基于语义分割的长白山火山岩性遥感数据集  

Lithology dataset of remote sensing image in Changbaishan volcano based on semantic segmentation

作　　者：李成范[1] 韩晶鑫 盘晓东[2,3] 刘岚 颜丽丽 康建红 刘学锋[1] 肖舟怡 LI ChengFan;HAN JingXin;PAN XiaoDong;LIU Lan;YAN LiLi;KANG JianHong;LIU XueFeng;XIAO ZhouYi(School of Computer Engineering and Science,Shanghai University,Shanghai 200444,China;National Observation and Research Station of Jilin Changbaishan Volcano,Jilin Earthquake Agency,Changchun 130117,China;Institute of Volcanology,China Earthquake Administration,Changchun 130117,China;School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China;National Observation and Research Station of Jilin Changbaishan Volcano,Institute of Geology,China Earthquake Administration,Beijing 100029,China)

机构地区：[1]上海大学计算机工程与科学学院,上海200444 [2]吉林省地震局吉林长白山火山国家野外科学观测研究站,长春130117 [3]中国地震局火山研究所,长春130117 [4]上海工程技术大学电子电气工程学院,上海201620 [5]中国地震局地质研究所吉林长白山火山国家野外科学观测研究站,北京100029

出　　处：《岩石学报》2025年第4期1442-1453,共12页Acta Petrologica Sinica

基　　金：上海市自然科学基金项目(22ZR1423200);吉林长白山火山国家野外科学观测研究站课题(NORSCBS23-02);吉林省科技发展计划项目(20230203132SF);国家重点研发计划项目(2021YFC3101604)联合资助.

摘　　要：火山岩性数据集是利用深度学习进行火山遥感岩性智能识别的关键和数据基础。当前,缺乏可信的火山岩性遥感数据集,制约了大区域、复杂地质环境下火山岩性智能识别的快速发展。本文在归纳和整合长白山火山岩性主要类型的基础上,以哨兵2(Sentinel-2)遥感图像为数据源,结合地质资料和野外核查制作了一个基于深度学习语义分割的长白山火山岩性遥感数据集。该数据集内容包含遥感图像、标签数据、说明文件,岩性类型覆盖玄武质火山岩、粗面质火山岩、碱流质火山岩、火山岩性混合堆积(碎屑堆积、火山泥流堆积、火山空落堆积);共计36张样本图像,单张图像尺寸为395像元×395像元,空间分辨率为10m。利用经典的深度卷积神经网络(deep convolution neural network,DCNN)DeepLab V3+模型对火山岩性数据集进行了测试和验证,实验结果表明本文数据集具有较强的火山岩性描述能力,鲁棒性和泛化性较好,总体准确率均高于88%;特征训练与提取过程中人为干扰较少,自动化水平较高。可为火山岩性智能识别提供数据基础,提高野外火山遥感岩性调查的准确性和效率。The volcanic lithology dataset is the key and data foundation for intelligent identification of volcanic lithology from remote sensing image by deep learning technology.Currently,the reliable remote sensing datasets of volcanic lithology are scarce,restricting the rapid development of intelligent identification for volcanic lithology in large-scale and complex geological environments.Based on summarizing and integrating the main types of lithology in Changbaishan volcano,in this paper,a novel lithology dataset of Sentinel-2 remote sensing image in Changbaishan volcano for deep learning semantic segmentation is constructed and presented assisting by geological data and field verification.The dataset includes remote sensing images,labeled data,and explanatory files,covering the lithology type of trachyte,basalt,pantellerite,and mixed volcanic rock deposits(i.e.,pyroclastic deposit,volcanic mudflow deposit,fallout tephra deposit).It has a total of 36 sample images of volcanic lithology,each sample image is 395×395 pixels with a spatial resolution of 10m.A traditional deep convolutional neural network(DCNN)(DeepLab V3+)model is used to test and validate the volcanic lithology identification on the constructed dataset.The experimental results show that our lithology dataset has strong ability to describe volcanic lithology,good robustness and generalization,and the overall accuracy is higher than 88%.There is less human interference in the deep learning and it automatic and intelligent trains and extracts the lithology feature from remote sensing image.The quality of volcanic lithology dataset is high.The dataset could provide a data foundation for intelligent identification of volcanic lithology,and further improve the accuracy and efficiency of volcanic lithology investigation from remote sensing image.

关 键 词：长白山火山 语义分割 岩性数据集 岩性识别 遥感图像 

分 类 号：P317[天文地球—固体地球物理学]