水土保持通报2025，Vol.45Issue(2)：190-200,210,12.DOI:10.13961/j.cnki.stbctb.2025.02.020
基于样本优化与深度特征提取的滑坡易发性评价
Landslide susceptibility evaluation based on sample optimization and deep feature extraction
摘要
Abstract
[Objective]The efficacy of the non-landslide sampling method and a model with excellent feature extraction in evaluating landslide susceptibility were explored,so as to provide theoretical support and scientific guidance for regional landslide prevention and control work.[Methods]A non-landslide sample optimization method based on Convolutional Auto-Encoder(CAE)was proposed,which was built based on the buffer sampling strategy.This method optimizes non-landslide samples by learning the features of landslide samples and using reconstruction errors.For the evaluation model,the Convolutional Block Attention Module(CBAM)was integrated into the Residual Network(ResNet)to construct the ResNet-CBAM landslide susceptibility model,which captures deeper,and more complex and representative features.Taking the Wanzhou District,Chongqing City in the Three Gorges reservoir area as the study area,12 influencing factors(e.g.,such as elevation)were selected.Four models,namely,SVM,DNN,CNN,and ResNet-CBAM,were used to compare and analyze the evaluation accuracy and results of buffer zone sampling versus CAE-based optimized sampling.[Results]Under the same evaluation model,the CAE-based optimization sampling strategy for non-landslide samples yielded higher reliability and accuracy.Under the same sampling strategy,the ResNet-CBAM model outperformed the other models in terms of accuracy,precision,recall,F1 score,and area under the curve(AUC)values.The evaluation results were similar across models,with higher and very higher susceptibility areas predominantly located in regions with lower vegetation cover and frequent human activity,such as along the Yangtze River.Moreover,the ResNet-CBAM model with CAE-based optimized sampling demonstrated superior prediction results and was more suitable for landslide susceptibility evaluation in this area.[Conclusion]Wanzhou District exhibits a high landslide susceptibility index,with numerous potential landslide risk zones identified within the area.The non-landslide sampling strategy and ResNet-CBAM evaluation model based on CAE optimization can effectively improve the accuracy of landslide susceptibility evaluations.关键词
滑坡易发性评价/非滑坡样本/卷积自编码器/残差网络/卷积注意力模块Key words
landslide susceptibility evaluation/non-landslide samples/convolutional auto-encoder/residual network/convolutional block attention module分类
地质学引用本文复制引用
徐金鸿,李清泉,韦春桃,赵芹..基于样本优化与深度特征提取的滑坡易发性评价[J].水土保持通报,2025,45(2):190-200,210,12.基金项目
重庆市自然科学基金面上项目"遥感知识图谱引导的耕地智能提取与监测"(CSTB2023NSCQ-MSX0781) （CSTB2023NSCQ-MSX0781）
