首页|基于Swin Transformer目标全景分割的三峡库首土质滑坡识别

基于Swin Transformer目标全景分割的三峡库首土质滑坡识别

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
[目的]滑坡识别是解决山区地质灾害隐患在哪里的关键。尤其人工智能是深度学习方法开始被广泛应用于目标识别领域，但对于多植被山区复杂环境下的滑坡隐患识别，存在着模型单一、精度较差等问题。[方法]故文章提出一种基于Swin Transformer(Shift Windows Transformer)作为骨干网络结合目标全景分割的智能识别方法，对三峡库首区域土质滑坡开展识别。将三峡库首的485处土质滑坡制作成样本集，并分为训练集和测试集。将训练集加载进Swin Transformer模型中进行训练，模型采用自注意力机制对训练集提取特征，构建特征图，测试集验证特征图的识别精度，保留识别精度最高的特征图。最终以此实现滑坡目标与背景区域的有效区分进而完成隐患识别，同时与DeepLab V3模型进行对比。[结果]结果显示:Swin Transformer模型在识别精度和识别速度上都要高于Deep-Lab V3模型，在三峡库首的试验中准确率可以达到83。55％，单张图片预测时间为0。18 s。[结论]结果表明:该方法能够在多植被山区复杂环境下快速识别土质滑坡，可为多植被山区的滑坡灾害调查提供参考。
Identification of soil landslides at the head of the Three Gorges Reservoir based on swin transformer target panoramic segmentation
[Objective]Landslide identification is the key to solve the problem of where the geological disaster hazards are in moun-tainous areas.Artificial intelligence,especially deep learning method,began to be widely used in the field of target recognition,but for landslide hazard recognition in complex environments in multi-vegetation mountainous areas,there are problems such as sin-gle model and poor accuracy.[Methods]Therefore,an intelligent recognition method based on Swin Transformer(Shift Windows Transformer)as backbone network combined with panoramic target segmentation is proposed in this paper to identify soil landslide in the head area of the Three Gorges Reservoir.The 485 soil landslides at the head of the Three Gorges Reservoir are made into a sample set and divided into a training set and a test set.The training set is loaded into the Swin Transformer model for training.The model adopts the self-attention mechanism to extract features from the training set and construct feature maps.Finally,this method can achieve effective differentiation between the landslide target and the background area,and then complete the potential hazards identification.At the same time,it is compared with DeepLab V3 model.[Results]The result show that the Swin Trans-former model is higher than the DeepLab V3 model in recognition accuracy and recognition speed,and the accuracy can reach 83.55％in the experiment at the head of the Three Gorges reservoir,and the prediction time of a single image is 0.18 s.[Con-clusion]The result show that the method can rapidly identify soil landslides in the complex environment of multi-vegetation moun-tainous areas,and can provide a reference for landslide hazard investigation of multi-vegetation mountainous areas.

the Three Gorges Reservoir Headsoil landslideSwin Transformerpanoramic segmentationhazard identificationlandslide

邓志勇、黄海峰、李清清、周红、张瑞、柳青、董志鸿

展开 >

三峡大学湖北长江三峡滑坡国家野外科学观测研究站,湖北宜昌 443002

三峡大学三峡库区地质灾害教育部重点实验室,湖北宜昌 443002

三峡大学湖北省水电工程智能视觉监测重点实验室

宜昌市地质环境监测站,湖北宜昌 443002

展开 >

三峡库首 土质滑坡 Swin Transformer 全景分割 隐患识别 滑坡

国家自然科学基金国家自然科学基金国家自然科学基金三峡库区地质灾害教育部重点实验室开放基金

U21A203142007237421074892020KDZ09

2024

10.13928/j.cnki.wrahe.2024.04.016

水利水电技术(中英文)
水利部发展研究中心

水利水电技术(中英文)

CSTPCD北大核心
影响因子：0.456
ISSN：1000-0860
年,卷(期)：2024.55(4)
  • 29