张萌萌，李少达，王潇，等.基于机器学习和编码器耦合的滑坡易发性评价［J］.河南理工大学学报（自然科学版）, doi: 10.16186/j.cnki.1673-9787.2023120035.

ZHANG M M，LI S D，WANG X，et al.Landslide susceptibility assessment Based on machine learning and encoders[J]. Journal of Henan Polytechnic University( Natural Science), doi: 10.16186/j.cnki.1673-9787.2023120035.

基于机器学习和编码器耦合的滑坡易发性评价(网络首发)

张萌萌¹，李少达¹，王潇²，李欣玥³，戴可人^1，4

（1. 成都理工大学 地球与行星科学学院，四川 成都 610059；2.成都大学 建筑与土木工程学院，四川 成都 610106；3.印度马轩德拉世界联合学院，印度 浦那 412108；4.成都理工大学 地质灾害防治与地质环境保护国家重点实验室，四川 成都 610059）

摘要:目的 为了在有限样本下，提升机器学习模型挖掘数据特征的能力，提高模型预测精度， 方法  选取四川省雅砻江和大渡河中下游省级水土流失重点预防区九龙县、康定市、泸定县和木里县为研究区，选择12个影响因子构建滑坡易发性评价指标体系，使用确定性系数（cer

tainty factor，CF）量化评价指标，对比逻辑回归（logistic regression，LR）和支持向量机（support

vector machine，SVM）模型，在表现最优的模型上添加降噪自编码器（denoising autoencoder，

DAE）和卷积自编码器（convolutional auto-encoders,CAE），并对比各模型提取数据特征。 结果 结果表明：与CF-LR模型和CF-SVM模型相比，CF-SVM模型的精确率（P）、F-measure、Kappa系数、总准确度（OA）和AUC相较于CF-LR模型的分别提高了31.9%，1.1%，17.1%，8.5%，8.6%；添加DAE编码器后，CF-SVM-DAE模型的召回率（R）、F-measure、Kappa系数和总准确度（OA）相比于CF-SVM模型分别提高了8.1%，5.8%，8.1%，4%；添加CAEs编码器后，CF-SVM-CAEs模型的召回率（R）、F-measure、Kappa 系数和总准确度（OA）相比于CF-SVM模型分别提高了0.4%，0.2%，0.2%，0.1%。 结论 选用的机器学习方法中，CF-SVM模型预测精度更高。在CF-SVM模型基础上添加DAE编码器比添加CAEs编码器鲁棒性更好，因此，CF-SVM-DAE模型在所有模型中表现最好，更适合当前研究区域。

关键词: 滑坡；易发性建模；LR；SVM；编码器

中图分类号：P642.22　　

doi: 10.16186/j.cnki.1673-9787.2023120035

基金项目: 国家自然科学基金资助项目（41801391）

收稿日期：2023-12-12

修回日期：2024-04-09

网络首发日期：2024-07-28

Landslide susceptibility assessment Based on machine learning and encoders (Online)

ZHANG Mengmeng¹，LI Shaoda¹，WANG Xiao²，LI Xinyue³，DAI Keren^1，4

（1. School of Earth and Planetary Science，Chengdu University of Technology，Chengdu 610059，Sichuan，China；2.School of Architecture and Civil Engineering，Chengdu University，Chengdu 610106，Sichuan，China；3.Mahindra United World College of India，Pune，MH 412108，India；4.State Key Laboratory of Geohazard Prevention and Geoenvironment Protection，Chengdu University of Technology，Chengdu 610059，Sichuan，China）

Abstract: Objectives To enhance the ability of machine learning models to extract data features with limited samples and improve the predictive accuracy of the models. Methods In this paper，Jiulong County，Kangding City，Luding County and Muli County，the key provincial erosion prevention areas in the middle and lower reaches of the Yalong and Dadu Rivers in Sichuan Province，were selected as the study area for landslide susceptibility evaluation，twelve influencing factors were selected to construct the landslide susceptibility evaluation index system，the coefficient of determination（CF）was used to quantify the evaluation index，and noise-reducing auto-encoders（DAEs） and convolutional auto-encoders（CAE） were added to the best-performing model by comparing the logistic regression（LR）and the support vector machine（SVM） models. Results By comparing with the CF-LR model，the CF-SVM model showed improvements of 31.9% in precision（P），1.1% in F-measure，17.1% in Kappa coefficient，8.5% in overall accuracy（OA），and 8.6% in AUC. The addition of the DAE encoder to CF-SVM increased recall（R） by 8.1%，F-measure by 5.8%，Kappa coefficient by 8.1%，and overall accuracy（OA）by 4% compared to the original CF-SVM. The addition of the CAE encoder to CF-SVM increased recall（R）by 0.4%，F-measure by 0.2%，Kappa coefficient by 0.2%，and overall accuracy（OA）by 0.1% compared to the original CF-SVM. Conclusions The CF-SVM

model has higher prediction accuracy among the selected machine learning methods. The DAE encoder added to the CF-SVM has better robustness compared to CAE encoder，thus the CF-SVM-DAE model performs the best among all models and is more suitable for the current study area.

Key words：Landslides；susceptibility modelling；LR；SVM；encoders

CLC: P642.22