A Stacking-SHAP ensemble method for landslide susceptibility prediction with high accuracy and interpretability

doi:10.11947/j.AGCS.2025.20250139

Abstract

Abstract:

Landslide susceptibility prediction and triggering factor analysis are crucial for developing scientifically effective landslide disaster prevention and control strategies. However, there is still a lack of landslide prediction models that can achieve both high prediction accuracy and interpretability. To address this, the present study proposes an interpretable enhanced ensemble learning method by constructing the Stacking-SHAP model, aimed at improving the accuracy of landslide susceptibility prediction and the reliability of triggering factor analysis. This model employs a Stacking ensemble framework, integrating various machine learning classifiers, such as XGBoost, CatBoost, LightGBM, logistic regression (LR), and random forest (RF), while ensuring prediction accuracy. Additionally, the shapley additive explanations (SHAP) algorithm is incorporated to enhance model interpretability. Experimental results demonstrate that the AUC value of the Stacking-SHAP model reaches 0.920, significantly outperforming individual classifier models, such as XGBoost (0.893), CatBoost (0.894), LightGBM (0.879), RF (0.859), and LR (0.794). More importantly, compared to SHAP-integrated single machine learning models, the Stacking-SHAP interpretable enhanced ensemble model shows superior overall performance in landslide triggering factor analysis, thereby improving the credibility of landslide causative factor analysis. Overall, this model combines high-accuracy prediction with highly reliable interpretation, offering an innovative approach to landslide susceptibility prediction and triggering factor analysis, with significant theoretical and practical value in landslide prevention and disaster reduction.

Key words: landslide susceptibility, geographical big data, Stacking algorithm, SHAP algorithm, landslide triggering factor analysis

CLC Number:

P237

Xin HUANG, Jian YE, Chengbing LIU, Qiuyu ZENG, Wanxin GUO, Zhikai GUO. A Stacking-SHAP ensemble method for landslide susceptibility prediction with high accuracy and interpretability[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54(10): 1826-1840.

Figures/Tables 16

Fig. 1

Fig. 2

Tab. 1

Fig. 3

Tab. 2

Fig. 4

Tab. 3

Fig. 5

Fig. 6

Tab. 4

Fig. 7

Fig. 8

Tab. 5

Fig. 9

Fig. 10

Fig. 11

References 29

[1]	FROUDE M J, PETLEY D N. Global fatal landslide occurrence from 2004 to 2016[J]. Natural Hazards and Earth System Sciences, 2018, 18(8): 2161-2181.
[2]	ZHANG Shuai, LI Can, PENG Jingyu, et al. Fatal landslides in China from 1940 to 2020: occurrences and vulnerabilities[J]. Landslides, 2023, 20(6): 1243-1264.
[3]	GUZZETTI F, REICHENBACH P, CARDINALI M, et al. Probabilistic landslide hazard assessment at the basin scale[J]. Geomorphology, 2005, 72(1/2/3/4): 272-299.
[4]	许强, 朱星, 李为乐, 等. “天-空-地”协同滑坡监测技术进展[J]. 测绘学报, 2022, 51(7): 1416-1436. DOI: . doi: 10.11947/j.AGCS.2022.20220320
	XU Qiang, ZHU Xing, LI Weile, et al. Technical progress of space-air-ground collaborative monitoring of landslide[J]. Acta Geodaetica et Cartographica Sinica, 2022, 51(7): 1416-1436. DOI: . doi: 10.11947/j.AGCS.2022.20220320
[5]	MARJANOVIĆ M, KOVAČEVIĆ M, BAJAT B, et al. Landslide susceptibility assessment using SVM machine learning algorithm[J]. Engineering Geology, 2011, 123(3): 225-234.
[6]	刘纪平, 梁恩婕, 徐胜华, 等. 顾及样本优化选择的多核支持向量机滑坡灾害易发性分析评价[J]. 测绘学报, 2022, 51(10): 2034-2045. DOI: . doi: 10.11947/j.AGCS.2022.20220326
	LIU Jiping, LIANG Enjie, XU Shenghua, et al. Multi-kernel support vector machine considering sample optimization selection for analysis and evaluation of landslide disaster susceptibility[J]. Acta Geodaetica et Cartographica Sinica, 2022, 51(10): 2034-2045. DOI: . doi: 10.11947/j.AGCS.2022.20220326
[7]	王世宝, 庄建琦, 樊宏宇, 等. 基于频率比与集成学习的滑坡易发性评价：以金沙江上游巴塘—德格河段为例[J]. 工程地质学报, 2022, 30(3): 817-828.
	WANG Shibao, ZHUANG Jianqi, FAN Hongyu, et al. Evaluation of landslide susceptibility based on frequency ratio and ensemble learning: taking the Batang-Dege section in the upstream of Jinsha river as an example[J]. Journal of Engineering Geology, 2022, 30(3): 817-828.
[8]	ZHOU Xinzhi, WEN Haijia, ZHANG Yalan, et al. Landslide susceptibility mapping using hybrid random forest with GeoDetector and RFE for factor optimization[J]. Geoscience Frontiers, 2021, 12(5): 101211.
[9]	GORSEVSKI P V, BROWN M K, PANTER K, et al. Landslide detection and susceptibility mapping using LiDAR and an artificial neural network approach: a case study in the Cuyahoga Valley National Park, Ohio[J]. Landslides, 2016, 13(3): 467-484.
[10]	TIEN BUI D, TUAN T A, KLEMPE H, et al. 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, 2016, 13(2): 361-378.
[11]	LOMBARDO L, CAMA M, CONOSCENTI C, et al. Binary logistic regression versus stochastic gradient boosted decision trees in assessing landslide susceptibility for multiple-occurring landslide events: application to the 2009 storm event in Messina (Sicily, southern Italy)[J]. Natural Hazards, 2015, 79(3): 1621-1648.
[12]	MERGHADI A, YUNUS A P, DOU Jie, et al. Machine learning methods for landslide susceptibility studies: a comparative overview of algorithm performance[J]. Earth-Science Reviews, 2020, 207: 103225.
[13]	AL-NAJJAR H A H, PRADHAN B. Spatial landslide susceptibility assessment using machine learning techniques assisted by additional data created with generative adversarial networks[J]. Geoscience Frontiers, 2021, 12(2): 625-637.
[14]	FANG Zhice, WANG Yi, PENG Ling, et al. A comparative study of heterogeneous ensemble-learning techniques for landslide susceptibility mapping[J]. International Journal of Geographical Information Science, 2021, 35(2): 321-347.
[15]	WANG Tianlong, ZHANG Keying, LIU Zhenghua, et al. Prediction and explanation of debris flow velocity based on multi-strategy fusion Stacking ensemble learning model[J]. Journal of Hydrology, 2024, 638: 131347.
[16]	HUANG Faming, CHEN Jiawu, LIU Weiping, et al. Regional rainfall-induced landslide hazard warning based on landslide susceptibility mapping and a critical rainfall threshold[J]. Geomorphology, 2022, 408: 108236.
[17]	LOCHE M, ALVIOLI M, MARCHESINI I, et al. Landslide susceptibility maps of Italy: lesson learnt from dealing with multiple landslide types and the uneven spatial distribution of the national inventory[J]. Earth-Science Reviews, 2022, 232: 104125.
[18]	CHA Y, SHIN J, GO B, et al. An interpretable machine learning method for supporting ecosystem management: application to species distribution models of freshwater macroinvertebrates[J]. Journal of Environmental Management, 2021, 291: 112719.
[19]	WEI Jing, WANG Zhihui, LI Zhanqing, et al. Global aerosol retrieval over land from Landsat imagery integrating Transformer and Google Earth Engine[J]. Remote Sensing of Environment, 2024, 315: 114404.
[20]	董佳奇, 胡冬梅, 闫雨龙, 等. 基于可解释性机器学习的城市O₃驱动因素挖掘[J]. 环境科学, 2023, 44(7): 3660-3668.
	DONG Jiaqi, HU Dongmei, YAN Yulong, et al. Revealing driving factors of urban O₃ based on explainable machine learning[J]. Environmental Science, 2023, 44(7): 3660-3668.
[21]	YANG Changlan, GUAN Xuefeng, XU Qingyang, et al. How can SHAP (shapley additive explanations) interpretations improve deep learning based urban cellular automata model?[J]. Computers, Environment and Urban Systems, 2024, 111: 102133.
[22]	CHU Wenhao, ZHANG Chunxiao, LI Heng, et al. SHAP-powered insights into spatiotemporal effects: unlocking explainable Bayesian-neural-network urban flood forecasting[J]. International Journal of Applied Earth Observation and Geoinformation, 2024, 131: 103972.
[23]	FENG Zezheng, JIANG Yifan, WANG Hongjun, et al. TrafPS: a shapley-based visual analytics approach to interpret traffic[J]. Computational Visual Media, 2024, 10(6): 1101-1119.
[24]	KANG Y, KIM M, KANG E, et al. Improved retrievals of aerosol optical depth and fine mode fraction from GOCI geostationary satellite data using machine learning over East Asia[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2022, 183: 253-268.
[25]	罗路广, 裴向军, 崔圣华, 等. 九寨沟地震滑坡易发性评价因子组合选取研究[J]. 岩石力学与工程学报, 2021, 40(11): 2306-2319.
	LUO Luguang, PEI Xiangjun, CUI Shenghua, et al. Combined selection of susceptibility assessment factors for Jiuzhaigou earthquake-induced landslides[J]. Chinese Journal of Rock Mechanics and Engineering, 2021, 40(11): 2306-2319.
[26]	张钟远, 邓明国, 徐世光, 等. 镇康县滑坡易发性评价模型对比研究[J]. 岩石力学与工程学报, 2022, 41(1): 157-171.
	ZHANG Zhongyuan, DENG Mingguo, XU Shiguang, et al. Comparison of landslide susceptibility assessment models in Zhenkang County, Yunnan Province, China[J]. Chinese Journal of Rock Mechanics and Engineering, 2022, 41(1): 157-171.
[27]	LUNDBERG S M, LEE S I. A unified approach to interpreting model predictions[C]//Proceedings of the 31st Conference on Neural Information Processing Systems. Long Beach: [s.n.], 2017.
[28]	HUANG Faming, MAO Daxiong, JIANG Shuihua, et al. Uncertainties in landslide susceptibility prediction modeling: a review on the incompleteness of landslide inventory and its influence rules[J]. Geoscience Frontiers, 2024, 15(6): 101886.
[29]	OTHMAN A A, GLOAGUEN R, ANDREANI L, et al. Improving landslide susceptibility mapping using morphometric features in the Mawat area, Kurdistan Region, NE Iraq: comparison of different statistical models[J]. Geomorphology, 2018, 319: 147-160.

条件因子	来源	数据采集时间	条件因子	来源	数据采集时间
高程	30 m GDEMV3	2000—2013年	地形湿度指数	30 m GDEMV3	2000—2013年
坡度	30 m GDEMV3	2000—2013年	水流功率指数	30 m GDEMV3	2000—2013年
坡向	30 m GDEMV3	2000—2013年	NDVI	https://www.nesdc.org.cn/	2020年
地形曲率	30 m GDEMV3	2000—2013年	多年平均降水量	https://www.resdc.cn/	2011—2020年
平面曲率	30 m GDEMV3	2000—2013年	土地覆盖类型	https://zenodo.org/	2020年
剖面曲率	30 m GDEMV3	2000—2013年	地层岩性	全国地质资料馆	2002年
地表粗糙度	30 m GDEMV3	2000—2013年	地貌类型	全国地质资料馆	2009年
地表切割深度	30 m GDEMV3	2000—2013年	道路距离	Open Street Map	2023年
地形起伏度	30 m GDEMV3	2000—2013年	河流距离	MapOpen Street	2023年
高程变异系数	30 m GDEMV3	2000—2013年	断层距离	国家地震科学数据中心	2002年

滑坡条件因子	VIF	滑坡条件因子	VIF
坡向	1.179	NDVI	1.737
平面曲率	1.470	多年平均降水量	3.810
剖面曲率	1.472	土地覆盖类型	1.822
地面粗糙度	2.320	地层岩性	1.593
地表切割度	1.919	地貌类型	1.335
高程变异系数	3.421	道路距离	1.996
地形湿度指数	2.962	河流距离	1.782
水流功率指数	2.144	断层距离	1.708

模型名称	模型原理	滑坡易发性预测优势
XGBoost	基于梯度提升决策树算法，通过目标函数优化模型性能，控制过拟合	具有较高的预测精度和泛化能力，可处理复杂非线性关系，在处理复杂地形和多变量数据上具有优势
CatBoost	基于梯度提升决策树框架，使用遗忘树处理类别特征，有效解决梯度和预测偏差问题	高效处理分类变量，提高模型精度，适合类别特征较多的滑坡数据集
LightGBM	基于梯度提升决策树算法，采用直方图分割方法加速训练	具有执行速度快、资源占用少的优势，适用于大规模滑坡数据的预测和分析
RF	基于决策树的集成模型，通过随机抽样生成多个训练子集，构建多棵树并结合输出结果	具有较高的预测精度，能够有效抵抗数据异常值和噪声，适合复杂环境下的滑坡预测
LR	广义线性模型，通过逻辑变换预测事件发生的概率	适用于滑坡易发性评价的二分类问题，能够快速评估滑坡影响因子对滑坡发生的影响

模型	Accuracy	AUC	Precision	Recall	F₁值	MCC	IOA
Stacking	0.896	0.920	0.901	0.890	0.895	0.791	0.895
XGBoost	0.893	0.893	0.896	0.891	0.894	0.787	0.893
CatBoost	0.894	0.894	0.897	0.891	0.894	0.788	0.894
LightGBM	0.879	0.879	0.891	0.866	0.878	0.759	0.879
RF	0.859	0.859	0.871	0.843	0.857	0.718	0.859
LR	0.711	0.794	0.718	0.697	0.708	0.422	0.711

指标	Stacking-SHAP	XGBoost-SHAP	CatBoost-SHAP	LightGBM-SHAP	RF-SHAP	LR-SHAP
特征贡献值标准差	0.020	0.335	0.348	0.020	0.022	0.233
滑坡样本解释方差	0.023	1.416	1.254	0.023	0.006	0.152
非滑坡样本解释方差	0.014	0.850	0.850	0.014	0.014	0.034
RMSE	0.323	0.326	0.325	0.347	0.375	0.537
诱因解释性能综合评分	0.997	0.257	0.276	0.969	0.938	0.556