly correlated and low-importance factors reduces redundancy and improves data quality, providing a solid foundation for model development.
A comparison of the landslide susceptibility maps(LSMs)generated by the two models shows that areas of high susceptibility closely correspond to observed landslide distributions. Extremely high and high susceptibility zones are mainly concentrated in two regions: the southeastern valley area far from the epicenter, and the zone extending from the BXF fault to the Niulan River along a west-east direction on the western side of the epicenter. These areas should be prioritized for post-earthquake monitoring, early warning, and emergency response.
Comparative analysis of the ROC curves and related statistical metrics indicates that the prediction accuracies of the AlexNetCBAM and RF models are 0.84 and 0.82, respectively, with corresponding AUC values of 0.91 and 0.90. These results demonstrate that the AlexNetCBAM model slightly outperforms the RF model in predicting coseismic landslides in the Ludian area. Correlation analysis of landslide susceptibility index(LSI)values between the two models reveals a linear relationship of LSICNN=0.82*LSIRF, indicating that the CNN-based model more effectively classifies non-landslide pixels into low-susceptibility zones and landslide pixels into high and very high susceptibility zones. Compared with the traditional RF model, the CNN model exhibits stronger predictive capability, highlighting the adaptability and advantages of deep learning methods in earthquake-induced landslide susceptibility assessment.
This study enables rapid and accurate identification of coseismic landslide-prone areas and provides a scientific basis for emergency response planning, including rescue team deployment and resource allocation, thereby helping to reduce casualties and property losses.

Key words: Ludian M_W7.8(M_S6.5)earthquake, coseismic landslide, Convolutional Neural Network, Random Forest, landslide susceptibility assessment

摘要：

地震滑坡作为一种破坏性极强的地质灾害, 不仅对灾区居民生命安全构成威胁, 还对当地基础设施和交通网络造成严重破坏。因此, 快速、准确地识别同震滑坡分布并评估其易发性, 对于决策者及时部署应急响应、合理配置资源及最大限度减少人员伤亡和财产损失具有至关重要的意义。文中以2014年 M_W6.2 鲁甸地震为例, 基于卷积神经网络(CNN)构建了AlexNetCBAM滑坡易发性评估模型, 并采用随机森林(RF)模型作为对比工具, 开展了鲁甸震区滑坡易发性评估。研究综合选取了高程、坡度、坡向、地震烈度、断层距离、地层岩性等10个影响因子, 并通过GINI指数定量分析了各影响因子对滑坡发生的贡献程度。结果表明, 相较于地貌特征和地质条件, 地震因素对鲁甸地震滑坡的控制作用最为显著。进一步比较2种模型预测的滑坡敏感性图(LSM), 发现滑坡高易发区的分布与实际滑坡分布情况基本一致, 鲁甸地震极高和高危险区主要集中在2个区域: 1)距震中最远的东南部河谷地区; 2)震中西侧沿EW向从BXF断层至牛栏河的区域。根据定量评价指标可知, AlexNetCBAM模型和RF模型的准确率分别为0.84和0.82, AUC值分别为0.91和0.90, 表明AlexNetCBAM模型在鲁甸地区同震滑坡预测中的表现略优于RF模型。

关键词: 鲁甸M_W6.2地震, 同震滑坡, 卷积神经网络, 随机森林, 易发性评估

WANG Wan-tong, MA Si-yuan, YAN Wu-jian, YUAN Ren-mao. LANDSLIDE SUSCEPTIBILITY ASSESSMENT OF THE 2014 LUDIAN EARTHQUAKE: COMPARATIVE STUDY BASED ON CONVOLUTIONAL NEURAL NETWORK AND RANDOM FOREST[J]. SEISMOLOGY AND GEOLOGY, 2026, 48(2): 497-519.

王莞瞳, 马思远, 严武建, 袁仁茂. 2014年 M_W6.2 鲁甸地震滑坡易发性评估:基于卷积神经网络与随机森林算法的对比[J]. 地震地质, 2026, 48(2): 497-519.

Figures/Tables 16

Fig. 1 Map showing the geographical location of study area(a), and the elevation, historical earthquakes, and tectonic setting of study area(b).

Fig. 2 Lithology distribution of the study area.

Fig. 3 The distribution of coseismic landslides triggered by the 2014 Ludian earthquake.

Fig. 4 The distribution characteristics of landslide in the study area.

Fig. 5 The distribution of the influencing factors in the study area.

Fig. 6 Diagram showing the principle of AlexNetCBAM model.

Fig. 7 Schematic diagram of the random forest model.

Table 1 VIF value and GI value of the influencing factors

影响因子	重要性(GINI)	VIF
与断层的距离	0.222	2.856
地震烈度	0.216	3.018
PGA	0.132	1.821
与河流的距离	0.111	3.405
NDVI	0.088	2.273
高程	0.074	3.194
坡度	0.052	1.250
地层岩性	0.039	2.131
坡向	0.037	1.031
土地利用	0.029	1.076

Fig. 8 The correlation coefficient matrix for the influencing factors.

Fig. 9 The predicted results of the occurrence probability of landslides based on two different models.

Fig. 10 Map of landslides susceptibility class for two different models using AlexNetCBAM.

Fig. 11 The statistical result of the predicted area and LND, LAD, landslide number under the landslides susceptibility class for two different models.

Fig. 12 The ROC curve and AUC value.

Table 2 Accuracy value, precision value, recall value and F1-score for two different models

模型	TP	TN	FN	FP	准确率	精确度	召回率	F1-score
AlexNetCBAM	2883	2441	726	284	0.841	0.910	0.799	0.851
RF	2941	2307	860	226	0.829	0.929	0.774	0.844

Fig. 13 The relationship of LSI between AlexNetCBAM and RF-based landslide susceptibility assessment results.

Fig. 14 Comparison of local landslide distribution in the occurrence probabilistic map.

References 58

[1]	杜鹏, 陈宁生, 伍康林, 等. 2024. 基于随机森林模型的藏东南地区滑坡易发性评价及主控因素分析[J]. 成都理工大学学报(自然科学版), 51(2): 328—344.
	DU Peng, CHEN Ning-sheng, WU Kang-lin, et al. 2024. Evaluation of landslide susceptibility in southeast Tibet based on a random forest model[J]. Journal of Chengdu University of Technology(Science & Technology Edition), 51(2): 328—344 (in Chinese).
[2]	侯沙沙. 2021. 注意力引导卷积神经网络的高分辨率遥感影像分类方法研究[D]. 西安: 西安科技大学.
	HOU Sha-sha. 2021. High resolution remote sensing image classification based on attention guided convolutional neural network[D]. Xi'an University of Science and Technology, Xi'an (in Chinese).
[3]	孔嘉旭, 庄建琦, 彭建兵, 等. 2023. 基于信息量和卷积神经网络的黄土高原滑坡易发性评价[J]. 地球科学, 48(5): 1711—1729.
	KONG Jia-xu, ZHUANG Jian-qi, PENG Jian-bing, et al. 2023. Evaluation of landslide susceptibility in Chinese Loess Plateau based on Ⅳ-RF and Ⅳ-CNN coupling models[J]. Earth Science, 48(5): 1711—1729 (in Chinese).
[4]	王世宝, 庄建琦, 郑佳, 等. 2022. 基于深度学习的CZ铁路康定—理塘段滑坡易发性评价[J]. 工程地质学报, 30(3): 908—919.
	WANG Shi-bao, ZHUANG Jian-qi, ZHENG Jia, et al. 2022. Landslide susceptibility evaluation based on deep learning along Kangding-Litang section of CZ railway[J]. Journal of Engineering Geology, 30(3): 908—919 (in Chinese).
[5]	吴玮莹, 许冲. 2018. 2014年鲁甸 M_W6.2 地震触发滑坡新编目[J]. 地震地质, 40(5): 1140—1148.
	WU Wei-ying, XU Chong. 2018. A new inventory of landslides triggered by the 2014 Ludian M_W6.2 earthquake[J]. Seismology and Geology, 40(5): 1140—1148 (in Chinese).
[6]	许冲. 2015. 利用同震滑坡分析2014年鲁甸地震震源性质与破裂过程[J]. 工程地质学报, 23(4): 755—759.
	XU Chong. 2015. Utilizing coseismic landslides to analyze the source and rupturing process of the 2014 Ludian earthquake[J]. Journal of Engineering Geology, 23(4): 755—759 (in Chinese).
[7]	许冲, 徐锡伟, 沈玲玲, 等. 2014. 2014年鲁甸 M_S6.5 地震触发滑坡编录及其对一些地震参数的指示[J]. 地震地质, 36(4): 1186—1203.
	XU Chong, XU Xi-wei, SHEN Ling-ling, et al. 2014. Inventory of landslides triggered by the 2 014 M_S6.5 Ludian earthquake and its implications on several earthquake parameters[J]. Seismology and Geology, 36(4): 1186—1203 (in Chinese).
[8]	杨硕, 李德营, 严亮轩, 等. 2021. 基于随机森林模型的乌江高陡岸坡滑坡地质灾害易发性评价[J]. 安全与环境工程, 28(4): 131—138.
	YANG Shuo, LI De-ying, YAN Liang-xuan, et al. 2021. Landslide susceptibility assessment in high and steep bank slopes along Wujiang River based on random forest model[J]. Safety and Environmental Engineering, 28(4): 131—138 (in Chinese).
[9]	Ai X, Sun B, Chen X. 2022. Construction of small sample seismic landslide susceptibility evaluation model based on Transfer Learning: A case study of Jiuzhaigou earthquake[J]. Bulletin of Engineering Geology and the Environment, 81(3): 116.
[10]	Amato G, Fiorucci M, Martino S, et al. 2023. Earthquake-triggered landslide susceptibility in Italy by means of Artificial Neural Network[J]. Bulletin of Engineering Geology and the Environment, 82(5): 160.
[11]	Breiman L. 2001. Random forests[J]. Machine Learning, 45(1): 5—32.
[12]	Chen X, Zhou Q, Liu C. 2015. Distribution pattern of coseismic landslides triggered by the 2014 Ludian, Yunnan, China M_W6.1 earthquake: Special controlling conditions of local topography[J]. Landslides, 12(6): 1159—1168.
[13]	Chen Z, Liang S, Ke Y, et al. 2019. Landslide susceptibility assessment using different slope units based on the evidential belief function model[J]. Geocarto International, 35(15): 1641—1664.
[14]	Cutler A, Cutler D R, Stevens J R. 2012. Random Forests [G]//Zhang C, Ma Y Q(eds). Ensemble Machine Learning: Methods and Applications. Springer, Berlin: 157—175.
[15]	Cutler D R, Edwards Jr T C, Beard K H, et al. 2007. Random forests for classification in ecology[J]. Ecology, 88(11): 2783—2792.
[16]	Dao D V, Jaffari A, Bayat M, et al. 2020. A spatially explicit deep learning neural network model for the prediction of landslide susceptibility[J]. Catena, 188:104451.
[17]	Dou J, Yunus A P, Merghadi A, et al. 2020. Different sampling strategies for predicting landslide susceptibilities are deemed less consequential with deep learning[J]. Science of the Total Environment, 720(10): 137320.
[18]	Dreyfus D, Rathje E M, Jibson R W. 2013. The influence of different simplified sliding-block models and input parameters on regional predictions of seismic landslides triggered by the Northridge earthquake[J]. Engineering Geology, 163: 41—54.
[19]	Fang Z, Wang Y, Peng L, et al. 2020. Integration of convolutional neural network and conventional machine learning classifiers for landslide susceptibility mapping[J]. Computers & Geosciences, 139:104470.
[20]	Fang Z, Wang Y, Van Westen C, et al. 2024. Landslide hazard spatiotemporal prediction based on data-driven models: Estimating where, when and how large landslide may be[J]. International Journal of Applied Earth Observation and Geoinformation, 126:103631.
[21]	Jessee A, Hamburger M, Allstadt K, et al. 2018. A global empirical model for near-real-time assessment of seismically induced landslides[J]. Journal of Geophysical Research: Earth Surface, 123(8): 1835—1859.
[22]	Jibson R W. 2011. Methods for assessing the stability of slopes during earthquakes: A retrospective[J]. Engineering Geology, 122(1-2): 43—50.
[23]	Kavzoglu T, Sahin E K, Colkesen I. 2013. Landslide susceptibility mapping using GIS-based multi-criteria decision analysis, support vector machines, and logistic regression[J]. Landslides, 11(3): 425—439.
[24]	Larsen I J, Montgomery D R, Korup O. 2010. Landslide erosion controlled by hillslope material[J]. Nature Geoscience, 3(4): 247—251.
[25]	Lee C T. 2014. Statistical seismic landslide hazard analysis: An example from Taiwan[J]. Engineering Geology, 182: 201—212.
[26]	Li B, Friedman J, Olshen R A, et al. 1984. Classification and Regression Trees(CART)[J]. Biometrics, 40(3): 358.
[27]	Meena S R, Ghorbanzadeh O, Blaschke T. 2019. A comparative study of statistics-based landslide susceptibility models: A case study of the region affected by the Gorkha earthquake in Nepal[J]. ISPRS International Journal of Geo-Information, 8(2): 94.
[28]	Mutlu B, Nefeslioglu H A, Sezer E A, et al. 2019. An experimental research on the use of recurrent neural networks in landslide susceptibility mapping[J]. ISPRS International Journal of Geo-Information, 8(12): 578.
[29]	Nhu V H, Hoang N D, Nguyen H, et al. 2020. Effectiveness assessment of Keras based deep learning with different robust optimization algorithms for shallow landslide susceptibility mapping at tropical area[J]. Catena, 188:104458.
[30]	Reichenbach P, Rossi M, Malamud B D, et al. 2018. A review of statistically-based landslide susceptibility models[J]. Earth-Science Reviews, 180: 60—91.
[31]	Rodríguez-Peces M J, Pérez-García J L, García-Mayordomo J, et al. 2011. Applicability of Newmark method at regional, sub-regional and site scales: Seismically induced Bullas and La Paca rock-slide cases(Murcia, SE Spain)[J]. Natural Hazards, 59(2): 1109—1124.
[32]	Sameen M I, Pradhan B, Lee S. 2020. Application of convolutional neural networks featuring Bayesian optimization for landslide susceptibility assessment[J]. Catena, 186:104249.
[33]	Shao X Y, Xu C. 2022. Earthquake-induced landslides susceptibility assessment: A review of the state-of-the-art[J]. Natural Hazards Research, 2(3): 172—182.
[34]	Stumpf A, Kerle N. 2011. Object-oriented mapping of landslides using Random Forests[J]. Remote Sensing of Environment, 115(10): 2564—2577.
[35]	Swets J A. 1988. Measuring the accuracy of diagnostic systems[J]. Science, 240(4857): 1285—1293.
[36]	Tanyas H, Rossi M, Alvioli M, et al. 2019. A global slope unit-based method for the near real-time prediction of earthquake-induced landslides[J]. Geomorphology, 327: 126—146.
[37]	Tharwat A. 2021. Classification assessment methods[J]. Applied Computing and Informatics, 17(1): 168—192.
[38]	Tien Bui D, Tuan T A, 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[J]. Landslides, 13(2): 361—378.
[39]	Wang F, Fan X, Yunus A P, et al. 2019. Coseismic landslides triggered by the 2018 Hokkaido, Japan(M_W6.6), earthquake: Spatial distribution, controlling factors, and possible failure mechanism[J]. Landslides, 16(8): 1551—1566.
[40]	Wang J, Jin W, Cui Y F, et al. 2018. Earthquake-triggered landslides affecting a UNESCO Natural Site: The 2017 Jiuzhaigou earthquake in the World National Park, China[J]. Journal of Mountain Science, 15(7): 1412—1428.
[41]	Wang W, He Z, Han Z, et al. 2020. Mapping the susceptibility to landslides based on the deep belief network: A case study in Sichuan Province, China[J]. Natural Hazards, 103(3): 3239—3261.
[42]	Wang Y, Fang Z, Hong H. 2019. Comparison of convolutional neural networks for landslide susceptibility mapping in Yanshan County, China[J]. Science of The Total Environment, 666: 975—993.
[43]	Wang Y, Fang Z, Wang M, et al. 2020. Comparative study of landslide susceptibility mapping with different recurrent neural networks[J]. Computers & Geosciences, 138:104445.
[44]	Wang Z, Zhao D, Wang J. 2010. Deep structure and seismogenesis of the north-south seismic zone in southwest China[J]. Journal of Geophysical Research: Solid Earth, 115(B12): 334.
[45]	Wei R, Ye C, Sui T, et al. 2022. Combining spatial response features and machine learning classifiers for landslide susceptibility mapping[J]. International Journal of Applied Earth Observation and Geoinformation, 107:102681.
[46]	Wu W, Xu C, Wang X, et al. 2020. Landslides triggered by the 3 August 2014 Ludian(China)M_W6.2 earthquake: An updated inventory and analysis of their spatial distribution[J]. Journal of Earth Science, 31(4): 853—866.
[47]	Wu X, Zhang Z, Zhang W, et al. 2021. A convolutional neural network based on grouping structure for scene classification[J]. Remote Sensing, 13(13): 2457.
[48]	Xu X, Xu C, Yu G, et al. 2015. Primary surface ruptures of the Ludian M_W6.2 earthquake, southeastern Tibetan plateau, China[J]. Seismological Research Letters, 86(6): 1622—1635.
[49]	Yang J, Huang X. 2021. The 30m annual land cover dataset and its dynamics in China from 1990 to 2019[J]. Earth System Science Data, 13(8): 3907—3925.
[50]	Yang X, Liu R, Yang M, et al. 2021. Incorporating landslide spatial information and correlated features among conditioning factors for landslide susceptibility mapping[J]. Remote Sensing, 13(11): 2166.
[51]	Yang Z, Xu C. 2022a. Efficient detection of earthquake-triggered landslides based on U-Net++: An example of the 2018 Hokkaido Eastern Iburi(Japan)M_W=6.6 earthquake[J]. Remote Sensing, 14(12): 2826.
[52]	Yang Z, Xu C, Shao X, et al. 2022b. Landslide susceptibility mapping based on CNN -3D algorithm with attention module embedded[J]. Bulletin of Engineering Geology and the Environment, 81(10): 412.
[53]	Yann L, Yoshua B, Geoffrey H. 2015. Deep learning[J]. Nature, 521(7553): 436—444.
[54]	Yao J, Qin S, Qiao S, et al. 2022. Application of a two-step sampling strategy based on deep neural network for landslide susceptibility mapping[J]. Bulletin of Engineering Geology and the Environment, 81(4): 148.
[55]	Yi Y, Zhang Z, Zhang W, et al. 2020. Landslide susceptibility mapping using multiscale sampling strategy and convolutional neural network: A case study in Jiuzhaigou region[J]. Catena, 195:104851.
[56]	Youssef A M, Pradhan B, Dikshit A, et al. 2022. Landslide susceptibility mapping using CNN -1D and 2D deep learning algorithms: Comparison of their performance at Asir Region, KSA[J]. Bulletin of Engineering Geology and the Environment, 81(4): 165.
[57]	Zhou J W, Lu P Y, Hao M H. 2015. Landslides triggered by the 3 August 2014 Ludian earthquake in China: Geological properties, geomorphologic characteristics and spatial distribution analysis[J]. Geomatics, Natural Hazards and Risk, 7(4): 1219—1241.
[58]	Zou Y, Qi S, Guo S, et al. 2022. Factors controlling the spatial distribution of coseismic landslides triggered by the M_W6.1 Ludian earthquake in China[J]. Engineering Geology, 296:106477.