摘要
针对传统滑坡易发性评价模型对复杂环境因子之间关系捕捉不足、对滑坡灾害点与周围环境之间的全局依赖关系建模能力不足等问题,构建了一种融合多注意力机制的轻量简洁的深度学习滑坡易发性评价模型CCTAF-LSM(compact CNN-Transformer with attention fusion for landslide susceptibility mapping)。该模型结合了卷积神经网络(CNN)的局部特征提取与Transformer的全局建模优势,并引入卷积块注意力机制(CBAM)和空间金字塔池化(SPP)模块,实现多尺度特征适配与增强。以四川省泸定县为研究区,选取高程、坡度、坡向等15个主控影响因子和归一化植被指数、土地覆盖类型、年均降水量等6个诱发影响因子作为模型输入开展了CCTAF-LSM模型实验。采用CNN、随机森林等模型进行了消融实验与模型效果对比。实验结果表明,CCTAF-LSM取得最高的准确率(85.98%)、精确率(84.07%)和F1分数(86.36%),滑坡易发性预测图更为精细,且模型结构简洁便于轻量化部署。四川省泸定县高和较高滑坡易发区约238.2 km^(2)(占全县11.0%),其分布与高程、道路、河流因子高度相关。本研究构建的模型在滑坡易发性评价中表现出明显的优势,可为滑坡灾害防治提供有效的技术支持。
This study proposes a lightweight deep learning model with multi-attention fusion,called the compact CNN-Transformer with attention fusion for landslide susceptibility mapping(CCTAF-LSM),to assess the susceptibility of areas to landslides.The aim is to solve the problems encountered by traditional models to this end,including an inability to represent the complex relationships among the relevant environmental factors,as well as the failure to model the global dependencies between the hazard points of landslides and the environment.The proposed model integrates the capability of the convolutional neural network(CNN)for local feature extraction with the advantage offered by the transformer in global modeling.Moreover,the authors introduce a convolutional block attention module(CBAM)and a spatial pyramid pooling(SPP)module to adapt and enhance multi-scale features.By using Luding County in Sichuan Province of China as the study area,data on 15 primary controlling factors,including the elevation,slope,and aspect,and six triggering factors,including the normalized difference vegetation index(NDVI),type of land cover,and annual average precipitation,were selected as the inputs to the CCTAF-LSM model to assess its performance.The CNN and random forest models were also used for ablation-related and comparative experiments.The results show that the CCTAF-LSM achieved the highest accuracy(85.98%),precision(84.07%),and F1-score(86.36%),and the predicted map of susceptibility to landslides generated by it contained fine spatial details.Moreover,it has a simple structure that is suitable for lightweight deployment.The authors found that zones in Luding County with high and very high susceptibility to landslides covered approximately 238.2 km^(2) of it(accounting for 11%of its total area),and their distribution was highly correlated with the elevation,as well as the presence of roads and bodies of rivers.The proposed model demonstrates sound performance in assessing the susceptibility of areas to landslides,and provides effective technical support for predicting them.
作者
李茂曈
明冬萍
李壮飞
马申奥
张杭
LI Maotong;MING Dongping;LI Zhuangfei;MA Shen’ao;ZHANG Hang(School of Artificial Intelligence,China University of Geosciences(Beijing),Beijing 100083,China;Hebei Key Laboratory of Geospatial Digital Twin and Collaborative Optimization,China University of Geosciences(Beijing),Beijing 100083,China;Frontier ScienceCenter for Deep-timeDigital Earth,China University of Geosciences(Beijing),Beijing 100083,China;State key Laboratory of Geological Processes and Mineral Resources,China University of Geosciences(Beijing),Beijing 100083,China)
出处
《成都理工大学学报(自然科学版)》
北大核心
2025年第6期1133-1150,共18页
Journal of Chengdu University of Technology: Science & Technology Edition
基金
国家自然科学基金面上基金项目(42371379)
中央高校基本科研业务费“深时数字地球前沿科学中心“深时数字地球”中央高校科技领军人才团队”项目(2652023001)
国家重点研发计划项目(2022YFB3903604)。
